Rethinking Indic AI from a Lens of Cultural Heritage Preservation APARNA MADVA, SHARATH SRIVATSA, SRINATH SRINIVASA, and TULIKA SAHA, International
arXiv:2607.06544v1 [cs.AI] 7 Jul 2026
Institute of Information Technology, Bengaluru, India As Artificial Intelligence (AI) makes inroads into different parts of the Indian subcontinent, there is significant interest in studying how AI impacts the linguistic and cultural foundations of this civilization. AI is seen as a “double-edged sword” where on the one hand, it can enable access and inclusion for a large population, on the other, it can homogenize worldviews and exclude underrepresented languages and worldviews. In this paper, we try to characterize this problem by addressing the extensive characteristic nature of Indian linguistics and the way they closely connect to cultural practices and worldview. We then perform a longitudinal survey of how Natural Language Processing (NLP) techniques have evolved in this space, tracing the historical development of Indic NLP, covering key milestones, methodological shifts, and resource creation efforts. In addition, the paper also examines the structural and sociolinguistic characteristics of Indian languages, such as rich morphology, complex scripts and grammar rules, diglossia, and large dialectal variation, and explains how these create unique challenges for building AI foundation models. We then discuss the growing role of Indic foundation models and analyze how these models address these long-standing resource and representation gaps. Finally, we propose a research direction called ‘Culture Sensing’, which re-imagines AI based on hermeneutic reasoning. Culture Sensing aims to address open problems such as ensuring equitable performance across low-resource languages and producing outputs that are culturally meaningful. By bringing together past work, current techniques, and emerging trends, this paper outlines research directions that can guide the next phase of Indic NLP and contribute to the development of more robust and inclusive Indic foundation models. CCS Concepts: • Computing methodologies → Natural language processing; • Applied computing → Arts and humanities; Digital libraries and archives; • Social and professional topics → Computing / technology policy. Additional Key Words and Phrases: Indic NLP, Cultural Heritage, Large Language Models, Algorithmic Homogenization, Culture Sensing
1
Introduction
Indic languages refer to the wide spectrum of languages spoken in the Indian subcontinent, including India, Nepal, Sri Lanka, Pakistan, Bhutan, Bangladesh, and other countries– that have a long linguistic and cultural history, collectively contributing to more than a fifth of the total world population. The Indian subcontinent has a long and rich linguistic history, hosting a wide variety of languages. Each language often uses a different script and has a rich body of literature. Just in India, the constitution recognizes 22 major literary languages as official languages of the country. In addition to this, there are about 121 other non-official, but major languages, and more than 19,000 minor languages, dialects, and creoles. Languages are also culturally significant across different regions of the subcontinent, often shaping their own unique regional identity. An innate hermeneutic diversity can be observed in the case of the Indian subcontinent. The population of the Indic subcontinent often holds diverse perspectives drawn from the sociocultural discourse. Figure 1 shows pairs of sentences, a source English sentence, and corresponding translations in the Indic languages Hindi and Kannada. In the first sentence, it can be seen that mentioning of the gender is inevitable while making a statement such as ‘I went to a movie yesterday with my friend’. From the second sentence, it can be observed that words that signify ‘identity’ rather than ‘ownership’ are used to convey the meaning of owning a house. Such minute details not only reflect the linguistic divergence but also the underlying hermeneutic differences. The integration of AI into the cultural and linguistic landscape of the Indian subcontinent is bound to significantly shape the development and progress of the region. For India, AI adaptation is driven by a dual necessity: the scope to leapfrog traditional infrastructure deficits [AIForAll 2018; Kalyanakrishnan et al. 2018] and the desire Authors’ Contact Information: Aparna Madva, [email protected]; Sharath Srivatsa, [email protected]; Srinath Srinivasa, sri@ iiitb.ac.in; Tulika Saha, [email protected], International Institute of Information Technology, Bengaluru, Bengaluru, Karnataka, India.
2
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Fig. 1. Translated Sentence Pairs Demonstrate the Innate Worldview of the Indic Subcontinent
to preserve, scale, and disseminate its immense cultural [Choudhary and Sukhvir 2025] and linguistic [Krishna et al. 2025] diversity. AI is projected to foster inclusion, economic growth, and also social good [Maheshwari [n. d.]]. India’s diverse sociocultural fabric can also benefit from using AI for the restoration and preservation of culture and heritage. It can enhance traditional values like Vasudhaiva Kutumbakam (meaning - the world is one family), ensuring focus on inclusive, community-centric technology. Apart from this, improvement in digital literacy and usage has resulted in an increase in Indian language content on the Internet across multiple platforms, including social media, news media, and others. Technological advances have been penetrating quickly into the Indian population, with the number of Indian language Internet users estimated to exceed 900 million as of the year 2025 [Indian Economy News 2025]. This advancement calls for language technology that enables people to access information in their native language. However, the potential of AI is faced with significant bottlenecks for adoption, primarily due to the digital divide and linguistic challenges. This is predominantly observed in the current-day Large Language Models (LLMs), that are prone to homogenization of hermeneutic interpretations due to lopsided representations of disparate worldviews in their training data, as well as lopsided representation of user base and their feedback that fine-tunes the model’s responses over time. The issue is amplified in the case of Indic languages since LLMs are shown to align disproportionately with the linguistic patterns of specific subpopulations [Sourati et al. 2025]. It has also been observed that an individual (or groups) exclusively receive negative outcomes in algorithmic decision making. For example, a job applicant may get rejected from a lot of job opportunities they apply to when all the companies make use of similar resume screening algorithms [Bommasani et al. 2022a]. The major reason for algorithmic bias is the lack of representation for not just Indic language text, but also the diverse worldviews in the Indian subcontinent. A majority of the LLMs are trained on translations from English language datasets that are collected mostly in urban contexts. When applied to Indic contexts, they fail to generalize to the different cultural nuances. Multiple initiatives have been focusing on Natural Language Processing for Indic languages– or Indic NLP for short, with the goal to build representative datasets in multiple Indic languages, and also to model the different linguistic features of Indic languages. Indic NLP has its own evolutionary history, somewhat paralleling mainstream research in NLP that is most prominently focused on English, and to a lesser extent, on other
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
3
European languages. A considerable amount of research focusing on language technology in Indian languages has been conducted for a long time and is continuously pursued at present. Research on cultural alignment of LLMs [Agarwal et al. 2026; Bhagat et al. 2026] is also gaining traction. These studies have contributed valuable resources, such as a corpus and models, to the Indic NLP community. It is important to ensure that both the hermeneutic and linguistic diversity is cohesively preserved. While there has been substantial development in Indic NLP pertaining to models, datasets, and benchmarks, representation of the multitude of worldviews is still lacking in current-day language models. To preserve the cultural diversity in the Indian subcontinent, it is essential to ensure that the corresponding hermeneutic plurality is represented in the models being built. In this paper, we present a longitudinal study of research on NLP for Indic languages. We provide an overview of the evolution of research in this area and the different paradigms explored. Due to the dynamically changing NLP landscape, we include studies published only till 2025 to the best of our knowledge. We also investigate prevailing challenges, including the attainment of equitable performance across low-resource languages and dialects, and the preservation of cultural and linguistic fidelity in generative outputs. We also introduce a new research direction, Culture Sensing, that rethinks the framework of AI based on hermeneutic reasoning. We substantiate this idea with the help of two different use cases where AI techniques are utilized to analyze lesser-known knowledge systems. Culture Sensing utilizes the vibrant multimodal knowledge systems in various native communities and represents the corresponding worldview in AI models. By this, the focus is on making AI models pluralistic and inclusive, thereby preserving lesser-known knowledge systems on the brink of extinction. By bringing together foundational research, contemporary methodologies, and emerging trends, this paper sets out strategic research directions intended to facilitate the advancement of robust, culturally inclusive Indic NLP. Section 2 provides the background of developments in the NLP field and distinguishing features of Indian languages to be considered for building efficient NLP systems. Section 3 reviews different Indic NLP works since their inception till the recent breakthroughs, considering the evolving approaches. Section 4 provides an overview of unique challenges for IndicNLP and discusses mitigation strategies. Section 5 discusses the proposed idea of Culture Sensing to motivate the need for culturally inclusive AI and provides directions towards its achievement.
2
Characteristics of Indic Languages
Languages in the Indian subcontinent have a rich linguistic heritage and a literary tradition that spans across millennia. Words in Indic languages have complex morphology, and the grammar for their organization includes elaborate rules. An understanding of the distinctive features of Indic languages is necessary for building meaningful language technology for them. This section explores the basic structure and essential features of Indic languages along with examples.
2.1
Akshara System of Indian Languages
Indic languages are predominantly phonetic in nature, where the alphabet is based on observing variegated forms of sounds emanating from the human vocal system comprising of throat, tongue, lips and nasal elements. The alphabet of a majority of Indic languages contains 33 consonants and 15 vowels. Some languages can also include additional 1-2 consonants and vowels. The consonants contain an inherent vowel and are usually divided into 25 structured consonants (shown in Figure 3a) and the remaining unstructured consonants (shown in Figure 3b). Consonants are further clustered into different segments depending on which part of the human vocal system is used to produce that class of sounds. An individual letter of the alphabet is called an akshara that consists of a vowel and 0 or more consonants. Each akshara has its own form and sound. The letters and words are pronounced exactly as they are written,
4
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Features of Indic Languages
Morphology
Akshara System
Grammar
Free Word Order
Inflection
Diglossia
Agglutination
Sandhi
Kaaraka
Derivation
Compounding
Regional
Communal
Fig. 2. Characteristics of Indic Languages
(a) Structured Consonants
(b) Unstructured Consonants Fig. 3. Consonants in Indic Languages
since the languages are phonetic. The number of written symbols can be more than the number of characters in the alphabet, as more than one consonant can be combined without an intervening vowel to form digraphs. Unlike English, where sounds are constructed by lexical concatenation of letters from a base alphabet, Indian languages construct syllables by modifying consonants with verbs and other consonants. Even though the alphabets of Indic languages are similar, the same is not reflected in the scripts used to write them. The written script has wide variability for different languages in India, as shown in Figure 4a. Differences can be observed even in the structure of graphemes and their combinations. The rules for the formation of graphemes vary within a script and also across scripts. For example, the Devanagari script used to write Hindi language includes a horizontal line called Shirorekha at the top of the word that connects every character of the word. But this style of writing is not used by every Indic language, and a few of them write graphemes without touching each other. The alpha syllabic writing system is used, where consonants and vowels are represented as
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
(a) Different Indic Language Scripts
•
5
(b) Agglutination Forming Different Words With One Root Word Fig. 4
a single unit, as shown in Figure 5a. Consonants represent a letter, while vowels can either appear in atomic form, not connected to consonants, or are dependent on the consonant and marked like a diacritic or some other form of modification. A vowel can appear to the left, right, top, or bottom of the consonant. A grapheme representing a combination of vowels is also represented similarly, and the supporting consonant can be a modified form of its original form, or it can be completely different as well. A large number of diacritics are also present, shown in Figure 5b.
(a) Letter Formation with Consonants and Vowels
(b) Different Diacritics used in Indic languages
Fig. 5. Inherent Features of Indic Languages
2.2
Formalism of Indian Languages: Panini’s Framework
The formal structures of most Indic languages are either directly based on, or are greatly influenced by the framework established by Panini, called Astadhyayi, written around the 6𝑡ℎ century BCE [Bharati and Sangal 1993]. This framework of structuring language also influences the way language is communicated and characteristics the resultant hermeneutic structures. In this section, we look at some key elements of the Paninian framework that are seen in almost all Indic languages. Morphemes are elemental meaning-bearing words in Indic languages, that can in turn, be considered to have two elements: the stem and the affix. Prefixes, suffixes, circumfuse, and infixes are different types of affixes. Prefixes precede the stem word, suffixes follow the stem word, and circumfuse morphemes precede and follow
6
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
the stem word. Affixes inserted into stem words are called Infixes. Stem words carry the primary meaning, and affixes give various kinds of additional meaning to the stem word [Neupane 2024]. Among the affixes are a set of case modifiers that are used to assign roles to specific words. This is enabled with the concept of kaaraka to depict the role of a given word in a sentence. A set of suffixes called vibhakti is usually used to express kaaraka. The number of vibhaktis in a language can be 7 or 8. This concept can be related to the idea of grammatical cases in languages like English and German. For example, the word raja (king) can be added with the suffix annu that represents dwitiya vibhakti to form the word rajanannu and the corresponding kaaraka as karma kaaraka, which refers to the object of an action. Hence, the resultant word rajanannu forms the object of a sentence it is part of. These role modifiers give the property of free word order to Indic languages, where the position of role-modified words in a sentence can be changed without modifying the overall meaning of the sentence. The concept of vibhakti has to be considered while designing for tasks such as Named Entity Recognition (NER) in Indic languages. It has been observed that the identification of an entity may depend on its suffix corresponding to an appropriate kaaraka in the case of the NER task for the Kannada language [M and Srinivasa 2023]. Another concept in Indic languages that gives it an agglutinative nature is Sandhi or morphophonemic rules that allow a word to be joined with other words or morphemes, resulting in a newly formed word with phonological changes and also possible orthographic variations. For example, joining the Kannada words mane (house) and inda (from) results in maneyinda (from house) following the Agama sandhi [Zydenbos 2020] rule. Compounding or samaasa is also observed frequently [Kulkarni et al. 2012]. Here, two or more meaningful words are joined together to output a (longer) word such that its meaning is derived from the underlying words. The produced word contains more than one stem. For example, the Sanskrit word vidyanipuna, meaning one who is sharp in the studies, is formed by two words vidya (education) and nipuna (expert), following the Tatpurusha samasa rule. These phenomena not only obscure word boundaries, but also modify the characters at the joining point of the words. This makes the task of word segmentation and tokenization complex in the case of Indic languages. Hybrid approaches for word tokenization, such as the one proposed by [Sandhan et al. 2022], have demonstrated improved performance using knowledge of the sandhi phenomenon. Bharati and Sangal [Bharati and Sangal 1993] describe a computational grammar formalism for Indian languages in general from Paninian Grammar, which was composed for Sanskrit, with the idea that Sanskrit can form the bridge language for translating between any two Indian languages. The authors show a compact formalism to parse simple and complex sentences and active and passive voices in sentences using the invariants appearing in verbal and noun parts of the sentence, while the sentence, on the whole, can follow free word order. The notion of Kaaraka relations is the core of the described Panian Framework. Kaaraka relations are syntactico-semantic (or semantico-syntactic) relations that link verbal parts of the sentences to noun parts by invariants appearing with root verbs and nouns in a free-word order sentence. The invariants in noun parts are vibhaktis or postpositions, and the invariants of verbal parts are tense-aspect-modality (TAM) elements. The invariants help capture the sentence’s gross meaning even when word order is changed. Moreover, a change in word order conveys secondary information like emphasis, etc., and does not affect the primary meaning of the sentence. The verbal group invariant forms impose or demand specific invariants in the noun group, and hence, the verbal group is the demand group, and the noun group is the source group. The invariants in demand and source are mainly prevalent in the spoken and written texts of Indian languages, which makes the Paninian Framework a suitable formalism for Indian languages. There are two components to establishing the kaaraka relations. The default Kaaraka chart, where three types of kaarakas or nominals, Karta, Karma, and Karana, are mapped to post positions or vibhaktis, and used to establish a relation when the verb group has basic TAM. The other map is the transformation rules, which map all possible TAMs in the verbal group to specific postpositions in the default kaaraka chart. One can use the kaaraka relation formalism to build a parser. The first stage is morphology analysis of a sentence, which involves
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
7
obtaining grammatical information about words from a lexicon or dictionary and local word grouping of noun groups and verbal groups. Noun groups contain postposition markers with nouns, and verbal groups contain verbs and their auxiliaries.
2.3
Diglossia in Indian languages
Diglossia refers to the difference in the way in which a language is spoken and used in colloquial settings as against the codified, formal variety that is often used in formal education [Khubchandani 1985]. This becomes an essential consideration while building language models since the training data for language models mostly belongs to the formal language variety. In a majority of Indic languages, the spoken (colloquial) style of the language differs significantly from the literary (formal) variety. Multiple dialects of a language exist, and these dialects can also exhibit distinguishing characteristics. Among multiple dialects, there can be differences in terms of phonology, morphology, lexicon, and syntax. The distinguishing factor for dialects can be regional or social. In some cases, the words can be entirely different between dialects. For example, the word for money is duddu in Bengaluru Kannada, whereas it is rokka in Dharwad Kannada. Similarly, some dialects of Hindi are Khari Boli, Braj, Bundeli, Marwari, Kumauni, and Garhwali. Some dialects of languages are also characterized by the community speaking it. For example, Soliga Kannada is used by the Soliga tribe and has some Tamil influence. Halakki Kannada is spoken by a tribe called Halakki Vokkaligas. A few members of the Gowda community living in the Kodagu district of Karnataka use the Arebhashe language. In addition to this, the inter-generational tacit knowledge is communicated orally using colloquial style languages, especially in the communities in rural areas. All of these variations are synchronic, and they occur within a single point in time. There also exist diachronic variations, when languages evolve through multiple historical periods. For example, Kannada language evolved through three periods: Halegannada, Nadugannada, and Hosagannada. Halegannada nurtured classical literature, and most of the popular literature belongs to either Nadugannada or Hosagannada. Many phonological variations can be observed between these language varieties. LLMs can be efficiently used to quantify the diachronic changes as demonstrated by [Hariharan and Mortensen 2025]. There is a need for detailed research in this area. Language models are generally trained using a corpus that uses a literary style of language. These language models hence contextually understand the formal style of the language. However, it’s not necessary that the same model demonstrates similarly when its input is with some dialect or colloquial variety of the same language. Building models that generalize for dialects is also challenging, as shown for the task of speech recognition in [Elfeky et al. 2018], and needs better strategies. An ideal language model should be inclusive and demonstrate competitive performance for different dialects of the same language.
3
Historical Context: Evolution of Indic NLP
Different paradigms of NLP have been utilized for processing Indic language data and for building models for Indic languages over time. The dynamically changing NLP landscape, right from rule-based processing techniques that used the structure of these languages to form processing rules to the recent transformer architecture, has been experimented with large quantities of data and evaluated for their performance in the case of Indic languages. We cover the research works on IndicNLP through the different techniques to the best of our knowledge in the following subsections.
3.1
Rule-based Indic NLP
Rule-based NLP is an approach for processing text using manually crafted, predefined linguistic rules and patterns. It relies on handcrafted rules built using expert-defined grammar and dictionaries, and pattern matching using
8
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
regular expressions to analyze the syntax and semantics. Although this approach was highly accurate for specific tasks, it was time-consuming, provided limited generalization, and was rigid. Early rule-based IndicNLP works mostly focused on Machine Translation (MT), parsing the Indic language text, and building lexical databases such as WordNet for Indian languages. Table 1 provides a comprehensive list of rule-based Indic NLP models and the tasks they addressed. 3.1.1 Machine Translation. The MT systems performed Indic-English, English-Indic, and Indic-Indic translation. Three types of strategies were used for the development of MT systems: • Direct translation • Interlingua strategy • Transfer strategy The direct translation strategy relied on resources using dictionaries, morphological analysis, and text processing systems for translation involving a specific source and target pair of languages without any intermediate representation. The transfer strategy first performed text analysis according to the structure of the source language. Following this, processing in the context of the target language was carried out to perform ’transfer’ at the lexical level, syntactic level, or at the semantic level. A target language independent universal representation was used by the Interlingua strategy for a given source language text. Any ambiguities were presumed to be resolved in this representation, and hence it should be usable to generate text in any target language. Anglabharti, an English-Indian language translation system, was proposed by [Sinha et al. 1995], which used a rule-based system using patterns with Context Free Grammar (CFG) like structure for the source language (English). With Hindi as the target language, a prototype system with about 50 rules for English to Hindi translation was reported to be able to translate most of the common sentences. Another machine translation system, Anusaaraka, proposed by [Bharati et al. 2003b] was able to perform translation for the Telugu, Kannada, Marathi, Bengali, and Punjabi to Hindi using human assistance. For preserving the information while translating from a source language to a target language, substitutibility and reversibility were maintained in the translated strings as explained by [Bharati et al. 2003a]. [Rao et al. 2000] proposed a framework for the syntactic transfer of compound complex sentences from English to Hindi using a transfer-based Machine Assisted Translation (MAT) system. Due to the inflectional nature of Hindi, this work used strategies for inflecting appropriate inflections and case markers as discussed in [Rao et al. 1998]. An interlingua-based approach for English to Hindi translation was proposed by [Dave et al. 2001], using Universal Networking Language (UNL) interlingua. This work investigates the ability of an interlingua-based approach for translation in handling cross-language divergences. 3.1.2 Parsing. In case of parsing, it is observed that the dependency parsing framework is better suited for free-word order languages, including Indic languages [Bharati et al. 2002a]. This approach was elaborated in an early study by [Bharati and Sangal 1990] in 1990, which proposed an overall parsing strategy for Indian languages using the kaaraka relations described in Paninian grammar. This work proposed a parsing system with three components: a morphological analyzer to extract roots of each word in the input sentence, along with other corresponding grammatical information, followed by a Local Word Grouper (LWG) [Bharati et al. 1991] that formed word groups using information based on adjoining words to reduce the complexity for the core parser [Bharati and Sangal 1993]. Dependency parsing analyses the words via the dependency relation between them rather than structuring the phrases hierarchically. Because of this, dependency analysis is more suited for Indian languages than CFG. [Begum et al. 2008] came up with an approach to perform dependency annotation on Indian languages following the Paninian framework. This study utilized the concept of kaaraka as they can be identified syntactically and can be used to understand the underlying semantics. An improved performance using data-driven dependency parsing was obtained by [Husain et al. 2009] following a modular cascaded approach in the case of the Hindi language. In this work, each intermediate layer or module
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
9
of the parser produced a linguistically valid partial parse. The final parse obtained using this method aimed at minimizing the adverse effects of long-range dependency and the non-projective nature in free word order languages. The two-stage approach for parsing was also used by [Bharati et al. 2009c] and [Bharati et al. 2009b] to propose a constraint-based hybrid approach for dependency parsing in free word order languages such as Hindi. This work incorporated hard and soft constraints (H-constraints and S-constraints) in the two-stage parsing approach. H-constraints, satisfied by any grammatical sentence, include lexical and structural knowledge of the language, such as structural constraints, lexicon, and other language-specific rules. S-constraints, on the other hand, were used as preferences that can be broken by a sentence. [Bharati et al. 2002c] proposed AnnCorra, a dependency-based tagging scheme for annotating corpora in Indian languages based on the Paninian grammatical model. The proposed generalized tag scheme aimed at efficiently annotating a corpus and developing treebanks accounting for the distinct syntactic structures in all Indian languages using a tagging scheme made up of notations, defaults, and tagsets. [Kanuparthi et al. 2012] proposed an algorithm for derivational morphological analysis of the Hindi language. This work utilized an existing inflectional morphological analyzer ([Bharati et al. 2002a]) for this work. They used a list of manually extracted derivational suffixes, which were used to form a set of rules for the analyzer, along with a list of words extracted from Hindi Wikipedia. These components were used by the proposed algorithm to output both inflectional analysis and derivational analysis for a given word.
3.1.3 Lexical Resource Creation. A rule-based automatic annotation approach to annotate a Hindi Treebank using the Paninian dependency framework was proposed by [Gupta et al. 2008]. This work extended the experiments in [Begum et al. 2008]. Here, dependency labels were marked only for inter-chunk relations where a chunk is a set of adjacent words. The robust formulation of rules for the automatic annotator anticipated a reduction in time and effort for manual annotators in dealing with a large corpus. A comparison of the annotator with a constraint-based parser explained in [Bharati et al. 2002b] showed that the annotator performed better than the parser for most of the dependency relations. It was also observed that the annotator helped in improving the results of the first parse of the parser and hence could also be used for post-processing. [Bharati et al. 2009a] extended this work to include intra-chunk annotation and included additional linguistic features for better identification of relations. The parser proposed in this work is a language-independent engine that takes a rule file for a specific language, as mentioned by [Gupta et al. 2008]. A subset of the dependency treebank proposed by [Begum et al. 2008] was utilized by [Bharati et al. 2008] to understand a few of the crucial cues of a language, which are essential for building a robust parser in the case of the Hindi language. Among the two techniques of dependency parsing, grammar-driven dependency parsing that rules out those parses that don’t satisfy the established constraints and data-driven dependency parsing that learns a probabilistic model for parsing, this work experimented with two different data-driven parsers. It was observed that the extended feature set for the parser with vibhakti labels increased the performance (here, vibhakti is a generic term for preposition, post-position, and suffix). AnnCorra and two more efforts, TransLexGram and Shabda-Sutra, aimed at creating lexical resources in Indian languages, were discussed by [Bharati et al. 2003c]. Transfer lexicon and grammar, or TransLexGram in short, attempted to produce a transfer lexicon and grammar from English to Hindi. This work generated bilingual dictionaries, parallel corpora for English-to-Indian-language translation, and simple transfer patterns that can be used directly by machine translation systems. Shabda Sutra explored the various semantic meanings of a polysemous word. [Kulkarni et al. 2010] proposed a WordNet for Sanskrit. The WordNet contained verbs in their root forms, following the technique explained in [Kulkarni and Bhattacharyya 2009] to efficiently store morphological information. [Bhingardive et al. 2014] proposed a semi-automatic approach to enhance this Sanskrit wordnet, using mapping from an existing bilingual Sanskrit English dictionary and the Princeton WordNet to populate the Sanskrit synsets.
10
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Table 1. NLP Research Works using Rule-based strategy for Indic Languages
Focused Task Model
Machine Translation
[Bharati and San- ✓ gal 1990] [Bharati et al. 1991] [Bharati and Sangal 1993] [Sinha et al. 1995] ✓ [Rao et al. 1998]
✓
[Dave et al. 2001]
✓
[Bharati et 2003b] [Begum et 2008] [Gupta et 2008] [Bharati et 2009a] [Bharati et 2008]
Parsing
Resource Creation
✓ ✓ ✓
al. ✓ al.
✓
✓
al.
✓
✓
al.
✓
al.
✓
[Husain et al. 2009] [Bharati et al. 2009c] [Bharati et al. 2003c] [Kulkarni et al. 2010] [Bhingardive et al. 2014]
✓
3.2
✓ ✓ ✓ ✓
Approach
Core parsing formulated as integer programming problem Local word grouping (verb and noun groups) used as a pre-parsing strategy Constraint based parsing using the Paninian framework Uses a ‘pseudo-target’ applicable to a group of Indic languages, later used for text generation Structural and Lexical transfer, tagged with semantic knowledge Interlingua-based translation using the Universal Networking Language (UNL) Human-machine aided MT via a languageindependent text image Syntactico-semantic dependency annotation based on the Paninian framework Automatic annotation using syntactic cues, following the Paninian dependency framework. Corpus-based parsing to identify inter and intrachunk dependency relations Data-driven dependency parsing enhanced with two features: conjoined vibhakti-label feature and minimal semantics Modular cascaded data-driven dependency parsing for Hindi Two stage constraint-based dependency parsing to identify inter and intra-clausal dependencies Collaborative development of lexical resources using crowd sourcing and open source tools Construction of Sanskrit wordnet using the expansion approach, based on Hindi wordnet Sanskrit wordnet extension using Princeton wordnet and Hindi wordnet
Corpus-based Indic NLP
Corpus-based or statistical NLP relies on the analysis of large, structured corpora to automatically derive linguistic patterns or statistical probabilities. Corpus-based NLP uses empirical data to perform tasks such as tagging, parsing, and translation. Unlike the standard practice that assumes a fixed sentence structure, statistical NLP was
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
11
used for morphological analysis of the morphologically rich and free word order Indic languages. Apart from this, research focused on hybrid parsing for Indic languages by integrating rule-based patterns with statistical models to overcome the drawbacks of the traditional constituency parsing technique. In the case of the Statistical Machine Translation (SMT), three approaches were used: word-based, phrasebased, and hierarchical phrase-based translation. While the input sentence was translated word by word and then arranged to get the target sentence in the first approach, each source and target sentence was divided into different phrases and aligned using patterns in the phrase-based approach. The hierarchical phrase-based translation approach used hierarchical phrases with recursive structures instead of simple phrases. A few studies also combined both rule-based translation and corpus-based translation approaches to achieve better results. In Example-based translation, the target sentence was formed from the source sentence using pre-translated examples. This used parallel corpora, which contained sentence pairs with a source sentence and its translation in the required language. Shata-Anuvadak by [Kunchukuttan et al. 2014] was the largest effort at the time and included a collection of phrase-based Statistical Machine Translation (SMT) systems for 110 language pairs. The study used the Indian Language Corpora Initiative (ILCI) parallel corpus of 11 Indian languages, containing roughly 50000 parallel sentences belonging to the health and tourism domains. An objective of the study was to understand the relation between the accuracy of translation and the language family involved. It also aimed to explore whether the shared characteristics of Indian languages can reduce the efforts and resources required for building technology. The provided results called for customized approaches for language family pairs, as the results of translation were better in the case of a few Indic languages, where a higher level of accuracy was seen compared to a few other Indic languages. Similar results were seen in experiments using increased corpus size, and while the translation performance was better in the case of a few languages, the improvement was minimal for other languages with rich morphology. The reordering on the source side helped languages such as Tamil, Telugu, Kannada, and Malayalam a little better than languages like Hindi, whereas the post-editing with transliteration improved the translation quality significantly for the language pairs with scripts derived from the Brahmi script. Incorporation of syntactic and morphological information has been shown to result in significant improvements in phrase-based SMT. For English-Hindi translation, reordering the English source sentence according to Hindi syntax and using a simple suffix separation program for Hindi can be effective [Ramanathan et al. 2008]. In English to Hindi translation, the case markers and suffixes in Hindi are predominantly determined by the combination of suffixes and semantic relations on the English side. Hence, the challenges posed by features of Hindi, such as free-word order and rich morphology for translation, are shown to be alleviated by augmenting the aligned corpus of English-Hindi with the correspondence of English suffixes and semantic relations with Hindi suffixes and case markers [Ramanathan et al. 2009]. Still, translation between language pairs with huge syntactic differences requires re-ordering, which in turn needs an understanding of the word order in the respective language. For English-Hindi translation, research by [Ramanathan et al. 2011] demonstrated that the translation quality can be significantly improved by performing clause-wise translation. [Patel et al. 2017] reported that using rules to transfer the structure of the source sentences before training and translation can lead to a better translation quality. Additionally, they showed that suffix separation can be used to tackle the morphological divergence between English and highly agglutinative Indian languages. Along with the model development efforts, various semantic resources were also created. A few of them are IndoWordNet [Bhattacharyya 2010], Bodo Wordnet [Sarma et al. 2010a], Tamil Wordnet [Rajendran et al. 2002], Kannada Wordnet [Sahoo and Vidyasagar 2003], Sanskrit Wordnet [Kulkarni et al. 2010], Assamese Wordnet [Sarma et al. 2010b], and Punjabi Wordnet [Narang et al. 2013]. Other resources, such as n-grams, were also proposed for Indic languages [Majumder and Mitra 2002].
12
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Multiple approaches for Part-of-Speech tagging were also proposed using techniques such as Hidden Markov Model (HMM) [Dandapat et al. 2004; Pandian and Geetha 2008] and Maximum Entropy Markov Model (MEMM) [Dalal et al. 2006, 2007]. Table 2 lists the corpus-based research works in IndicNLP.
3.3
Indic NLP using Deep Learning
The emergence of Deep Learning (DL) revolutionized Indic NLP by moving the focus from manual feature engineering that was common in the statistical NLP era to automatic representation learning. Characterized by the use of embeddings and attention mechanisms to capture the semantic details that were missed by the previous models, the usage of deep learning models for Indic languages was driven by the need to handle complex morphology, script dissimilarity, and the deficiency of huge parallel corpora for regional language pairs. Also, the dense vector representation used by these models addressed the issue of the formation of the Out-of-Vocabulary (OOV) words in Indic languages, often due to inaccurate sub-word level representation. DL was used for tasks such as Neural Machine Translation (NMT), generating Multilingual representations, and language generation. Summary of the Indic NLP works using deep learning is provided in Table 3. 3.3.1 Embedding Representations and Indic Language Models. Pre-trained embeddings (monolingual and crosslingual) for 14 Indic languages were generated by [Saurav et al. 2020] using both contextual and non-contextual approaches. IndicFT [Kakwani et al. 2020] developed a set of pre-trained word embeddings for 11 Indic languages trained on the massive IndicCorp dataset [Kakwani et al. 2020]. The core of IndicFT’s strategy relies on the FastText model, which utilizes character n-grams (subword information) rather than treating whole words as indivisible atomic units. This method was explicitly chosen because Indian languages are morphologically rich and agglutinative, meaning words often change form with complex suffixes. By breaking words down into smaller character chunks (n-grams) during training, IndicFT could generate embeddings for rare or unseen words by summing the representations of their sub-parts, allowing the model to effectively capture semantic similarities between words that share common roots. To address the need for robust Natural Language Understanding (NLU) across Indic languages, the IndicBERT [Kakwani et al. 2020] model was developed using the compact ALBERT architecture trained on the 8.9 billion token IndicCorp dataset. Distinct from the character n-gram approach used in IndicFT, IndicBERT employed a SentencePiece [Kudo and Richardson 2018] tokenizer with a large shared vocabulary of 200,000 tokens, specifically designed to accommodate the diverse scripts and wide lexical variety of the 12 supported languages (11 Indian languages plus English). Despite having significantly fewer parameters than massive multilingual baselines like multilingual BERT [Devlin et al. 2019] (mBERT) and XLM-R [Conneau et al. 2019] (12M vs. 110M+), IndicBERT demonstrated superior performance on several IndicGLUE [Kakwani et al. 2020] benchmark tasks. Complementing model development, [Karthika et al. 2025] presented a comprehensive intrinsic evaluation of tokenization specifically for 17 Indian languages. This work contrasted the efficacy of Byte Pair Encoding (BPE) and Unigram Language Model (ULM), finding that ULM provides superior morphological alignment, adhering more closely to the linguistic segmentation required for morphologically rich and agglutinative Indic languages. The study further highlighted the critical impact of pre-processing; tokenizers trained on normalized corpora (standardizing Unicode and script-specific characters) consistently yielded lower fertility scores, indicating significantly more efficient segmentation. Additionally, they demonstrated that cluster-based training, grouping languages by typological similarity, significantly reduced the Word Fragmentation Rate (WFR) for underrepresented languages compared to joint training strategies, thereby mitigating the dominance of high-resource languages in the shared vocabulary. The design of multilingual vocabularies presents significant trade-offs. While IndicBERT adopted a 200k vocabulary to ensure broad coverage, [Karthika et al. 2025] observed that increasing vocabulary size from 32k to 256k consistently improves intrinsic metrics such as fertility and fragmentation rate. However, this work
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
Table 2. Research works using the corpus-based NLP techniques for Indic languages
Paper
Language(s)
Approach Parsing
[Dandapat et al. 2004] [Singh et al. 2005] [Dalal et al. 2006] [Dalal et al. 2007] [Gorla et al. 2008] [Pandian and Geetha 2008] [Bharati et al. 2009c] [Bharati et al. 2009b] [V. et al. 2009] [Kumar et al. 2010] [Ramasamy and Žabokrtskỳ 2011]
Bengali Hindi Hindi Hindi Hindi Tamil
Hidden Markov Model Hidden Markov Model Maximum Entropy Markov Model Maximum Entropy Markov Model Normalized Conditional Mutual Information (NCMI) Hidden Markov Model
Hindi Hindi Tamil Sanskrit Tamil
Constraint based hybrid parsing Constraint based Two-stage parsing Conditional Random Fields Automatic segmentation of compounds Dependency parsing Machine Translation (MT) or Transliteration
[Vijayanand et al. 2002] [Ramanathan et al. 2008] [Ramanathan et al. 2009] [Chaudhury et al. 2010] [Ahsan et al. 2010] [Visweswariah et al. 2011] [Ramanathan et al. 2011] [Kunchukuttan and Bhattacharyya 2012] [Sankaran et al. 2012] [Kunchukuttan et al. 2014] [Kunchukuttan et al. 2015] [Patel et al. 2017] [Tholpadi et al. 2017]
Bengali, Assamese
Example-based MT
English, Hindi
Using syntactic and morphological information
English, Hindi
Using case markers and inflections
English, Hindi
Shallow parsing, deep parsing
English, Hindi English, Hindi, Urdu
Coupling rule-based and statistical MT Source side sentence reordering
English, Hindi
Clause-based translation with reordering constraints
Multilingual
Source side phrase reordering
English, Hindi Multilingual
Hierarchical phrase-based MT Phrase-based statistical MT
Multilingual
Parallel transliteration corpora
Multilingual Multilingual
Using pre-ordering and suffix separation Comparable corpora-based translation correspondence induction Semantic Processing
[Reddy et al. 2009] [Srirampur and Chandibhamar 2014]
Hindi Multilingual
Semantic category labeling Statistical morphology analysis
13
14
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
identified a diminishing return in cross-lingual alignment: vocabulary overlap between languages increased up to 128k but decreased at 256k, suggesting that larger vocabularies may begin to include arbitrary, non-shared tokens. Furthermore, the study revealed promising zero-shot transfer capabilities; multilingual tokenizers trained on high-resource languages were able to effectively segment unseen, extremely low-resource dialects like Awadhi, Bhojpuri, and Magahi, achieving fertility scores (e.g., 1.3–1.4) comparable to those of high-resource languages. To address the limitations of massive multilingual models in representing Indian languages, [Khanuja et al. 2021] proposed Multilingual Representations for Indian Languages (MuRIL), which employs a cased WordPiece vocabulary of 197,285 tokens generated specifically from upsampled Indian language corpora. Unlike standard multilingual models that often lowercase input, MuRIL preserved case to retain accent information, which is semantically significant in many Indic languages. The vocabulary generation process explicitly utilized upsampling to ensure that low-resource languages are adequately represented, preventing high-resource languages from dominating the subword inventory. Empirical analysis demonstrated that MuRIL achieves a significantly lower fertility ratio (average subwords per word) across all 17 supported languages compared to mBERT. High fertility ratios, common in mBERT due to its vocabulary being approximately 78% Latin script, often lead to the over-fragmentation of Indic words into meaningless characters, thereby degrading semantic understanding. In contrast, MuRIL allocated a substantially higher proportion of its vocabulary to Indic scripts (e.g., Devanagari, Bengali, Tamil), ensuring that native words are tokenized into meaningful subword units rather than arbitrary character sequences. Furthermore, MuRIL addressed the prevalence of code-mixing and transliteration in the Indian digital landscape by explicitly including transliterated data (native languages written in Latin script) during pre-training. While mBERT’s vocabulary didn’t provide sufficient coverage for transliterated forms, MuRIL’s tokenizer could effectively recognize and segment transliterated terms, contributing to its superior performance on tasks involving informal text and code-switching. Recent experimental evidence underscores the significant performance advantage of language-specific pretraining over cross-lingual transfer learning for complex morphological tasks. [Dasari et al. 2023] provided a comprehensive evaluation of Telugu morphological analysis and demonstrated that a monolingual Transformer model (BERT-Te), trained from scratch on a dedicated corpus of roughly 8 million Telugu sentences, consistently outperformed massive multilingual baselines including mBERT, XLM-R, and IndicBERT. This finding challenges the prevailing assumption that multilingual models are sufficient for low-resource languages, particularly those with agglutinative morphology. The performance gap is most evident in the extraction of fine-grained grammatical features. The study reported that the monolingual BERT-Te achieved an F1 score of 0.778 for Gender tagging, significantly surpassing IndicBERT (0.527) and XLM-R (0.624). Similarly, for the category of Person, the monolingual model achieved an F1 score of 0.704, whereas IndicBERT struggled with a score of 0.475. The authors attributed this disparity to “concentrated language instruction”; while multilingual models dilute their capacity across hundreds of languages, the monolingual architecture captures the intricate inflectional nuances and structural complexities specific to the target language. Consequently, for tasks requiring deep morphological understanding, domain-specific training provides a quantifiable advantage over generalized multilingual representations. 3.3.2 Sequence Modeling Tasks. Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and Transformer architectures were widely used for handling the sequential nature of text. These models performed better at ‘remembering’ long-term dependencies in a sentence compared to the previous modeling approaches. This capability made them better suited for free-word order Indic languages. LSTMs were widely used for POS tagging and Named Entity Recognition (NER) tasks. POS tagging leveraged multilingual Transformer models to overcome data scarcity. Addressing the lack of resources for extremely low-resource languages, [Kumar et al. 2024a] presented the first Universal Dependencies
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
15
(UD)-compliant POS tagging datasets for Angika and Magahi, alongside a new parallel dataset for Bhojpuri. This work highlighted the critical challenge of sub-optimal tokenization in multilingual models like MuRIL and XLM-R, where over-fragmentation of words in low-resource languages leads to poor tagging performance. To mitigate this, this work proposed novel “Look-back” and “Look-back-with-score” techniques, which utilized the tag of the first sub-word token or the token with the maximum logit score to represent the entire word. Experiments demonstrated that Indic-specific models such as MuRIL outperform massive multilingual models in zero-shot settings due to reduced language interference, and that cross-lingual transfer from Hindi is highly effective when alignment errors are minimized. Complementing the development of taggers, recent interpretability studies have investigated how transformer models encode POS and morphological properties. [Aravapalli et al. 2024] introduced INDICSENTEVAL, a benchmark for probing linguistic properties across six Indic languages. Unlike standard tagging tasks, this work utilized POS information (such as subject/object number and verb attributes) as diagnostic probing tasks to assess the internal representations of nine multilingual models. The findings from this study reinforce the superiority of Indic-specific models (e.g., MuRIL, IndicBERT) in capturing semantic POS properties compared to universal models. However, a perturbation analysis revealed a counter-intuitive insight: universal models (e.g., InfoXLM [Chi et al. 2021], mT5 [Xue et al. 2020]) often exhibit greater robustness to input noise, such as the dropping of nouns and verbs, suggesting that while Indic-specific models are more accurate for clean text, they may rely more heavily on specific lexical cues and word order. A significant innovation in recent Transformer-based architectures is the move away from explicit, rule-based stemming towards “implicit stemming" mechanisms embedded within subword tokenization. [Kakwani et al. 2020] demonstrated this with IndicBERT, a multilingual model pre-trained on large-scale corpora spanning 11 Indian languages. Unlike traditional pipelines that require a dedicated pre-processing step to strip suffixes using fixed linguistic rules, IndicBERT utilized the SentencePiece tokenizer to decompose words into statistically frequent subword units (e.g., splitting a fused word into root and affix components). This approach allowed the model to learn morphological representations and root-affix relationships directly from data, effectively bypassing the need for handcrafted stemmers while handling Out-of-Vocabulary (OOV) terms through subword composition.
3.4
Foundation Models
The development of Natural Language Understanding (NLU) for Indic languages has advanced through three distinct phases: initial resource creation, architectural specialization for linguistic nuances, and rigorous, humancentric evaluation. This progression is evident across key NLU tasks such as Named Entity Recognition (NER), Question Answering (QA), Sentence Retrieval, Natural Language Inference (NLI), and Paraphrase Detection. [Kakwani et al. 2020] established the first comprehensive baseline with the IndicNLP Suite. This work introduced IndicBERT v1 (an ALBERT-based model) and the IndicGLUE benchmark. For NER, the model was evaluated on the silver-standard WikiAnn dataset, achieving an F1 score of 64.47, but trailing the mBERT baseline of 73.24 due to limited model capacity. In Question Answering, the suite utilized a cloze-style multiple-choice format, where the model achieved an accuracy of 41.91, demonstrating the feasibility of knowledge extraction in Indic languages. The work also benchmarked cross-lingual sentence retrieval using the CVIT-Mann Ki Baat dataset (Precision@10: 27.12) and introduced coreference-focused inference via the Winograd NLI task (Accuracy: 56.34 on Hindi). Paraphrase detection was evaluated on the Amritha dataset, with the model scoring 93.11 accuracy on Hindi, setting a strong initial precedent for classification tasks. Addressing the limitations of general multilingual models, [Khanuja et al. 2021] introduced MuRIL, a BERTbased model augmented with translation and transliteration data to handle code-mixing and script diversity. This specialization led to state-of-the-art results on the XTREME benchmark. In NER, MuRIL achieved an F1 score
16
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Table 3. Indic NLP Works using Deep Learning
Paper / Initiative
Core Approach
Unique Features
[Saurav et al. 2020] IndicFT [Kakwani et al. 2020]
Generated both contextual and non• Covered 14 Indic languages contextual pre-trained embeddings Utilized the FastText model which relies on character n-grams (subword infor- • Covered 11 Indic languages trained on the IndicCorp dataset mation) rather than whole words • Explicitly handles morphologically rich and agglutinative structures by generating embeddings for rare or unseen words via sub-parts IndicBERT [Kak- Employed a compact ALBERT architecwani et al. 2020] ture using a SentencePiece tokenizer • Supported 12 languages (11 Indian + English) with a large shared vocabulary of • Outperformed massive baselines (mBERT, XLM-R) on IndicGLUE despite having significantly fewer parame200,000 tokens ters (12M vs. 110M+) [Karthika et al. Provided an intrinsic evaluation of tok2025] enization, contrasting Byte Pair Encod- • Evaluated 17 Indian languages ing (BPE) and Unigram Language Model • Demonstrated zero-shot transfer capabilities to segment unseen dialects (Awadhi, Bhojpuri, Magahi) (ULM) MuRIL [Khanuja Employs a cased WordPiece vocabulary et al. 2021] of 197,285 tokens generated from up- • Preserved text case to retain semantically significant accent information sampled Indian language corpora • Explicitly included transliterated data during pretraining to handle code-mixing and code-switching • Achieved a lower fertility ratio compared to mBERT by allocating more space to native Indic scripts BERT-Te [Dasari Trained a language-specific, monolinet al. 2023] gual Transformer model from scratch • Consistently outperformed massive multilingual baselines on Telugu morphological analysis on roughly 8 million Telugu sentences • Achieved significantly higher F1 scores for finegrained grammatical features like Gender and Person tagging via "concentrated language instruction" INDICSENTEVAL Introduced a diagnostic probing bench[Aravapalli et al. mark using POS information to assess • Evaluated 6 Indic languages 2024] the internal representations of 9 multi- • Reinforced the superiority of Indic-specific models in capturing semantic POS properties, while revealing lingual models that universal models (e.g., InfoXLM, mT5) showed greater robustness to input noise
of 77.2, significantly outperforming global baselines like mBERT and XLM-R on structural tasks. For Question Answering, it demonstrated superior reading comprehension on extractive datasets like XQuAD (Hindi F1: 73.9) and TyDiQA (Avg F1: 75.4). The model also excelled in cross-lingual tasks, achieving high accuracy in Natural Language Inference (IndicXNLI: 74.1) and Paraphrase Detection (72.4), leveraging its transliteration-aware pretraining to bridge semantic gaps across languages.
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
17
The most recent phase, led by [Doddapaneni et al. 2023], prioritized evaluation integrity with the IndicXTREME benchmark. This work introduced IndicBERT v2, trained on the massive IndicCorp v2. A key shift was the use of human-verified datasets like Naamapadam for NER, where IndicBERT v2 scored 72.4, beating global baselines but trailing MuRIL. In Question Answering, the model matched the strong MuRIL baseline on the manually curated IndicQA dataset (F1: 47.7) while vastly outperforming XLM-R. IndicBERT v2 showed its strongest gains in Sentence Retrieval, scoring 69.4 on the high-quality FLORES dataset, a 14.5-point improvement over prior models. However, in "strict" semantic tasks like Paraphrase Detection on the new IndicXParaphrase dataset, it scored 56.9, again highlighting that while larger corpora improve general alignment, specific pretraining strategies remain crucial for resolving fine-grained semantic nuances. Bharat Parallel Corpus Collection (BPCC), the largest publicly available parallel corpus for Indic languages, was created by [Gala et al. 2023]. In addition to this, the work also introduced an n-way parallel benchmark covering 22 Indian languages, featuring diverse domains, Indian-origin content, and source-original test sets. IndicTrans2, the first model to support all 22 languages, was also created as a part of this work. The lack of parallel training data, robust benchmarks, and translation models spanning all 22 languages was overcome with this work. 3.4.1 Indic NLP using Generative AI. Recently, several groups have been leading efforts to build generative LLMs that cater to the diverse linguistic landscape of India. A few notable initiatives are Sarvam AI1 , BharatGen2 , and AI4Bharat3 among others. BharatGen has introduced pioneering works, including [DocBodh 2025] - a GenAI suite for Indic Document understanding, [Param 2025] - a text model trained on India-centric data, and can understand and generate human-like text in multiple Indian languages and dialects. Sarvam AI has released powerful models, including [Sarvam-M 2025] - a multilingual, hybrid-reasoning, text-only model built on Mistral-Small4 , following Sarvam-1 - a language model with 2-Billion parameters specifically optimized for Indian languages. Other noteworthy models are [Sarvam Translate 2025] - providing comprehensive coverage of all scheduled Indian languages with formal translation style, and [Mayura 2025] - designed to convert text between English and Indian languages while preserving meaning and context. In addition to this, initiatives such as Bhashini5 are launched by the government with an aim to bridge the language gap and enable access to digital services and content in regional languages. Bhashini provides translation services across more than 36 languages and also hosts a unified hub6 meant for discovering, exploring, and contributing AI models and datasets tailored for India’s diverse linguistic - technological scenario. The development of foundation and instruction-tuned models for Indic languages has accelerated since late 2023. Some of them are Gajendra [AI 2024], Airavata [Gala et al. 2024], Krutrim LLM [Kallappa et al. 2025], Tamil Llama [Balachandran 2023], Bharat GPT [CoRover.ai 2024], and Navarasa [Lab 2024]. Table 4 lists the generative AI works using Indic languages. The training datasets for prominent Indic language models and initiatives often come from a combination of large-scale web scraping, synthetic data generation [Manoj et al. 2025], curated translation pairs [Karthika et al. 2026], and crowdsourced speech [Bhashadaan 2025; Javed et al. 2024]. To the best of our knowledge, the representation of knowledge from indigenous communities is little when compared to that from the conventional sources. Directions towards bridging this gap are discussed in Section 5.
1 https://www.sarvam.ai/ 2 https://bharatgen.com/ 3 https://ai4bharat.iitm.ac.in/ 4 https://mistral.ai/news/mistral-small-3-1 5 https://bhashini.gov.in/ 6 https://bhashini.gov.in/vatika
18
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Table 4. Indic NLP Foundation Models
Model / Initiative
Core Approach
IndicNLP Suite Established the first [Kakwani et al. comprehensive NLU 2020] baseline using IndicBERT v1 and the IndicGLUE benchmark MURIL [Khanuja Specialized BERTet al. 2021] based model augmented with translation and transliteration data BPCC & Indic- Created the Bharat Trans2 [Gala Parallel Corpus Colet al. 2023] lection (BPCC)—the largest public parallel corpus for Indic languages—and developed the IndicTrans2 mode BharatGen Introduced a suite of specialized Indiacentric generative models Sarvam
Bhashini
3.5
Released a series of foundational text and specialized language models
Key Results • Evaluated NER on the silver-standard WIKIANN dataset (F1: 64.47) • Benchmarked tasks like Cloze-style Question Answering, crosslingual sentence retrieval (CVIT-MANN KI BAAT), and Winograd NLI • Achieved state-of-the-art results on the XTREME benchmark • Outperformed global baselines in structural tasks like NER (F1: 77.2) and extractive QA (XQUAD Hindi F1: 73.9; TyDi-QA Avg F1: 75.4) • Introduced an n-way parallel benchmark featuring diverse domains and Indian-origin content • IndicTrans2 became the first translation model to natively support all 22 scheduled Indian languages
• DocBodh [DocBodh 2025]: A GenAI suite for Indic document understanding • Param [Param 2025]: A text model trained on India-centric data capable of understanding and generating regional text and dialects • Sarvam-1 [Sarvam 1 2024]: A 2-Billion parameter language model optimized for Indian languages • Sarvam-M [Sarvam-M 2025]: A multilingual, hybrid-reasoning text model built on Mistral-Small • Sarvam Translate [Sarvam Translate 2025]: Open-weights model for formal translation across 22 scheduled languages • Mayura [Mayura 2025]: Designed to convert text between English and Indian languages while preserving context
Government-led initiative serving as a uni- • Provides translation services across more than 36 languages to bridge the digital regional language gap fied digital ecosystem hub
Speech Recognition for Indic Languages
Research in Automatic Speech Recognition (ASR) for Indic Languages has achieved significant progress, paralleling the works in NLP with large foundational models like Whisper [Radford et al. 2022]. ASR faces similar challenges
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
19
Table 5. Key Projects and Models in Indic ASR
Project / Model Name
Description
IndicWav2Vec [Javed et al. 2021]
Using Wav2Vec-like model architecture along with unlabeled data to pre-train the model. Achieved state-of-the-art results on 9 Indic languages, including on very low-resource Indic languages such as Sinhala and Nepali An initiative focused on building an end-to-end toolkit for ASR using speech data in 23 Indic languages. Contributed ASR models in 18 Indic Languages, along with post-processing utilities such as Punctuation restoration and Inverse Text Normalization A large-scale dataset mined from public archives (e.g., All India Radio), containing over 6,400 hours of labeled audio across 12 Indian languages Kathbath, a large scale ASR dataset for 12 Indian languages used to create a benchmark for ASR, language identification, speaker identification, and speaker verification across the Indic language family A dataset designed to capture diverse linguistic variations, accents, and code-switching, essential for dialectal ASR evaluation. Contains 2,043 hours of transcribed audio spanning 105 languages from 28 states and 3 union territories, many of these languages are being represented at this scale for the first time Streaming speech recognition supporting 22 Indian languages, featuring low-latency decoding and support for code-mixed speech
Vakyansh [Chadha et al. 2022]
Shrutilipi [Bhogale et al. 2022] IndicSUPERB [Javed et al. 2022]
VAANI [Pulikodan et al. 2026]
Saaras (V3) [Saaras V3 2026]
to the ones explained in Section 4 in addition to code mixing [Thara and Poornachandran 2018], multilingual speech, and noisy recording. Table 5 lists the recent works in ASR for Indic languages in brief. A critical bottleneck identified in recent studies for Indic ASR is the tokenization process, which disproportionately impacts low-resource languages compared to others. While high-resource languages benefit from extensive token sets in pre-trained models, low-resource languages suffer from limited vocabulary coverage. This deficiency forces the tokenizer to fragment words into numerous small subwords or characters, leading to increased sequence lengths and substantially slower auto-regressive inference speeds. To address these inefficiencies, [Tripathi et al. 2025] proposed a novel tokenization strategy that augments the existing Whisper vocabulary with language-specific Byte Pair Encoding (BPE) tokens. These tokens were derived from datasets representative of common Indian language sequences to reduce token fertility. The pre-trained model architecture was modified by expanding the dimension of the final token head layer, initializing new random weights for the added tokens while freezing the original weights to preserve multilingual capabilities. Through ablation studies, an optimal threshold of 250 additional tokens per language was identified. This targeted vocabulary expansion significantly compressed token sequences; for example, token counts for Hindi and Malayalam sentences were reduced by approximately 30% and 60%, respectively. Empirical evaluations demonstrated that this reduction not only accelerated inference speeds outperforming optimized baselines but also consistently improved the Word Error Rate (WER) across model variants. While these Indic ASR models are increasingly capable of addressing cultural nuances, they are far from being robust. Substantial progress is still required to bridge the gap between the performance in mainstream languages and the real linguistic diversity. The focus of future research must be on curation of representative
20
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
datasets and design of models that inherently capture the regional variations, rather than fixing them into rigid, monolingual frameworks. A low-cost and easy-to-implement approach for ASR using colloquial speech is discussed in Section 5.2.1.
4
Challenges for Indic NLP
While Indic NLP research has made significant progress in recent times, persistent challenges pose challenges for building representative and inclusive language models for Indic languages. These challenges exist in the case of a majority of Indic languages and also other related languages, and are not observed in languages like English, which are used to build mainstream NLP systems. In this section, we discuss these challenges in detail and also discuss relevant research works that focused on finding solutions using specific cues such as the structure of Indic languages, the difference between dialects, etc. Challenges for Indic NLP
Complex Morphology
Grammar Rules
Diglossia
Lack of Resources
Fig. 6. Main Challenges for Research in Indic NLP
4.1
Complex Morphology and Grammar Rules
Some of the distinct features of morphology and grammar in the case of Indic languages, such as the kaaraka rules along with the free word order, pose significant challenges when it comes to building language models for these languages. While there is a growing social and commercial need for the availability of language technologies in Indic languages, the scale and diversity of these languages make the design of inclusive language models difficult. Additionally, a few of the language features also need elaborate modeling. For example, compound verbs are verb-verb collocations where the polar verb adds the semantics and the vector verb acts as a modifier [Slade 2016]. Processing a sentence with such verbs requires a computational model to acquire both morphological and pragmatic understanding. Building language models that can understand and represent such fine nuances with precision is a complex task. [Bhattacharyya et al. 2019] discussed a list of challenges for Indic language computing and suggested some pointers for overcoming them. A few of the proposed solutions are a unified representation of Indic languages using a superset of sounds, interactive crowd sourcing for resource creation, and generic acoustic models and generic language models across various languages. In this section, we discuss the related studies that focus on using NLP techniques to model the linguistic features of Indic languages. Indic languages are highly inflectional, and one root word can have multiple morphological variants. The agglutinative nature also makes morpheme boundaries hard to distinguish. Even though a set of rules can be used to identify the stem of a given word, this process requires linguistic expertise, is time-consuming, and languagespecific. Researchers have explored the application of corpus-based techniques for morphological processing. [Kulkarni and Srinivasa 2013] proposed an improved indexing technique for identifying grammatically modified versions of the same root word. A fuzzy matching-based indexing mechanism inspired by prefix trees was used to index and retrieve different morphological forms of a term. This solution can be adapted for other similar languages with minimum modifications due to the similarities in their Unicode encoding and morphological behaviors. Another study by [Bhat 2012] found unsupervised morphological segmentation algorithms to perform well for the problem of morpheme boundary detection in Kannada. [Ramanathan and Rao 2003] proposed a domain-independent and computationally inexpensive stemmer for the Hindi language. A suffix list was developed and used in this work. In the case of the Bengali language, [Majumder et al. 2007] explored the clustering-based
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
21
approach, where equivalence classes of root words and their morphological variants were identified from a lexicon using string distance measures. Stemming performance using the approach provided results comparable to the standard rule-based stemmers and was observed to be effective for languages that are primarily suffixing. Unsupervised learning for acquisition of morphological knowledge from a text corpus was experimented with by [Sharma et al. 2008] for the Assamese language. This approach was used to build a lexicon and subsequently used for morphological analysis. A formulation for derivational morphology, augmented Finite State Automata, was proposed along with a formalism for a lexicon writer for specifying a lexicon by [Sengupta and Chaudhuri 1996]. This approach was also proposed to be useful as a building block for a spelling corrector in the case of the Bangla language. A finite state machine was used to develop a prototype morphological analyzer for the Kannada language [Vikram and Urs 2007]. Efficient morphological processing along with syntactic cues significantly improved the overall performance when incorporated into tasks such as SMT [Ramanathan et al. 2008] and data-driven dependency parsing [Ambati et al. 2010; Malladi and Mannem 2013]. Another study found that augmenting the phrase table with all possible forms of a verb can improve the overall accuracy of a phrase-based MT system in the case of highly inflected languages [Gandhe et al. 2011]. Morphological analysis has also been integrated with the task of tagging in multiple studies. Using a stemmer along with a morphological analyzer, [Shrivastava et al. 2005] performed Part-of-Speech (POS) tagging in the Hindi language. [Shrivastava and Bhattacharyya 2008] showed that an HMM-based POS tagger was able to achieve reasonably good accuracy when stemming was performed as preprocessing. To overcome the lack of standardized tagsets, [Sankaran et al. 2008] proposed a common POS tagset hierarchical framework for eight Indic languages.
4.2
Diglossia
Most of the Indic NLP applications focus on the formal language varieties. Along with this, the corpora used for training of Indic LLMs are dialect-neutral. Studies are being conducted to model the relation between the formal variety of a language and its dialects. [Mishra and Bali 2010] listed rules for phonology transfer from standard Hindi to a number of its prominent dialects, Bundeli, Bagheli, Kanauji and Awadhi. For Punjabi language and its dialects, Malwai and Doabi, a conversion system was proposed by [Singh and Singh 2015] using bilingual dictionaries and morphological conversion rules. In the case of the Tamil language, an approach for conversion from spoken dialectal language to standard written Tamil utilized Finite State Transducer (FST) [Marimuthu and Devi 2014]. A framework for capturing synchronic variation by modeling the underlying diachronic variation was proposed by [Choudhury et al. 2006]. This work used an ordered set of rewrite rules representing phonological changes that occurred during the evolution of dialects to derive current dialectal forms of the Bangla language from their classical counterparts. Corpora collected in informal settings with unscripted, impromptu conversations can reveal the vibrant nature of languages and also reveal their characteristic patterns [Dash and Chaudhuri 2001]. Human in the loop approach can be helpful in collating dialect-rich corpus from informal sources as shown by [Srivatsa et al. 2024]. This work utilized a small speech corpus collected from a rural community radio to build a colloquial text corpus. With the widespread availability of language models, there is a growing need to build dialect-rich corpora that can be used in turn to incorporate diverse knowledge into the language models.
4.3
Lack of Resources
The lack of resources in Indic NLP refers to the multifaceted scarcity that spans data, computation, and linguistic expertise. While the Indian subcontinent is data-rich in terms of human interactions and knowledge systems, the digital footprint is disproportionately small compared to English. LLMs primarily rely on large amounts of high-quality data along with rich representations in the form of word embeddings. These word embeddings
22
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
capture the contextual semantics of words and tokens, and are learned from huge text corpora with millions of words. However, the availability of such huge datasets in the case of Indic languages is limited, and this presents a challenge for the development of Indic language technologies. The largest publicly available Indic language datasets are of the order of billions of tokens7 , whereas it is much higher for languages like English. Also, Indic languages have adopted multiple different encoding standards, such as ISCII [Indian Script Code for Information Interchange 2009], Unicode [UNICODE STANDARD For Indic Scripts [n. d.]], and font standards like ISFOC [ISFOC Standard for Fonts [n. d.]]. These multiple standards have different software dependencies, and hence, uniform adaptation becomes difficult. In addition to this, there is no standard for transliteration among Indic languages and from Indic languages to other languages. Studies have proposed multiple approaches for Indic language transliteration [Diwakar et al. 2010; Joshi et al. 2018; Kunchukuttan et al. 2015; Srivastava and Bhat 2013; Surana and Singh 2008]. However, there can still be variation in the representation of words. For example, the word for ’Thank you’ in Hindi can be written as dhanyawad, dhanywad, dhanyavad, or dhanyvad. Large, labeled corpora are critically needed for Indic languages, and active research efforts are underway to address this gap. Enabling Minority Language Engineering (EMILLE) [McEnery et al. 2000], Linguistic Data Consortium for Indian Languages (LDC-IL) [Choudhary 2021] are a few of the initiatives aimed at creating a corpus in Indic languages. Studies have also experimented with various approaches to overcome the lack of data. For the task of translation, [Philip et al. 2021] used iterative alignment to provide a large-scale sentence-aligned dataset using publicly available websites such as press releases by the government. For translation between Assamese and other Indic languages, a comparison between statistical and neural MT was made by [Baruah et al. 2021] to improve the translation performance in the case of the Assamese language. [Kumar et al. 2024b] contributed a Universal Dependency compliant parallel dataset for POS tagging in three Indic languages- Angika, Magahi, and Bhojpuri. Submissions to the low-resource Indic language translation task as part of the Conference on Machine Translation (WMT) [Pakray et al. 2025] revealed important insights. A few of them are(1) a strong correlation exists between translation performance and the data size (2) the effect of data quality and model optimization can outbalance the data quantity (3) asymmetry in translation quality from one direction to the other (4) models that perform well on one metric also perform well on other metrics of translation quality The shared linguistic features of Indic languages have also been useful in building language resources for low-resource languages. Studies have utilized resources of a related language with more resources to overcome the limitations for low-resource languages [Singh et al. 2008]. The similarity in scripts and sentence structure is shown to be effective in building language models for low-resource languages [Khemchandani et al. 2021], rather than direct training or using English as the pivot language. Even for translation, it is found that using a parent multilingual neural MT model and fine-tuning it on a low-resource language pair of interest performs better than standard neural MT [Goyal et al. 2020]. Apart from the data quantity standpoint, the lack of specialized resources for Indic languages leads to technical inefficiencies. The issue of subword fragmentation, where a standard tokenizer breaks a single word into multiple tiny, meaningless fragments, is very much relevant in the case of Indic languages. This causes an increase in computation cost and latency. Hence, an English sentence with ‘n’ tokens might have significantly more than fragments in the case of the Indic languages, also making it slower, more expensive, and meaningless. An approach was proposed by [Brahma et al. 2025], incorporating sandhi splitting to enhance the subword tokenization. It was shown that handling dependent vowels by forming a cohesive unit with other characters instead of occurring as a single unit leads to a reduction in fertility scores while maintaining performance in the language modeling task. 7 https://huggingface.co/datasets?language=language:hi&sort=largest
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
23
Techniques such as cross-lingual transfer [Pawar et al. 2023] and back-translation [Das et al. 2025] have been adapted to mitigate the data scarcity issue. Crowd-sourcing initiatives like Bhashadaan8 are working on manually collecting data to train and to empower the linguistic diversity in Indic languages.
5
Language and Culture
Sections 3 and 4 outlined the historical evolution and predominant challenges of Indic NLP, translating these insights into pluralistic AI requires new conceptual frameworks. To nudge the field to move forward, we propose Culture Sensing - the process of understanding the worldviews and hermeneutics of a population to conduct meaningful communication with them, and also to enrich our own understanding of the world. Culture Sensing seeks to gather knowledge from various native discourses, regardless of the medium used, such as speech, text, and more. Unscripted, spontaneous data curation techniques are preferred to emphasize authenticity. The focus is on preserving the numerous credible hermeneutic schools and their associated methods of inquiry, as well as enhancing foundation models accordingly. In this section, we discuss the relation between language and culture, and how the inherent worldview in foundation models reflects the underlying culture. We also provide a demonstration of the Culture Sensing approach reflected in two different applications. Language and Culture have a synergistic relation. Language acts as the primary vehicle through which cultural values, histories, and social systems are conserved and communicated. This relationship is complex in the context of the Indian subcontinent due to the immense linguistic diversity and vibrant cultural heritage. Language mirrors the intricate social structures and value systems in this region. For example, Hindi words Tu, Tum, and Aap all mean ‘you’, but refer to distinct levels of formality or relationship. The cultural emphasis on family structures is also clearly visible in the language usage. For example, a majority of the Indic languages have distinguishing names for paternal uncle and maternal uncle. A few of the concepts precisely elaborated in Indic languages, such as Dharma and Karma carry inherent philosophical ideas and are untranslatable into non-Indic languages without significant loss of meaning. The powerful connection between culture and language has also influenced political boundaries (Karnataka state for Kannada language speakers, Tamil Nadu for Tamil language speakers), and different linguistic groups share a unique cultural legacy and literary tradition. A majority of the subcontinent’s history, in the form of folklore, oral epics, and proverbs, is preserved through intergenerational transmission. These oral traditions are prone to being erased when the corresponding dialect ceases to exist. In recent times, English serves as a lingua franca and has been intertwined such that ‘Indian English’ incorporates the local syntax and loanwords. The common practice of code-switching (Hinglish, Kanglish) represents a modern hybrid identity that combines traditional values with global influences. It can be said that a language expresses the collective beliefs, perceptions, and value systems held by the members of the same social group, which constitute the culture of that society. With the rapid adaptation of LLMs to different language processing tasks, it is essential to inquire into the underlying cultural representation. Foundation Models have demonstrated exceptional performance in language generation and reasoning, openended question answering, and multimodal understanding. While LLMs are applied across a broad range of verticals, it is necessary to understand the extensive societal consequences they entail. The paradigm shift in AI brought about by LLMs also comes with intrinsic biases such as misrepresentation, underrepresentation, and overrepresentation, and extrinsic harms that can be representational harms, abuse, and performance disparities [Bommasani et al. 2022b]. The downstream applications created by fine-tuning the foundation models have often displayed performance disparity in the case of minority social groups [Blodgett and O’Connor 2017; Koenecke et al. 2020]. Linguistic homogenization can be clearly identified in LLM-assisted writing. AI-enabled writing assistants have been linked to reduced linguistic diversity and alteration of individual stylistic elements while amplifying 8 https://bhashadaan.bhashini.co.in/bhashadaan/en/home
24
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
dominant characteristics or biases [Sourati et al. 2025]. Repeated usage of such assistants can result in normalized grammar, simplification of dialects, and erosion of sociolinguistic markers in writing. LLMs have also been shown to erase culturally specific linguistic markers while retaining the semantic meaning, especially in professional writing [Kumar Navneet et al. 2026]. LLMs also often demonstrate cultural homogenization, where they internalize dominant cultural assumptions such as the prioritization of Western values, usage of internet slang, and globally dominant political views. A cross-cultural study [Agarwal et al. 2025] with participants from India and the United States showed that AI models such as writing assistants homogenize writing towards Western norms, decreasing subtle cultural differences. This study showed that among the two groups, Indians were more likely to write like Americans while using AI assistants, but they needed to put more effort than Americans to gain a similar productivity boost. Here, AI doesn’t just change the written content but also the more internalized elements of cultural expression. Language models are also prone to follow similar writing styles and similar ‘neutral’ tones, and can underperform [Marco et al. 2024] when used for creative writing tasks. Contrasted with the rich oral storytelling traditions with non-linear, layered components, LLMs follow a similar structure almost all the time. It can be said that homogenization is circular; lack of diversity in model training leads to lack of diverse outputs, which in turn guides the user expectations and standards. The lack of cultural, epistemic, and linguistic inclusivity in LLMs is alarming. While these models are being used in different domains, including education, technology, healthcare, etc., the inherent narrative or worldview in these LMs should be understood in-depth. Especially for the Indian subcontinent, with thousands of language dialects and grouped multilingual identities, it is essential to represent the diversity in the AI ecosystem. In this section, we discuss how the language used to train the LLMs can be used to understand the embedded cultural cues.
5.1
Worldview of LLMs
A majority of the LLMs are trained using data extracted from conventional sources, including Wikipedia, news websites, and other publicly available datasets. For example, mBERT is trained on Wikipedia articles of the top 100 languages with the largest Wikipedia.9 These datasets sourced from such mainstream platforms are widely available as labeled gold standard corpora and are frequently used as benchmark datasets. Otherwise, extracting good-quality training corpora followed by relevant processing and labeling for the corresponding downstream task or domain is an expensive task. It requires sufficient time, computational resources, and annotators with linguistic knowledge for hand-labeling the dataset. As a result of this, a majority of the training and fine-tuning experiments are dependent on a very few public datasets that were built either by crowdsourcing or by specific vendors. This is likely to lead to standardization of algorithms due to more and more language models being trained on data from similar sources, finally leading to homogenization of the model outputs [Creel and Hellman 2022]. Language models that are trained on snapshots of the internet are more likely to mimic the views of crowd workers who were involved in the creation of data. These LMs have the least representation for a few demographic groups (for example, senior citizens, people without a college degree) that were excluded from the crowd sourcing process [Santurkar et al. 2023]. A likely outcome of this can be that a social group with minimal representation in the dataset might be isolated in the language model system. The LLMs fine-tuned with human feedback are often seen demonstrating a worldview aligned to a particular group of people in response to subjective queries. A response from these models contains an embedded narrative, and it can be of great influence when used in information retrieval and decision-making tasks. For example, when it comes to political ideology, Chat GPT has expressed left-libertarian aligned views [Ceron et al. 2024; Hartmann et al. 2023; Motoki et al. 2023]. With a sufficiently large corpus of social media interactions of a political group, 9 https://github.com/google-research/bert/blob/master/multilingual.md#data-source-and-sampling
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
25
GPT-style models can even be fine-tuned to inquire worldviews of the particular community [Jiang et al. 2022]. In addition to this, the ideological stance of LLMs reflects the deep-rooted worldview of their creators. They are also influenced by the geopolitical region of their creation and the language they were created in [Buyl et al. 2025]. With their increasing use in Information Retrieval (IR), it is essential to consider the origin of these models along with the worldview they reflect. A widely used approach for building language models for low-resource languages involves fine-tuning of multilingual foundation models. Even though this approach seems to help in overcoming the resource scarcity issue, these language models internally ‘think’ in English while processing other languages. A research work [Wendler et al. 2024] showed that English is used as the internal pivot language in multilingual models trained on unbalanced, English-dominated corpora. This work demonstrated an abstract ‘concept space’ in the intermediate layers of the transformer model architecture, where a proximity of the embeddings to English tokens rather than the input language was observed. This finding instills English as a pivot language in a semantic sense rather than a lexical one. This kind of internal representation bias inside transformer models poses difficulty for Indic NLP by revealing the hidden English dominance. The Anglocentric bias could incline the model to certain linguistic elements, such as lexicon and grammar. Efforts such as Indic-TunedLens [Panchal et al. 2026] have proposed a framework for interpretability for Indic languages by learning shared affine transformations. LLMs play a key role in documenting and disseminating diverse knowledge systems. They can reflect the embedded value systems and worldviews of different cultures. Due to this capability, they can be used to build local dialect datasets and community-focused language models, to preserve diverse scripts, and also for educational purposes. The current necessity is ‘culture sensing’ where LLMs not only process or generate data, but also retain the writing style, storytelling pattern, and the innate worldview.
5.2
Culture Sensing
AI foundation models can operate using multiple languages. But they do not contain the lived experience of the language speakers that reflects the correlated cultural meaning. The interplay between language and culture has long been studied to understand how the perceptions, beliefs, and values of people are encoded linguistically [Kramsch 2014]. Integration of AI into indigenous knowledge frameworks coming from communities needs the negotiation of needs and motivations, and establishment of relevant meanings [Lewis et al. 2024]. It is the need of the hour to utilize the enormous capacity of AI towards understanding and preserving the cultural and linguistic diversity. Underrepresentation of the Indic languages and the diverse cultural traditions is an understudied topic. Culture Sensing aims to amend the current-day foundation models based on hermeneutic reasoning. Culture Sensing approaches narrative diversity in AI using the abundant knowledge in indigenous communities that can be available mostly in the form of speech. In this section, we briefly discuss our proposed approach for culture sensing along with a few future directions. 5.2.1 Culture Sensing for Oral Community Knowledge. Oral data and storytelling are characteristic knowledgesharing practices in indigenous, close-knit communities. A thorough analysis of such oral knowledge can uncover the fundamental worldview and the value system of the community in focus. Indigenous worldviews provide a holistic outlook and emphasize the relationship between humankind and nature as symbiotic. On the contrary, the mainstream or conventional worldview stems from the modern industrial advancements and is reductionist, individualistic, and places humankind in a predominant position compared to nature. A majority of the formal literature encapsulates the conventional worldview, while the native worldview is mostly present in the lived experience of the community and is context-dependent. The stark contrast between these two worldviews has resulted in the isolation of the indigenous knowledge in the mainstream systems, including AI applications.
26
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Fig. 7. Reference Architecture for Culture Sensing
The native communities are at a disadvantage due to their limited digital presence, ultimately leading to the marginalization of their pluralistic discourse. Bringing AI into the picture can lead to beneficial outcomes due to the trainability of the AI models. [Aparna et al. 2024] proposed an architecture using NLP, ASR, and Retrieval Augmented Generation (RAG) to address the problem of preserving and disseminating oral community knowledge. This architecture also aims to reveal the underlying native worldviews and also understand how they deviate from the mainstream worldview [Srivatsa et al. 2026]. Here, an ASR pipeline is designed to convert a speech corpus to a searchable text corpus. The speech corpus is made of audio files collected from rural communities in the state of Karnataka, India. The audio files contain natural speech from the rural community members. Following this, the text corpus is utilized in multiple NLP applications such as keyword search and NER. This work also compares the community worldview with the mainstream worldview using corpora extracted from the corresponding groups. This is done using semantic analysis of embedding neighborhoods and RAG. A general approach for Culture Sensing is shown in Figure 7. This pipeline can be adapted for different use cases. As part of this work, we have developed two applications: Graama Kannada [Srivatsa et al. 2024], and Parichaya [Srivatsa et al. 2025]. The Graama Kannada application supports audio search and uses an audio corpus collected from the Namma Halli Radio10 organization. This application aims to tackle two challenges: handling noisy, dialectal, and small speech corpora and maintaining low application development cost. Using a fuzzy search technique of n-gram texts, this application performs keyword search efficiently in the case of the low-resource language Kannada. The ASR model used in this work demonstrates a comparable WER to output the audio transcripts that support the keyword search functionality. 10 https://blog.janastu.org/covid-19-campaign-namma-halli-radio/
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
27
Parichaya is an application for the management of a rural colloquial audio corpus on the topic of sandalwood cultivation. It contains two interfaces to enable users to interact with the audio content. The first interface enables browsing of the content using frequently occurring keywords and the corresponding context words, supporting a Keywords in Context (KWIC) analysis. This helps to gather an understanding of the unique aspects of information in the speech corpus. It also provides a surface-level interpretation of the corpus content. The second is a question-answering interface that provides an answer relevant to the input question along with a summary. The users can also listen to the audio fragments that contain the answers to the question. Both of these applications provides access to lesser-known corpora and enable discourse analysis to understand the crucial insights. Unlike the mainstream knowledge sources that exhibit the conventional worldview, these applications disclose some alternate narratives and diverging worldviews. Table 6. Future Directions for Culture Sensing
Data
• Collect data from unconventional sources such as community radios, local community media centers, and public broadcasting • Ensure representation for multiple dialects rather than a ‘uniform’ language style • Prioritize unscripted, spontaneous, ethical, and natural data collection techniques
Model
• Adapt existing foundation models to native language datasets by fine-tuning • To counteract diglossia, extend language modeling to dialects of a language using corresponding corpora • Utilize post-training techniques such as Reinforcement Learning with Human Feedback (RLHF), and inference techniques such as RAG to understand the embedded worldview • Employ human-in-the-loop style verification for downstream applications
Users
• Incentivize knowledge digitization by rural community members using easily usable platforms such as community radios • More community-driven resource creation efforts • Ethical data collection efforts to promote respect the cultural distinction of closeknit, sensitive communities
These studies clearly showcase that the well-established, homogeneous worldview embedded in most of the datasets that are used for training the LLMs differs substantially from the numerous unique, underrepresented native knowledge systems. AI can be used to preserve such indigenous worldviews by representing the plural, heterogeneous discourses. This early study demonstrates the lack of hermeneutic diversity in foundation models. Further work in this direction can help strengthen both the cultural and linguistic representation in LLMs and the underlying narrative diversity. A few suggestions in this regard, positioned across three verticals: data, model, and users, are listed in Table 6.
6
Conclusions
The rapid growth of the Web and AI has brought out numerous applications. Especially in areas such as NLP, different paradigms have been proposed that have demonstrated exemplary performance. Deep learning has further incentivized the language modeling task by introducing foundation models with state-of-the-art performance. These developments have become both boon and bane in the case of Indic languages. On one hand, Indic NLP is becoming more efficient with time, and more LLMs for Indic languages have been coming up.
28
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
However, this rapid growth has also resulted in the homogenization of hermeneutics, linguistic biases, and the risk of loss of representation of multiple worldviews. In this paper, we present an overview of Natural Language Processing (NLP) in the context of Indic languages. We discuss the unique characteristics of these languages and how they differ from English. Additionally, we address the challenges that these distinctive features present when applying different NLP techniques to model Indic languages. Finally, we provide a comprehensive overview of the historical development of Indic NLP, highlighting the evolution of NLP techniques over time. Ultimately, we propose Culture Sensing to enable foundation models to represent diverse and heterogeneous worldviews. Culture Sensing aims to integrate the vast amount of native knowledge available in various forms, including speech and text, into the framework of AI. This initiative is focused on ensuring that the pluralistic and vibrant worldviews are included. It addresses the urgent need for inclusivity in today’s AI models. By providing a generic research direction in this area, we hope to inspire future projects that include diverse narratives from different underrepresented regions.
Ethics and Privacy Statement In developing the Culture Sensing framework, we have prioritized the preservation of linguistic and hermeneutic diversity and the mitigation of algorithmic biases that would otherwise homogenize worldviews or misrepresent the lesser-known Indic languages. This work recognizes that linguistic and cultural data are important components and require high levels of stewardship. We emphasize that all data mentioned in this work are derived from publicly available, ethically sourced corpora, and this work respects the cultural nuances and intellectual authority of the represented communities. We acknowledge the potential misrepresentation of cultural narratives associated with automatic analysis and recommend human-in-the-loop validation for all downstream use cases. Generative AI Usage Acknowledgment. During the preparation of this work, the authors used Generative AI tools to assist with language refinement, clear phrasing, and image editing. All content generated by these tools was thoroughly reviewed, edited, and validated by the authors.
References Dhruv Agarwal, Mor Naaman, and Aditya Vashistha. 2025. AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1117, 21 pages. doi:10.1145/3706598.3713564 Dhruv Agarwal, Anya Shukla, Sunayana Sitaram, and Aditya Vashistha. 2026. Fluent but Foreign: Even Regional LLMs Lack Cultural Alignment. arXiv:2505.21548 [cs.CL] https://arxiv.org/abs/2505.21548 Arafat Ahsan, Prasanth Kolachina, Sudheer Kolachina, Dipti Misra, and Rajeev Sangal. 2010. Coupling Statistical Machine Translation with Rule-based Transfer and Generation. In Proceedings of the 9th Conference of the Association for Machine Translation in the Americas: Research Papers. Association for Machine Translation in the Americas, Denver, Colorado, USA. https://aclanthology.org/2010.amta-papers.6/ Bhabha AI. 2024. Gajendra: 7B Hindi-Hinglish-English instruction finetuned model. Retrieved May 13, 2026 from https://huggingface.co/ BhabhaAI/Gajendra-v0.1 AIForAll 2018. National Strategy for Artificial Intelligence. Accessed at https://www.niti.gov.in/sites/default/files/2023-03/National-Strategyfor-Artificial-Intelligence.pdf. [Online; accessed 01-May-2026]. Bharat Ram Ambati, Samar Husain, Joakim Nivre, and Rajeev Sangal. 2010. On the role of morphosyntactic features in Hindi dependency parsing. In Proceedings of the NAACL HLT 2010 First Workshop on Statistical Parsing of Morphologically-Rich Languages. 94–102. M. Aparna, Sharath Srivatsa, G. Sai Madhavan, T. B. Dinesh, and Srinath Srinivasa. 2024. AI-Based Assistance for Management of Oral Community Knowledge in Low-Resource and Colloquial Kannada Language. In Big Data Analytics in Astronomy, Science, and Engineering, Shelly Sachdeva and Yutaka Watanobe (Eds.). Springer Nature Switzerland, Cham, 3–16. Akhilesh Aravapalli, Mounika Marreddy, Radhika Mamidi, Manish Gupta, and Subba Reddy Oota. 2024. IndicSentEval: How Effectively do Multilingual Transformer Models encode Linguistic Properties for Indic Languages? arXiv preprint arXiv:2410.02611 (2024). Abhinand Balachandran. 2023. Tamil-Llama: A New Tamil Language Model Based on Llama 2. arXiv:2311.05845 [cs.CL] https://arxiv.org/ abs/2311.05845
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
29
Rupjyoti Baruah, Rajesh Kumar Mundotiya, and Anil Kumar Singh. 2021. Low resource neural machine translation: Assamese to/from other indo-aryan (indic) languages. Transactions on Asian and Low-Resource Language Information Processing 21, 1 (2021), 1–32. Rafiya Begum, Samar Husain, Arun Dhwaj, Dipti Misra Sharma, Lakshmi Bai, and Rajeev Sangal. 2008. Dependency Annotation Scheme for Indian Languages. In Proceedings of the Third International Joint Conference on Natural Language Processing: Volume-II. https: //aclanthology.org/I08-2099/ Kirti Bhagat, Shaily Bhatt, Athul Velagapudi, Aditya Vashistha, Shachi Dave, and Danish Pruthi. 2026. TALES: A Taxonomy and Analysis of Cultural Representations in LLM-generated Stories. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). Association for Computing Machinery, New York, NY, USA, Article 984, 26 pages. doi:10.1145/3772318.3790519 Akshar Bharati, Vineet Chaitanya, Amba P. Kulkarni, and Rajeev Sangal. 2003a. Language Access: An Information Based Approach. arXiv:cs/0308019 [cs.CL] https://arxiv.org/abs/cs/0308019 Akshar Bharati, Vineet Chaitanya, Amba P. Kulkarni, Rajeev Sangal, and G Umamaheshwara Rao. 2003b. Anusaaraka: Overcoming the Language Barrier in India. arXiv:cs/0308018 [cs.CL] https://arxiv.org/abs/cs/0308018 Akshar Bharati, Vineet Chaitanya, and Rajeev Sangal. 1991. Local word grouping and its relevance to Indian languages. Frontiers in Knowledge Based Computing (KBCS90), VP Bhatkar and KM Rege (eds.), Narosa Publishing House, New Delhi (1991), 277–296. Akshar Bharati, Vineet Chaitanya, Rajeev Sangal, and Brendan Gillon. 2002a. Natural Language Processing: A Paninian Perspective. (07 2002). Akshar Bharati, Mridul Gupta, Vineet Yadav, Karthik Gali, and Dipti Misra Sharma. 2009a. Simple parser for Indian languages in a dependency framework. In Proceedings of the Third Linguistic Annotation Workshop (LAW III). 162–165. Akshar Bharati, Samar Husain, Bharat Ambati, Sambhav Jain, Dipti Sharma, and Rajeev Sangal. 2008. Two semantic features make all the difference in parsing accuracy. Proc. of ICON 8 (2008). Akshar Bharati, Samar Husain, Dipti Misra Sharma, and Rajeev Sangal. 2009b. Two stage constraint based hybrid approach to free word order language dependency parsing. In Proceedings of the 11th International Conference on Parsing Technologies (IWPT’09). 77–80. Akshar Bharati, Samar Husain, Meher Vijay, Kalyan Deepak, Dipti Misra Sharma, and Rajeev Sangal. 2009c. Constraint based hybrid approach to parsing indian languages. In Proceedings of the 23rd Pacific Asia Conference on Language, Information and Computation. Waseda University, 614–621. Akshar Bharati and Rajeev Sangal. 1990. A Karaka Based Approach to Parsing of Indian Languages. In COLING 1990 Volume 3: Papers presented to the 13th International Conference on Computational Linguistics. https://aclanthology.org/C90-3005/ Akshar Bharati and Rajeev Sangal. 1993. Parsing free word order languages in the Paninian framework. In Proceedings of the 31st Annual Meeting on Association for Computational Linguistics (Columbus, Ohio) (ACL ’93). Association for Computational Linguistics, USA, 105–111. doi:10.3115/981574.981589 Akshar Bharati, Rajeev Sangal, Vineet Chaitanya, Amba Kulkarni, Dipti Misra Sharma, and KV Ramakrishnamacharyulu. 2002c. Anncorra: building tree-banks in indian languages. In COLING-02: The 3rd Workshop on Asian Language Resources and International Standardization. Akshar Bharati, Rajeev Sangal, and T Papi Reddy. 2002b. A constraint based parser using integer programming. Proc. of ICON 274 (2002). Akshar Bharati, Dipti Sharma, Vineet Chaitanya, Amba Kulkarni, Rajeev Sangal, and Durgesh Rao. 2003c. LERIL: Collaborative Effort for Creating Lexical Resources. CoRR cs.CL/0308020 (08 2003). doi:10.48550/arXiv.cs/0308020 Bhashadaan 2025. Bhashadaan: One platform, Multiple crowdsourcing initiatives. Accessed at https://bhashadaan.bhashini.co.in/bhashadaan/ en/home. [Online; accessed 20-June-2026]. Suma Bhat. 2012. Morpheme Segmentation for Kannada Standing on the Shoulder of Giants. In Proceedings of the 3rd Workshop on South and Southeast Asian Natural Language Processing, WSSANLP@COLING 2012, Mumbai, India, December 8, 2012, Virach Sornlertlamvanich and Abbas Malik (Eds.). The COLING 2012 Organizing Committee, 79–94. https://aclanthology.org/W12-5007/ Pushpak Bhattacharyya. 2010. IndoWordNet. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10), Nicoletta Calzolari, Khalid Choukri, Bente Maegaard, Joseph Mariani, Jan Odijk, Stelios Piperidis, Mike Rosner, and Daniel Tapias (Eds.). European Language Resources Association (ELRA), Valletta, Malta. https://aclanthology.org/L10-1635/ Pushpak Bhattacharyya, Hema Murthy, Surangika Ranathunga, and Ranjiva Munasinghe. 2019. Indic language computing. Commun. ACM 62, 11 (2019), 70–75. Sudha Bhingardive, Tanuja Ajotikar, Irawati Kulkarni, Malhar Kulkarni, and Pushpak Bhattacharyya. 2014. Semi-Automatic Extension of Sanskrit Wordnet using Bilingual Dictionary. In Proceedings of the Seventh Global Wordnet Conference, Heili Orav, Christiane Fellbaum, and Piek Vossen (Eds.). University of Tartu Press, Tartu, Estonia, 324–329. https://aclanthology.org/W14-0145/ Kaushal Santosh Bhogale, Abhigyan Raman, Tahir Javed, Sumanth Doddapaneni, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. Effectiveness of Mining Audio and Text Pairs from Public Data for Improving ASR Systems for Low-Resource Languages. arXiv:2208.12666 [cs.CL] https://arxiv.org/abs/2208.12666 Su Lin Blodgett and Brendan O’Connor. 2017. Racial Disparity in Natural Language Processing: A Case Study of Social Media African-American English. arXiv:1707.00061 [cs.CY] https://arxiv.org/abs/1707.00061 Rishi Bommasani, Kathleen A. Creel, Ananya Kumar, Dan Jurafsky, and Percy S Liang. 2022a. Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?. In Advances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal,
30
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
D. Belgrave, K. Cho, and A. Oh (Eds.), Vol. 35. Curran Associates, Inc., 3663–3678. https://proceedings.neurips.cc/paper_files/paper/2022/ file/17a234c91f746d9625a75cf8a8731ee2-Paper-Conference.pdf Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Ré, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. 2022b. On the Opportunities and Risks of Foundation Models. arXiv:2108.07258 [cs.LG] https://arxiv.org/abs/2108.07258 Maharaj Brahma, N J Karthika, Atul Singh, Devaraj Adiga, Smruti Bhate, Ganesh Ramakrishnan, Rohit Saluja, and Maunendra Sankar Desarkar. 2025. MorphTok: Morphologically Grounded Tokenization for Indian Languages. arXiv:2504.10335 [cs.CL] https://arxiv.org/abs/2504.10335 Maarten Buyl, Alexander Rogiers, Sander Noels, Guillaume Bied, Iris Dominguez-Catena, Edith Heiter, Iman Johary, Alexandru-Cristian Mara, Raphaël Romero, Jefrey Lijffijt, and Tijl De Bie. 2025. Large Language Models Reflect the Ideology of their Creators. arXiv:2410.18417 [cs.CL] https://arxiv.org/abs/2410.18417 Tanise Ceron, Neele Falk, Ana Barić, Dmitry Nikolaev, and Sebastian Padó. 2024. Beyond prompt brittleness: Evaluating the reliability and consistency of political worldviews in llms. Transactions of the Association for Computational Linguistics 12 (2024), 1378–1400. Harveen Singh Chadha, Anirudh Gupta, Priyanshi Shah, Neeraj Chhimwal, Ankur Dhuriya, Rishabh Gaur, and Vivek Raghavan. 2022. Vakyansh: ASR Toolkit for Low Resource Indic languages. arXiv:2203.16512 [cs.CL] https://arxiv.org/abs/2203.16512 Sriram Chaudhury, Ankitha Rao, and Dipti M Sharma. 2010. Anusaaraka: An expert system based machine translation system. In Proceedings of the 6th International Conference on Natural Language Processing and Knowledge Engineering(NLPKE-2010). 1–6. doi:10.1109/NLPKE.2010. 5587789 Zewen Chi, Li Dong, Furu Wei, Nan Yang, Saksham Singhal, Wenhui Wang, Xia Song, Xian-Ling Mao, Heyan Huang, and Ming Zhou. 2021. InfoXLM: An Information-Theoretic Framework for Cross-Lingual Language Model Pre-Training. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Kristina Toutanova, Anna Rumshisky, Luke Zettlemoyer, Dilek Hakkani-Tur, Iz Beltagy, Steven Bethard, Ryan Cotterell, Tanmoy Chakraborty, and Yichao Zhou (Eds.). Association for Computational Linguistics, Online, 3576–3588. doi:10.18653/v1/2021.naacl-main.280 Anand Choudhary and Singh Sukhvir. 2025. Leveraging AI for the recreation and restoration of ancient Indian costumes and accessories. Tekstilec 68, 3 (2025), 210–226. Narayan Choudhary. 2021. LDC-IL: The Indian repository of resources for language technology. Language Resources and Evaluation 55 (01 2021). doi:10.1007/s10579-020-09523-3 Monojit Choudhury, Monjur Alam, Sudeshna Sarkar, and Anupam Basu. 2006. A rewrite rule based model of bangla morphophonological change. Proc. of ICCPB (2006), 64–71. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised Cross-lingual Representation Learning at Scale. CoRR abs/1911.02116 (2019). arXiv:1911.02116 http://arxiv.org/abs/1911.02116 CoRover.ai. 2024. Indigenous Sovereign AI Language Models. Retrieved May 13, 2026 from https://bharatgpt.ai/ Kathleen Creel and Deborah Hellman. 2022. The Algorithmic Leviathan: Arbitrariness, Fairness, and Opportunity in Algorithmic DecisionMaking Systems. Canadian Journal of Philosophy 52, 1 (2022), 26–43. doi:10.1017/can.2022.3 Aniket Dalal, Kumar Nagaraj, Uma Sawant, and Sandeep Shelke. 2006. Hindi Part-of-Speech Tagging and Chunking : A Maximum Entropy Approach. (01 2006). Aniket Dalal, Kumar Nagaraj, U Swant, Sandeep Shelke, and Pushpak Bhattacharyya. 2007. Building feature rich pos tagger for morphologically rich languages: Experience in Hindi. ICON (2007). Sandipan Dandapat, Sudeshna Sarkar, and Basu Anupam. 2004. A Hybrid Model for Part-of-Speech Tagging and its Application to Bengali. 169–172. Sudhansu Bala Das, Samujjal Choudhury, Dr Tapas Kumar Mishra, and Dr Bidyut Kr Patra. 2025. Investigating the Effect of Backtranslation for Indic Languages. In Proceedings of the First Workshop on Natural Language Processing for Indo-Aryan and Dravidian Languages, Ruvan Weerasinghe, Isuri Anuradha, and Deshan Sumanathilaka (Eds.). Association for Computational Linguistics, Abu Dhabi, 152–165. https://aclanthology.org/2025.indonlp-1.18/
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
31
Priyanka Dasari, Abhijith Chelpuri, Nagaraju Vuppala, Mounika Marreddy, Parameshwari Krishnamurthy, and Radhika Mamidi. 2023. Transformer-based Context Aware Morphological Analyzer for Telugu. In Proceedings of the Third Workshop on Speech and Language Technologies for Dravidian Languages. 25–32. NILADRI SEKHAR Dash and BIDYUT BARAN Chaudhuri. 2001. Why do we need to develop corpora in Indian languages. In the International Working Conference on Sharing Capability in Localisation and Human Language Technologies SCALLA-2001. Bangalore. Shachi Dave, Jignashu Parikh, and Pushpak Bhattacharyya. 2001. Interlingua-based English–Hindi machine translation and language divergence. Machine Translation 16 (2001), 251–304. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805 [cs.CL] https://arxiv.org/abs/1810.04805 Sapan Diwakar, Pulkit Goyal, and Rohit Gupta. 2010. Transliteration among indian languages using wx notation. In Proceedings of the conference on natural language processing 2010. Saarland University Press, 147–150. DocBodh 2025. DocBodh: AI-native document understanding for India. Accessed at https://bharatgen.com/products/docbodh/. [Online; accessed 01-March-2026]. Sumanth Doddapaneni, Rahul Aralikatte, Gowtham Ramesh, Shreya Goyal, Mitesh M Khapra, Anoop Kunchukuttan, and Pratyush Kumar. 2023. Towards leaving no indic language behind: Building monolingual corpora, benchmark and models for indic languages. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 12402–12426. Mohamed G. Elfeky, Pedro Moreno, and Victor Soto. 2018. Multi-Dialectical Languages Effect on Speech Recognition: Too Much Choice Can Hurt. Procedia Computer Science 128 (2018), 1–8. doi:10.1016/j.procs.2018.03.001 1st International Conference on Natural Language and Speech Processing. Jay Gala, Pranjal A. Chitale, Raghavan AK, Varun Gumma, Sumanth Doddapaneni, Aswanth Kumar, Janki Nawale, Anupama Sujatha, Ratish Puduppully, Vivek Raghavan, Pratyush Kumar, Mitesh M. Khapra, Raj Dabre, and Anoop Kunchukuttan. 2023. IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages. arXiv:2305.16307 [cs.CL] https://arxiv.org/abs/2305.16307 Jay Gala, Thanmay Jayakumar, Jaavid Aktar Husain, Aswanth Kumar M, Mohammed Safi Ur Rahman Khan, Diptesh Kanojia, Ratish Puduppully, Mitesh M. Khapra, Raj Dabre, Rudra Murthy, and Anoop Kunchukuttan. 2024. Airavata: Introducing Hindi Instruction-tuned LLM. arXiv:2401.15006 [cs.CL] https://arxiv.org/abs/2401.15006 Ankur Gandhe, Rashmi Gangadharaiah, Karthik Visweswariah, and Ananthakrishnan Ramanathan. 2011. Handling verb phrase morphology in highly inflected Indian languages for Machine Translation. In Proceedings of 5th international joint conference on natural language processing. 111–119. Jagadeesh Gorla, Anil Kumar Singh, Rajeev Sangal, Karthik Gali, Samar Husain, and Sriram Venkatapathy. 2008. A Graph Based Method for Building Multilingual Weakly Supervised Dependency Parsers. In Advances in Natural Language Processing, Bengt Nordström and Aarne Ranta (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 148–159. Vikrant Goyal, Sourav Kumar, and Dipti Misra Sharma. 2020. Efficient neural machine translation for low-resource languages via exploiting related languages. In Proceedings of the 58th annual meeting of the association for computational linguistics: student research workshop. 162–168. Mridul Gupta, Vineet Yadav, Samar Husain, and Dipti M Sharma. 2008. A Rule Based Approach for Automatic Annotation of a Hindi Tree-bank. In Proc. Of the 6th International Conference on Natural Language Processing (ICON-08), CDAC Pune, India. Ananth Hariharan and David Mortensen. 2025. Transformer-Enabled Diachronic Analysis of Vedic Sanskrit: Neural Methods for Quantifying Types of Language Change. arXiv:2512.05364 [cs.CL] https://arxiv.org/abs/2512.05364 Jochen Hartmann, Jasper Schwenzow, and Maximilian Witte. 2023. The political ideology of conversational AI: Converging evidence on ChatGPT’s pro-environmental, left-libertarian orientation. arXiv:2301.01768 [cs.CL] https://arxiv.org/abs/2301.01768 Samar Husain, Phani Gadde, Bharat Ambati, Dipti Misra Sharma, and Rajeev Sangal. 2009. A Modular Cascaded Approach to Complete Parsing. In 2009 International Conference on Asian Language Processing. 141–146. doi:10.1109/IALP.2009.37 Indian Economy News 2025. India’s internet users to exceed 900 million in 2025, driven by Indic languages. https://www.ibef.org/news/indias-internet-users-to-exceed-900-million-in-2025-driven-by-indic-languages. [Online; accessed 20-May-2026]. Indian Script Code for Information Interchange 2009. Character Encoding Standard For Indian Languages. https://egovstandards.gov.in/sites/default/files/2021-07/Character%20Encoding-%20Standard%20Ver1.0.pdf. ISFOC Standard for Fonts [n. d.]. ISFOC Standard for Fonts. https://cdac.in/index.aspx?id=mlc_gist_isfoc. Tahir Javed, Kaushal Santosh Bhogale, Abhigyan Raman, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. IndicSUPERB: A Speech Processing Universal Performance Benchmark for Indian languages. arXiv:2208.11761 [cs.CL] https://arxiv.org/abs/2208.11761 Tahir Javed, Sumanth Doddapaneni, Abhigyan Raman, Kaushal Santosh Bhogale, Gowtham Ramesh, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2021. Towards Building ASR Systems for the Next Billion Users. arXiv:2111.03945 [cs.CL] https: //arxiv.org/abs/2111.03945 Tahir Javed, Janki Nawale, Eldho George, Sakshi Joshi, Kaushal Bhogale, Deovrat Mehendale, Ishvinder Sethi, Aparna Ananthanarayanan, Hafsah Faquih, Pratiti Palit, Sneha Ravishankar, Saranya Sukumaran, Tripura Panchagnula, Sunjay Murali, Kunal Gandhi, Ambujavalli
32
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
R, Manickam M, C Vaijayanthi, Krishnan Karunganni, Pratyush Kumar, and Mitesh Khapra. 2024. IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages. In Findings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 10740–10782. doi:10.18653/v1/2024.findings-acl.639 Hang Jiang, Doug Beeferman, Brandon Roy, and Deb Roy. 2022. CommunityLM: Probing Partisan Worldviews from Language Models. arXiv:2209.07065 [cs.SI] https://arxiv.org/abs/2209.07065 Akshat Joshi, Kinal Mehta, Neha Gupta, and Varun Kannadi Valloli. 2018. Indian language transliteration using deep learning. In 2018 IEEE Recent Advances in Intelligent Computational Systems (RAICS). IEEE, 103–107. Divyanshu Kakwani, Anoop Kunchukuttan, Satish Golla, Gokul N.C., Avik Bhattacharyya, Mitesh M. Khapra, and Pratyush Kumar. 2020. IndicNLPSuite: Monolingual Corpora, Evaluation Benchmarks and Pre-trained Multilingual Language Models for Indian Languages. In Findings of the Association for Computational Linguistics: EMNLP 2020, Trevor Cohn, Yulan He, and Yang Liu (Eds.). Association for Computational Linguistics, Online, 4948–4961. doi:10.18653/v1/2020.findings-emnlp.445 Aditya Kallappa, Palash Kamble, Abhinav Ravi, Akshat Patidar, Vinayak Dhruv, Deepak Kumar, Raghav Awasthi, Arveti Manjunath, Himanshu Gupta, Shubham Agarwal, Kumar Ashish, Gautam Bhargava, and Chandra Khatri. 2025. Krutrim LLM: Multilingual Foundational Model for over a Billion People. arXiv:2502.09642 [cs.CL] https://arxiv.org/abs/2502.09642 Shivaram Kalyanakrishnan, Rahul Alex Panicker, Sarayu Natarajan, and Shreya Rao. 2018. Opportunities and Challenges for Artificial Intelligence in India. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society (New Orleans, LA, USA) (AIES ’18). Association for Computing Machinery, New York, NY, USA, 164–170. doi:10.1145/3278721.3278738 Nikhil Kanuparthi, Abhilash Inumella, and Dipti Misra Sharma. 2012. Hindi derivational morphological analyzer. In Proceedings of the Twelfth Meeting of the Special Interest Group on Computational Morphology and Phonology (Montreal, Canada) (SIGMORPHON ’12). Association for Computational Linguistics, USA, 10–16. NJ Karthika, Maharaj Brahma, Rohit Saluja, Ganesh Ramakrishnan, and Maunendra Sankar Desarkar. 2025. Multilingual Tokenization through the Lens of Indian Languages: Challenges and Insights. arXiv preprint arXiv:2506.17789 (2025). N J Karthika, Keerthana Suryanarayanan, Jahanvi Purohit, Ganesh Ramakrishnan, Jitin Singla, and Anil Kumar Gourishetty. 2026. Samasāmayik: A Parallel Dataset for Hindi-Sanskrit Machine Translation. arXiv:2603.24307 [cs.CL] https://arxiv.org/abs/2603.24307 Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, Dilip Kumar Margam, Pooja Aggarwal, Rajiv Teja Nagipogu, Shachi Dave, Shruti Gupta, Subhash Chandra Bose Gali, Vishnu Subramanian, and Partha Pratim Talukdar. 2021. MuRIL: Multilingual Representations for Indian Languages. ArXiv abs/2103.10730 (2021). https://api.semanticscholar.org/CorpusID:232290691 Yash Khemchandani, Sarvesh Mehtani, Vaidehi Patil, Abhijeet Awasthi, Partha Talukdar, and Sunita Sarawagi. 2021. Exploiting language relatedness for low web-resource language model adaptation: An Indic languages study. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 1312–1323. Lachman M. Khubchandani. 1985. Diglossia Revisited. Oceanic Linguistics Special Publications 20 (1985), 199–211. http://www.jstor.org/ stable/20006723 Allison Koenecke, Andrew Nam, Emily Lake, Joe Nudell, Minnie Quartey, Zion Mengesha, Connor Toups, John Rickford, Dan Jurafsky, and Sharad Goel. 2020. Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences 117 (03 2020), 201915768. doi:10.1073/pnas.1915768117 Claire Kramsch. 2014. Language and culture. AILA review 27, 1 (2014), 30–55. D. Manoj Krishna, Vadini Gupta, Kiran Kumari, and Kumari Nidhi. 2025. Impact of AI- Powered Translation Tools: Upholding Indian Linguistic Diversity. Journal of Digital Sociohumanities (2025). https://api.semanticscholar.org/CorpusID:279233826 Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226 (2018). Amba Kulkarni, Soma Paul, Malhar Kulkarni, Anil Kumar Nelakanti, and Nitesh Surtani. 2012. Semantic processing of compounds in indian languages. In Proceedings of COLING 2012. 1489–1502. Malhar Kulkarni and Pushpak Bhattacharyya. 2009. Verbal Roots in the Sanskrit Wordnet. In Sanskrit Computational Linguistics, Gérard Huet, Amba Kulkarni, and Peter Scharf (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 328–338. Malhar Kulkarni, Chaitali Dangarikar, Irawati Kulkarni, Abhishek Nanda, and Pushpak Bhattacharyya. 2010. Introducing sanskrit wordnet. In Proceedings on the 5th global wordnet conference (GWC 2010), Narosa, Mumbai. 287–294. Sumant Kulkarni and Srinath Srinivasa. 2013. TrieIR: Indexing and Retrieval Engine for Kannada Unicode Text. In Digital Libraries: Social Media and Community Networks, Shalini R. Urs, Jin-Cheon Na, and George Buchanan (Eds.). Springer International Publishing, Cham, 21–24. https://doi.org/10.1007/978-3-319-03599-4_3. Anil Kumar, Vipul Mittal, and Amba Kulkarni. 2010. Sanskrit Compound Processor. In Sanskrit Computational Linguistics, Girish Nath Jha (Ed.). Springer Berlin Heidelberg, Berlin, Heidelberg, 57–69. Sanjeev Kumar, Preethi Jyothi, and Pushpak Bhattacharyya. 2024a. Part-of-speech Tagging for Extremely Low-resource Indian Languages. In Findings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
33
for Computational Linguistics, Bangkok, Thailand, 14422–14431. doi:10.18653/v1/2024.findings-acl.857 Sanjeev Kumar, Preethi Jyothi, and Pushpak Bhattacharyya. 2024b. Part-of-speech tagging for extremely low-resource Indian languages. In Findings of the Association for Computational Linguistics: ACL 2024. 14422–14431. Satyam Kumar Navneet, Joydeep Chandra, and Yong Zhang. 2026. When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26). ACM, 1–7. doi:10.1145/3772363.3799085 Anoop Kunchukuttan and Pushpak Bhattacharyya. 2012. Partially modelling word reordering as a sequence labelling problem. In Proceedings of the Workshop on Reordering for Statistical Machine Translation. 47–54. Anoop Kunchukuttan, Abhijit Mishra, Rajen Chatterjee, Ritesh Shah, and Pushpak Bhattacharyya. 2014. Shata-Anuvadak: Tackling Multiway Translation of Indian Languages. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC‘14), Nicoletta Calzolari, Khalid Choukri, Thierry Declerck, Hrafn Loftsson, Bente Maegaard, Joseph Mariani, Asuncion Moreno, Jan Odijk, and Stelios Piperidis (Eds.). European Language Resources Association (ELRA), Reykjavik, Iceland, 1781–1787. https://aclanthology.org/L141355/ Anoop Kunchukuttan, Ratish Puduppully, and Pushpak Bhattacharyya. 2015. Brahmi-Net: A transliteration and script conversion system for languages of the Indian subcontinent. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, Matt Gerber, Catherine Havasi, and Finley Lacatusu (Eds.). Association for Computational Linguistics, Denver, Colorado, 81–85. doi:10.3115/v1/N15-3017 Telugu LLM Lab. 2024. Navarasa: Collection of Gemma finetuned 7B/ 2B Indic Navarasa models. Retrieved May 13, 2026 from https: //huggingface.co/collections/Telugu-LLM-Labs/navarasa Jason Lewis, Hēmi Whaanga, and Ceyda Yolgörmez. 2024. Abundant intelligences: placing AI within Indigenous knowledge frameworks. AI & SOCIETY 40 (10 2024), 2141–2157. doi:10.1007/s00146-024-02099-4 Aparna M and Srinath Srinivasa. 2023. Active learning for Named Entity Recognition in Kannada. TechRxiv 2023, 1127 (2023). arXiv:https://www.techrxiv.org/doi/pdf/10.36227/techrxiv.24580582.v1 doi:10.36227/techrxiv.24580582.v1 Deepak Maheshwari. [n. d.]. AI’s Global South Pivot: Equity, Ethics and Ecology. ([n. d.]). Prasenjit Majumder and Mandar Mitra. 2002. N-gram: a language independent approach to IR and NLP. https://api.semanticscholar.org/ CorpusID:6496350 Prasenjit Majumder, Mandar Mitra, Swapan K Parui, Gobinda Kole, Pabitra Mitra, and Kalyankumar Datta. 2007. YASS: Yet another suffix stripper. ACM transactions on information systems (TOIS) 25, 4 (2007), 18–es. Deepak Kumar Malladi and Prashanth Mannem. 2013. Statistical morphological analyzer for hindi. In Proceedings of the Sixth International Joint Conference on Natural Language Processing. 1007–1011. Guduru Manoj, Neel Prabhanjan Rachamalla, Ashish Kulkarni, Gautam Rajeev, Jay Piplodiya, Arul Menezes, Shaharukh Khan, Souvik Rana, Manya Sah, Chandra Khatri, and Shubham Agarwal. 2025. BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages. arXiv:2511.10338 [cs.CL] https://arxiv.org/abs/2511.10338 Guillermo Marco, Julio Gonzalo, M.Teresa Mateo-Girona, and Ramón Del Castillo Santos. 2024. Pron vs Prompt: Can Large Language Models already Challenge a World-Class Fiction Author at Creative Text Writing?. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 19654–19670. doi:10.18653/v1/2024.emnlp-main.1096 K Marimuthu and Sobha Lalitha Devi. 2014. Automatic conversion of dialectal Tamil text to standard written Tamil text using FSTs. In Proceedings of the 2014 Joint Meeting of SIGMORPHON and SIGFSM. 37–45. Mayura 2025. Making complex information accessible with Mayura. Accessed at https://www.sarvam.ai/blogs/blog-mayura. [Online; accessed 01-March-2026]. Anthony McEnery, Paul Baker, Rob Gaizauskas, and Hamish Cunningham. 2000. EMILLE: building a corpus of South Asian languages. In Proceedings of the International Conference on Machine Translation and Multilingual Applications in the new Millennium: MT 2000. University of Exeter, UK. https://aclanthology.org/2000.bcs-1.11/ Diwakar Mishra and Kalika Bali. 2010. Hindi dialects phonological transfer rules for verb root 𝐶 e𝑙 e. In 13th oriental COCOSDA-2010 conference in coordination with International Committee for the Co-ordination and Standardization of Speech Databases and Assessment Techniques held at Kathmandu. Nepal. 24–25. Fabio Motoki, Valdemar Pinho Neto, and Victor Rangel. 2023. More Human than Human: Measuring ChatGPT Political Bias. SSRN Electronic Journal (01 2023). doi:10.2139/ssrn.4372349 Ashish Narang, RK Sharma, and Parteek Kumar. 2013. Development of Punjabi WordNet. CSI transactions on ICT 1, 4 (2013), 349–354. Ramnath Neupane. 2024. Morphological Processes in English and Sanskrit: A Cross-Linguistic Study. Vox Batauli 9, 01 (2024), 72–79. Partha Pakray, Reddi Krishna, Santanu Pal, Advaitha Vetagiri, Sandeep Dash, Arnab Kumar Maji, Saralin A Lyngdoh, Lenin Laitonjam, Anupam Jamatia, Koj Sambyo, et al. 2025. Findings of WMT 2025 shared task on low-resource indic languages translation. In Proceedings of the Tenth Conference on Machine Translation. 532–553.
34
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Mihir Panchal, Deeksha Varshney, Mamta, and Asif Ekbal. 2026. Indic-TunedLens: Interpreting Multilingual Models in Indian Languages. doi:10.48550/arXiv.2602.15038 S. Pandian and T.V Geetha. 2008. Morpheme based Language Model for Tamil Part-of-Speech Tagging. Polibits 38 (12 2008), 19–25. doi:10.17562/PB-38-2 Param 2025. Powering India’s AI with Contextual Understanding. Accessed at https://bharatgen.com/products/bharat-gen-text-model/. [Online; accessed 01-March-2026]. Raj Patel, Prakash Pimpale, and Sasikumar Mukundan. 2017. MTIL17: English to Indian Langauge Statistical Machine Translation. (08 2017). doi:10.48550/arXiv.1708.07950 Siddhesh Pawar, Pushpak Bhattacharyya, and Partha Talukdar. 2023. Evaluating Cross Lingual Transfer for Morphological Analysis: a Case Study of Indian Languages. In Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morphology, Garrett Nicolai, Eleanor Chodroff, Frederic Mailhot, and Çağrı Çöltekin (Eds.). Association for Computational Linguistics, Toronto, Canada, 14–26. doi:10.18653/v1/2023.sigmorphon-1.3 Jerin Philip, Shashank Siripragada, Vinay P Namboodiri, and CV Jawahar. 2021. Revisiting low resource status of indian languages in machine translation. In Proceedings of the 3rd ACM India Joint International Conference on Data Science & Management of Data (8th ACM IKDD CODS & 26th COMAD). 178–187. Sujith Pulikodan, Abhayjeet Singh, Agneedh Basu, Nihar Desai, Pavan Kumar J, Pranav D Bhat, Raghu Dharmaraju, Ritika Gupta, Sathvik Udupa, Saurabh Kumar, Sumit Sharma, Visruth Sanka, Dinesh Tewari, Harsh Dhand, Amrita Kamat, Sukhwinder Singh, Shikhar Vashishth, Partha Talukdar, Raj Acharya, and Prasanta Kumar Ghosh. 2026. VAANI: Capturing the language landscape for an inclusive digital India. arXiv:2603.28714 [eess.AS] https://arxiv.org/abs/2603.28714 Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356 [eess.AS] https://arxiv.org/abs/2212.04356 Sankaravelayuthan Rajendran, Selvaraj Arulmozi, B Kumara Shanmugam, S Baskaran, and S Thiagarajan. 2002. Tamil wordnet. In Proceedings of the first international global WordNet conference. Mysore, Vol. 152. 271–274. Ananthakrishnan Ramanathan, Pushpak Bhattacharyya, Karthik Visweswariah, Kushal Ladha, and Ankur Gandhe. 2011. Clause-Based Reordering Constraints to Improve Statistical Machine Translation. In Proceedings of 5th International Joint Conference on Natural Language Processing, Haifeng Wang and David Yarowsky (Eds.). Asian Federation of Natural Language Processing, Chiang Mai, Thailand, 1351–1355. https://aclanthology.org/I11-1152/ Ananthakrishnan Ramanathan, Hansraj Choudhary, Avishek Ghosh, and Pushpak Bhattacharyya. 2009. Case markers and Morphology: Addressing the crux of the fluency problem in English-Hindi SMT. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, Keh-Yih Su, Jian Su, Janyce Wiebe, and Haizhou Li (Eds.). Association for Computational Linguistics, Suntec, Singapore, 800–808. https://aclanthology.org/P09-1090/ Ananthakrishnan Ramanathan, Jayprasad Hegde, Ritesh M. Shah, Pushpak Bhattacharyya, and Sasikumar M. 2008. Simple Syntactic and Morphological Processing Can Help English-Hindi Statistical Machine Translation. In Proceedings of the Third International Joint Conference on Natural Language Processing: Volume-I. https://aclanthology.org/I08-1067/ Ananthakrishnan Ramanathan and Durgesh Rao. 2003. A Lightweight Stemmer for Hindi. https://api.semanticscholar.org/CorpusID:46993455 Loganathan Ramasamy and Zdeněk Žabokrtskỳ. 2011. Tamil dependency parsing: results using rule based and corpus based approaches. In International Conference on Intelligent Text Processing and Computational Linguistics. Springer, 82–95. Durgesh Rao, Pushpak Bhattacharya, and Radhika Mamidi. 1998. Natural language generation for English to Hindi human-aided machine translation. VIVEK-BOMBAY- 11 (1998), 32–39. Durgesh Rao, Kavitha Mohanraj, Jayprasad Hegde, Vivek Mehta, and Parag Mahadane. 2000. A practical framework for syntactic transfer of compound-complex sentences for English-Hindi machine translation. In Proceedings of KBCS, Vol. 2000. Citeseer. Siva Reddy, Abhilash Inumella, Rajeev Sangal, and Soma Paul. 2009. All Words Unsupervised Semantic Category Labeling for Hindi. In Proceedings of the International Conference RANLP-2009, Galia Angelova and Ruslan Mitkov (Eds.). Association for Computational Linguistics, Borovets, Bulgaria, 365–369. https://aclanthology.org/R09-1066/ Saaras V3 2026. Introducing Saaras V3. Built for the Way India Speaks. Accessed at https://www.sarvam.ai/blogs/asr. [Online; accessed 01-June-2026]. Kalyanamalini Sahoo and V Eshwarchandra Vidyasagar. 2003. Kannada WordNet-A lexical database. In TENCON 2003. Conference on Convergent Technologies for Asia-Pacific Region, Vol. 4. IEEE, 1352–1356. Jivnesh Sandhan, Rathin Singha, Narein Rao, Suvendu Samanta, Laxmidhar Behera, and Pawan Goyal. 2022. TransLIST: A Transformer-Based Linguistically Informed Sanskrit Tokenizer. arXiv:2210.11753 [cs.CL] https://arxiv.org/abs/2210.11753 Baskaran Sankaran, Kalika Bali, Tanmoy Bhattacharya, Pushpak Bhattacharyya, Girish Nath Jha, S Rajendran, K Saravanan, et al. 2008. Designing a common POS-tagset framework for Indian languages. In Proceedings of the 6th workshop on Asian language resources. Baskaran Sankaran, Majid Razmara, and Anoop Sarkar. 2012. Kriya - An end-to-end Hierarchical Phrase-based MT System. The Prague Bulletin of Mathematical Linguistics 97 (05 2012). doi:10.2478/v10108-012-0004-y
Rethinking Indic AI from a Lens of Cultural Heritage Preservation
•
35
Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. Whose Opinions Do Language Models Reflect? arXiv:2303.17548 [cs.CL] https://arxiv.org/abs/2303.17548 Shikhar Kr Sarma, Biswajit Brahma, Moromi Gogoi, and Mane Bala Ramchiary. 2010a. A wordnet for Bodo language: Structure and development. In Global Wordnet Conference (GWC10), Mumbai, India. Shikhar Kr Sarma, R Medhi, M Gogoi, Utpal Saikia, et al. 2010b. Foundation and structure of developing an Assamese WordNet. In Proceedings of 5th international conference of the global WordNet Association. Sarvam 1 2024. Sarvam 1: The first Indian language LLM. Accessed at https://www.sarvam.ai/blogs/sarvam-1. [Online; accessed 01-June-2026]. Sarvam-M 2025. Sarvam-M: Explorations in Post Training and Inferencing Optimizations for a Hybrid Indic LLM. Accessed at https: //www.sarvam.ai/blogs/sarvam-m/. [Online; accessed 01-March-2026]. Sarvam Translate 2025. Sarvam-Translate: an open-weights model, supports translation across 22 Indian languages with structured long-form text. Accessed at https://www.sarvam.ai/blogs/sarvam-translate. [Online; accessed 01-March-2026]. Kumar Saurav, Kumar Saunack, Diptesh Kanojia, and Pushpak Bhattacharyya. 2020. “A Passage to India”: Pre-trained Word Embeddings for Indian Languages. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages (SLTU) and Collaboration and Computing for Under-Resourced Languages (CCURL), Dorothee Beermann, Laurent Besacier, Sakriani Sakti, and Claudia Soria (Eds.). European Language Resources association, Marseille, France, 352–357. https://aclanthology.org/2020.sltu-1.49/ Probal Sengupta and BB Chaudhuri. 1996. Morphological processing of Indian languages for lexical interaction with application to spelling error correction. Sadhana 21, 3 (1996), 363–380. Utpal Sharma, Jugal K Kalita, and Rajib K Das. 2008. Acquisition of morphology of an Indic language from text corpus. ACM Transactions on Asian Language Information Processing (TALIP) 7, 3 (2008), 1–33. Manish Shrivastava, Nitin Agrawal, Bibhuti Mohapatra, Smriti Singh, and Pushpak Bhattacharya. 2005. Morphology based natural language processing tools for Indian languages’. In Proceedings of the 4th Annual Inter Research Institute Student Seminar in Computer Science, IIT, Kanpur, India, April. Manish Shrivastava and Pushpak Bhattacharyya. 2008. Hindi POS tagger using naive stemming: harnessing morphological information without extensive linguistic knowledge. In International Conference on NLP (ICON08), Pune, India. Akshay Singh, Sushma Bendre, and Rajeev Sangal. 2005. HMM based chunker for Hindi. In Companion Volume to the Proceedings of Conference including Posters/Demos and tutorial abstracts. Arvinder Singh and Parminder Singh. 2015. Punjabi dialects conversion system for Malwai and Doabi dialects. Indian Journal of Science and Technology 8, 27 (2015), 1–6. Anil Kumar Singh, Kiran Pala, and Harshit Surana. 2008. Estimating the Resource Adaption Cost from a Resource Rich Language to a Similar Resource Poor Language.. In LREC. R.M.K. Sinha, K. Sivaraman, A. Agrawal, R. Jain, R. Srivastava, and A. Jain. 1995. ANGLABHARTI: a multilingual machine aided translation project on translation from English to Indian languages. In 1995 IEEE International Conference on Systems, Man and Cybernetics. Intelligent Systems for the 21st Century, Vol. 2. 1609–1614 vol.2. doi:10.1109/ICSMC.1995.538002 Benjamin Slade. 2016. Compound verbs. The University of Utah (2016). Zhivar Sourati, Farzan Karimi-Malekabadi, Meltem Ozcan, Colin McDaniel, Alireza Ziabari, Jackson Trager, Ala Tak, Meng Chen, Fred Morstatter, and Morteza Dehghani. 2025. The shrinking landscape of linguistic diversity in the age of large language models. arXiv preprint arXiv:2502.11266 (2025). Saikrishna Srirampur and Ravi Chandibhamar. 2014. Statistical Morph Analyzer (SMA++) for Indian Languages. 103–109. doi:10.3115/v1/W145312 Rishabh Srivastava and Riyaz Ahmad Bhat. 2013. Transliteration systems across indian languages using parallel corpora. In Proceedings of the 27th Pacific Asia conference on language, information, and computation (PACLIC 27). 390–398. Sharath Srivatsa, M. Aparna, Srinath Srinivasa, and T. B. Dinesh. 2026. Safeguarding Plurality: The Digital Preservation of Diverse Worldviews. In Big Data Analytics in Astronomy, Science, and Engineering, Shelly Sachdeva, Yutaka Watanobe, and Subhash Bhalla (Eds.). Springer Nature Switzerland, Cham, 268–282. Sharath Srivatsa, Aparna M, Sai Madhavan G, and Srinath Srinivasa. 2024. Knowledge Management Framework Over Low Resource Indian Colloquial Language Audio Contents. In Proceedings of the 7th Joint International Conference on Data Science & Management of Data (11th ACM IKDD CODS and 29th COMAD) (Bangalore, India) (CODS-COMAD ’24). Association for Computing Machinery, New York, NY, USA, 553–557. doi:10.1145/3632410.3632483 Sharath Srivatsa, Aparna M, Malavika V, Samarth P, and Srinath Srinivasa. 2025. Parichaya: Rural Colloquial Knowledge AI Interface. In Proceedings of the 8th International Conference on Data Science and Management of Data (12th ACM IKDD CODS and 30th COMAD) (CODS-COMAD ’24). Association for Computing Machinery, New York, NY, USA, 391–394. doi:10.1145/3703323.3704271 Harshit Surana and Anil Kumar Singh. 2008. A more discerning and adaptable multilingual transliteration mechanism for indian languages. In Proceedings of the Third International Joint Conference on Natural Language Processing: Volume-I. S Thara and Prabaharan Poornachandran. 2018. Code-Mixing: A Brief Survey. In 2018 International Conference on Advances in Computing, Communications and Informatics (ICACCI). 2382–2388. doi:10.1109/ICACCI.2018.8554413
36
•
Aparna Madva, Sharath Srivatsa, Srinath Srinivasa, and Tulika Saha
Goutham Tholpadi, Chiranjib Bhattacharyya, and Shirish Shevade. 2017. Corpus-based translation induction in indian languages using auxiliary language corpora from Wikipedia. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP) 16, 3 (2017), 1–25. Kumud Tripathi, Raj Gothi, and Pankaj Wasnik. 2025. Enhancing whisper’s accuracy and speed for indian languages through prompt-tuning and tokenization. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5. UNICODE STANDARD For Indic Scripts [n. d.]. UNICODE STANDARD For Indic Scripts. https://www.unicode.org/L2/L2003/03102-indicov.pdf. Dhanalakshmi V., Padmavathy P., Anand Kumar M., Soman K.P., and Rajendran S. 2009. Chunker for Tamil. In 2009 International Conference on Advances in Recent Technologies in Communication and Computing. 436–438. doi:10.1109/ARTCom.2009.191 K. Vijayanand, S.I. Choudhury, and P. Ratna. 2002. VAASAANUBAADA: automatic machine translation of bilingual Bengali-Assamese news texts. In Language Engineering Conference, 2002. Proceedings. 183–188. doi:10.1109/LEC.2002.1182307 T. N. Vikram and Shalini R. Urs. 2007. Development of Prototype Morphological Analyzer for he South Indian Language of Kannada. In Asian Digital Libraries. Looking Back 10 Years and Forging New Frontiers, Dion Hoe-Lian Goh, Tru Hoang Cao, Ingeborg Torvik Sølvberg, and Edie Rasmussen (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 109–116. Karthik Visweswariah, Rajakrishnan Rajkumar, Ankur Gandhe, Ananthakrishnan Ramanathan, and Jiri Navratil. 2011. A Word Reordering Model for Improved Machine Translation. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, Regina Barzilay and Mark Johnson (Eds.). Association for Computational Linguistics, Edinburgh, Scotland, UK., 486–496. https://aclanthology. org/D11-1045/ Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. 2024. Do Llamas Work in English? On the Latent Language of Multilingual Transformers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 15366–15394. doi:10.18653/v1/2024.acl-long.820 Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mT5: A massively multilingual pre-trained text-to-text transformer. CoRR abs/2010.11934 (2020). arXiv:2010.11934 https://arxiv.org/abs/2010.11934 Robert J Zydenbos. 2020. A Manual of Modern Kannada. CrossAsia E-Publishing. https://doi.org/10.11588/xabooks.736.