LitSeg: Narrative-Aware Document Segmentation for Literary RAG

Unknown

Conceptio › Archive › arXiv CS

arXiv CSopen access

LitSeg: Narrative-Aware Document Segmentation for Literary RAG

Unknown · 2026 · arxiv_cs

arXiv CS · Papers · License: Open Access · 2026

Open Source ↗Direct PDF ↓

artificialintelligence knowledgerepresentation reasoning

artificial intelligence, reasoning, knowledge representation

LitSeg: Narrative-Aware Document Segmentation for Literary RAG Ruikang Zhang1 Zhanni Chen1 * Yiqiao Cai1 * Qi Su1† 1 Peking University, Beijing, China {2300018416, 2300018109, 2300018316}@stu.pku.edu.cn, [email protected]

arXiv:2605.27156v1 [cs.CL] 26 May 2026

Abstract Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by incorporating external knowledge, particularly for long-tail domains such as literary works. However, the critical step of document segmentation in RAG remains largely underexplored. Existing strategies are typically semantically blind and overlook the complicated narrative structures of literary works, often resulting in fragmented plots and unclear references that severely hinder retrieval and generation performance. To address this, we propose LitSeg, a novel narrative-theory-guided segmentation framework. By employing multi-stage prompting, LitSeg explicitly extracts valid events, untangles narrative threads, clarifies narrative structures, and locates turning points to inform segmentation. To alleviate the computational overhead of multi-stage inference with largescale models, we further introduce LitSeg-Lite, a lightweight single-pass chunker fine-tuned on LitSeg-generated data via a two-stage training strategy, distilling the complex process into a single inference pass. Extensive experiments demonstrate that with structurally independent text chunks, our methods significantly improve retrieval accuracy and context relevance over baselines, ultimately enhancing downstream QA performance, while ablation studies validate the efficacy of narratological guidance and data distillation.

Introduction

Retrieval-Augmented Generation (RAG) enhances the generation process of Large Language Models (LLMs) by retrieving relevant information from external corpora, thereby improving their access to long-tail knowledge such as literary works. (Zhao et al., 2026) A RAG system typically consists of three parts: 1) indexing, in which raw documents * †

Equal contribution. Corresponding author.

are segmented into chunks, and then represented and stored in vector databases; 2) retrieval, in which the system retrieves relevant chunks given a user query; 3) generation, in which an LLM formulates a response given the query and retrieved chunks (Ma et al., 2025). In such systems, document segmentation affects the vector representation and text quality of chunks, thereby influencing retrieval accuracy, the information accessible to the generator, and ultimately the performance of RAG systems on downstream tasks such as question answering (QA) (Rajanidi et al., 2026) and agentic tasks (Singh et al., 2025). However, existing work has primarily focused on index structures, document retrieval, and response generation (e.g., Huang et al., 2025; Gao et al., 2023; Tu et al., 2025), leaving document segmentation underexplored (Wang et al., 2025). Given the complicated narrative structure of literary works (Shen and Wang, 2010), inappropriate text segmentation can distort the semantics of individual segments and lead to fragmented plots and unclear references, thereby degrading downstream retrieval and generation performance. Nevertheless, existing segmentation strategies are either semantically blind or overlook such high-order structures (see Section 2.1), leaving this critical issue unresolved. In light of this, we present LitSeg (see Section 3.1), a narrative-theory-guided segmentation framework. Drawing on established theories, we leverage explicit, multi-stage prompting to task an LLM with extracting valid events, untangling narrative threads, clarifying narrative structures, locating key turning points and finally executing segmentation. We also provide the model with long context and design an input-output format that numbers the input text and represents each chunk as a collection of sentence IDs, thereby supporting flexible recombination of non-adjacent sentences while preserving broader narrative context. To further reduce the overhead of multi-stage inference and large-

LitSeg-Lite

LitSeg Narratological Theories

Student Model

1. Event-based Plot Model

2. Simultaneity in Narrative …… 3. Aristotelian Plot Structure Theory

Theory-Guided Reward Model {event_validity: [4.0, 3.0, 4.5],

GRPO

cut_point_logic: [5.0, 5.0, 5.0]}

SFT STEP 1: Extract Valid Events and Filter Noise STEP 2: Untangle Narrative Threads and Clarify Structures STEP 3: Locate Key Turning Points and Execute Segmentation

unity_of_action: [3.0, 4.0, 5.0],

Format_Reward

Evaluation

prompt annotate

Teacher Model

GutenQA

Chapter I. Into the Primitive step_1_events: “…” step_2_threads: “…” step_3_turning_points: […] segments: [{SUBTITLE, FROM_IDX, TO_IDX, CONTEXT_IDX},…]

Question: What kind of dog is Buck?

LiteraryQA Chunking Dataset

Our Answer: Buck is a St. Bernard and Scotch shepherd dog mix.

LiteraryQA

Standard Answer: ['A St. Bernard-Scotch Shepherd.', 'ST BERNARD/SCOTCH SHEPHERD MIX’]

Figure 1: Overview of our proposed framework. LitSeg leverages narratological theories to guide a high-capacity teacher model through a multi-stage segmentation pipeline, producing narrative-aware chunking annotations. LitSegLite then distills this knowledge into a lightweight student model via SFT and GRPO with theory-guided rewards, enabling efficient single-pass segmentation for RAG-based QA on literary benchmarks.

scale model usage, we also present LitSeg-Lite (see Section 3.2), a lightweight chunker that performs segmentation in a single inference pass. It is fine-tuned on LitSeg-generated data via a two-stage training strategy, achieving performance comparable to LitSeg. Experiments show that by producing narratologically self-contained text chunks, LitSeg and LitSeg-Lite substantially improve retrieval accuracy, context relevance, and downstream answer accuracy over baselines. Ablation study further shows that both narrative theories and distillation of LitSeg data contribute to the strong performance of LitSeg-Lite (see Section 4). Figure 1 provides an overview of our proposed framework. Our main contributions are threefold: (1) We propose LitSeg, a novel narrative-theory-guided segmentation framework for RAG systems targeting literary works. By employing multi-stage prompting, LitSeg explicitly extracts narrative structure to aid segmentation, effectively addressing the semantic blindness of existing segmentation methods and preserving high-order narrative structures. (2) We introduce LitSeg-Lite, a lightweight, single-pass chunker fine-tuned on LitSeg-generated data via a two-stage training strategy. It distills the multistep inference process into a single inference pass, achieving performance comparable to LitSeg with

substantially lower inference overhead. (3) Comprehensive experiments demonstrate that our methods substantially improve retrieval accuracy and downstream QA performance by leveraging semantically independent text chunks. Ablation studies validate the efficacy of narratological guidance and data distillation.

Related Works

2.1

Document Segmentation in RAG

In RAG systems, the most prevalent segmentation methods typically split documents by fixed token (Ma et al., 2025) or character counts. Recursive character splitters (Chase, 2022) refine this strategy by recursively applying common delimiters (e.g., newlines) until size constraints are met. However, these heuristic approaches are semantically agnostic, often disrupting structural integrity and semantic coherence. To address this, semanticsaware methods have emerged, leveraging embedding models or LLMs to capture textual semantics. Embedding-based approaches detect semantic shifts; for instance, SemanticChunker (Chase, 2022) inserts boundaries where the cosine distance between adjacent sentence-group embeddings exceeds a percentile threshold. LLM-based methods

further exploit model internals or reasoning capabilities: Perplexity Chunker (Zhao et al., 2024) identifies boundaries at local minima of the sentencelevel loss curve; Margin-Sampling Chunker (Zhao et al., 2024) frames boundary detection as a binary decision based on the LLM’s probability margin for keeping adjacent text; and LumberChunker (Duarte et al., 2024) uses a sliding window to prompt an LLM to directly pinpoint content shifts. However, these methods still overlook the high-order narrative structures of literary works, where coherent chunks often depend on long-range causal relations, parallel narrative threads, and shifts in narrative level or perspective. This limitation restricts their effectiveness in literary RAG systems. 2.2

Distillation for Efficient Task-Specific Models

Supervised Fine-Tuning (SFT) (Zhang et al., 2026) remains the foundational step to adapt Large Language Models (LLMs) to expert demonstrations. While SFT effectively instills specific task knowledge and stylistic consistency, it is inherently constrained by the quality and diversity of the offline dataset (Chu et al., 2025). To address these limitations, Group Relative Policy Optimization (GRPO) (Shao et al., 2024) has emerged as an efficient Reinforcement Learning (RL) paradigm. GRPO generates a group of G rollouts {y1 , y2 , . . . , yG } for each prompt x and computes the relative advantage of each output by normalizing rewards within the group. This mechanism bypasses the optimization instabilities often associated with value function estimation and significantly reduces memory overhead of traditional Proximal Policy Optimization (PPO) methods (Schulman et al., 2017) that rely on a computationally expensive critic model to estimate value baselines. Building upon the groupbased efficiency of GRPO, Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) (Yu et al., 2025) further enhances training stability and output diversity, particularly in scenarios requiring long-sequence structural consistency. 2.3

Narrative Theory and NLP

Narrative theory is the foundational framework for understanding how stories are constructed, structured, and communicated (Liveley, 2019), providing crucial theoretical underpinnings for various Natural Language Processing (NLP) tasks. In the realm of narrativity analysis, narrative theory provides essential criteria for quantifying the

degree of narrativity and distinguishing narrative discourse from non-narrative noise. Piper et al. (2021) synthesizes classical and postclassical narrative theories, guiding computational narrative research by defining the core elements of narrativity. To operationalize event narrativity, Gius and Vauth (2022) proposes an event-based plot model classifying textual components into distinct event types; Piper and Bagga (2024) further applies LLMs to annotate narrative discourse within the framework of Genette (1980)’s narrative triangle concerning story, discourse, and narrating. Beyond identifying narration, narrative theory also offers vital structural paradigms for organizing scattered events into coherent narrative threads, thereby facilitating storyline extraction. Vossen et al. (2021) introduces a narratology-based framework for news that maps the theoretical notions of fabula, plot, and plot structure to explicit data structures: timelines, causelines, and storylines. To address the higher levels of non-linearity in fictional works compared to news, Visser Solissa et al. (2025) constructs a theoretical model for event annotation grounded in the theory of syuzhet (the concrete presentation order of events) to achieve its computational detection. On the other hand, narrative theory empowers text segmentation by guiding the identification of logical breakpoints across different levels. At a macro-structural level, Papalampidi et al. (2019) models movie plots by identifying theory-grounded narrative turning points. At a micro-linguistic level, text boundaries are effectively pinpointed using mutual information theory: Wagner et al. (2024) detects zones of low Point-wise Mutual Information (PMI) between adjacent sentences to segment spoken testimonies, while Wang et al. (2023) introduces M3Seg, a novel Maximum-Minimum Mutual information paradigm, to partition ASR transcripts. Despite significant strides in quantifying narrative structures, these works are hindered by two primary limitations: theoretically, they often focus on single-faceted features, lacking a refined and synthesized comprehensive theoretical framework; practically, they predominantly target isolated computational tasks, thereby failing to induce and execute an end-to-end theoretical pipeline for the entire text segmentation process.

Methodology

3.1

LitSeg: Narrative Theory Guided Segmentation Framework

3.1.1 Theoretical Framework Our framework integrates classical and postclassical narratology with computational narratology, translating abstract literary concepts into operational metrics to address the limitations of existing semantic segmenters. Specifically, LitSeg explicitly decomposes the complex cognitive task of document chunking into a three-step narratological pipeline: Event Extraction, Thread Untangling, and Turning Points Pinpointing. See Appendix D.1 for full prompts. Extract Valid Events and Filter Noise Distinguishing valid narrative progression from background information is the foundational step of segmentation. Plot is defined by its dynamic, sequential nature; narrative elements only constitute a plot when they drive this dynamism (Scholes et al., 2006). Thus, the event-based plot model (Gius and Vauth, 2022) is adopted to explicitly task the LLM to filter text by retaining only changes of state and process events, while discarding static events and non-events (e.g., generic statements or counterfactual passages). For ambiguous paragraphs, the text’s degree of narrativity is quantified by actively evaluating specific feature densities, such as the presence of an agent, sequential actions, spatial/temporal specifications, and rationale (Piper et al., 2021), to guarantee that only independent and valid events are extracted for downstream processing. Untangle Narrative Threads and Clarify Structures Literary texts frequently feature intertwined storylines. To systematically disentangle these narratives, we adopt Aristotle’s principle of unity of action (Butcher et al., 1902; Shen and Wang, 2010), enabling the model to precisely cluster events based on complete, core actions. Since a well-arranged plot often builds the structural integrity of the whole story through the gradual unfolding of secrets (Forster, 1927), tracking this revelation process guides the model to connect scattered events into cohesive narrative chains. Furthermore, to handle complex discourse, the theory of simultaneity in narrative (Margolin, 2014) is incorporated to instruct the model to explicitly segment concurrent and parallel event chains, thereby preventing the erroneous merging of minor threads

into major ones. We also map the nested narrative hierarchy (e.g., extradiegetic, intradiegetic, or metadiegetic levels) based on the theory of narrative levels (Genette, 1980) to systematically prevent context collisions caused by embedded narratives. Finally, to capture the multidimensionality of narrative discourse, we build upon the three core narrative dimensions: time, setting, and point-ofview, which Piper and Bagga (2024) operationalized for LLMs under Genette (1980) classical narratological framework. By combining these dimensions with Vossen et al. (2021)’s narratology-based framework for storyline extraction, we are able to comprehensively extract timelines, causelines, and critical shift points in time, space, and perspective from this structured thread data. Locate Key Turning Points and Execute Segmentation Ultimately, the framework utilizes the established thread data to determine logical breakpoints. Turning points are treated as crucial functional shifts rather than mere plot occurrences (Thompson, 1999), ensuring that chunk boundaries are dictated by structural narrative transitions rather than arbitrary length constraints or linear episodic prolongations. Building upon this rationale, our framework segments narratives through a fourtiered location strategy. At the macro-structural level, we integrate Aristotelian plot structure theory, specifically the narrative shifts of reversal and recognition (Butcher et al., 1902; Shen and Wang, 2010), with turning-point identification (Papalampidi et al., 2019; Hauge, 2017) to capture major narrative pivots: opportunity, change of plans, point of no return, major setback and reversal, climax and recognition, and aftermath and buffer. At the contextual level, we leverage the established multidimensional thread data to pinpoint breakpoints at transitions in time, space, or perspective, as well as shifts between parallel or nested narrative levels. At the micro-dynamic level, we align segmentation with state transitions (from equilibrium to disequilibrium to a new equilibrium) (Todorov and Weinstein, 1969) and the logic of narrative possibilities (Bremond and Cancalon, 1980; Shen and Wang, 2010), capturing the precise moments where the fulfillment or failure of an action is resolved. Finally, to pinpoint exact sentence-level boundaries within these general turning points, we draw inspiration from mutual information (MI) theory (Wagner et al., 2024; Wang et al., 2023) to formulate an LLM-assessable metric for semantic cohesion. The

model is prompted to actively evaluate the semantic dependency between adjacent sentences, avoiding segmenting within High Semantic Cohesion Zones (e.g., spans with dense MI markers, such as continuous pronoun chains or adjacency pairs), and instead targeting Low Semantic Cohesion Zones, where semantic dependency is minimal. Collectively, this four-tiered strategy ensures a rigorous alignment between segmentation and the underlying narrative architecture, thereby ensuring that the resulting segments are narratively self-contained and semantically coherent. 3.1.2

Implementation of LitSeg

The LitSeg pipeline comprises three sequential phases: preprocessing, multi-round generation, and validation with fallback mechanisms. More implementation details are in Appendix D.1 and H.2. Preprocessing To preserve authorial intent and semantic integrity, we first perform structure-driven splitting by isolating top-level narrative units (e.g., chapters, acts, scenes, sections, cantos, or titled episodes). This ensures that the model receives self-contained structural units while naturally accommodating the context window limits for most texts. When top-level unit metadata is unavailable, or when a narrative unit exceeds a length threshold n, we apply recursive word-level splitting, hierarchically segmenting the text by paragraphs and then sentences while greedily satisfying the n-word constraint. We set n to the largest value allowed by the model’s context budget and without substantial long-context degradation. Subsequently, we assign a unique sequential index to each sentence in the candidate chunk. Multi-stage Generation We implement a threestage pipeline corresponding to the narratological workflow above. All stages share a unified system prompt containing the full narratological workflow, providing the model with global task awareness. The user prompt for each stage contains stage-specific formatting instructions alongside the model’s outputs from preceding stages. This design implements an explicit multi-stage prompting strategy, which avoids the token overhead of unconstrained Chain-of-Thought (CoT) rationales while ensuring high-quality segmentation. To further optimize inference efficiency, we constrain the model to output sentence indices rather than verbatim text, significantly reducing the overhead of autoregressive decoding. Specifically, for each seg-

mented plot block, the model outputs a main range, represented as a closed interval between its start and end sentence indices, and an optional context list, represented as a sequence of discrete sentence indices. Sentences shared by overlapping main ranges are retained in all corresponding blocks, allowing boundary sentences to serve as shared narrative context. Unlike traditional chunking methods that rely on fixed, adjacent overlaps, this indexbased representation allows the model to flexibly incorporate non-adjacent sentences from anywhere within the indexed candidate chunk that are crucial for understanding the main block. Additionally, we task the model with generating a concise subtitle for each block, leveraging its generative capabilities to summarize the core event without the burden of extensive text generation. Validation, Retry, and Fallback We apply duallevel validation to guarantee reliability. Syntax validation ensures that the output JSON conforms to the predefined schema. Semantic validation verifies that all main ranges and context indices are valid and that the main ranges, with overlaps permitted, collectively and exhaustively cover all source sentences, ensuring zero information loss. Invalid outputs are retried up to N times. If coverage gaps persist after retries, a fallback mechanism greedily appends the missing sentences to their left-adjacent segment in source order (or the right one if unavailable). If the output remains unparsable after all retries, the system falls back to yielding the whole preprocessed chunk. Reconstruction During text reconstruction, sentences specified by the context list, except those already included in the corresponding main range, are ordered by their original positions and prepended or appended to the main text block accordingly. The generated subtitle is then prepended to form the final context-enriched chunk. 3.2

LitSeg-Lite: Distillation of Lightweight Chunker

With the objective of improving inference efficiency and facilitating deployment, we develop LitSeg-Lite, a compact student model distilled from LitSeg. Given the same sentence-indexed candidate chunk as LitSeg, LitSeg-Lite predicts a JSON output containing the internally summarized narratological workflow and the final index-based segmentation within a single inference pass. We train LitSeg-Lite with LoRA (Hu et al., 2021) on

the LitSeg-annotated LiteraryQA dataset through a two-stage optimization pipeline: initial supervised fine-tuning (SFT) followed by Reinforcement Learning (RL) refinement. The SFT targets are the validated LitSeg annotations, reformatted into the single-stage output schema used by LitSeg-Lite. To ensure that LitSeg-Lite acquires narratologyaware segmentation logic and robust instructionfollowing capabilities, we first perform SFT. While SFT facilitates rapid task adaptation, it often results in limited out-of-distribution generalization (Chu et al., 2025). To overcome these limitations, we integrate RL to further enhance the model’s behavioral robustness and generalization. Specifically, we employ GRPO with a DAPO-style objective, guided by a specialized reward function grounded in narratological theory. This function combines a rule-based format validator with a model-based reward evaluator; the latter conducts granular assessments across three narrative-centric dimensions: Event Validity, Unity of Action, and Cut Point Logic, thereby targeting common weaknesses of the student model (see Appendix D.3 for the comprehensive scoring rubrics). The model-based scores are first averaged equally across the three dimensions and then over predicted segments to form a narratological reward, which is combined with the rule-based format reward through a weighted sum as the final optimization objective. This reward design explicitly aligns the student model with narratological criteria for structural analysis. See Appendix H.3.2 for training configurations and reward details. While LitSeg achieves high-fidelity results through a sophisticated three-stage prompting pipeline, such a multi-turn process introduces significant computational overhead. For practical deployment, LitSeg-Lite performs segmentation with a single model call, eliminating LitSeg’s multistage prompting overhead while maintaining competitive performance. This offers a practical balance between narrative depth and operational efficiency. For prompts for LitSeg-Lite, see Appendix D.2. For inference configurations, see Appendix H.3.1.

Experiments

In this section, we comprehensively evaluate LitSeg and LitSeg-Lite on two narrative QA benchmarks, investigating their retrieval and generation performance (Subsection 4.1), assessing the intrin-

sic quality and structural independence of their text segments (Subsection 4.2), and validating the contribution of core design choices through ablation studies (Subsection 4.3). Details of our RAG pipeline are provided in Appendix H.1. 4.1

Retrieval and Generation Performance

Datasets and Baselines We employ two benchmarks for evaluation. LiteraryQA (Bonomo et al., 2025), a long-document narrative QA dataset focusing on diverse literary works (e.g., novels, narrative poetry, and drama), with all reported results evaluated on its official test split. However, as LiteraryQA lacks gold evidence-span annotations for precise retrieval evaluation, and to assess the generalization performance of LitSeg-Lite, which is trained on LitSeg annotations of the LiteraryQA training set in our experiments, we further adopt GutenQA (Duarte et al., 2024), a retrieval-focused benchmark featuring “needle-in-a-haystack” factual QA. Since LitSeg primarily serves as a highcapacity teacher for generating segmentation annotations rather than as a deployable chunker, we use LiteraryQA to report both teacher and student performance, while using GutenQA mainly to evaluate the cross-dataset generalization of the deployable LitSeg-Lite model. We compare our approaches against heuristic, embedding-based, and LLM-based segmentation baselines. Detailed baseline configurations are provided in Appendices G and H.5. Evaluation Metrics and Human Judgment We comprehensively assess retrieval and generation using lexical and LLM-based automatic metrics (See Appendices E and H.4 for metric definitions and implementation details). While reliable for structured, information-centric domains with well-defined evidence boundaries, these metrics systematically penalize answers containing enriched, contextually appropriate information not included in ground truths, as they only measure surface-level literal or factual matching against ground truths (Gerrits et al., 2026), hindering their validity on nuanced literary QA tasks. For instance, on GutenQA, our richer narrative context prompts the generator to produce highly accurate and informative, even indepth answers. This creates a systematic discrepancy with the concise, factoid-style reference labels, which are referenced by automated metrics in isolation from the broader context, thereby underestimating our method’s performance. Similar

phenomena have been observed in literary translation (Zhang et al., 2025) and other expert knowledge tasks (Szymanski et al., 2025). To obtain a reliable assessment, we therefore conduct a human pair-wise evaluation on GutenQA. Annotators compare LitSeg’s outputs against reference answers, with both judged against the original text within a ±500-character context window. The primary criterion is factual accuracy, followed by informativeness. See Appendix A for the full annotation guidelines and inter-annotator agreement. Furthermore, our manual audit uncovers five categories of inherent flaws in the existing benchmarks that further undermine automated evaluation: 1) ambiguous or erroneous questions; 2) incomplete ground truth; 3) literal extractions misinterpreting narrative subtext; 4) narrow reference answers, and 5) entity and granularity mismatches. Detailed case studies are provided in Appendix C. Table 1: Generation results on LiteraryQA and GutenQA. EM = Exact Match; R-L = ROUGE-L; Mtr = METEOR; AA = Answer Accuracy; WR = Pairwise Win Rate. Best results in bold. Method

GutenQA

LiteraryQA

R-L

Mtr

RAG Baselines Token Recursive Char Perplexity + Merge Margin-Sampling + Merge LumberChunker + Merge

0.583 0.552 0.603 0.618 0.652 0.594 0.584 0.588

0.066 0.078 0.072 0.079 0.065 0.080 0.076 0.077

0.240 0.253 0.235 0.253 0.205 0.255 0.250 0.252

0.248 0.259 0.245 0.259 0.216 0.262 0.255 0.257

0.296 0.298 0.280 0.302 0.247 0.299 0.299 0.303

0.418 0.415 0.373 0.423 0.308 0.418 0.435 0.438

Ours LitSeg-Lite LitSeg

* —

0.079 0.081

0.258 0.256

0.263 0.261

0.311 0.309

0.448 0.450

Table 2: Retrieval results on GutenQA and LiteraryQA. CR = Context Relevance; MRR = Mean Reciprocal Rank; H@k = Hit@k. Best results in bold.

Results and Analysis Tables 1 and 2 report the main results on both benchmarks. Our proposed methods consistently achieve strong performance across generation and retrieval metrics, validating the effectiveness of our narrative-theory-guided segmentation framework. Notably, significant gains are observed in Context Relevance, where LitSeg-Lite outperforms the best baselines by substantial margins (+0.045 on GutenQA; +0.053 on LiteraryQA), demonstrating that narratological boundaries effectively preserve plot-complete passages. This segmentation advantage is further reflected in retrieval performance, evidenced by the highest MRR (0.683) and substantial improvements in Hit@k metrics (e.g., 0.813 on H@5 for GutenQA), indicating that our method not only retrieves the correct chunks but also ranks them significantly higher. Consequently, this precise retrieval and prioritization of relevant context greatly empowers the generator to utilize necessary information, confirmed by LitSeg achieving the highest Answer Accuracy on LiteraryQA (0.450), surpassing the best baseline (0.438). On GutenQA, Litseg consistently outperforms all alternative metrics in pair-wise evaluations (e.g., securing a 65.2% win rate against Margin-Sampling). Furthermore, the distillation process proves highly effective, as LitSeg-Lite not only matches but slightly exceeds LitSeg’s Context Relevance (0.805 vs. 0.799) and achieves comparable generation scores (e.g., Answer Accuracy of 0.448 vs. 0.450). This confirms that LitSeg-Lite successfully internalizes the narrative-aware segmentation logic of LitSeg in a single, efficient pass using a compact model. To qualitatively illustrate how our framework preserves narrative integrity and enhances downstream QA performance, we provide case studies and detailed textual analyses in Appendix B. See Appendix F for more metrics on GutenQA. 4.2

Method

GutenQA

LitQA

CR MRR H@1 H@2 H@3 H@5 H@20

RAG Baselines Token 0.885 0.572 0.478 0.589 0.654 0.728 Recursive Char 0.871 0.614 0.531 0.631 0.689 0.748 Perplexity 0.861 0.555 0.463 0.573 0.640 0.702 + Merge 0.889 0.599 0.514 0.619 0.679 0.734 Margin-Sampling 0.781 0.383 0.309 0.392 0.446 0.510 + Merge 0.877 0.593 0.509 0.605 0.668 0.732 LumberChunker 0.885 0.633 0.552 0.652 0.708 0.761 + Merge 0.893 0.656 0.574 0.674 0.736 0.785 Ours LitSeg-Lite LitSeg

0.728 0.748 0.702 0.734 0.510 0.732 0.761 0.785

0.654 0.664 0.665 0.724 0.564 0.714 0.739 0.752

0.938 0.683 0.600 0.701 0.759 0.813 0.813 — — — — — — —

0.805 0.799

Segment Quality

To establish that our narratological-theory-guided framework yields text chunks with superior structural and thematic independence, we conduct an intrinsic evaluation of the segmentation quality. Specifically, we measure the boundary semantic transition by calculating the cosine similarity between the sentences immediately preceding and succeeding each predicted boundary. As illustrated in Table 3, LitSeg-Lite achieves the lowest average boundary similarity across both benchmarks (0.4025 on GutenQA and 0.3699

Table 3: Average semantic similarity of adjacent sentences across predicted segment boundaries. Lower values indicate better segmentation independence; best results are in bold.

Table 6: Ablation study on segmentation quality (average boundary semantic similarity). Best results are in bold. Method

Method

GutenQA

LiteraryQA

RAG Baselines Token Recursive Char Perplexity + Dynamic Merge Margin-Sampling + Dynamic Merge LumberChunker + Dynamic Merge

0.4536 0.4382 0.4257 0.4378 0.4365 0.4468 0.4135 0.4096

0.4047 0.4219 0.4064 0.4179 0.4148 0.4276 0.3805 0.3761

Ours LitSeg LitSeg-Lite

— 0.4025

0.3768 0.3699

LitSeg-Lite w/o FT +w/o Theo.S.1 +w/o Theo.S.2 +w/o Theo.S.3 +w/o Theory

on LiteraryQA). This empirical drop in crossboundary similarity indicates a sharper semantic transition, demonstrating that LitSeg successfully identifies profound plot transitions and narrative boundaries, thereby rendering the chunks highly self-contained and informative for retrieval. 4.3

Ablation Studies

Table 4: Ablation study on LiteraryQA and GutenQA generation. Best results in bold. Method

LitSeg-Lite w/o FT +w/o Theory +w/o Theo.S.1 +w/o Theo.S.2 +w/o Theo.S.3

GutenQA

LiteraryQA

R-L

Mtr

* 0.593 0.584 0.553 0.564 0.548

0.079 0.077 0.073 0.074 0.076 0.078

0.258 0.253 0.250 0.248 0.250 0.255

0.263 0.256 0.255 0.252 0.254 0.259

0.311 0.306 0.301 0.298 0.300 0.308

0.448 0.446 0.438 0.441 0.441 0.445

Table 5: Ablation study on GutenQA and LiteraryQA retrieval. Best results in bold. Method

GutenQA

LitQA

CR MRR H@1 H@2 H@3 H@5 H@20 LitSeg-Lite 0.938 0.683 0.600 0.701 0.759 0.813 w/o FT 0.927 0.646 0.550 0.670 0.737 0.797 +w/o Theory 0.936 0.632 0.536 0.657 0.720 0.783 +w/o Theo.S.1 0.927 0.634 0.532 0.656 0.728 0.794 +w/o Theo.S.2 0.922 0.628 0.527 0.647 0.719 0.792 +w/o Theo.S.3 0.925 0.625 0.531 0.648 0.709 0.774

0.813 0.796 0.783 0.794 0.792 0.774

CR 0.805 0.787 0.764 0.791 0.781 0.771

To validate the contributions of distilled knowledge and narratological theory, we ablate LitSegLite into five variants: w/o FT (retains the theory prompt but removes fine-tuning), +w/o Theo.S.1/2/3 (each ablates one specific theorydriven stage beyond w/o FT), and +w/o Theory (replaces the entire narratological pipeline with a generic prompt beyond w/o FT). See Appendix D.4 for detailed prompt templates.

GutenQA

LiteraryQA

0.4025 0.4350 0.4349 0.4385 0.4403 0.4324

0.3699 0.3989 0.3941 0.4012 0.4031 0.3991

As shown in Tables 4 and 5, both distillation and the multi-step narratological pipeline are indispensable. Removing fine-tuning or skipping any individual theory-driven stage consistently degrades performance, with retrieval metrics suffering the most pronounced declines. On GutenQA, MRR drops from 0.683 to 0.646 (w/o FT), 0.634 (+w/o Theo.S.1), 0.628 (+w/o Theo.S.2), and 0.625 (+w/o Theo.S.3), with the fully generic prompt (+w/o Theory) yielding a comparable 0.632. A similar pattern emerges on LiteraryQA, where Context Relevance degrades from 0.805 to 0.787 (w/o FT) and further to 0.764 (+w/o Theory), indicating that narratological knowledge is the primary driver of retrieval quality. These retrieval deficits cascade to downstream generation: Answer Accuracy on LiteraryQA decreases from 0.448 to 0.438 (+w/o Theory), while pair-wise evaluation on GutenQA is consistently defeated by the full model, securing a win rate of 58.4%(+w/o Theory). Ablating one component even incurs degraded performance compared to +w/o Theory, confirming that the holistic integration of both distillation and theory-guided pipelines are indispensable for robust document retrieval and high-fidelity generation. Beyond downstream performance, we conduct an intrinsic evaluation of segmentation boundary quality using average boundary semantic similarity (Table 6), where lower values indicate more semantically distinct cuts. Both fine-tuning and narratological prompting reduce this similarity across benchmarks (e.g., from 0.4324 to 0.4025 on GutenQA and from 0.3991 to 0.3699 on LiteraryQA), confirming that our framework is essential for regularizing the model to produce selfcontained, semantically coherent text segments.

Conclusion

In this work, we address the critical challenge of document segmentation for literary RAG systems, where existing methods overlook higher-order se-

mantics and severely disrupt narrative coherence, resulting in suboptimal performance in QA systems. We introduce LitSeg, a narrative-theoryguided framework employing multi-stage prompting to explicitly extract narrative structures, complemented by a novel index-based output schema that reduces autoregressive overhead and enables flexible text reconstruction. To mitigate the computational cost of this multi-stage pipeline, we further present LitSeg-Lite, a lightweight single-pass chunker distilled via a two-stage SFT and RL strategy. Extensive experiments on GutenQA and LiteraryQA demonstrate that our proposed methods substantially outperform baselines, yielding significant gains in context relevance, retrieval accuracy, and downstream answer accuracy. Ablation studies confirm the substantial contributions of both narratological guidance and knowledge distillation; notably, distillation not only transfers the teacher’s segmentation logic but also ensures near-perfect output stability. Together, LitSeg and LitSeg-Lite bridge the gap between narrative theory and practical RAG deployment, providing a robust and efficient foundation for understanding literary texts.

Limitations While our framework demonstrates strong performance on English literary texts, its current evaluation is confined to English. Generalizing LitSeg to multilingual contexts remains an avenue for future research. Furthermore, dedicated benchmarks for the literary domain remain relatively scarce. Despite achieving significant empirical gains across evaluated metrics, our evaluation is, as discussed previously, partially constrained by the inherent quality flaws prevalent in existing datasets. These benchmark-level limitations prevent the quantitative improvements from fully reflecting our method’s true capabilities. A detailed qualitative analysis of these flawed scenarios is presented in Appendix C. Developing more robust benchmarks for literary texts thus remains another vital direction for future research.

References Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Transla-

tion and/or Summarization, pages 65–72, Ann Arbor, Michigan. Association for Computational Linguistics. Steven Bird and Edward Loper. 2004. NLTK: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, pages 214–217, Barcelona, Spain. Association for Computational Linguistics. Tommaso Bonomo, Luca Gioffré, and Roberto Navigli. 2025. LiteraryQA: Towards effective evaluation of long-document narrative QA. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 34086–34107, Suzhou, China. Association for Computational Linguistics. Claude Bremond and Elaine D Cancalon. 1980. The logic of narrative possibilities. New Literary History, 11(3):387–411. Samuel Henry Butcher and 1 others. 1902. The poetics of Aristotle. Macmillan. Harrison Chase. 2022. LangChain. Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V. Le, Sergey Levine, and Yi Ma. 2025. Sft memorizes, rl generalizes: A comparative study of foundation model post-training. Preprint, arXiv:2501.17161. André V. Duarte, João DS Marques, Miguel Graça, Miguel Freire, Lei Li, and Arlindo L. Oliveira. 2024. LumberChunker: Long-form narrative document segmentation. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6473– 6486, Miami, Florida, USA. Association for Computational Linguistics. Edward Morgan Forster. 1927. Aspects of the Novel. Harcourt, Brace. Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1762–1777, Toronto, Canada. Association for Computational Linguistics. Gérard Genette. 1980. Narrative discourse: An essay in method, volume 3. Cornell University Press. Kyo Gerrits, Rik van Noord, and Ana Guerberof Arenas. 2026. Creativity bias: How machine evaluation struggles with creativity in literary translations. Preprint, arXiv:2605.13596. Evelyn Gius and Michael Vauth. 2022. Towards an event based plot model. a computational narratology approach. Journal of Computational Literary Studies, 1(1). Michael Hauge. 2017. Storytelling Made Easy: Persuade and Transform Your Audiences, Buyers, And Clients-Simply, Quickly, and Profitably. BookBaby.

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. CoRR, abs/2106.09685. Haoyu Huang, Yongfeng Huang, Yang Junjie, Zhenyu Pan, Yongqiang Chen, Kaili Ma, Hongzhi Chen, and James Cheng. 2025. Retrieval-augmented generation with hierarchical knowledge. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 6044–6060, Suzhou, China. Association for Computational Linguistics. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. Genevieve Liveley. 2019. Narratology. Oxford University Press. Yusong Ma, Hongxuan Nie, Chao Chen, Jiujie Zhang, Jiali Jiang, Bisheng Wang, and Yuqin Xia. 2025. A survey of retrieval-augmented generation (rag) for large language models. In 2025 International Conference on Trustworthy Big Data and Artificial Intelligence (ICTBAI), pages 7–13. Uri Margolin. 2014. Simultaneity in Narrative, pages 777–786. De Gruyter, Berlin, München, Boston. Pinelopi Papalampidi, Frank Keller, and Mirella Lapata. 2019. Movie plot analysis via turning point identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1707–1717. Andrew Piper and Sunyam Bagga. 2024. Using large language models for understanding narrative discourse. In Proceedings of the 6th Workshop on Narrative Understanding, pages 37–46. Andrew Piper, Richard Jean So, and David Bamman. 2021. Narrative theory for computational narrative understanding. In Proceedings of the 2021 conference on empirical methods in natural language processing, pages 298–311. Saikrishna Rajanidi, M. Anbazhagan, and G. R. Ramya. 2026. Rag in specialized domains: A survey of qa chatbots. In Data Science and Applications, pages 352–369, Cham. Springer Nature Switzerland. Robert Scholes, James Phelan, and Robert Leland Kellogg. 2006. The nature of narrative: Revised and expanded. OUP USA. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. Preprint, arXiv:1707.06347.

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300. Dan Shen and Yali Wang. 2010. Western Narratology: Classical and Post-Classical. Peking University Press. In Chinese. Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic retrieval-augmented generation: A survey on agentic rag. Preprint, arXiv:2501.09136. Annalisa Szymanski, Noah Ziems, Heather A. EicherMiller, Toby Jia-Jun Li, Meng Jiang, and Ronald A. Metoyer. 2025. Limitations of the llm-as-a-judge approach for evaluating llm outputs in expert knowledge tasks. In Proceedings of the 30th International Conference on Intelligent User Interfaces, IUI ’25, page 952–966, New York, NY, USA. Association for Computing Machinery. Kristin Thompson. 1999. Storytelling in the new Hollywood: Understanding classical narrative technique. Harvard University Press. Tzvetan Todorov and Arnold Weinstein. 1969. Structural analysis of narrative. In NOVEL: A forum on fiction, volume 3, pages 70–76. JSTOR. Yiteng Tu, Weihang Su, Yujia Zhou, Yiqun Liu, and Qingyao Ai. 2025. Robust fine-tuning for retrieval augmented generation against retrieval defects. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’25, page 1272–1282, New York, NY, USA. Association for Computing Machinery. Noa Visser Solissa, Andreas van Cranenburgh, and Federico Pianzola. 2025. Event detection between literary studies and nlp: A survey, a narratological reflection, and a case study. Piek Vossen, Tommaso Caselli, and Roxane Segers. 2021. A narratology-based framework for storyline extraction. Computational Analysis of Storylines: Making Sense of Events, 125:125–140. Eitan Wagner, Renana Keydar, Amit Pinchevski, and Omri Abend. 2024. Automatic topic-guided segmentation of holocaust survivor testimonies. Journal of Computational Literary Studies, 2(1). Ke Wang, Xiutian Zhao, Yanghui Li, and Wei Peng. 2023. M3seg: A maximum-minimum mutual information paradigm for unsupervised topic segmentation in asr transcripts. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7928–7934.

Zhitong Wang, Cheng Gao, Chaojun Xiao, Yufei Huang, Shuzheng Si, Kangyang Luo, Yuzhuo Bai, Wenhao Li, Tangjian Duan, Chuancheng Lv, Guoshan Lu, Gang Chen, Fanchao Qi, and Maosong Sun. 2025. Document segmentation matters for retrievalaugmented generation. In Findings of the Association for Computational Linguistics: ACL 2025, pages 8063–8075, Vienna, Austria. Association for Computational Linguistics.

thorough guideline was provided to maintain consistency across all evaluations. For each sample, the tri-binary preference (Win/Loss/Tie) of each of the five annotators was recorded. We then computed the average win and loss frequencies across all five annotators for each instance as the overall win rate of our strategy.

Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, and 16 others. 2025. Dapo: An open-source llm reinforcement learning system at scale. Preprint, arXiv:2503.14476.

A.2

Ran Zhang, Wei Zhao, and Steffen Eger. 2025. How good are LLMs for literary translation, really? literary translation evaluation with humans and LLMs. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 10961– 10988, Albuquerque, New Mexico. Association for Computational Linguistics. Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Guoyin Wang, and Fei Wu. 2026. Instruction tuning for large language models: A survey. ACM Comput. Surv., 58(7). Jihao Zhao, Zhiyuan Ji, Yuchen Feng, Pengnian Qi, Simin Niu, Bo Tang, Feiyu Xiong, and Zhiyu Li. 2024. Meta-chunking: Learning text segmentation and semantic completion via logical perception. Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, Jie Jiang, and Bin Cui. 2026. Retrieval-augmented generation for ai-generated content: A survey. Data Science and Engineering, 11(1):1–29.

Details of Pairwise Evaluation

A.1

Expert Information and Annotation Procedure

We recruited five volunteers to participate in the evaluation. All volunteers are graduate students majoring in English Literature, ensuring they possess the necessary domain expertise, linguistic sensitivity, and critical capacity to judge the nuances of the text. The evaluation was conducted under strict quality control to ensure objectivity and minimize bias. We made sure all annotators completed the tasks independently; no communication, discussion, or collaboration was permitted among the volunteers during the entire annotation process. A

Guideline for Pairwise Evaluation

Thank you for participating in this evaluation. As a literary professional, your scholarly expertise is invaluable in assessing the accuracy, depth, and nuance of these generated responses. Please use the following guidelines to inform your judgment. I. Evaluation Principle Core Principle: "Factual Accuracy as the Baseline, Informativeness as the Deciding Factor." Please rely entirely on the provided Original Context as the ultimate textual evidence. The Reference Answer is provided merely as an auxiliary guide and may contain interpretive or factual flaws; please apply your own scholarly judgment and do not blindly trust it. II. Evaluation Dimensions and Criteria 1. Dimension 1: Factual Accuracy and Textual Fidelity This dimension serves as the absolute baseline. A response that meets any "Fail" condition is deemed factually invalid and fails the correctness check. 1.1 Textual Fidelity vs. Unwarranted Invention: a) Pass: The response is strictly anchored in the provided text and its claims are fully traceable to the source. b) Fail: The response invents unmentioned plot points, fabricates character actions, or engages in baseless subjective extrapolation beyond the textual evidence. 1.2 Decoding Narrative Subtext and True Intention: a) Pass: The response successfully penetrates characters' surface-level discourse (e.g., deception, pretexts, politeness) to accurately reveal the underlying narrative motives and dramatic irony. b) Fail: The response relies superficially on literal readings, mistakenly accepting character deceit or posturing as truth, thereby distorting the narrative logic. 1.3 Critical Engagement with Flawed Questions: a) Pass: When confronted with questions containing factual distortions or ambiguous referents, the response actively corrects the flawed premise using textual evidence. b) Fail: The response succumbs to the query's flawed assumptions, resulting in an invalid or logically contradictory reading of the text. 2. Dimension 2: Informativeness and Interpretive Depth Assuming both candidate responses meet the baseline of factual accuracy, this dimension serves to distinguish the superior reading. 2.1 Horizontal Breadth: Comprehensive Coverage of Parallel Facts: Superior: For complex events driven by concurrent motivations or parallel actions, the response synthesizes a comprehensive mosaic of textual clues. This outperforms reductive answers that extract only a single, isolated fact. 2.2 Longitudinal Breadth: Closure of Sequential Narrative Arcs: Superior: The response demonstrates an ability to synthesize information across the text, capturing the complete narrative arc from its initial setup to its definitive resolution. This outperforms fragmented answers that prematurely truncate an ongoing action. 2.3 Precision: Specificity of Textual Detail:

Superior: Supplying a more precise, fine-grained proper noun or specific descriptive detail than the other response demonstrates a higher-resolution close reading and warrants a tie-breaking win. III. Scoring Rules Each pairwise comparison is evaluated as a zero-sum game ( awarding 1 point total per pair): 1. Win (Winner: 1.0, Loser: 0.0): a) Superiority in Accuracy: One candidate response passes the baseline check, while the other exhibits one or more "Fail" conditions. b) Superiority in Informativeness: Both candidate responses pass the baseline accuracy check, but the winning response meets a "Superior" condition by demonstrating greater narrative breadth or more precise textual detail. 2. Tie (0.5 points each): Both candidate responses meet a " Fail" condition, or both are highly homogeneous in their accuracy and informativeness.

A.3

Inter-Annotator Agreement

See Fig 2. 1.0 =0.6109 Ag=74.85%

=0.5581 Ag=71.31%

=0.5512 Ag=70.85%

=0.5281 Ag=69.38%

=0.5228 Ag=69.00%

=0.5504 Ag=70.77%

=0.5001 Ag=67.54%

=0.4938 Ag=67.00%

=0.4873 Ag=66.62%

=0.4902 Ag=66.77%

=0.6109 Ag=74.85%

=0.5581 Ag=71.31%

=0.5228 Ag=69.00%

=0.5512 Ag=70.85%

=0.5504 Ag=70.77%

=0.4938 Ag=67.00%

0.6

Cohen's kappa

Annotator

0.8

0.4

0.2

=0.5281 Ag=69.38%

=0.5001 Ag=67.54%

=0.4873 Ag=66.62%

=0.4902 Ag=66.77%

A3 Annotator

0.0

Figure 2: Inter-annotator agreement heatmap for the human pair-wise evaluation on GutenQA (N = 1300). This figure reports the pairwise Cohen’s kappa (κ) and raw agreement percentages (Ag) among five independent annotators (A1, A2, A3, A4, and A5). The scores demonstrate a reliable consensus, validating the robustness of the human judgment used to mitigate the inherent limitations of automated metrics in complex literary QA tasks.

Case Study on Baseline Chunking Failure Modes

To provide a deeper qualitative understanding of how our framework preserves narrative integrity and enhances downstream question-answering (QA) performance, we present three representative case studies from our benchmarks. These examples contrast baseline segments against our distilled model, LitSeg-Lite, illustrating three critical failure modes of baseline chunking methods: premature

event termination, mid-action truncation, and ambiguous pronoun reference. B.1

Overcoming Premature Event Termination and Lack of Macro-Event Perspective

The case study on Barchester Towers by Anthony Trollope (illustrated in Fig 3a) exemplifies the failure mode of premature event termination. When answering a question regarding Mrs. Bold’s reaction to Mr. Slope’s proposal, traditional segmentation methods impose a rigid boundary that ends prematurely right after Eleanor expresses her verbal disgust (“. . . unable to bear much more of it.”). By terminating the segment early, the baseline fails to capture the macro-perspective of the event— specifically, its physical resolution. Deprived of the full event horizon, the baseline QA model can only infer an answer based solely on her verbal refusal, delivering a fundamentally incomplete and incorrect response. Conversely, despite being a highly compact model, LitSeg-Lite successfully internalizes macro-event boundaries by encapsulating the entire proposal arc within a single chunk. By preserving the definitive physical resolution where she slaps him (“. . . dealt him a box on the ear. . . ”), LitSeg-Lite equips the downstream generator with the necessary context to formulate a precise and complete answer. B.2

Eliminating Mid-Action Truncation and Omission of Crucial Information

As demonstrated in the analysis of The Vampyre by John Polidori in Fig 3b, non-narratological approaches frequently suffer from mid-action truncation, which breaks the textual sequence in the middle of a continuous plotline. When querying how Aubrey attempts to warn his sister, the baseline chunker cuts the text abruptly at “but in vain.”, entirely omitting the crucial information that immediately follows: the acquisition of “pen and paper” and the core action of “writing a letter”. This severe information bottleneck leaves the generator in an information vacuum, forcing the downstream LLM to hallucinate a highly vivid but entirely fabricated sequence regarding the trampling of a portrait. LitSeg-Lite effectively mitigates this issue by respecting the sequential integrity of continuous actions. LitSeg-Lite keeps the setup, the intent, and the final action structurally intact within the same context window, thereby ensuring clean and complete retrieval of the golden evidence.

【Std Ans: She slaps him.】

Q: What does Mrs. Bold do when Mr. Slope proposes? Other Method Retrieves:

Our Method Retrieves:

'Ah! Eleanor, will it not be sweet with the Lord's assistance, to travel hand in hand through this mortal valley which his mercies will make pleasant to us, till hereafter we shall dwell together at the foot of his throne?‘ ...'Ah! Eleanor—’ 'My name, Mr Slope, is Mrs Bold,' said Eleanor, who, though determined to hear out the tale of his love, was too much disgusted by his blasphemy to be able to bear much more of it.

... ‘Ah! Eleanor, will it not be sweet with the Lord’s assistance, to travel hand in hand through this mortal valley which his mercies will make pleasant to us, till hereafter we shall dwell together at the foot of his throne?‘ ...’Ah! Eleanor—’ ‘My name, Mr Slope, is Mrs Bold,’ said Eleanor, who, though determined to hear out the tale of his love, was too much disgusted by his blasphemy to be able to bear much more of it. ‘Sweetest angel, be not so cold,’ said he, and as he said it the champagne broke forth, and he contrived to pass his arm around her waist. …She sprang from him as she would have jumped from an adder, …and then, quick as thought, she raised her little hand and dealt him a box on the ear with such right good will, that it sounded among the trees like a miniature thunder-clap. …

Other Method Answers:

Our Method Answers:

Clifford's own death will prove his innocence.

The Judge's death, which occurred in a manner similar to his ancestor's and without the suspicious circumstances that seemed to have orchestrated Clifford's uncle's death, will prove Clifford's innocence.

à WRONG ANSWER

à CORRECT ANSWER from The House of the Seven Gables by Nathaniel Hawthorne

(c) Case study on The House of the Seven Gables. Our method correctly links narrative evidence while baseline method misunderstands the text without sufficient context.

Figure 3: Qualitative comparison of text retrieval and QA results (Continued). Figures (a), (b), and (c) demonstrate how LitSeg preserves narrative integrity and surpasses baseline methods in downstream performance.

B.3 Resolving Ambiguous Pronoun Reference and Context Deprivation

Case Study on Benchmark Quality Issues

C.1 Ambiguous Phrasing or Factual Errors in Questions Finally, the text from The House of the Seven Gables by Nathaniel Hawthorne, displayed in Fig 3c, highlights a pervasive linguistic vulnerability in baseline text chunking methods: ambiguous pronoun reference resulting from context deprivation. As marked by the dashed circle in the figure, the baseline segment begins abruptly with an orphaned possessive pronoun (“His own death. . . ”). Because the upstream context containing the pronoun’s actual antecedent (the Judge) is completely cut off, the baseline QA model suffers from reference misalignment. Lacking the necessary contextual tracking to resolve the anaphora, the model misattributes the pronoun to Clifford, culminating in a logically inverted and erroneous deduction. LitSeg-Lite elegantly circumvents this limitation by dynamically anchoring its boundaries around major narrative shifts. By preserving the preceding context that introduces the correct referent, LitSeg-Lite maintains proper referential chains, allowing the reader model to accurately synthesize the multi-layered plotline and correctly establish Clifford’s innocence.

As highlighted in Fig 4, Some questions in the dataset are ambiguous, flawed or incorrect, making them unanswerable. On one hand, the benchmark introduces severe linguistic ambiguity by employing unresolvable referents and incomplete semantic predicates. Queries containing indeterminate pronouns (e.g., "he") or isolated definite descriptions (e.g., "the dog") fail to provide the localized context necessary for proper coreference resolution, leaving any evaluation of a model’s reading comprehension capacity completely invalid. On the other hand, the dataset incorporates overt factual errors and misleading premises that directly invert or distort the logical relationships explicitly stated in the source text. For example, by truncating the phrasal verb "break out" to "break," the question severely distorts the character’s physical action. This is further compounded by blatant narrative contradictions, such as asking how certain characters convinced a counterparty to pay them in gold, when the transaction is exactly opposite. Forcing language models to navigate structurally

Flaw 1. Ambiguous Phrasing or Factual Errors in Questions Question: What type of dog is he trained to be in Canada?

Who is “he”? Question: Who stole the dog?

Which dog? Question: What did Buck have to break to win the wager?

“Break”? Or “break out”? [Original Text] “Buck can start a thousand pounds.” “And break it out? and walk off with it for a hundred yards?” demanded Matthewson... "And break it out, and walk off with it for a hundred yards”… Question: How did Red Pierre and his companion convince McGuire to give them clothes and pay them in gold?

Factual error!!! It should be Red Pierre and his companion paying McGuire gold! [Original Text] Jacqueline took one of them and Pierre the other under his left arm... “…here's what the clothes are worth to us.” And into the quaking hands of McGuire he poured a chinking stream of gold pieces.

Figure 4: An illustration of dataset flaws resulting from ambiguous phrasing and factual errors in questions.

broken and factually misleading queries severely compromises the integrity of the benchmark, penalizing high-performing models that detect these contradictions, while artificially rewarding models that succumb to hallucination or accept the query’s flawed premises, thereby undermining the validity of the evaluation metric. C.2

provided answers fail to capture the full scope of valid responses. This flaw manifests as a severe under-reporting of facts, where questions with multiple, equally correct answers are falsely restricted to a single target answer in the dataset. By omitting alternative valid ground truths, this deficiency creates a misleading evaluation bottleneck, unfairly punishing models that successfully extract complete and accurate information from the source text, hindering the system’s ability to learn comprehensive multi-perspective reasoning. C.3

Literal Extractions Misinterpreting Narrative Subtext Flaw 3. Literal Extractions Misinterpreting Narrative Subtext Question: Why does Elizabeth's father express concern about her acceptance of Mr. Darcy despite his wealth? Answer: Elizabeth's father is worried because he believes Mr. Darcy is a proud and unpleasant man. [Original Text] Have you any other objection,” said Elizabeth, “than your belief of my indifference?” “None at all. We all know him to be a proud, unpleasant sort of man; but this would be nothing if you really liked him…. But let me advise you to think better of it. I know your disposition, Lizzy. I know that you could be neither happy nor respectable, unless you truly esteemed your husband, unless you looked up to him as a superior. Your lively talents would place you in the greatest danger in an unequal marriage. You could scarcely escape discredit and misery.

Better Answer: Because he believes she dislikes Darcy, warning that her lively disposition requires a husband she truly esteems to avoid the misery and discredit of an unequal marriage.

Incomplete Ground-Truth Coverage Flaw 2. Incomplete Ground-Truth Coverage Question: What is the name of the dog that had to be shot? Answer: Dave

Fact: (1) Dave (2) Dub Question: What does Reinhard do to impress Elisabeth? Answer: Writes fairy tales.

Fact: (1) Writes fairy tales. (2) Teaches her botany. (3) ……

Figure 5: An illustration of the incomplete ground-truth coverage flaw within the dataset

As illustrated in Fig 5, the dataset also suffers from incomplete ground-truth coverage, where the

Figure 6: An illustration of dataset flaws stemming from literal extractions misinterpreting narrative subtext.

As demonstrated in Fig 6, the dataset fails to understand the connotations behind the text, but relies on superficial keyword matching to construct ground-truth answers, resulting in a severe reduction of literary meaning. In the first case, the dataset’s literal answer mistakes a father’s voiced observation for his actual motive, failing to decode that his true anxiety stems from his daughter’s emotional indifference and the possibility of an unequal marriage. In the second case, the dataset naively treats a character’s deceptive verbal pretense of "delivering a message" as a literal fact, completely blind to the narrative subtext and subsequent dramatic irony where the "message" is a fatal ambush

Flaw 3. Literal Extractions Misinterpreting Narrative Subtext (Continued) Question: Why did Pierre ask for McGurk at Gaffney's place? Answer: Because he had a message specifically meant for McGurk, which he did not want to pass through someone else. [Original Text] 'Where's McGurk?' ...'What do you want with him?' 'Got a message for him.' 'Tell it to me, and I'll pass it along.' Pierre met the eye of the other and smiled faintly. 'Not this message.' 'Oh,' said the other, and then shouted: 'McGurk!' …The door opened and framed McGurk. He did not start, seeing Pierre. He said: "None of the rest of them had the guts even to bring me the message, eh?"..."That's not the message," answered a voice which Pierre did not recognize as his own. "Out with it, then." "It's in the leather on my hip." And he went for his gun...he scooped up the gun with his left and twisted. That movement made the third shot of McGurk fly wide and Pierre fired from the floor and saw a spasm of pain contract the face of the outlaw.

Better Answer: Pierre asked for McGurk under the pretense of delivering a message, but his true intention was to confront and shoot him with the gun hidden on his hip.

Figure 6: An illustration of dataset flaws stemming from literal extractions misinterpreting narrative subtext (Continued).

with a hidden firearm. By evaluating models based on these literal readings, the dataset heavily penalizes advanced language systems capable of deep narrative synthesis, ironic interpretation, and intent modeling, thereby limiting its utility for complex literary reading comprehension tasks. C.4

Incomplete Reference Answers with Insufficient Narrative Coverage

As demonstrated in Fig 7, the text constructs Ryder’s motivations across two distinct phases: first, his letter explicitly states a dual purpose (revealing Pierre’s illegitimate birth and seeing him before dying); second, upon meeting, he makes his ultimate concrete request (burial in a specific plot). Despite the significant narrative span between these events, a valid reading comprehension evaluation requires cross-contextual synthesis. By fragmenting this continuous arc, the incomplete ground truth unfairly penalizes models capable of multi-hop evidence integration. C.5

Entity Name Variations and Granularity Mismatches

As illustrated in Fig 8, the dataset enforces a single, rigid ground-truth string, failing to account for the

Flaw 4. Reference Answers Lacking Narrative Breadth Question: What is the main reason Martin Ryder is reaching out to Pierre? Answer: Because he fears dying alone and unattended, with his sons waiting to take his savings and leave him to rot, hoping that Pierre might come before he passes away. [Original Text] (1) 'Morgantown, R.F.D. No. 4. SON PIERRE:...So I'm writing to you, Pierre, part to tell you what you ought to know; part because I got a sort of crazy idea that maybe you could get down here to me before I go out. (2) Son, you didn‘t come none too soon. ...Here’s all I ask: ...And down in the middle of Morgantown is the buryin‘-ground. I’ve ridden past it a thousand times an‘ watched a corner plot, where the grass grows quicker than it does anywheres else in the cemetery. Pierre, I’d die plumb easy if I knew I was goin‘ to sleep the rest of time in that place."

Our Answer: The main reason Martin Ryder is reaching out to Pierre is to reveal that Pierre is his illegitimate son and to ask Pierre to help him arrange his burial in a specific corner plot at the Morgantown graveyard in his dying moments.

Figure 7: An illustration of dataset flaws due to incomplete reference answers with insufficient narrative coverage.

Flaw 5. Entity Name Variations and Granularity Mismatches Question: Who steals Buck from Judge Miller's home? Answer: The gardener's assistant.

Our Answer: Manuel, a gardener's helper who owed money to a stranger, steals Buck from Judge Miller's home. Question: Who was Eva? Answer: Daughter of a Jewish financier that Tancred met.

Our Answer: Eva was the daughter of Besso, …. Question: What defeats Zog? Answer: The forces of good.

Our Answer: King Anko defeats Zog by catching him in his coils and crushing his magic power with fear, causing him to turn into a mass of jelly-like pulp and die.

Figure 8: An illustration of dataset flaws rooted in entity name variations and granularity mismatches.

multi-dimensional ways a correct answer can be expressed, and consequently penalizing semantically valid responses. This rigidity becomes particularly problematic when dealing with entity name variations, where a single narrative character is naturally referenced across a text via diverse linguistic expressions, such as proper names, titles, or descriptive phrases. For instance, a model extracting the precise proper noun "Manuel" or "Besso" is falsely penalized simply because the ground truth records the functional equivalents "the gardener’s assistant" or "a Jewish financier." This problem is further exacerbated when these variations shift across levels of narrative scale, introducing profound granularity mismatches where valid responses operate at differing tiers of abstraction. Depending on the depth of text synthesis, an event may be legitimately framed macroscopically through broad thematic forces and collective factions, or microscopically through individual agents and localized actions. As demonstrated in the text, attributing an antagonist’s defeat to "the forces of good" versus the physical intervention of "King Anko" represents a divergence in granularity rather than accuracy, yet both assertions remain factually grounded. By failing to bridge these parallel axes of linguistic variation and narrative scale, the evaluation metric introduces substantial noise, systematically penalizing advanced models that capture high-precision, fine-grained entities.

Full Prompts

D.1

Prompt for LitSeg

System Prompt Shared Across Stages 1–3 # Role: Narratology Segmentation Expert ## Profile: - Language: English - Description: You are a professional literary editor and Natural Language Processing expert deeply versed in structuralism and narratology. Your core task is to perform fine-grained plot segmentation on literary texts. ## Core Competencies 1. Fine-grained plot segmentation preserving original literary aesthetics. 2. Extracting valid dynamic events and effectively filtering out static or non-narrative noise. 3. Untangling complex multi-line, parallel, and nested narrative structures based on "Unity of Action" and internal story logic. 4. Pinpointing logical breakpoints by synthesizing macrostructural shifts (spatial, temporal, tension-based, etc.) with micro-level evaluations of semantic dependency. ## Primary Objectives 1. Perform logical, narratology-driven plot segmentation on

literary texts. 2. Enable readers to independently read and understand each segmented snippet out of its full context. 3. Output the exact analytical and segmentation results in strict JSON format. ## Constraints 1. Role Consistency: Don't break character or bypass the narratological framework under any circumstance. 2. Narrative-Driven Granularity: The recommended maximum length for any single segment is 100 sentences. This is a guidance threshold rather than an absolute hard cap. When narrative coherence clearly requires it, a segment may exceed this limit. NEVER unnaturally merge distinct narrative blocks just to hit a sentence count. Abandon any mathematical division mindset. Breakpoints must be dictated organically by narratological theories. 3. Conditional Contextual Overlap: A small overlap of key sentences between adjacent plot blocks is allowed ONLY to ensure narrative continuity. This is NOT mandatory. If a natural narrative breakpoint is clear and coherent, make a clean, hard cut. Do not blindly copy previous sentences just for the sake of overlapping. 4. Descriptive Subtitle: Extract a highly concise subtitle that summarizes the core action or reversal event for each plot block. 5. Full Coverage and Order: You must segment the entire chapter in its original order. Do not skip any sentences or paragraphs. Every sentence must belong to at least one segment, i.e., covered in the index range between "from_idx" and "to_idx" of at least one segment. 6. Indexing Rule: The "from_idx" of the first segment MUST be 1. The "to_idx" of the last segment MUST be the last index of the chapter. Indices are 1-based and inclusive like [1] in the original text. 7. Context Indices: For sentences crucial for understanding a segment but outside its main range, include their indices in the "context_idx" field. Sentences cannot ONLY exist in "context_idx"; they must be part of some segment's main range. ## Narratological Workflow When determining logical breakpoints for plot segmentation, you must execute the following three-step narratological framework sequentially: ### Step 1: Extract Valid Events and Filter Noise 1. Focus on Dynamic Changes: Treat the plot as a sequence of dynamic events. Ignore static character portraits or pure scenery descriptions. Extract an event ONLY when characters or objects undergo dynamic changes. 2. Filter by Event Type: Retain "Changes of state" ( physical or mental state changes) and "Process events " (actions or happenings without state change, such as talking, thinking, and feeling) as core events. Discard "Stative events" (static physical or mental states) and "Non-events" (generic statements, counterfactuals, questions) to aggressively reduce noise. 3. Handling Ambiguous Paragraphs: For borderline paragraphs , gauge narrativity by checking for: a situation, an agent, one or more sequential actions, a potential object, a spatial location, a temporal specification, and a rationale. A higher density of these features indicates a valid, independent event. ### Step 2: Untangle Narrative Threads and Clarify Structures Group the events extracted in Step 1 into clear narrative threads using the following criteria: 1. Apply "Unity of Action": Group events by a complete core action, abandoning the "same character" stereotype. Separate unrelated actions performed by the same character; merge joint actions advanced by different characters. 2. Track the "Revelation of Secrets": Use the gradual unfolding of author-planted secrets as a tight internal logic to connect scattered events into a cohesive narrative chain. 3. Clarify Parallel & Minor Threads: Explicitly identify and segment concurrent event chains (e.g., A1 -> B1 -> A2 -> B2). Never merge minor threads into major ones; preserve their structural independence.

4. Identify Nested Hierarchy: Classify the thread's narrative level: [Extradiegetic] (story introduction) , [Intradiegetic] (core main plot), or [Metadiegetic] (a story within a story). 5. Establish Basic Data: For each thread, map the timeline, causeline, and critical shift points in time, space, and perspective. ### Step 3: Locate Key Turning Points and Execute Segmentation Use the data established in Step 2 to locate logical breakpoints. Apply these markers organically, segmenting the text only when shifts clearly and naturally occur: 1. Macro-Structural & Plot Tension: Track tension through the overarching arc (Exposition -> Predicament -> Extrication). Consider segmenting the text at the boundaries or transitions between these major driving events: - Opportunity: The introductory event that triggers the story after the background is set. - Change of Plans: The event where the main goal is defined and action intensifies. - Point of No Return: The event pushing characters to fully commit to their goal. - Major Setback & Reversal: The event where the situation completely falls apart, marking a formal reversal of fortune (peripeteia). This generally serves as a strong signal for segmentation. - Climax & Recognition: The final resolution and peak tension, often accompanied by a cognitive shift from ignorance to truth (anagnorisis). - Aftermath & Buffer: Post-climax events that resolve the predicament and lower tension, smoothly transitioning into the final Extrication phase. 2. Contextual Shifts: Segment at definitive leaps in time, space, or perspective. Cut when the narrative switches between parallel threads or crosses nested narrative levels. 3. Micro-Dynamic Shifts (State & Action): Segment when a situation's fundamental state transitions ( equilibrium -> disequilibrium -> new equilibrium) or at the precise moment an action's outcome (success/ failure) is revealed. 4. Sentence-Level Boundary Pinpointing: After narrowing down a general turning point, actively evaluate the semantic dependency between adjacent sentences to pinpoint the exact breakpoint: - High Semantic Cohesion Zone (Avoid Segmenting): Within the scope of the potential turning point, identify sentences with strong semantic dependency or interlocking logic. Examples include adjacency pairs (e.g., question/answer, attack/defense), continuous pronoun chains, or an immediate physical/emotional reaction to a specific action. Avoid executing a cut within these sequences. - Low Semantic Cohesion Zone (Execute the Cut): Within the scope of the potential turning point, locate the exact sentence where the semantic dependency drops to its lowest level. Make the cut right before the sentence that can semantically stand alone--one that initiates a new topic, phase, or action, and requires minimal context from the immediately preceding sentence to be fully understood. ## Output Specifications 1. JSON Format: You must output STRICT JSON format. Escape all internal double quotes within JSON string fields using a backslash. 2. Content Requirement: For each segmented plot block, output a concise subtitle summarizing the core action , and the index ranges based on the constraints. The exact JSON structure will be defined in the user prompt based on the specific Step being executed. ## Initialization: As a Narratology Segmentation Expert, you must strictly follow the Constraints and the Narratological Workflow. To initiate the analysis, await the provision of the target Step and the Chapter Text Content. Upon receipt, output the segmentation results exclusively in the requested STRICT JSON format.

Stage-1 User Prompt You are executing ONLY Step 1 of the Workflow: Extract Valid Events and Filter Noise. [Chapter Text Content] {chapter_content} Return STRICT JSON only (no markdown): { "step_1_events": "A string briefly listing the core dynamic events extracted (filtering out noise)." }

Stage-2 User Prompt You are executing ONLY Step 2: Untangle Narrative Threads and Clarify Structures. Use both the chapter text and Step 1 result as input context. [Chapter Text Content] {chapter_content} [Step 1 Result JSON] {step_1_result} Return STRICT JSON only (no markdown): { "step_2_threads": "A string briefly describing the narrative threads and their nested hierarchy, based on unity of action, secrets, and key shifts in time, space, or perspective." }

Stage-3 User Prompt You are executing ONLY Step 3: Locate Key Turning Points and Execute Segmentation. Use the chapter text, Step 1 result, and Step 2 result as input. [Chapter Text Content] {chapter_content} [Step 1 Result JSON] {step_1_result} [Step 2 Result JSON] {step_2_result} Return STRICT JSON only (no markdown): { "chapter_title": "The chapter title or identifier", "step_3_turning_points": [ { "type": "The type of turning point, do not force all predefined types.", "description": "A string briefly describing the specific event or shift that constitutes this turning point, explicitly justified by a macro/micro narrative transition or a drop in semantic cohesion." } ], "segments": [ { "subtitle": "The subtitle of the plot block", "from_idx": "Integer start index, with optional overlap with previous segment if necessary. First segment must start with index 1.", "to_idx": "Integer end index, with optional overlap with next segment if necessary. Last segment must end with index {last_index}.", "context_idx": "List of indices of sentences that provide necessary context, if necessary." } ] }

D.2

Prompt for LitSeg-Lite

System Prompt Single-Stage # Role: Narratology Segmentation Expert ## Profile: - Language: English - Description: You are a professional literary editor and Natural Language Processing expert deeply versed in structuralism and narratology. Your core task is to perform fine-grained plot segmentation on literary texts. ## Core Competencies 1. Fine-grained plot segmentation preserving original literary aesthetics. 2. Extracting valid dynamic events and effectively filtering out static or non-narrative noise. 3. Untangling complex multi-line, parallel, and nested narrative structures based on "Unity of Action" and internal story logic. 4. Pinpointing logical breakpoints by synthesizing macrostructural shifts (spatial, temporal, tension-based, etc.) with micro-level evaluations of semantic dependency ## Primary Objectives 1. Perform logical, narratology-driven plot segmentation on literary texts. 2. Enable readers to independently read and understand each segmented snippet out of its full context. 3. Output the exact analytical and segmentation results in strict JSON format. ## Constraints 1. Role Consistency: Don't break character or bypass the narratological framework under any circumstance. 2. Narrative-Driven Granularity: The recommended maximum length for any single segment is 100 sentences. This is a guidance threshold rather than an absolute hard cap. When narrative coherence clearly requires it, a segment may exceed this limit. NEVER unnaturally merge distinct narrative blocks just to hit a sentence count. Abandon any "mathematical division" mindset. Breakpoints must be dictated organically by narratological theories. 3. Conditional Contextual Overlap: A small overlap of key sentences between adjacent plot blocks is allowed ONLY to ensure narrative continuity. This is NOT mandatory. If a natural narrative breakpoint is clear and coherent, make a clean, hard cut. Do not blindly copy previous sentences just for the sake of overlapping. 4. Descriptive Subtitle: Extract a highly concise subtitle that summarizes the core action or reversal event for each plot block. 5. Full Coverage and Order: You must segment the entire chapter in its original order. Do not skip any sentences or paragraphs. Every sentence must belong to at least one segment, i.e., covered in the index range between "from_idx" and "to_idx" of at least one segment. 6. Indexing Rule: The "from_idx" of the first segment MUST be 1. The "to_idx" of the last segment MUST be the last index of the chapter. Indices are 1-based and inclusive like [1] in the original text. 7. Context Indices: For sentences crucial for understanding a segment but outside its main range, include their indices in the "context_idx" field. Sentences cannot ONLY exist in "context_idx"; they must be part of some segment's main range. ## Narratological Workflow When determining logical breakpoints for plot segmentation, you must execute the following three-step narratological framework sequentially: ### Step 1: Extract Valid Events and Filter Noise 1. Focus on "Dynamic" Changes: Treat the plot as a sequence of dynamic events. Ignore static character portraits or pure scenery descriptions. Extract an event ONLY when characters or objects undergo dynamic changes. 2. Filter by Event Type: Retain "Changes of state" ( physical or mental state changes) and "Process events " (actions or happenings without state change, such

as talking, thinking, and feeling) as core events. Discard "Stative events" (static physical or mental states) and "Non-events" (generic statements, counterfactuals, questions) to aggressively reduce noise. 3. Handling Ambiguous Paragraphs: For borderline paragraphs , gauge narrativity by checking for: a situation, an agent, one or more sequential actions, a potential object, a spatial location, a temporal specification, and a rationale. A higher density of these features indicates a valid, independent event. ### Step 2: Untangle Narrative Threads and Clarify Structures Group the events extracted in Step 1 into clear narrative threads using the following criteria: 1. Apply "Unity of Action": Group events by a complete core action, abandoning the "same character" stereotype. Separate unrelated actions performed by the same character; merge joint actions advanced by different characters. 2. Track the "Revelation of Secrets": Use the gradual unfolding of author-planted secrets as a tight internal logic to connect scattered events into a cohesive narrative chain. 3. Clarify Parallel & Minor Threads: Explicitly identify and segment concurrent event chains (e.g., A1 -> B1 -> A2 -> B2). Never merge minor threads into major ones; preserve their structural independence. 4. Identify Nested Hierarchy: Classify the thread's narrative level: [Extradiegetic] (story introduction) , [Intradiegetic] (core main plot), or [Metadiegetic] (a story within a story). 5. Establish Basic Data: For each thread, map the timeline, causeline, and critical shift points in time, space, and perspective. ### Step 3: Locate Key Turning Points and Execute Segmentation Use the data established in Step 2 to locate logical breakpoints. Apply these markers organically, segmenting the text only when shifts clearly and naturally occur: 1. Macro-Structural & Plot Tension: Track tension through the overarching arc (Exposition -> Predicament -> Extrication). Consider segmenting the text at the boundaries or transitions between these major driving events: - Opportunity: The introductory event that triggers the story after the background is set. - Change of Plans: The event where the main goal is defined and action intensifies. - Point of No Return: The event pushing characters to fully commit to their goal. - Major Setback & Reversal: The event where the situation completely falls apart, marking a formal reversal of fortune (peripeteia). This generally serves as a strong signal for segmentation. - Climax & Recognition: The final resolution and peak tension, often accompanied by a cognitive shift from ignorance to truth (anagnorisis). - Aftermath & Buffer: Post-climax events that resolve the predicament and lower tension, smoothly transitioning into the final Extrication phase. 2. Contextual Shifts: Segment at definitive leaps in time, space, or perspective. Cut when the narrative switches between parallel threads or crosses nested narrative levels. 3. Micro-Dynamic Shifts (State & Action): Segment when a situation's fundamental state transitions ( equilibrium -> disequilibrium -> new equilibrium) or at the precise moment an action's outcome (success/ failure) is revealed. 4. Sentence-Level Boundary Pinpointing: After narrowing down a general turning point, actively evaluate the semantic dependency between adjacent sentences to pinpoint the exact breakpoint: - High Semantic Cohesion Zone (Avoid Segmenting): Within the scope of the potential turning point, identify sentences with strong semantic dependency or interlocking logic. Examples include adjacency pairs (e.g., question/answer, attack/defense), continuous pronoun chains, or an immediate physical/emotional reaction to a specific action. Avoid executing a cut within these sequences. - Low Semantic Cohesion Zone (Execute the Cut): Within

the scope of the potential turning point, locate the exact sentence where the semantic dependency drops to its lowest level. Make the cut right before the sentence that can semantically stand alone--one that initiates a new topic, phase, or action, and requires minimal context from the immediately preceding sentence to be fully understood. ## Output Specifications 1. JSON Format: You must output STRICT JSON format. Escape all internal double quotes within JSON string fields using a backslash. 2. Content Requirement: For each segmented plot block, output a concise subtitle summarizing the core action , and the index ranges based on the constraints. ## Initialization: As a Narratology Segmentation Expert, you must strictly follow the Constraints and the Narratological Workflow. To initiate the analysis, await the provision of the Chapter Text Content. Upon receipt, execute all three workflow steps internally and output the segmentation results exclusively in the requested STRICT JSON format. Return STRICT JSON only (no markdown): {

}

"step1": { "step_1_events": "A string briefly listing the core dynamic events extracted (filtering out noise)." }, "step2": { "step_2_threads": "A string briefly describing the narrative threads and their nested hierarchy, based on unity of action, secrets, and key shifts in time, space, or perspective." }, "step3": { "chapter_title": "The chapter title or identifier", "step_3_turning_points": [ { "type": "The type of turning point, do not force all predefined types.", "description": "A string briefly describing the specific event or shift that constitutes this turning point, explicitly justified by a macro/micro narrative transition or a drop in semantic cohesion." } ], "segments": [ { "subtitle": "The subtitle of the plot block", "from_idx": "Integer of the index where the segment starts, with optional overlap with previous segment if necessary. The first segment must start with index 1.", "to_idx": "Integer of the index where the segment ends, with optional overlap with next segment if necessary. The last segment must end with index { last_index}.", "context_idx": "List of indices of sentences that provide necessary context for understanding this segment, if necessary." } ] }

D.3

Prompt for Reward Model

Reward Model System Prompt # Role and Task You are a Senior Narratologist and NLP Evaluation Expert. Your task is to evaluate the provided segment against the Original Passage based on three primary dimensions: Event Validity, Event Phase & Unity of Action, and Cut Point Logic. Evaluate each chunk. # Evaluation Dimensions and Criteria

## Dimension 1: Event Validity - Evaluation Scope: Focuses on two core areas: dynamic event extraction and fidelity to the original text. Dynamic events include "Changes of state" (physical or mental state changes) and "Process events" ( actions or happenings without state change, such as talking, thinking, and feeling). Assess whether the segment accurately captures an event. Summaries must be precise, avoiding subjective hallucinations or distortion of facts. - Core Pain Points: Deviating from the dynamic narrative logic of the source; composed entirely of static description, with no narrative events; subtitles that are too vague (e.g., "The Island"); subtitles that focus on a minor detail while ignoring the main plot movement; misalignment between the title and the actual text. Scoring Criteria: - 5 (Excellent): The entire segment revolves around a clear , dynamic core event that is self-contained, featuring a complete narrative function within the chunk. Static descriptions (setting, character psychology, background, etc.) are naturally preserved as narrative support and are seamlessly integrated to serve the progression of the plot. The subtitle perfectly captures the core of the event. - 4 (Good): The segment contains a clear dynamic event that is largely self-contained. Although the proportion of static description or exposition within the block is relatively high, the primary narrative axis remains easily identifiable, and the dynamic movement is not obscured. The subtitle fidelity is acceptable. - 3 (Passing): The segment contains at least one minor valid event (e.g., a few lines of dialogue or a micro -action) but captures only a fragment or a partial slice of a larger event. Lengthy static descriptions or background information heavily dilute the sense of dynamic momentum in both length and perception. The subtitle fidelity is acceptable. - 2 (Poor): The vast majority of the segment consists of static settings or chatter unrelated to the current plot progression. The narrative is completely fragmented, severely lacking dynamic change and leaving the reader with a sense of narrative stagnation. The subtitle may contain noticeable factual distortions. There is a significant deviation from narrative logic. - 1 (Failure): 100% - Special Notes: Assign a score of 5.0 to any chunk that contains the entirety of the author's prologue/ monologue, background introduction, or copyright information, regardless of the above scoring criteria. ## Dimension 2: Event Phase and Unity of Action - Evaluation Scope: Focuses exclusively on the scope/focus of events within the segment. Assess whether the segment strictly adheres to the "Unity of Action" (a single phase or a single thread) and whether it merges multiple actions that should have remained independent. - Core Pain Points: "Stitching" together unrelated parallel threads, actions of irrelevant characters, or multilevel/nested events into a single segment that could have been split into independent phases. Scoring Criteria: - 5 Points (Excellent): Perfect focus on a single phase. Strictly locked into a pure action phase where the cause and effect of the event are highly cohesive. - 4 Points (Good): Clear primary action. Focuses on a single event segment, occasionally including a very minor prelude (one or two sentences) to the next action phase without feeling bloated. - 3 Points (Passing): Cluttered multi-phase actions. A clear main thread exists, but the block contains two or more consecutive events that could have been clearly divided into independent segments, causing the focus to lose its singularity. - 2 Points (Poor): Forced stitching of unrelated threads.

Unrelated parallel threads or different narrative levels (e.g., reality mixed with long flashbacks) are forced into the same block. - 1 Point (Failure): Complete fragmentation of events. A total "mishmash" where small fragments within the block have no temporal or logical connection in terms of action. ## Dimension 3: Cut Point Logic and Overlapping Connectivity - Evaluation Scope: Focuses on the "placement of the cut" between each chunk and the neighboring one(s): , and the "contextual bridging mechanism." Assess whether the segmentation breaks a tight causal chain and whether "Overlap" is reasonably utilized to ensure the segment is independently readable. - Critical Deficiencies: - Front-end Fragmentation: Setting a chunk boundary directly within a high semantic cohesion region without "Overlap" buffer effectively "beheads" the context, severely compromising semantic integrity and rendering the segment unintelligible in isolation. - Back-end Truncation: Setting a chunk boundary that prematurely severs the causal link prevents the eventual outcome or resolution of the event from being included, resulting in an incomplete narrative arc. Scoring Criteria: - 5 Points (Excellent): The cut points falls precisely at a natural transition zone of low semantic cohesion. If cutting within a tight action sequence, a highly logical Overlap mechanism is used to keep the causal chain intact, making the segment fully readable and self-contained on its own. - 4 Points (Good): Natural cut and basic independence. The cut position is logical. Minor contextual dependency exists and the Overlap mechanism isn't fully utilized , or the Overlap is slightly redundant but doesn't hinder independent reading. - 3 Points (Passing): Lacks buffering and depends on context. A "hard cut" was made between sentences with some causal connection (without reasonable Overlap buffering). When read out of context, the reader must exert effort to "fill in the blanks" to follow the flow. - 2 Points (Poor): Illogical cut point and difficult to read independently. Directly cuts through a zone of extremely high semantic cohesion (e.g., continuous causal chains or compact actions) without any bridging buffer, leaving the segment nonsensical without the preceding text, or suspended without the following text. - 1 Point (Failure): Purely mechanical/blind cutting. An absurd physical break that completely destroys basic semantic meaning. - Special Notes: If the overlap buffer contains sentences entirely from within the current chunk, it constitutes an improper implementation of the mechanism. Consequently, a **0.5-point penalty** will be deducted from the final score. If the overlap buffer is utilized effectively and provides crucial supplementary context that allows the chunk to be fully understood in isolation, an **additional 0.5 points** will be awarded. # Output Format Output must be in valid JSON. For each dimension, return a list containing the score for each individual chunk. Example (suppose there are 4 chunks in the segment): { "dimension1": [1.5, 1.0, 1.0, 1.0], "dimension2": [1.0, 1.5, 1.0, 1.0], "dimension3": [1.0, 1.0, 1.5, 1.0] }

Reward Model User Prompt Template Segment Pipeline Prompt: {segment_pipeline_prompt}

Original Passage: {original_passage} Segment Pipeline: {segment_pipeline}

}

]

}

System Prompt for Ablation w/o Event Extraction

D.4

Prompt for Ablation Studies

System Prompt for Ablation w/o Theory-guided Segmentation # Role: Semantic Segmentation Expert ## Profile: - Language: English - Description: You are an expert in breaking down texts into segments. Your sole task is to split the text based on semantic content.

# Role: Narratology Segmentation Expert ## Profile: - Language: English - Description: You are a professional literary editor and Natural Language Processing expert deeply versed in structuralism and narratology. Your core task is to perform fine-grained plot segmentation on literary texts.

## Core Competencies 1. Detecting semantic transitions and boundaries. 2. Producing segments that can be read and understood clearly. 3. Ensuring every sentence is assigned to a segment, with no omissions and no reordering.

## Core Competencies 1. Fine-grained plot segmentation preserving original literary aesthetics. 2. Untangling complex multi-line, parallel, and nested narrative structures based on "Unity of Action" and internal story logic. 3. Pinpointing logical breakpoints by synthesizing macrostructural shifts (spatial, temporal, tension-based, etc.) with micro-level evaluations of semantic dependency.

## Primary Objectives 1. Perform a semantic segmentation of the provided text. 2. For each segment, output: - A concise subtitle that captures its main idea or subject. - The inclusive sentence index range it covers. 3. Output the segmentation in strict JSON format.

## Primary Objectives 1. Perform logical, narratology-driven plot segmentation on literary texts. 2. Enable readers to independently read and understand each segmented snippet out of its full context. 3. Output the exact analytical and segmentation results in strict JSON format.

## Constraints 1. **Full Coverage and Original Order**: Segment the entire text in its original sentence order. Do not skip any sentence. Every sentence must belong to at least one segment's main index range. 2. **Indexing Rule**: Indices are 1-based and inclusive. The "from_idx" of the first segment must be 1. The " to_idx" of the last segment must equal the last index of sentences in the text. Indices are like [1] in the original text. 3. **Context Indices**: If a segment requires a sentence from outside its main range to be fully comprehensible, you may list that sentence's index in "context_idx". However, a sentence cannot exist * only* in "context_idx"; every sentence must appear as a main member of exactly one segment (i.e., within its "from_idx"-"to_idx" range). 4. **No External Knowledge**: Segment based only on the semantic content of the text itself. Do not use any external knowledge or assumptions about the text. 5. **Semantic Clarity**: Each segment should represent a clear and coherent meaning. 6. **Segment Length**: The recommended maximum length for any single segment is 100 sentences. This is a guidance threshold rather than an absolute hard cap. When necessary, a segment may exceed this limit. Do not artificially merge or split to reach a target length. 7. **Subtitle**: Provide a very short, descriptive subtitle for each segment.

## Constraints 1. Role Consistency: Don't break character or bypass the narratological framework under any circumstance. 2. Narrative-Driven Granularity: The recommended maximum length for any single segment is 100 sentences. This is a guidance threshold rather than an absolute hard cap. When narrative coherence clearly requires it, a segment may exceed this limit. NEVER unnaturally merge distinct narrative blocks just to hit a sentence count. Abandon any "mathematical division" mindset. Breakpoints must be dictated organically by narratological theories. 3. Conditional Contextual Overlap: A small overlap of key sentences between adjacent plot blocks is allowed ONLY to ensure narrative continuity. This is NOT mandatory. If a natural narrative breakpoint is clear and coherent, make a clean, hard cut. Do not blindly copy previous sentences just for the sake of overlapping. 4. Descriptive Subtitle: Extract a highly concise subtitle that summarizes the core action or reversal event for each plot block. 5. Full Coverage and Order: You must segment the entire chapter in its original order. Do not skip any sentences or paragraphs. Every sentence must belong to at least one segment, i.e., covered in the index range between "from_idx" and "to_idx" of at least one segment. 6. Indexing Rule: The "from_idx" of the first segment MUST be 1. The "to_idx" of the last segment MUST be the last index of the chapter. Indices are 1-based and inclusive like [1] in the original text. 7. Context Indices: For sentences crucial for understanding a segment but outside its main range, include their indices in the "context_idx" field. Sentences cannot ONLY exist in "context_idx"; they must be part of some segment's main range.

Return STRICT JSON only (no markdown): {

"chapter_title": "The chapter title or identifier", "segments": [ { "subtitle": "The subtitle of the plot block", "from_idx": "Integer of the index where the segment starts, with optional overlap with previous segment if necessary. The first segment must start with index 1.", "to_idx": "Integer of the index where the segment ends, with optional overlap with next segment if necessary. The last segment must end with index { last_index}.", "context_idx": "List of indices of sentences that provide necessary context for understanding this segment, if necessary."

## Narratological Workflow When determining logical breakpoints for plot segmentation, you must execute the following two-step narratological framework sequentially: ### Step 1: Untangle Narrative Threads and Clarify Structures Identify and extract the narrative threads directly from the provided text using the following criteria: 1. Apply "Unity of Action": Group events by a complete core action, abandoning the "same character" stereotype.

Separate unrelated actions performed by the same character; merge joint actions advanced by different characters. 2. Track the "Revelation of Secrets": Use the gradual unfolding of author-planted secrets as a tight internal logic to connect scattered events into a cohesive narrative chain. 3. Clarify Parallel & Minor Threads: Explicitly identify and segment concurrent event chains (e.g., A1 -> B1 -> A2 -> B2). Never merge minor threads into major ones; preserve their structural independence. 4. Identify Nested Hierarchy: Classify the thread's narrative level: [Extradiegetic] (story introduction) , [Intradiegetic] (core main plot), or [Metadiegetic] (a story within a story). 5. Establish Basic Data: For each thread, map the timeline, causeline, and critical shift points in time, space, and perspective. ### Step 2: Locate Key Turning Points and Execute Segmentation Use the data established in Step 1 to locate logical breakpoints. Apply these markers organically, segmenting the text only when shifts clearly and naturally occur: 1. Macro-Structural & Plot Tension: Track tension through the overarching arc (Exposition -> Predicament -> Extrication). Consider segmenting the text at the boundaries or transitions between these major driving events: - Opportunity: The introductory event that triggers the story after the background is set. - Change of Plans: The event where the main goal is defined and action intensifies. - Point of No Return: The event pushing characters to fully commit to their goal. - Major Setback & Reversal: The event where the situation completely falls apart, marking a formal reversal of fortune (peripeteia). This generally serves as a strong signal for segmentation. - Climax & Recognition: The final resolution and peak tension, often accompanied by a cognitive shift from ignorance to truth (anagnorisis). - Aftermath & Buffer: Post-climax events that resolve the predicament and lower tension, smoothly transitioning into the final Extrication phase. 2. Contextual Shifts: Segment at definitive leaps in time, space, or perspective. Cut when the narrative switches between parallel threads or crosses nested narrative levels. 3. Micro-Dynamic Shifts (State & Action): Segment when a situation's fundamental state transitions ( equilibrium -> disequilibrium -> new equilibrium) or at the precise moment an action's outcome (success/ failure) is revealed. 4. Sentence-Level Boundary Pinpointing: After narrowing down a general turning point, actively evaluate the semantic dependency between adjacent sentences to pinpoint the exact breakpoint: - High Semantic Cohesion Zone (Avoid Segmenting): Within the scope of the potential turning point, identify sentences with strong semantic dependency or interlocking logic. Examples include adjacency pairs (e.g., question/answer, attack/defense), continuous pronoun chains, or an immediate physical/emotional reaction to a specific action. Avoid executing a cut within these sequences. - Low Semantic Cohesion Zone (Execute the Cut): Within the scope of the potential turning point, locate the exact sentence where the semantic dependency drops to its lowest level. Make the cut right before the sentence that can semantically stand alone--one that initiates a new topic, phase, or action, and requires minimal context from the immediately preceding sentence to be fully understood. ## Output Specifications 1. JSON Format: You must output STRICT JSON format. Escape all internal double quotes within JSON string fields using a backslash. 2. Content Requirement: For each segmented plot block, output a concise subtitle summarizing the core action , and the index ranges based on the constraints. ## Initialization: As a Narratology Segmentation Expert, you must strictly

follow the Constraints and the Narratological Workflow. To initiate the analysis, await the provision of the Chapter Text Content. Upon receipt, execute the two workflow steps internally and output the segmentation results exclusively in the requested STRICT JSON format. Return STRICT JSON only (no markdown): {

}

"step1": { "step_1_threads": "A string briefly describing the narrative threads and their nested hierarchy, based on unity of action, secrets, and key shifts in time, space, or perspective." }, "step2": { "step_2_turning_points": [ { "type": "The type of turning point, do not force all predefined types.", "description": "A string briefly describing the specific event or shift that constitutes this turning point, explicitly justified by a macro/micro narrative transition or a drop in semantic cohesion." } ], "chapter_title": "The chapter title or identifier", "segments": [ { "subtitle": "The subtitle of the plot block", "from_idx": "Integer of the index where the segment starts, with optional overlap with previous segment if necessary. The first segment must start with index 1.", "to_idx": "Integer of the index where the segment ends, with optional overlap with next segment if necessary. The last segment must end with index { last_index}.", "context_idx": "List of indices of sentences that provide necessary context for understanding this segment, if necessary." } ] }

System Prompt for Ablation w/o Thread Untangling # Role: Narratology Segmentation Expert ## Profile: - Language: English - Description: You are a professional literary editor and Natural Language Processing expert deeply versed in structuralism and narratology. Your core task is to perform fine-grained plot segmentation on literary texts. ## Core Competencies 1. Fine-grained plot segmentation preserving original literary aesthetics. 2. Extracting valid dynamic events and effectively filtering out static or non-narrative noise. 3. Pinpointing logical breakpoints by synthesizing macrostructural shifts (spatial, temporal, tension-based, etc.) with micro-level evaluations of semantic dependency. ## Primary Objectives 1. Perform logical, narratology-driven plot segmentation on literary texts. 2. Enable readers to independently read and understand each segmented snippet out of its full context. 3. Output the exact analytical and segmentation results in strict JSON format. ## Constraints 1. Role Consistency: Don't break character or bypass the narratological framework under any circumstance. 2. Narrative-Driven Granularity: The recommended maximum length for any single segment is 100 sentences. This is a guidance threshold rather than an absolute hard cap. When narrative coherence clearly requires it, a

segment may exceed this limit. NEVER unnaturally merge distinct narrative blocks just to hit a sentence count. Abandon any "mathematical division" mindset. Breakpoints must be dictated organically by narratological theories. 3. Conditional Contextual Overlap: A small overlap of key sentences between adjacent plot blocks is allowed ONLY to ensure narrative continuity. This is NOT mandatory. If a natural narrative breakpoint is clear and coherent, make a clean, hard cut. Do not blindly copy previous sentences just for the sake of overlapping. 4. Descriptive Subtitle: Extract a highly concise subtitle that summarizes the core action or reversal event for each plot block. 5. Full Coverage and Order: You must segment the entire chapter in its original order. Do not skip any sentences or paragraphs. Every sentence must belong to at least one segment, i.e., covered in the index range between "from_idx" and "to_idx" of at least one segment. 6. Indexing Rule: The "from_idx" of the first segment MUST be 1. The "to_idx" of the last segment MUST be the last index of the chapter. Indices are 1-based and inclusive like [1] in the original text. 7. Context Indices: For sentences crucial for understanding a segment but outside its main range, include their indices in the "context_idx" field. Sentences cannot ONLY exist in "context_idx"; they must be part of some segment's main range. ## Narratological Workflow When determining logical breakpoints for plot segmentation, you must execute the following two-step narratological framework sequentially: ### Step 1: Extract Valid Events and Filter Noise 1. Focus on "Dynamic" Changes: Treat the plot as a sequence of dynamic events. Ignore static character portraits or pure scenery descriptions. Extract an event ONLY when characters or objects undergo dynamic changes. 2. Filter by Event Type: Retain "Changes of state" ( physical or mental state changes) and "Process events " (actions or happenings without state change, such as talking, thinking, and feeling) as core events. Discard "Stative events" (static physical or mental states) and "Non-events" (generic statements, counterfactuals, questions) to aggressively reduce noise. 3. Handling Ambiguous Paragraphs: For borderline paragraphs , gauge narrativity by checking for: a situation, an agent, one or more sequential actions, a potential object, a spatial location, a temporal specification, and a rationale. A higher density of these features indicates a valid, independent event. ### Step 2: Locate Key Turning Points and Execute Segmentation Use the valid events extracted in Step 1 to locate logical breakpoints. Apply these markers organically, segmenting the text only when shifts clearly and naturally occur: 1. Macro-Structural & Plot Tension: Track tension through the overarching arc (Exposition -> Predicament -> Extrication). Consider segmenting the text at the boundaries or transitions between these major driving events: - Opportunity: The introductory event that triggers the story after the background is set. - Change of Plans: The event where the main goal is defined and action intensifies. - Point of No Return: The event pushing characters to fully commit to their goal. - Major Setback & Reversal: The event where the situation completely falls apart, marking a formal reversal of fortune (peripeteia). This generally serves as a strong signal for segmentation. - Climax & Recognition: The final resolution and peak tension, often accompanied by a cognitive shift from ignorance to truth (anagnorisis). - Aftermath & Buffer: Post-climax events that resolve the predicament and lower tension, smoothly transitioning into the final Extrication phase. 2. Contextual Shifts: Segment at definitive leaps in time, space, or perspective. Cut when the narrative crosses macro settings.

3. Micro-Dynamic Shifts (State & Action): Segment when a situation's fundamental state transitions ( equilibrium -> disequilibrium -> new equilibrium) or at the precise moment an action's outcome (success/ failure) is revealed. 4. Sentence-Level Boundary Pinpointing: After narrowing down a general turning point, actively evaluate the semantic dependency between adjacent sentences to pinpoint the exact breakpoint: - High Semantic Cohesion Zone (Avoid Segmenting): Within the scope of the potential turning point, identify sentences with strong semantic dependency or interlocking logic. Examples include adjacency pairs (e.g., question/answer, attack/defense), continuous pronoun chains, or an immediate physical/emotional reaction to a specific action. Avoid executing a cut within these sequences. - Low Semantic Cohesion Zone (Execute the Cut): Within the scope of the potential turning point, locate the exact sentence where the semantic dependency drops to its lowest level. Make the cut right before the sentence that can semantically stand alone--one that initiates a new topic, phase, or action, and requires minimal context from the immediately preceding sentence to be fully understood. ## Output Specifications 1. JSON Format: You must output STRICT JSON format. Escape all internal double quotes within JSON string fields using a backslash. 2. Content Requirement: For each segmented plot block, output a concise subtitle summarizing the core action , and the index ranges based on the constraints. ## Initialization: As a Narratology Segmentation Expert, you must strictly follow the Constraints and the Narratological Workflow. To initiate the analysis, await the provision of the Chapter Text Content. Upon receipt, execute the two workflow steps internally and output the segmentation results exclusively in the requested STRICT JSON format. Return STRICT JSON only (no markdown): {

}

"step1": { "step_1_events": "A string briefly listing the core dynamic events extracted (filtering out noise)." }, "step2": { "step_2_turning_points": [ { "type": "The type of turning point, do not force all predefined types.", "description": "A string briefly describing the specific event or shift that constitutes this turning point, explicitly justified by a macro/micro narrative transition or a drop in semantic cohesion." } ], "chapter_title": "The chapter title or identifier", "segments": [ { "subtitle": "The subtitle of the plot block", "from_idx": "Integer of the index where the segment starts, with optional overlap with previous segment if necessary. The first segment must start with index 1.", "to_idx": "Integer of the index where the segment ends, with optional overlap with next segment if necessary. The last segment must end with index { last_index}.", "context_idx": "List of indices of sentences that provide necessary context for understanding this segment, if necessary." } ] }

System Prompt for Ablation w/o Turning Points Pinpointing # Role: Narratology Segmentation Expert ## Profile: - Language: English - Description: You are a professional literary editor and Natural Language Processing expert deeply versed in structuralism and narratology. Your core task is to perform fine-grained plot segmentation on literary texts. ## Core Competencies 1. Fine-grained plot segmentation preserving original literary aesthetics. 2. Extracting valid dynamic events and effectively filtering out static or non-narrative noise. 3. Untangling complex multi-line, parallel, and nested narrative structures based on "Unity of Action" and internal story logic. 4. Executing text segmentation to establish clear start and end sentence boundaries for each narrative block. ## Primary Objectives 1. Perform logical, narratology-driven plot segmentation on literary texts. 2. Enable readers to independently read and understand each segmented snippet out of its full context. 3. Output the exact analytical and segmentation results in strict JSON format. ## Constraints 1. Role Consistency: Don't break character or bypass the narratological framework under any circumstance. 2. Narrative-Driven Granularity: The recommended maximum length for any single segment is 100 sentences. This is a guidance threshold rather than an absolute hard cap. When narrative coherence clearly requires it, a segment may exceed this limit. NEVER unnaturally merge distinct narrative blocks just to hit a sentence count. Abandon any "mathematical division" mindset. Breakpoints must be dictated organically by narratological theories. 3. Conditional Contextual Overlap: A small overlap of key sentences between adjacent plot blocks is allowed ONLY to ensure narrative continuity. This is NOT mandatory. If a natural narrative breakpoint is clear and coherent, make a clean, hard cut. Do not blindly copy previous sentences just for the sake of overlapping. 4. Descriptive Subtitle: Extract a highly concise subtitle that summarizes the core action or reversal event for each plot block. 5. Full Coverage and Order: You must segment the entire chapter in its original order. Do not skip any sentences or paragraphs. Every sentence must belong to at least one segment, i.e., covered in the index range between "from_idx" and "to_idx" of at least one segment. 6. Indexing Rule: The "from_idx" of the first segment MUST be 1. The "to_idx" of the last segment MUST be the last index of the chapter. Indices are 1-based and inclusive like [1] in the original text. 7. Context Indices: For sentences crucial for understanding a segment but outside its main range, include their indices in the "context_idx" field. Sentences cannot ONLY exist in "context_idx"; they must be part of some segment's main range. ## Narratological Workflow When determining logical breakpoints for plot segmentation, you must execute the following three-step narratological framework sequentially: ### Step 1: Extract Valid Events and Filter Noise 1. Focus on "Dynamic" Changes: Treat the plot as a sequence of dynamic events. Ignore static character portraits or pure scenery descriptions. Extract an event ONLY when characters or objects undergo dynamic changes. 2. Filter by Event Type: Retain "Changes of state" ( physical or mental state changes) and "Process events " (actions or happenings without state change, such as talking, thinking, and feeling) as core events. Discard "Stative events" (static physical or mental

states) and "Non-events" (generic statements, counterfactuals, questions) to aggressively reduce noise. 3. Handling Ambiguous Paragraphs: For borderline paragraphs , gauge narrativity by checking for: a situation, an agent, one or more sequential actions, a potential object, a spatial location, a temporal specification, and a rationale. A higher density of these features indicates a valid, independent event. ### Step 2: Untangle Narrative Threads and Clarify Structures Group the events extracted in Step 1 into clear narrative threads using the following criteria: 1. Apply "Unity of Action": Group events by a complete core action, abandoning the "same character" stereotype. Separate unrelated actions performed by the same character; merge joint actions advanced by different characters. 2. Track the "Revelation of Secrets": Use the gradual unfolding of author-planted secrets as a tight internal logic to connect scattered events into a cohesive narrative chain. 3. Clarify Parallel & Minor Threads: Explicitly identify and segment concurrent event chains (e.g., A1 -> B1 -> A2 -> B2). Never merge minor threads into major ones; preserve their structural independence. 4. Identify Nested Hierarchy: Classify the thread's narrative level: [Extradiegetic] (story introduction) , [Intradiegetic] (core main plot), or [Metadiegetic] (a story within a story). 5. Establish Basic Data: For each thread, map the timeline, causeline, and critical shift points in time, space, and perspective. ### Step 3: Execute Segmentation Use the data established in Step1 and Step 2 to directly execute the text segmentation. ## Output Specifications 1. JSON Format: You must output STRICT JSON format. Escape all internal double quotes within JSON string fields using a backslash. 2. Content Requirement: For each segmented plot block, output a concise subtitle summarizing the core action , and the index ranges based on the constraints. ## Initialization: As a Narratology Segmentation Expert, you must strictly follow the Constraints and the Narratological Workflow. To initiate the analysis, await the provision of the Chapter Text Content. Upon receipt, execute all three workflow steps internally and output the segmentation results exclusively in the requested STRICT JSON format. Return STRICT JSON only (no markdown): {

"step1": { "step_1_events": "A string briefly listing the core dynamic events extracted (filtering out noise)." }, "step2": { "step_2_threads": "A string briefly describing the narrative threads and their nested hierarchy, based on unity of action, secrets, and key shifts in time, space, or perspective." }, "step3": { "chapter_title": "The chapter title or identifier", "segments": [ { "subtitle": "The subtitle of the plot block", "from_idx": "Integer of the index where the segment starts, with optional overlap with previous segment if necessary. The first segment must start with index 1.", "to_idx": "Integer of the index where the segment ends, with optional overlap with next segment if necessary. The last segment must end with index { last_index}.", "context_idx": "List of indices of sentences that provide necessary context for understanding this segment, if necessary." }

]

E.4

Evaluation Metric Overview

E.1

Exact Match

METEOR (Banerjee and Lavie, 2005) aligns unigrams between the prediction and the reference using exact, stem, and synonym matches. Let Pm and Rm denote unigram precision and recall after alignment. METEOR first computes a weighted harmonic mean,

}

Exact Match (EM) measures whether a prediction exactly matches a reference answer after normalization. Let norm(·) denote lowercasing and punctuation-stripping normalization. For a prediction â and a set of references A, EM is defined as EM(â, A) = max I[norm(â) = norm(a)] . (1) a∈A

The final EM score is the average over all examples. E.2

Fmean =

10Pm Rm , Rm + 9Pm

Penalty = γ

F1 =

2P R . P +R

(3)

E.5

(8)

Context Relevance

Context Relevance1 evaluates whether the retrieved contexts are pertinent to the user input. This is done via two independent LLM-as-a-Judge prompts that each rate the relevance on a scale of 0, 1, or 2. The ratings are then converted to a [0, 1] scale and averaged to produce the final score. Given a question q and retrieved contexts c, each judge assigns a score rj ∈ {0, 1, 2}, where 0 means the retrieved contexts are not relevant to the query at all, 1 means the contexts are partially relevant, and 2 means the contexts are completely relevant. The normalized context relevance score for one example is 2

CtxRel(q, c) =

1 X rj . 2 2

(9)

j=1

ROUGE-L

ROUGE-L (Lin, 2004) measures longest-commonsubsequence overlap. Let LCS(â, a) be the length of the longest common subsequence between the prediction â and a reference answer a. ROUGE-L precision and recall are LCS(â, a) , |â|

RLCS =

LCS(â, a) . |a| (4)

The ROUGE-L F-score is ROUGE-L =

(7)

where m is the number of matched unigrams and c is the number of contiguous matched chunks. The final score is

When multiple references are available, we use the maximum F1 over references.

PLCS =

c θ

METEOR = (1 − Penalty)Fmean .

The token-level F1 score is

(6)

and then applies a fragmentation penalty,

F1 computes token-level overlap between the prediction and the reference answer. For a prediction â and a reference a, let T (â) and T (a) denote their normalized token multisets, and let m be the number of overlapping tokens. Precision and recall are m m , R= . (2) P = |T (â)| |T (a)|

E.3

METEOR

(1 + β 2 )PLCS RLCS , RLCS + β 2 PLCS

where β controls the relative weight of recall.

(5)

Higher scores indicate that the contexts are more closely aligned with the user’s query. The final score is averaged over examples. E.6

Answer Accuracy

Answer Accuracy1 measures the agreement between a model’s response and a reference ground truth for a given question. This is done via two distinct LLM-as-a-Judge prompts that each return a rating (0, 2, or 4). The metric converts these ratings into a [0, 1] scale and then takes the average of the two scores from the judges. Given a question 1 Available at: https://docs.ragas.io/en/stable/ concepts/metrics/available_metrics/nvidia_ metrics/

q, prediction â, and reference answer a, each judge assigns a score sj ∈ {0, 2, 4}, where 0 means the response is inaccurate or does not address the same question as the reference, 2 means the response partially aligns with the reference, and 4 means the response exactly aligns with the reference. The normalized answer accuracy score for one example is 2 1 X sj AnsAcc(q, â, a) = . (10) 2 4 j=1

The two prompts use distinct templates to ensure robustness: one compares the response with the reference directly, while the other swaps the roles of the response and reference. If both ratings are valid, the final score is the average; otherwise, it takes the valid one. Higher scores indicate that the model’s answer closely matches the reference. The final score is averaged over examples. E.7

For GutenQA, we additionally report Mean Reciprocal Rank (MRR) as a retrieval metric. MRR evaluates the ranking quality based on the gold text span. Let gi be the gold span for example i, and let di,j be the j-th chunk in the reranked list Di sorted by reranker score. We define the rank of the first relevant chunk as:

MRR =

1 N

i=1

1 . ranki

Before matching, both the gold span and retrieved chunks are lowercased and stripped of whitespace and newline characters. We report Hit@1, Hit@2, Hit@3, Hit@5, and Hit@20. Hit@k measures whether required evidence is present.

Automated Evaluation Results on GutenQA

See Tables 7 and 8. Table 7: Automated evaluation of generation results on GutenQA. GutenQA EM

R-L

Mtr

RAG Baselines Token Recursive Char Perplexity + Merge Margin-Sampling + Merge LumberChunker + Merge

0.0427 0.0420 0.0390 0.0383 0.0257 0.0327 0.0343 0.0380

0.5389 0.5371 0.5327 0.5340 0.4959 0.5255 0.5276 0.5256

0.5018 0.4999 0.4962 0.4963 0.4640 0.4896 0.4887 0.4889

0.5592 0.5584 0.5535 0.5570 0.5103 0.5492 0.5538 0.5527

0.6396 0.6246 0.6129 0.6268 0.5154 0.6128 0.6229 0.6262

Ours LitSeg

0.0320

0.5174

0.4791

0.5525

0.6252

Table 8: Automated evaluation of ablation study on GutenQA dataset. Method

LitSeg w/o FT +w/o Theory +w/o Theo.S.1 +w/o Theo.S.2 +w/o Theo.S.3

(12)

Before matching, both the gold span and retrieved chunks are lowercased and stripped of whitespace and newline characters. MRR measures how effectively the system places relevant evidence at the top of the ranked list, heavily penalizing systems that require scanning many chunks. E.8

i=1

(11)

1 where ranki = ∞ (and thus rank = 0) if no chunk i in the list contains the gold span. MRR is then calculated as the average of the reciprocal ranks across the N examples: N X

1 X Hit@k = I [∃d ∈ Di,k s.t. gi ⊆ d] . (13) N

Method

Mean Reciprocal Rank

ranki = min{j | gi ⊆ di,j },

for example i, and let Di,k be the top-k reranked chunks. Hit@k is

Hit@k

For GutenQA, we additionally report Hit@k as a retrieval metric. Each question is associated with a gold text span that must be contained in the retrieved evidence. We sort reranked chunks by reranker score and check whether the gold span appears in the top-k chunks. Let gi be the gold span

GutenQA EM

R-L

Mtr

0.0320 0.0323 0.0307 0.0273 0.0317 0.0310

0.5174 0.5191 0.5202 0.5107 0.5154 0.5154

0.4791 0.4801 0.4830 0.4716 0.4752 0.4775

0.5525 0.5519 0.5511 0.5447 0.5479 0.5458

0.6252 0.6259 0.6262 0.6200 0.6286 0.6195

Baseline Overview

Token splitter. The token splitter divides text into fixed-length token windows with overlap. It is a length-based baseline and does not model semantics. Recursive-character splitter. The recursivecharacter splitter creates chunks according to a target character length. It recursively chunks and falls back to smaller separators, including paragraph breaks, newlines, punctuation, and whitespace.

Perplexity chunker. Perplexity Chunker (Zhao et al., 2024) computes language-model loss for each sentence and identifies local minima in the sentence-level loss curve as potential boundaries. Let si denote the i-th sentence. We compute the average token-level cross-entropy loss: 1 X L(si ) = CE (pθ (t | t<i ), t) , (14) |si | t∈s i

where pθ is the language-model distribution. A sentence is selected as a boundary if it is a local minimum and the loss drop is sufficiently large: L(si ) < L(si−1 ) ∧ L(si ) < L(si+1 ) ∧ (∆i−1 ≥ τ ∨ ∆i+1 ≥ τ ) ,

(15)

where ∆j = L(sj ) − L(si ) and τ = 0.5 by default. In implementation, the selected minimum sentence is included in both neighboring chunks, which creates a one-sentence overlap at the detected boundary. This method is sensitive to distributional or stylistic changes but does not explicitly encode narratological criteria. Margin-sampling chunker. Margin-Sampling Chunker (Zhao et al., 2024) treats each candidate boundary as a binary decision between splitting and keeping adjacent text together. Given the current chunk L and the next sentence R, the model is prompted to choose between option “1” for split and option “2” for keep. Margin-sampling decision prompt This is a text chunking task. You are a text analysis expert. Please group two related paragraphs together and separate unrelated paragraphs based on the logical structure and semantic content of the provided sentences. Choose one chunking method from the following two options according to the above requirements: 1. Split "{left} {right}" into "{left}" and "{right}"; 2. Keep "{left} {right}" unsplit in its original form; Please answer 1 or 2.

Let p1 and p2 denote the model probabilities assigned to option “1” and option “2”, respectively. The boundary score is score(L, R) = p2 − p1 .

(16)

The chunker keeps L and R together when score(L, R) > θ and splits otherwise. The threshold θ is initialized to 0 and updated as the moving average of the most recent five scores: 1 X θt = u, |Ht | ≤ 5, (17) |Ht | u∈Ht

where Ht contains the most recent margin scores.

LumberChunker. LumberChunker (Duarte et al., 2024) follows the idea of prompting an LLM to identify content shifts. A sliding window of sentence-level units is formatted as a numbered list and presented to the model as paragraphs. The model is asked to return the first ID whose content clearly changes compared with previous units. LumberChunker prompt You will receive as input an English document with paragraphs identified by 'ID XXXX: <text>'. Task: Find the first paragraph (not the first one) where the content clearly changes compared to the previous paragraphs. Output: Return the ID of the paragraph with the content shift as in the exemplified format: 'Answer: ID XXXX'. Additional Considerations: Avoid very long groups of paragraphs. Aim for a good balance between identifying content shifts and keeping groups manageable. Document: {numbered_units}

Let it be the current start position and let Wt = {uit , . . . , ujt } be the current window, whose total length is bounded by 550 words. The model returns an index b̂t inside the window, and the next chunk is Ct = [uit , . . . , uit +b̂t −1 ]. (18) The start position is then updated to it+1 = it + b̂t . Invalid responses are retried up to a predefined times. If all retries fail, the chunker falls back to splitting after the first unit in the current window. Dynamic merge in baselines. For perplexity, margin-sampling, and LumberChunker, we also evaluate a dynamic-merge variant. Dynamic merge greedily merges adjacent chunks when their combined length is no more than 200 words. Given chunks Ci and Ci+1 , the merge condition is |Ci |word + |Ci+1 |word ≤ 200.

(19)

This post-processing step reduces extremely short chunks that may be harmful for retrieval.

Implementation Details

This appendix describes our implementation details. H.1 H.1.1

RAG Pipeline Pipeline Overview

Our RAG system follows a classic yet robust retrieve–rerank–generate architecture. Specifically, the pipeline consists of the following stages:

Reranker instruction

1. Text loading. 2. Text splitting. The target chunker segments each chapter into chunks. 3. Dense indexing. Chunks are encoded by a fixed embedding model and stored in vector databases. 4. Hybrid retrieval. Dense similarity search and sparse search retrieve candidate chunks under book-level metadata filters.

Given a literary question, retrieve relevant literature text chunks that answer the question.

Generation system prompt You are a helpful and precise assistant for answering questions about literature works. Your task is to answer the question based on information provided. If you cannot find the answer in the provided information, say you don't know. Answer directly to the question without explanation or additional information.

Generation system prompt benchmark-specific suffixes

5. Reranking. A fixed reranker scores the retrieved candidates and selects the top chunks. 6. Answer generation. A fixed generator produces a concise answer based on the reranked evidence.

LiteraryQA: Answer in one phrase or one sentence, as concise as possible. GutenQA: Answer in one sentence, and be concise.

Non-RAG generation user prompt template

H.1.2

RAG Implementation Details

Table 9 summarizes the RAG configuration. All components are shared and kept unchanged across chunkers. See Table 10 for generator configuration.

Book Title: {title} Question: {query}

RAG generation user prompt template

Table 9: RAG pipeline configuration. Book Title: {title} Component

Configuration

Question: {query}

Variable component Vector store Embedding model Dense retrieval Dense top-k Sparse retrieval Sparse top-k Candidate fusion Reranker Reranker top-k Generator Generator mode

Text splitter (chunker) Chroma Qwen/Qwen3-Embedding-8B Similarity search 18 BM25 2 Union of dense and sparse candidates Qwen/Qwen3-Reranker-8B 5 Qwen/Qwen3.5-9B Non-thinking mode

References: Chunk: {doc}

Table 10: RAG pipeline generator configuration. Parameter

Value

Model Thinking mode Temperature TopP TopK MinP Presence Penalty Repetition Penalty

Qwen/Qwen3.5-9B Disabled 1.0 0.95 20 0.0 1.5 1.0

H.2

See Table 11. Table 11: LitSeg chunker configuration. Parameter

Value

Model Sentence tokenizer Input chunk size Maximum retries Thinking mode Temperature TopP TopK MinP Presence Penalty Repetition Penalty

Qwen/Qwen3.6-27B-FP8 NLTK (Bird and Loper, 2004) 25,000 words 20 Disabled 1.0 0.95 20 0.0 1.5 1.0

H.3 Embedding and query instruction Represent the literature text chunk for literary question answering:

LitSeg Chunker

LitSeg-Lite Chunker

H.3.1 Inference Details See Table 12. H.3.2 Training Details See Table 13, 14, 15 and 16.

Table 12: LitSeg-Lite chunker configuration. Parameter

Value

Base model Sentence tokenizer Input chunk size Maximum retries Temperature TopP TopK MinP

Qwen/Qwen3-4B-Instruct-2507 NLTK 25,000 words 20 0.7 0.8 20 0

Table 13: LitSeg-Lite SFT configuration.

Table 16: Rule-based format reward for LitSeg-Lite GRPO training. Condition

Reward

Description

Valid output

0.0

Invalid JSON

−1.0

The output is parsable JSON, contains all required fields, and passes segment validation. The output cannot be parsed as JSON after basic cleanup. The output JSON misses any of step1, step2, or step3. The step3 segmentation fails validation, e.g., invalid index ranges, invalid context indices, or sentence coverage violations.

Missing required −0.8 fields Invalid step3 seg- −0.6 ments

H.4

Evaluation Metrics

See Table 17 and 18. Parameter

Value

Base model Precision PEFT method LoRA rank r LoRA alpha α LoRA dropout LoRA bias Target modules Epochs Learning rate Per-device batch size Gradient accumulation Optimizer Random seed

Qwen/Qwen3-4B-Instruct-2507 fp16 LoRA 16 32 0 none q, k, v, o, gate, up, down projections 3 2 × 10−4 1 8 adamw_torch_fused 42

Table 17: BERTScore metric configuration. Parameter

Value

Model Baseline rescaling

FacebookAI/roberta-large Enabled

Table 18: RAGAS metrics configuration. Parameter

Value

Judge model Thinking mode Temperature Seed

Qwen/Qwen3.5-9B Disabled 0.0 42

Table 14: LitSeg-Lite GRPO configuration. Parameter

Value

Base model Loss type Reward weights Generations per prompt (G) Iterations per batch Epochs Learning rate Per-device batch size Gradient accumulation Sampling temperature Optimizer Random seed

Qwen/Qwen3-4B-Instruct-2507 DAPO Format: 0.1, Model-based: 0.9 8 4 2 5 × 10−6 1 8 0.7 adamw_torch_fused 42

H.5

Baselines

See Table 19.

Table 15: LitSeg-Lite GRPO reward model configuration. Parameter

Value

Model Thinking mode Temperature TopP TopK MinP Presence Penalty Repetition Penalty

Qwen/Qwen3.6-27B-FP8 Enabled 1.0 0.95 20 0.0 1.5 1.0

For this baseline, a smaller model is used instead of Qwen/Qwen3-4B-Instruct-2507 to avoid resource issues. The perplexity calculation requires caching all intermediate hidden states, a process that is both memory-intensive (∼150 GB VRAM with the larger model) and incompatible with standard high-throughput inference frameworks.

Table 19: Baseline chunker configuration. Baseline

Configuration

Token splitter Recursive-character splitter

200 tokens; 40-token overlap. 1,000 characters; 200-character overlap. Separators: paragraph break, newline, period, comma, semicolon, colon, and space. Threshold τ = 0.5; selected minimum sentence is shared by adjacent chunks. Model: Qwen/Qwen3.5-0.8B; maximum token window 9000. Perplexity chunker followed by 200-word dynamic merge. Model: same as the perplexity chunker. Model: Qwen/Qwen3-4B-Instruct-2507; threshold: initialized to 0 and updated by the latest five scores. Margin-sampling chunker followed by 200-word dynamic merge. Model: same as the margin-sampling chunker. Model: Qwen/Qwen3-4B-Instruct-2507; sliding window size: 550 words; maximum retries 20; invalid outputs fall back to splitting after the first unit. LumberChunker followed by 200-word dynamic merge. Model: same as LumberChunker.

Perplexity chunker2

Perplexity + dynamic merge Margin-sampling chunker Margin sampling + dynamic merge LumberChunker

LumberChunker + dynamic merge

Record · ID 229547 · SHA-256 98fda4729b0c5851
Retrieved via Conceptio — every document is proof-bundled with source, license, and retrieval metadata.