Emotion correction for weibo users via weakly supervised learning and semantic understanding.

Emotion correction for weibo users via weakly supervised learning and semantic understanding - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice iScience . 2026 May 27;29(6):116101. doi: 10.1016/j.isci.2026.116101 Emotion correction for weibo users via weakly supervised learning and semantic understanding Ganglong Duan 1 , Zhengjuan Chen 1, 2, ∗ , Yanying Shang 1 , Hongquan Xue 1 Author information Article notes Copyright and License information 1 School of Economics and Management, Xi’an University of Technology, Xi’an 710054, China ∗ Corresponding author 2250520075@stu.xaut.edu.cn 2 Lead contact Received 2026 Feb 25; Revised 2026 Apr 11; Accepted 2026 May 8; Collection date 2026 Jun 19. © 2026 The Author(s) This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). PMC Copyright notice PMCID: PMC13235523  PMID: 42256311 Summary With the fast pace of modern society, individuals are facing increasing psychological stress and emotional challenges. As an important platform for public emotional expression, Weibo contains extensive textual data reflecting users’ emotional states. This paper proposes an emotion correction approach for Weibo users that integrates weakly supervised learning with semantic understanding. The approach combines emotion recognition, emotion-cause classification, and personalized content recommendation, where the emotional polarity of Weibo texts is first identified, negative posts are then classified into corresponding emotion-cause categories, and positively oriented content related to users’ emotion causes is subsequently recommended to provide targeted emotional support. Experiments on a real-world Weibo dataset show that the proposed approach achieves accuracies of 96.2% and 83.5% in emotion recognition and emotion-cause classification, respectively, while also delivering strong recommendation performance. These results demonstrate the potential of the proposed model for supporting users’ emotional correction and contributing to mental health. Subject areas: neuroscience, cognitive neuroscience, machine learning, computer modeling Graphical abstract Open in a new tab Highlights • Emotion correction framework forming a closed loop from recognition to intervention • Eight-category cause lexicon supporting low-cost and interpretable classification • Cause-matching and polarity-screening dual filter for targeted intervention Neuroscience; Cognitive neuroscience; Machine learning; Computer modeling Introduction In today’s fast-paced society, psychological stress and emotional problems are becoming increasingly prevalent, with a growing number of people experiencing negative emotions such as anxiety, depression, and loneliness. Traditional psychological counseling services are limited by manpower, geographic constraints, and social stigma associated with mental illness, making large-scale preventive screening and real-time intervention difficult to achieve. 1 The widespread use of social media has provided the public with diverse channels for emotional expression, among which Weibo has become an important platform for individuals to voluntarily share their emotions. 2 In recent years, natural language processing (NLP) has achieved significant progress and has been widely applied in various domains, such as sentiment analysis, 3 personality detection, 4 and content generation. 5 These technologies enable machines to understand and analyze human language at scale, providing powerful tools for the large-scale analysis of user-generated textual data. Among these applications, sentiment analysis has attracted considerable attention due to its ability to identify emotional polarity and understand public opinion in online environments. Weibo posts contain rich emotional information, and performing sentiment analysis on such data is of great research value. Emotions are closely related to human health and play a crucial role in psychological well-being. Studies have shown that persistent negative emotional states and difficulties in emotion regulation are closely associated with various psychological problems, such as depression and anxiety. 6 Medical studies have further demonstrated that individuals with depression exhibit significant differences in emotional dynamics compared to healthy populations, and emotional features can serve as an important basis for automatic depression detection. 7 From a broader perspective, the widespread diffusion of negative emotions on social platforms, often characterized as emotional contagion, may also have significant social impacts, influencing public opinion and shaping the overall online environment. 8 Therefore, addressing public emotional issues is of great importance for both individual well-being and social stability. However, existing studies in sentiment analysis have primarily focused on emotional polarity recognition, with limited attention to the underlying causes of negative emotions, thereby making it difficult to address users’ emotional issues at their root. Meanwhile, in this era of information overload, recommender systems have emerged as effective tools for personalized information filtering, widely applied in personalized recommendation domains such as e-commerce and news. Yet, conventional recommender systems typically focus on user interests while neglecting their latent emotional needs. If sentiment analysis and recommendation technologies could be integrated—so that users exhibiting negative emotions receive personalized recommendations of emotionally corrective content—this would offer a novel perspective for psychological health intervention. To address the aforementioned challenges, this study proposes an “emotion-corrective recommendation” framework featuring a three-stage architecture: (1) recognition—identifying negative users through a sentiment analysis model; (2) classification—extracting the causes behind users’ negative emotions using an emotion-cause classification model; and (3) recommendation—performing content matching based on the identified causes to achieve targeted emotional recommendations. The framework follows the logical progression of “emotion recognition-cause analysis-emotional intervention,” thereby forming a closed-loop process from emotion detection to regulation. The specific research content mainly consists of four aspects: data construction, emotion recognition, emotion-cause modeling, and the design of an emotion-corrective recommendation mechanism. First, user-generated textual data are collected and preprocessed from the Weibo platform to construct an experimental dataset for sentiment analysis and emotion-cause research. Second, an emotion recognition model is developed to identify the emotional tendencies of user posts and to filter texts exhibiting prominent negative emotional characteristics. Subsequently, from the perspective of the causes of negative emotions, weakly supervised signals and semantic representation methods are introduced to model and classify emotion causes in Weibo texts, and the dominant emotion causes are further summarized and analyzed at the user level. Finally, based on the results of emotion recognition and emotion-cause analysis, an emotion-corrective content recommendation mechanism is constructed. By matching users’ emotion causes with positively oriented content, targeted emotional interventions for users with negative emotions are achieved, forming a complete research workflow from emotion detection and cause analysis to emotional regulation. The main contributions of this paper are as follows. (1) We construct large-scale and diverse Weibo text datasets with corresponding data collection and preprocessing strategies, providing a reusable data foundation for social media sentiment analysis and emotion-cause modeling. (2) We develop and implement an emotion recognition model tailored to Weibo texts, enabling accurate identification of users with negative emotions and supporting subsequent emotion-cause analysis and corrective recommendation. (3) We propose an emotion-cause modeling method by integrating weakly supervised signals with semantic representations, effectively improving the generalization ability of fine-grained emotion-cause recognition under limited annotation conditions. (4) We design a content recommendation mechanism that integrates emotional polarity and emotion-cause consistency, transforming emotion analysis results into actionable positive content recommendations and establishing a closed-loop framework for emotion correction. The main sections of this paper are organized as follows: Section “ RESULTS ” presents the experimental setup and analyzes the evaluation results; Section “ DISCUSSION ” discusses the findings in a broader context, including comparisons with related work and implications for future research; Section “ STAR METHODS ” provides detailed descriptions of the proposed emotion-corrective recommendation framework and the method details. Results Dataset Data collection In sentiment analysis and emotion correction tasks, the selection of datasets directly affects model performance and the reliability of research conclusions. Existing Weibo datasets are mostly constructed at the post level—each user corresponds to only one Weibo post—which fails to capture users’ overall emotional states over extended periods. The objective of this study is to identify users’ overall emotional tendencies and further achieve emotion correction by analyzing the causes of their emotions. Therefore, the data must be user-centered, covering multiple Weibo posts from each user over a certain period, to more accurately capture their long-term emotional characteristics. Based on this requirement, existing public datasets are insufficient; therefore, self-constructed datasets were built. This study constructs three datasets in total. (1) Weibo User Dataset: This dataset serves as the core data foundation of the study and is used across experiments in all three modules. It consists of Weibo posts published by users over approximately one year and is used to construct users’ overall emotional profiles. The final dataset includes 651,949 posts from 11,128 Weibo users collected between January and October 2025, with each user contributing at least 10 posts. (2) Emotion-Cause Lexicon Corpus Dataset: This dataset is used for emotion-cause modeling in the second module and provides corpus support for constructing the emotion-cause lexicon and generating weak supervision signals. It is organized around multiple representative emotion-related themes, with each theme containing approximately 50,000 related Weibo posts, resulting in a total size of about 400,000 posts. (3) Recommendation Candidate Dataset: This dataset is used to construct candidate content for the emotion-corrective recommendation module and serves as the source corpus for delivering positive emotional content to users. It consists of 2,545 Weibo posts collected based on different thematic keywords. Data preprocessing After data collection, the raw Weibo text must be normalized and cleaned to ensure input quality and consistency for the models. The preprocessing pipeline includes the following steps. (1) Data Cleaning Weibo posts often contain a large amount of noise irrelevant to semantics, such as HTML tags, emojis, URLs, and special symbols. Using regular expressions and Python’s built-in string processing functions, such irrelevant characters were systematically removed. (2) Chinese Word Segmentation Since Chinese text is not naturally space-delimited, the Jieba segmentation tool (precise mode) was used to segment cleaned Weibo posts into word tokens, ensuring accurate semantic segmentation and improved embedding learning. (3) Stopword Removal Functional words contribute little to semantic representation. Based on the Harbin Institute of Technology (HIT) stopword list, these high-frequency but semantically weak tokens were removed to enhance the model’s ability to capture core information. For the Emotion-Cause Lexicon Corpus Dataset, two additional processing steps were performed: deduplication and class balancing. Duplicate posts generated during multi-page scraping were removed using the pandas.drop_duplicates() function to ensure data uniqueness. To mitigate bias caused by class imbalance, 50,000 samples were randomly selected from each cause category, resulting in a lexicon corpus with a uniform class distribution. Evaluation metrics Metrics for emotion recognition and classification modules To comprehensively evaluate the performance of the emotion recognition and emotion-cause classification modules, four classical metrics were employed: accuracy, precision, recall, and F1-score. Additionally, a confusion matrix heatmap was introduced to visualize classification results. (1) Accuracy Accuracy represents the proportion of correctly predicted samples among all samples: A c c u r a c y = T P + T N + F P + F N T P + T N (Equation 1) where TP (true positive) refers to samples correctly predicted as positive, TN (true negative) to samples correctly predicted as negative, FP (false positive) to negative samples misclassified as positive, and FN (false negative) to positive samples misclassified as negative. (2) Precision Precision measures the proportion of TPs among all samples predicted as positive: P r e c i s i o n = T P T P + F P (Equation 2) A higher precision indicates a lower false-positive rate and greater purity of predictions. (3) Recall Recall measures the proportion of correctly predicted positive samples among all actual positives: R e c a l l = T P T P + F N (Equation 3) A higher recall means the model can identify more positive samples. (4) F1-Score The F1-score is the harmonic mean of precision and recall, providing a balanced measure of both: F 1 = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l (Equation 4) When precision and recall differ greatly, the F1-score provides a more objective assessment of overall model performance. (5) Confusion Matrix The confusion matrix visually represents the model’s classification performance across categories, typically structured as: [ T P F P F N T N ] (Equation 5) The diagonal elements represent correctly classified samples, while off-diagonal values indicate misclassifications. Visualizing this matrix as a heatmap allows intuitive identification of category-specific weaknesses. Metrics for recommendation module To evaluate the performance of the recommendation module, we adopt three ranking-oriented metrics, including Precision@k, normalized discounted cumulative gain (NDCG@k), and mean average precision (MAP@k). To assess recommendation quality, a three-point rating scale (0–2) is adopted to evaluate each recommended Weibo post. The scoring criteria are defined as follows. 0 (Irrelevant/Unhelpful): The content is completely unrelated to the user’s emotional cause, or although related, it expresses negativity that may intensify the user’s negative emotions. 1 (Partially Relevant, Limited Help): The content has some connection to the user’s emotional cause, but it is not sufficiently positive, or its effect on emotional regulation is limited. 2 (Relevant and Positive, Effective): The content is clearly related to the user’s emotional cause, and it conveys a positive, constructive message that can effectively alleviate the negative emotions associated with that cause. (1) Precision@k Precision@k measures the proportion of relevant items among the top-k recommended items, reflecting the recommendation accuracy at position: P r e c i s i o n @ k = 1 k ∑ i = 1 k r e l i (Equation 6) where rel i ∈{0,1} denotes whether the i -th item is relevant. (2) NDCG@k (normalized discounted cumulative gain) NDCG@k evaluates ranking quality by considering both relevance scores and positions of recommended items, with values closer to 1 indicating better ranking performance: N D C G @ k = D C G @ k I D C G @ k (Equation 7) D C G @ k = ∑ i = 1 k 2 r e l i − 1 l o g 2 ( i + 1 ) (Equation 8) where IDCG @ k is the ideal discounted cumulative gain under perfect ranking, and rel i is the relevance score of the i -th item. (3) MAP@k (mean average precision) MAP@k is the mean of average precision across multiple users or recommendation lists, which comprehensively assesses overall recommendation performance over the test set: M A P @ k = 1 Q ∑ q = 1 Q A P q @ k (Equation 9) A P q @ k = 1 m q ∑ i = 1 k P r e c i s i o n @ i · r e l i (Equation 10) where Q is the total number of users or recommendation lists, m q is the number of relevant items for user q , and AP q @ k is the average precision at k for user q . Experimental setup and parameter settings All experiments were conducted on a workstation equipped with an Intel Core i5-1135G7 CPU and 16 GB of RAM. The models were implemented in Python 3.8 under the Windows 11 operating system using the PyTorch (version 2.4.1) and Keras (version 2.10.0) deep learning frameworks. Different models used in various stages of this study were configured with detailed parameter settings, as listed later in discussion. (1) Word2Vec Model During the word embedding training phase, a Word2Vec model based on the Skip-Gram architecture was employed to construct pre-trained word vectors. The parameters are shown in Table 1 . (2) BiLSTM Model Parameters Table 1. Word2Vec model parameter settings Parameter Value vector_size 100 Windows 5 min_count 2 Negative 15 Epochs 10 Open in a new tab In the sentiment analysis stage, a BiLSTM model was adopted, with its embedding layer initialized using the pre-trained Word2Vec embeddings. The parameters are shown in Table 2 . (3) TF-IDF Model Parameters Table 2. BiLSTM model parameter settings Parameter Value input_length 100 embedding_dim 100 lstm_units 128 SpatialDropout1D (rate) 0.2 Dropout 0.3 recurrent_dropout 0.3 Optimizer Adam learning_rate 0.001 batch_size 128 Epochs 10 Open in a new tab In the emotion cause classification module, TF-IDF was used to extract keywords for lexicon construction. The parameters are shown in Table 3 . (4) BERT Model Parameters Table 3. TF-IDF model parameter settings Parameter Value token_pattern r”(?u)\b\w{2,}\b” max_df 0.8 min_df 2 ngram_range (1,1) smooth_idf True sublinear_tf True top_n 500 Open in a new tab Note: The token_pattern r”(?u)\b\w{2,}\b” is a regular expression used for tokenization, which extracts tokens consisting of at least two alphanumeric characters, in order to avoid introducing a large number of uninformative single-character tokens. In the emotion cause classification stage, a BERT pre-trained language model (PLM) was employed. The model was trained on weakly labeled data and fine-tuned on a small labeled dataset. The base model used in this study is BERT-base-Chinese. The parameters are shown in Table 4 . Table 4. BERT model parameter settings Parameter Value max_len 128 batch_size 16 learning_rate 2e−5 (training)/1e−5 (fine-tuning) weight_decay 0.01 Epochs 15 (training)/10 (fine-tuning) Optimizer AdamW Open in a new tab Experimental results and analysis This study centers on three core modules: sentiment analysis, negative emotion cause classification, and emotion correction. These modules are sequentially connected—each module’s output serves as the input for the next—forming a complete workflow from overall emotion detection to cause classification and finally emotion-corrective recommendation. The model architecture, module structures, and experimental workflow are illustrated in Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 . Figure 1. Open in a new tab Overall architecture of the RCR model Figure 2. Open in a new tab Structure of the sentiment analysis model Figure 3. Open in a new tab Construction process of the negative emotion causes lexicon Figure 4. Open in a new tab Structure of the BERT-based emotion-cause classification model Figure 5. Open in a new tab Experimental workflow of the recommendation model Figure 6. Open in a new tab Process of content recommendation Sentiment analysis module The objective of the first module is to determine the overall emotional tendency of Weibo users, i.e., to identify whether a user is “negative.” The model employs pre-trained Word2Vec embeddings as input and a BiLSTM for modeling. To enhance generalization and robustness, the training corpus integrates multiple public Weibo sentiment datasets, including weibo_senti_100k.csv and other open-source corpora from GitHub, totaling 255,560 samples. The dataset was split into 70% training, 15% validation, and 15% test sets. The sentiment analysis performance of the model is presented in Table 5 . Table 5. Sentiment analysis performance Class Precision Recall F1-score Negative 0.9521 0.9733 0.9626 Positive 0.9726 0.9509 0.9616 Open in a new tab Accuracy: 96.2%. The evaluation is conducted on a held-out test set comprising 15% of the full dataset. The results indicate that the model achieves high accuracy and stable classification performance in the sentiment analysis task. On the basis of validating the overall performance of the BiLSTM model, further comparative experiments are conducted to evaluate its relative advantages in sentiment classification. Specifically, several mainstream text sentiment classification methods are selected for comparison, including traditional machine learning-based logistic regression (LR) and support vector machine (SVM) models, as well as deep learning-based TextCNN and BERT models. The F1 score for positive and negative sentiment classes is used as the evaluation metric. As shown in Figure 7 , traditional machine learning models achieve reasonable performance in sentiment classification tasks; however, their overall effectiveness is constrained by limited feature representation capability. In contrast, deep learning models demonstrate clear advantages in capturing contextual semantic information and recognizing sentiment patterns in Weibo texts. The TextCNN and BERT models achieve competitive performance. Their performance is slightly inferior to that of the BiLSTM model. Therefore, we further examine the computational complexity of the three models. Figure 7. Open in a new tab Performance comparison of sentiment analysis models The evaluation is conducted on a held-out test set comprising 15% of the full dataset. Performance is measured using the F1-score for both positive and negative classes. We analyze their parameter size (Params), floating-point operations (FLOPs), and actual training time. The BiLSTM model contains approximately 2.23M parameters with about 23M FLOPs per sample. The TextCNN model increases the model capacity to 5.27M parameters and 1.85 GFLOPs, while the BERT-based model includes about 102M parameters with approximately 22 GFLOPs. The corresponding training times for BiLSTM, TextCNN, and BERT are 8.28 h, 9.61 h, and 30.03 h, respectively. Considering both classification performance and computational efficiency, the BiLSTM model provides the best balance between effectiveness and efficiency. Negative emotion cause classification module In the second module, the goal is to classify the negative Weibo posts of the previously identified users into eight emotion cause categories. A hybrid method combining lexicon-based weak supervision and BERT modeling was used. (1) Baseline Model Comparison To evaluate the effectiveness of the proposed hybrid approach, a set of comparative experiments was conducted. The baseline methods and the method proposed in this paper are defined as follows. ①Pure Lexicon Method: Classification is performed solely based on keyword matching using the constructed negative emotion-cause lexicon. ②Pure BERT Method: A standard pretraining-fine-tuning paradigm is adopted, where a pretrained BERT model is directly fine-tuned on 280 manually annotated samples. ③Lexicon + TextCNN: The emotion-cause lexicon is used to generate pseudo labels for unlabeled Weibo texts, and a TextCNN model is trained on the resulting weakly supervised dataset to perform emotion-cause classification. ④Lexicon + BiLSTM: Pseudo labels generated from the emotion-cause lexicon are employed to train a BiLSTM-based model for emotion-cause classification under a weakly supervised setting. ⑤Lexicon + BERT (Proposed Method): The emotion-cause lexicon is first used to generate pseudo labels for large-scale unlabeled Weibo texts to perform weakly supervised pretraining of BERT, followed by fine-tuning with a small amount of manually annotated data to enhance semantic understanding and generalization ability. All models were evaluated on the same test set containing 520 samples, with the “other” category excluded to focus on the recognition of primary emotion causes. The performance comparison of different methods is presented in Table 6 , and the corresponding bar chart visualizing these results is shown in Figure 8 . Table 6. Performance comparison of different methods on the emotion cause classification task Model Precision Recall F1-score Pure Lexicon Method 0.7934 0.7673 0.7686 Pure BERT Method 0.6839 0.6827 0.6745 Lexicon + TextCNN 0.7845 0.7423 0.7469 Lexicon + BiLSTM 0.7793 0.7635 0.7634 Lexicon + BERT (Proposed Method) 0.8382 0.8346 0.8318 Open in a new tab The evaluation is performed on a test dataset consisting of n = 520 Weibo posts, covering eight emotion-cause categories with 65 samples per category. Figure 8. Open in a new tab Performance of different models on emotion cause classification The evaluation is performed on a test dataset consisting of n = 520 Weibo posts, covering eight emotion-cause categories with 65 samples per category. Precision, Recall, and F1-score are used as evaluation metrics. As shown in the results, the pure lexicon-based method can achieve reasonable performance without manual annotation; however, due to its reliance on keyword matching and lack of contextual semantic modeling, its discriminative ability is limited. When neural models such as TextCNN and BiLSTM are introduced and trained on lexicon-generated pseudo labels, semantic modeling capability is enhanced, but performance does not consistently surpass the pure lexicon method, likely due to noise in the pseudo labels. The pure BERT method, which relies solely on a small annotated dataset, also shows limited performance improvement because of insufficient training data. In contrast, the proposed Lexicon + BERT approach achieves the best performance across all evaluation metrics. By combining weakly supervised pretraining with limited supervised fine-tuning, the proposed method effectively leverages both lexicon-based prior knowledge and BERT’s strong semantic modeling capability, leading to more accurate emotion-cause classification under low-resource conditions. (2) Comparison with Mainstream ECPE Models To further assess the performance level of the proposed hybrid model in the field of emotion-cause analysis, it was compared with mainstream ECPE models. Although ECPE is a more complex sequence-labeling task, while this study focuses on document-level classification, the two tasks are comparable in their core ability of cause recognition. Several high-performing ECPE models from recent years were selected, and their macro-F1 performance on the cause clause recognition subtask was compared. ①GAT-ECPE 9 : A graph attention network incorporating external knowledge, which constructs inter-clause dependency and probability graphs to achieve multi-level aggregation of semantic and structural features. ②SEG 10 : A semantic-structure-guided emotion model integrating emotion type and semantic structure information to capture relationships between emotion and cause clauses. ③HSE-TTLS 11 : A multi-task framework with a hierarchical shared encoder and task-specific Transformer layers, decomposing the extraction task into emotion extraction and emotion-cause extraction subtasks. By comparison, as shown in Figure 9 , the proposed lexicon-based weak supervision and BERT modeling approach achieves a macro-F1 of 0.8318, surpassing the above ECPE models in cause recognition performance. In addition, the proposed method also achieves the highest Precision (0.8382) and Recall (0.8346) among the compared models, indicating its consistently superior performance across all three evaluation metrics. This result indicates that the proposed approach, while maintaining a simpler task formulation and lower annotation requirements, achieves comparable or superior performance in cause recognition compared with current state-of-the-art ECPE extraction models. This fully demonstrates the effectiveness and practical potential of the proposed method in user emotion-cause analysis. (3) Final Model Performance Analysis Figure 9. Open in a new tab Performance comparison of the proposed method and mainstream ECPE models Proposed method results are from our test set ( n = 520). Results for GAT-ECPE, SEG, and HSE-TTLS are cited from original publications (different corpora). Performance is evaluated in terms of Precision, Recall, and F1-score. This section further analyzes the performance of the final model. Figure 10 presents the confusion matrix of the final negative-emotion cause classification results. Figure 10. Open in a new tab Confusion matrix of the final negative-emotion cause classification results The confusion matrix is generated based on a test dataset consisting of n = 520 samples across eight emotion-cause categories. The detailed evaluation results of the model on the test set are summarized in Table 7 . Table 7. Evaluation of the negative emotion-cause classification model Class Precision Recall F1-score Health problems 0.9104 0.9385 0.9242 Appearance anxiety 0.8356 0.9385 0.8841 Academic problems 0.7792 0.9231 0.8451 Family issues 0.7679 0.6615 0.7107 Romantic entanglement 0.7917 0.8769 0.8321 Social difficulties 0.8776 0.6615 0.7544 Financial stress 0.8143 0.8769 0.8444 Workplace dilemmas 0.9286 0.8000 0.8595 Macro avg 0.8382 0.8346 0.8318 Open in a new tab Overall Accuracy: 83.5%. The evaluation is conducted on a test dataset consisting of n = 520 Weibo posts across eight emotion-cause categories, with 65 samples per category. From these detailed results, it can be observed that the model achieves high precision and recall across most categories. In particular, categories with clearer thematic features, such as health problems, appearance anxiety, and workplace dilemmas, exhibit the best performance. Overall, the model demonstrates a strong ability to accurately distinguish between different types of emotional causes. Emotion-corrective recommendation module In the third module, the objective of the experiment is to recommend Weibo posts that are both highly relevant to the user’s identified negative emotion causes and exhibit positive emotional tendencies, thereby achieving the goal of emotion correction. The list of emotion causes identified for each user in the previous module is ranked by frequency of occurrence, where a higher frequency indicates a stronger negative emotional influence on the user. Accordingly, the system adopts a weighted allocation mechanism, dynamically determining the number of recommended posts based on the proportional frequency of each emotion cause. For example, for the user “AllSheWrote,” the system identified the following negative emotion causes and their frequencies: social difficulties (9), romantic entanglement (7), health problems (4), financial stress (3), academic problems (2), workplace dilemmas (2), and family issues (1). Based on the weight distribution, within a total of ten recommendations, the system allocated four posts related to Social Difficulties, two posts related to Romantic Entanglement, and one post each for the other four causes, while the least frequent cause (family issues) was not assigned any recommendation. This design reflects the model’s principle of prioritizing the user’s core emotional conflicts. The specific recommendation results for this user are presented in Table 8 (long texts are partially omitted for brevity). Table 8. Emotion-corrective recommendations for user “AllSheWrote” Negative Emotion Cause Recommended Content Social difficulties Daily Q&A: Teacher, hello. For someone who is naturally quiet and reserved, how can they change their personality to become more talkative? This would help them leave a stronger impression on others. How can they improve? Answer: People are products of their environment; personality develops alongside environmental influences … Personally, I think if it doesn’t affect daily life or work, there’s no need to change one’s personality … If your personality is naturally like that of a tech-savvy engineer like Ma Huateng, there’s no need to force yourself to become someone with the personality of Jack Ma or Liu Qiangdong. Find a platform and work where you can leverage your talents, focus on it, and stick with it. You’ll achieve great things. What your personality is doesn’t matter much; success primarily depends on persistence, perseverance, and seizing opportunities. #WorkplaceStories Social difficulties Share more with your good friends and keep your enthusiasm alive. Good friends are the light of life, my friends. Social difficulties Sometimes, when hanging out with friends, things can get a little awkward. But I’ve come to realize that true friendships can actually grow out of these moments. The key is to let go of inhibitions and just be yourself. Social difficulties Online socializing sometimes makes me feel distant, but I’ve learned to express myself sincerely and gradually found a way of communication that suits me. Whether online or offline, sincerity is always the core of connection. Romantic entanglement The purpose of meeting is to release the deep affection that builds up in my heart when we’re apart. When we’re not together, I focus on living well and eagerly anticipate our next reunion. #EmotionalQuotes #Women #Emotions #SpringLifeRecords Romantic entanglement #LoveIsNotANanthem# Love is not an antidote; it’s a flower with thorns sharper than roses … Love isn’t a necessity. True love means I’m already whole, but with you, I feel even better … True love is a mutual journey, not a tangled entanglement. If love isn’t there, live well anyway—merely meeting in this world is enough. #100 Ways to Express Love##Weibo Interest Creation Project##Emotions##Emotional Confessions##Life Musings##Emotional Quotes#Emotional Quotes Super Topic Health problems When you have nothing, health is your trump card; When you have everything, health is your ace in the hole. Life is really about keeping yourself happy; First and foremost, you must keep yourself smiling. In this lifetime, nothing belongs to you except your health. #HealthyLiving Financial stress Consume rationally, buy only what you need, avoid following trends, and prevent waste to cultivate sound spending habits. Academic problems #GraduateEntranceExamCountdown#280 days left ∼ Daily Motivation for Test-Takers: Time rewards every person who strives quietly and diligently! #GraduateExamFightOn##StudyCheckIn#GraduateExamSuperTalk#GraduateExamSuccess Workplace dilemmas What truths did you only grasp after becoming a leader? Your position dictates your perspective—where you sit often shapes how you think … The essence of the superior-subordinate relationship is fundamentally a partnership. In the real workplace, many people are extremely “shrewd,” weighing benefits and drawbacks with crystal clarity. If you’re a weak leader or lack strong resources, they often won’t put their heart or effort into working with you. Being a leader (manager) isn’t always easy—it requires both proactively seeking resources and making appropriate compromises. #RisingStarProgram #WorkplaceStories #OfficeGossip Open in a new tab As shown in Table 8 , the recommendation system is able to match thematically relevant and positively oriented content to different types of users’ negative emotion causes. For instance, for the cause Social Difficulties, the system recommends content that encourages adjustments in social mindset and interpersonal cognition; for Health Problems and Financial Stress, the recommended posts tend to promote a more positive attitude toward life and rational consumption behaviors. Overall, the recommendation results exhibit a certain degree of emotional positivity and thematic consistency, suggesting the potential effectiveness of the proposed emotion correction mechanism. To further evaluate the recommendation performance, experiments are conducted on the annotated dataset, and the results are reported in Table 9 . The baseline methods and our proposed method are defined as follows. (1) Sentiment-Based: This method generates recommendations by randomly selecting content from the positive candidate pool after sentiment filtering. (2) Similarity-Based: This method ranks candidate posts based on similarity, without applying sentiment filtering. (3) Hybrid (Proposed): The proposed method combines sentiment filtering with similarity-based matching to generate recommendations. Table 9. Performance comparison of different recommendation methods Model Precision@10 NDCG@10 MAP@10 Sentiment-based 0.0980 0.5996 0.5336 Similarity-based 0.4420 0.7390 0.5544 Hybrid (proposed) 0.6820 0.8769 0.7701 Open in a new tab The evaluation is conducted on n = 50 users, where each user is associated with a recommendation list of length 10. Recommendation performance is measured using Precision@k, NDCG@k, and MAP@k, averaged across all users. As shown in Table 9 , the recommendation dataset is constructed based on topic-related keywords, and therefore, the overall performance of all methods appears relatively strong compared with general recommendation scenarios. The proposed Hybrid model consistently outperforms the baseline approaches across all evaluation metrics, including Precision@10, NDCG@10, and MAP@10. Compared with the Sentiment-Based method, the proposed model significantly improves recommendation relevance, indicating the importance of incorporating emotion-cause information. In addition, compared with the Similarity-Based method, the Hybrid model demonstrates better performance in both ranking quality and overall effectiveness, suggesting that integrating emotional polarity and cause consistency contributes to more accurate and meaningful recommendations. Overall, the results validate the effectiveness of combining sentiment filtering and similarity modeling in supporting emotion correction through more accurate and meaningful recommendations. Discussion This study addresses the problem of identifying and intervening in negative emotions among social media users by proposing an integrated emotion-corrective recommendation framework. To address the limitations of existing studies, the proposed approach combines sentiment analysis, emotion-cause modeling, and recommendation strategies into a unified pipeline. Specifically, the framework follows a three-stage process of “recognition-classification-recommendation,” forming a closed-loop mechanism from emotion perception to emotional intervention. The results demonstrate that the proposed framework can effectively capture users’ emotional tendencies and reveal their underlying causes, thereby enabling more targeted and positive emotional guidance. Comparison with related work Emotion correction From a psychological perspective, emotion correction refers to helping individuals adjust maladaptive emotional response patterns through specific interventions, thereby establishing healthier emotional experiences and modes of expression. At present, systematic research on emotion correction remains limited. Miu et al. 12 conducted a meta-analysis and found that individuals with childhood trauma are more prone to difficulties in emotion regulation, tend to suppress or ruminate on negative emotions, and consequently face a higher risk of psychological disorders. Troy et al. 13 proposed an integrative theoretical framework of emotion regulation, suggesting that psychological resilience results from the interaction among emotion regulation strategies, adversity characteristics, and cultural factors, and emphasizing the central role of emotion regulation in coping with adversity. Stellern et al. 14 also demonstrated through meta-analysis that individuals with substance use disorders experience greater difficulties in emotional control—particularly in impulse management and the use of regulation strategies—than healthy individuals. Overall, existing studies have been primarily situated within the field of clinical psychology, focusing on theoretical models of self-emotion regulation or analyses of variable relationships. In addition, within research that applies computational techniques to mental health analysis, many studies focus on detecting or identifying specific psychological problems from user data. For example, Xu et al. 15 proposed ABAFnet, an attention-driven acoustic feature fusion network that integrates four types of acoustic features and incorporates a type-adaptive CNN, an LSTM-Attention mechanism, and a dynamic weight fusion strategy to improve depression detection performance. Zhu et al. 16 introduced a multimodal model based on image captioning and cross-attention mechanisms, which enhances visual representations through semantic image descriptions and OCR information and fuses textual and visual features via multi-head cross-attention to accomplish anxiety recognition. However, these studies remain at the detection stage and offer limited exploration of how to provide effective and automated follow-up interventions for identified users. Therefore, whether in theory-oriented psychological research or computational approaches focused on mental-state recognition, studies on concrete emotion correction strategies that can be implemented in specific contexts are still relatively scarce. In clinical psychology, cognitive behavioral therapy (CBT) is recognized as one of the most effective forms of psychological intervention. Proposed by Beck, 17 , 18 CBT is grounded in the principle of identifying and modifying irrational cognitive patterns and negative automatic thoughts to alter emotional responses and behavioral outcomes. This provides an important theoretical foundation for the present study. Building on the psychological concept of emotion correction, this study extends and applies it by constructing a computational model based on social media data. Unlike previous research relying on users’ self-regulation, the proposed RCR closed-loop system proactively delivers positively oriented content. This approach transforms emotion correction from a theoretical construct into a computational and automated intervention, offering a novel pathway for psychological health support in non-clinical settings. Text sentiment analysis Text sentiment analysis is a key NLP task that aims to determine the emotional polarity of text through automated analysis of lexical, syntactic, and semantic information. Current mainstream approaches are based on deep learning, using neural networks to automatically extract semantic features, capture affective signals, and model contextual dependencies for emotion classification. Recent research has continuously sought to improve the accuracy of emotional polarity classification. For example, Onan 19 proposed a convolutional recurrent neural network architecture combining BiLSTM and GRU with an intra-group enhancement mechanism to optimize feature extraction. Talaat 20 integrated pre-trained BERT models with BiLSTM and BiGRU structures for sentiment classification, demonstrating that models incorporating BiGRU significantly outperform traditional and single BERT baselines. Ouyang et al. 21 introduced the ASHGAT model based on a hypergraph attention mechanism, which constructs word-level hypergraphs that integrate syntactic and semantic relations, thereby enhancing the modeling of associations between aspect words and sentiment words. Meng et al. 22 proposed a multi-message passing framework based on heterogeneous graphs, which integrates multimodal features including text, audio, and visual signals and models interactions among speakers through graph-based message passing, for emotion category identification in conversational contexts. Although deep learning-based sentiment analysis has achieved remarkable progress in emotion classification, user-generated content contains not only emotional categories but also the underlying causes that trigger these emotions. In real-world applications, understanding the underlying causes and contextual events behind emotions is of greater practical significance, particularly for emotion regulation and psychological intervention. Nevertheless, research in this area remains relatively limited. In recent years, with the continuous advancement of information extraction tasks, emotion cause extraction (ECE) has gradually gained increasing attention. 23 The task of ECE was first introduced by Lee et al., 24 whose goal was to identify the underlying causes behind emotions expressed in text. They constructed the first Chinese emotion cause corpus and proposed a rule-based detection method that relied on linguistic cues. However, their approach required prior emotion annotation before cause extraction, thereby overlooking the interdependence between emotions and their causes. To address this issue, Xia et al. 25 proposed the emotion-cause pair extraction (ECPE) task, which enables simultaneous extraction of emotions and corresponding causes from documents without pre-labeling. ECPE is typically divided into two stages: individual extraction of emotional and causal clauses via multi-task learning, followed by emotion-cause pairing and filtering. While ECPE has substantially advanced research in emotion cause analysis, several challenges remain. First, ECPE is inherently a joint extraction task with high modeling complexity, making unified representation difficult. Second, the task requires simultaneous annotation of emotional and causal clauses, which is highly labor-intensive; moreover, public datasets remain limited in scale and diversity. Third, many state-of-the-art ECPE models depend on complex model architectures and training strategies, which impose high computational costs during training and deployment. Even the best-performing ECPE models on public datasets still exhibit considerable room for improvement in accuracy and robustness, particularly when dealing with complex or implicit emotional-causal relationships. This study not only extends the focus of sentiment analysis from traditional polarity classification to the exploration of the underlying causes of negative emotions, but also proposes a lightweight and practical classification approach that reduces dependence on large-scale manual annotation. Pre-trained language models and weakly supervised learning In recent years, PLMs have represented a major breakthrough in NLP. Trained through self-supervised learning on large-scale corpora, these models capture rich semantic and syntactic representations. Through fine-tuning, knowledge from PLMs can be effectively transferred to downstream tasks, enabling rapid adaptation to diverse application domains. BERT, proposed by Devlin et al., 26 adopts a bidirectional Transformer architecture and performs deep semantic modeling through masked language modeling and next sentence prediction. It has been widely applied in various NLP tasks, such as sentiment analysis, named entity recognition (NER), and dialogue generation. For instance, Wan et al. 27 proposed ECR-BERT, which integrates the OCC emotion model’s cognitive rules with BERT for sentiment analysis of emergency-related social media texts. Li et al. 28 proposed a self-distillation-based multi-level semantic enhancement BERT model for Chinese NER, which introduced multi-granularity semantic features and a reweighted distillation mechanism to improve recognition performance. Lv et al. 29 proposed a transformer-based dialogue generation model for Chinese text, which incorporated representations of the large-scale pre-trained Chinese BERT model NEZHA (a Chinese BERT-based model) into character-level embeddings via a word-phrase fusion embedding mechanism to enhance semantic understanding, and additionally optimized sparse Softmax to boost generation capabilities, thus effectively elevating the quality of Chinese dialogue generation. Despite their success, pre-trained models still face challenges such as high annotation costs and limited domain adaptability in practical applications. To alleviate these issues, weakly supervised learning has been proposed as a viable solution. This paradigm aims to exploit incomplete, inexact, or limited supervision—often through rules, lexicons, pseudo-labels, or teacher models—to generate indirect training signals and reduce manual labeling costs. 30 A growing body of research has explored neural model training under weakly or semi-supervised conditions. For example, Lien et al. 31 proposed a generalized weak supervision (GWS) framework for neural information retrieval, which iteratively re-labels data to improve ranking performance without manual annotations. Xie et al. 32 introduced a class-aware pseudo-labeling (CAP) method for weakly supervised multi-label image classification, which adaptively sets class-specific thresholds to generate pseudo-labels and jointly trains with true labels, improving classification accuracy. You et al. 33 developed a semi-supervised contrastive learning framework for medical image segmentation, which enhances model stability and performance under limited annotation conditions through variance reduction and stratified sampling. The combination of deep learning and weak supervision has become an important direction for addressing data scarcity. Compared with image data, where visual features and augmentation techniques are more intuitive, textual data pose greater challenges due to semantic complexity and label noise sensitivity. 34 However, integrating PLMs within a weakly supervised framework leverages their deep semantic understanding to mitigate pseudo-label noise. For example, Wallaart et al. 35 proposed a weakly supervised Sentence-based Aspect Category and Sentiment Classification (SB-ASC) method, which utilizes Sentence-BERT to generate high-quality pseudo-labels and jointly trains them with a BERT model for aspect and sentiment prediction. Building on this idea, the present study integrates weakly supervised learning with pre-trained language models to address the dual challenges of limited annotation and textual noise, and develops a negative emotion-cause classification model that achieves strong semantic representation and domain adaptability even under low-resource conditions. Recommendation algorithms Recommender systems were originally developed to help users efficiently locate content matching their needs. Their core principle is to predict user preferences for unseen items by analyzing users’ historical behaviors, interests, and item attributes. Such systems have been widely applied across e-commerce, social media, and other online platforms, attracting considerable attention from researchers in recent years. With the deepening of research, scholars have gradually begun integrating emotional factors into recommendation systems. Abbasi-Moud et al. 36 proposed a semantic clustering-based sentiment analysis approach for tourism recommendation, which analyzes emotional tendencies in user reviews before generating personalized recommendations that combine contextual and destination features. Choudhary et al. 37 developed SARWAS, a deep ensemble learning method that fuses sentiment analysis outputs with user ratings and models them using multilayer neural networks to achieve more accurate and personalized product recommendations. Darraz et al. 38 combined lexicon-based and deep learning approaches by integrating TextBlob and MLP models with DeepMF collaborative filtering to improve hotel recommendation accuracy. Although such emotion-aware recommender systems can align recommendations with users’ emotional tendencies, they primarily capture attitudinal preferences toward content or products rather than addressing users’ underlying psychological and emotional states. In essence, these systems remain interest-driven. More critically, consistently recommending content that matches users’ current negative emotions may exacerbate these emotions—a phenomenon termed the “emotion amplification” effect. Mechanistically, this is similar to the “preference amplification” effect, where users mainly consume recommended content, leading to a narrowing of interests, reinforcement of preferences, content homogenization, and the formation of “filter bubbles.” 39 For instance, when a user experiences heartbreak, continuously recommending sad quotes or breakup-related content may deepen negative emotions rather than alleviate them. Hence, emotionally congruent recommendations are not necessarily emotionally beneficial. To address this issue, Liu et al. 40 proposed a novel emotion-based personalized music recommendation (EPMR) framework built on LSTM networks. The framework incorporates a “care factor” as a bias parameter, ensuring that recommended music is slightly more positive than the user’s current mood, thereby promoting emotional uplift. Inspired by this work, the present study introduces the concept of emotion-corrective recommendation, aiming to achieve emotion correction through Weibo content recommendation. However, in contrast to the EPMR framework, this study advances emotion modeling by not only identifying users’ emotional states but also extracting the underlying causes of their negative emotions from textual data. This enables genuinely targeted interventions, allowing the system to “treat the cause rather than the symptom” by recommending positive Weibo content tailored to each user’s specific emotional cause. Overall, although existing studies have made substantial progress in emotion-related research from multiple perspectives, several limitations remain. First, research on emotion correction is still largely confined to theoretical frameworks in psychology, with the limited exploration of computational and automated intervention methods in real-world scenarios. Second, most sentiment analysis studies focus primarily on emotion polarity recognition, while the underlying causes of emotions are often overlooked, limiting their practical value in emotional intervention. Third, although pre-trained language models and weakly supervised learning have shown strong potential in addressing data scarcity and improving semantic representation, their integration for fine-grained emotion-cause analysis remains insufficiently explored. Finally, existing emotion-aware recommender systems mainly emphasize aligning recommendations with user preferences or emotional tendencies, but lack mechanisms to actively guide users toward positive emotional states, and may even lead to emotion amplification effects. Therefore, there remains a research gap in developing an integrated framework that can simultaneously identify users’ emotional states, uncover the underlying causes of negative emotions, and provide targeted emotion-corrective recommendations. In the emotion recognition stage, the model achieves high classification accuracy (96.2%), suggesting that the combination of Word2Vec embeddings and BiLSTM architecture is capable of capturing contextual emotional features in semantically diverse, noisy, and context-dependent Weibo texts. This provides a reliable foundation for subsequent modules. In the emotion-cause classification stage, the hybrid approach integrating lexicon-based weak supervision and BERT demonstrates clear advantages over both purely rule-based and fully supervised methods, particularly in balancing semantic representation with limited annotated data. As emotion-cause identification is still relatively challenging and less explored, the performance of this module suggests that incorporating lexicon-based weak supervision with BERT-based contextual modeling provides an effective way to capture fine-grained emotional triggers. In the recommendation stage, the proposed hybrid strategy that combines sentiment filtering with similarity-based matching significantly outperforms baseline methods in Precision@k, NDCG@k, and MAP@k, and both qualitative case analysis and quantitative evaluation confirm its ability to provide users with targeted recommendations of emotionally corrective Weibo content that is closely related to their underlying emotional causes. Taken together, these findings suggest that the three modules function not only as independent components but also as an integrated pipeline, in which the robustness and reliability of upstream results directly support the effectiveness of downstream emotional intervention. The proposed research demonstrates significant theoretical and practical value. From a theoretical perspective, this study extends traditional sentiment analysis from emotion polarity recognition to emotion-cause modeling and intervention-oriented applications, offering a perspective that connects emotion understanding with emotional regulation. It also moves beyond conventional recommender systems that focus primarily on user preferences by incorporating emotion-cause information into the recommendation process, thereby introducing an emotion-aware and correction-oriented modeling paradigm. In addition, by integrating methods from NLP, deep learning, and recommender systems, this study illustrates how interdisciplinary approaches can be effectively combined to address complex social and behavioral problems, providing a useful reference for related cross-domain research. From a practical perspective, the proposed method enables a non-intrusive and low-cost intervention mechanism by leveraging social media data, allowing emotional guidance to be embedded into users’ everyday information consumption. In practice, many users are reluctant to actively seek psychological support, while the proposed framework can identify potential negative emotional states without requiring explicit user participation and deliver emotionally corrective content through recommendation. By modeling the underlying causes of users’ emotional expressions, the system provides cause-aware positive content in a more natural and less intrusive manner, helping to reduce psychological resistance. Furthermore, by mitigating the causes of negative emotional content, the proposed approach can help limit the spread of negative emotions in social networks and contribute to a healthier online environment, with potential applications in mental health support and public health services. Limitations of the study Despite these contributions, several limitations remain. The ultimate effectiveness of the recommendations—i.e., whether users’ emotional states improve—requires actual user feedback, which has not yet been collected, and thus the real-world impact of the emotion correction intervention has not been directly evaluated. In addition, factors such as individual differences in emotional expression, linguistic metaphors, and contextual dependency may affect the model’s generalization ability. Furthermore, the current framework mainly relies on textual data and does not fully incorporate multimodal information, which may limit its applicability in more complex social media environments. Future research can be conducted in several directions. Online experiments or emotion-scale questionnaires can be designed to collect users’ emotional changes before and after receiving recommendations, thereby enabling real-world validation of the model’s effectiveness in emotion correction. Moreover, integrating multimodal information such as images, emojis, and comments may further improve the model’s ability to recognize complex emotional states and enhance the precision of intervention strategies. Finally, the proposed correction strategy can be combined with dialogue systems to enable real-time interaction and dynamic adjustment of recommendations based on user feedback, ultimatelyachieving a personalized emotion correction mechanism characterized by “dynamic understanding-instant intervention.” Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Zhengjuan Chen ( 2250520075@stu.xaut.edu.cn ). Materials availability This study did not generate new unique reagents. Data and code availability • The datasets generated and analyzed during this study have been deposited at GitHub and are publicly available as of the date of publication at: GitHub: https://github.com/zj521Q/WeiboEmotionCorrection . • All original code has been deposited at GitHub and is publicly available as of the date of publication at: GitHub: https://github.com/zj521Q/WeiboEmotionCorrection . • Any additional information required to reanalyze the data reported in this paper is available from the Lead contact upon request. Acknowledgments This research was funded by the Shaanxi Provincial Special Research Program for Philosophy and Social Sciences (grant nos. 2025HZ1275 and 2025HZ1457), the Youth Fund for Humanities and Social Sciences Research, Ministry of Education of China (grant no. 25YJC630113), the Humanities Project under the General Scientific Research Program of the Shaanxi Provincial Department of Education (grant no. 25JK0162), and the Shaanxi Provincial Research Project on Teacher Education Reform and Teacher Development (grant no. SJS2025ZK074). Author contributions Conceptualization, G.L.D. and Z.J.C.; methodology, G.L.D. and Z.J.C.; investigation, Z.J.C.; data curation, Z.J.C., Y.Y.S. and H.Q.X.; formal analysis, Z.J.C.; visualization, Z.J.C.; writing – original draft, Z.J.C.; writing – review and editing, Z.J.C., Y.Y.S. and H.Q.X.; funding acquisition, G.L.D. and Y.Y.S.; resources, G.L.D. and Z.J.C.; project administration, G.L.D. and Z.J.C.; supervision, G.L.D. All authors have read and approved the final version of the manuscript. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCE SOURCE IDENTIFIER Deposited data Weibo Emotion Correction Dataset This paper GitHub: https://github.com/zj521Q/WeiboEmotionCorrection Source code This paper GitHub: https://github.com/zj521Q/WeiboEmotionCorrection Software and algorithms bert-base-chinese Hugging Face https://huggingface.co/bert-base-chinese text2vec-base-chinese Hugging Face https://huggingface.co/shibing624/text2vec-base-chinese Python 3.8 Python Software Foundation https://www.python.org PyTorch 2.4.1 PyTorch https://pytorch.org Keras 2.10.0 Keras https://keras.io Open in a new tab Experimental model and study participant details This study did not involve the recruitment of human participants, the use of animal models, or the collection of new biological samples. All data used in this study were obtained from publicly available Weibo posts. Method details Overall architecture To effectively achieve precise emotional intervention and positive guidance for Weibo users, this study constructs an integrated emotion regulation framework—namely, a three-stage RCR(Recognition–Classification–Recommendation) architecture. As illustrated in Figure 1 , the proposed emotion correction model consists of three core modules. The leftmost module performs emotion recognition through a hierarchical deep learning model; the middle module focuses on emotion cause classification, built upon a self-constructed emotion cause lexicon and a pre-trained language model; and the rightmost module implements recommendation, which relies on the results of the first two modules. Additionally, the recommendation module reuses parts of the model constructed in the emotion recognition stage for implementation. The modules are tightly connected in both data flow and functional objectives, forming a progressive mechanism that moves from recognition to classification to recommendation. User sentiment analysis model based on Word2Vec and BiLSTM The ultimate objective of this study is to achieve emotion correction through content recommendation. However, such recommendation is not intended for all users but specifically for those exhibiting emotional distress. Therefore, it is necessary to first identify the target users with negative emotions through sentiment analysis. Accordingly, sentiment analysis serves as a critical foundation for the subsequent stages and must demonstrate high reliability. Considering the semantic diversity and contextual dependence of Weibo posts, traditional machine-learning approaches struggle to capture deep semantic relationships. Therefore, we adopt a well-established deep learning approach with strong performance in sentiment analysis to perform polarity classification, constructing a Word2Vec+BiLSTM model. The overall structure of the model is illustrated in Figure 2 . During model training, the embedding layer is initialized with pre-trained Word2Vec embeddings to enhance semantic representation capability. Next, SpatialDropout1D is applied to randomly drop a portion of word embeddings to reduce overfitting. The core BiLSTM component captures bidirectional contextual dependencies and extracts global semantic representations. A single LSTM unit can be expressed as follows: f t = σ ( W f · [ h t − 1 , x t ] + b f ) (Equation 11) i t = σ ( W i · [ h t − 1 , x t ] + b i ) (Equation 12) c ˜ t = tanh ( W c · [ h t − 1 , x t ] + b C ) (Equation 13) c t = f t · c t − 1 + i t · c ˜ t (Equation 14) o t = σ ( W o · [ h t − 1 , x t ] + b o ) (Equation 15) h t = o t · tanh ( c t ) (Equation 16) where f t , i t , o t denote the forget, input, and output gates respectively; c ˜ t is the candidate memory cell; c t is the current cell state; h t is the hidden state; σ represents the sigmoid function; tanh is the hyperbolic tangent function; W f , W i , W C , W o are weight matrices; b f , b i , b C , b o are bias terms. In BiLSTM, the forward LSTM generates, the backward LSTM generates, and the final output is their concatenation: h t = [ h → t ; h ⃖ t ] (Equation 17) Subsequently, GlobalMaxPooling1D is used to extract the most salient emotional features from the BiLSTM outputs. Finally, a dense layer with a sigmoid activation outputs sentiment probabilities. The model uses binary cross-entropy as the loss function and applies L2 regularization to improve generalization and prevent overfitting. User negative emotion-cause classification model based on lexicon-based weak supervision and BERT In the emotion recognition module, the target users of this study have already been identified. However, merely determining that a user’s emotional tendency is negative is not sufficient to support personalized emotion correction. It is also essential to understand why users experience negative emotions, which is the key to realizing individualized emotion correction. Firstly, considering that different groups face diverse emotional problems 41 and that intervention strategies vary across emotional issue types, it is necessary to perform fine-grained classification of negative emotions and categorize users more precisely according to their emotional causes. Secondly, when expressing certain types of distress, users often use specific keywords that not only reflect the thematic content of the text but also reveal the underlying emotional causes. These keywords thus serve as crucial cues for classification. Meanwhile, the classification results should be interpretable and maintainable to ensure transparency and scalability in content recommendation. To simultaneously meet the requirements for fine-grained classification, feature–cause association, and interpretability, this study develops a method combining lexicon-based weak supervision and BERT modeling. Lexicon-based approaches are commonly used in sentiment polarity classification due to their simplicity in design, high computational efficiency, and flexibility in lexicon maintenance and expansion. 42 Building on the fundamental concept of lexicon-based sentiment analysis, this study introduces innovations and extensions in both lexicon construction and application. First, instead of using the traditional “positive–negative sentiment lexicon”, this study constructs a dedicated negative emotion-cause lexicon based on users’emotional sources. Specifically, common causes of negative emotions are categorized into eight major classes, and corresponding cause-specific lexicons are constructed for each category to support the preliminary classification of users’emotional causes. Second, relying solely on lexicon matching has inherent limitations, such as difficulties in handling semantic ambiguity and capturing contextual dependencies. Therefore, this study integrates the lexicon-based method with a pre-trained language model. Specifically, lexicon matching is first used to generate initial pseudo-labels for large-scale data, which are then used as weak supervision signals to train a BERT model, followed by fine-tuning on a small set of manually annotated data. In this framework, the lexicon is not used as the final classifier. Instead, it serves as a weak supervision mechanism to generate large-scale pseudo-labeled data for training the pre-trained language model. The final emotion-cause classification is performed by the BERT model. (1) Lexicon Pseudo-Label Generation ①Construction of a Multi-Dimensional Emotion Cause Lexicon Drawing on a systematic review of cutting-edge research in psychology, this study established the core dimensions of emotional cause classification. Authoritative studies by Valkenburg et al. 43 and Jost et al. 44 have identified key sources of psychological stress among contemporary youth, including social relationships, appearance, romantic experiences, health, and academic performance. However, since Weibo users are not limited to adolescents, focusing solely on these emotional sources would be incomplete. Based on empirical observations of Weibo corpora, this study expanded and refined these dimensions accordingly. Through this integrative process, the causes of negative emotions were ultimately categorized into eight major types: health problems, appearance anxiety, academic problems, family issues, romantic entanglement, social difficulties, financial stress and workplace dilemmas. These categories encompass the most common sources of negative emotions encountered in daily life. The lexicon construction process consists of two main stages: automatic keyword extraction and manual refinement, as illustrated in Figure 3 . (a) Automatic Extraction and Preliminary Filtering TF-IDF Keyword Extraction: A TF-IDF weighting method with smooth IDF and sublinear TF scaling is adopted to measure the importance of words within a document: T F − I D F ( t , d , D ) = ( 1 + log ( t f t , d ) ) × ( log ( ( 1 + N ) ( 1 + d f ( t ) ) ) + 1 ) (Equation 18) where t denotes a term, d represents a single document, D refers to the document collection, N is the total number of documents in the corpus, f t , d indicates the frequency of term t in document d , and df ( t ) denotes the number of documents containing term t . Words with high TF-IDF values are selected as candidate keywords. Part-of-Speech Filtering: Using jieba.posseg, only nouns (n), organization names (nt), other proper nouns (nz, excluding person and place names), verbs (v), and adjectives (a) are retained. Function words and irrelevant tokens are removed. (b) Manual Correction and Expansion Remove verbs with no emotional significance and general words that apply to multiple categories. During automatic screening, the matching results were restricted to words with two or more characters to reduce noise; therefore, it is necessary to manually add single-character keywords and semantically important but low-frequency terms. Retain 40–80 high-quality keywords per category to balance precision and coverage. ②Lexicon-Based Emotion Cause Classification Algorithm After constructing the lexicon, negative Weibo posts are classified by matching segmented tokens with category-specific keyword sets. The algorithm counts keyword hits per category and assigns the category with the highest match as the emotion cause. Since some users may not explicitly express the cause of their negative emotions in their posts, if a post does not match any keywords in all categories, it is classified as “Other” to avoid misclassification caused by forced categorization. Mathematically, let the Weibo i ’s token set be T i = { t i 1 , t i 2 ,…},and the emotion cause categories are C = { c 1 , c 2 , …, c m },where each category c corresponds to a keyword set K c = { κ c 1 , κ c 2 ,…}. The match count is: c o u n t c ( i ) = ∑ κ ∈ K c | { t : t = k , t ∈ T i } | = ∑ j = 1 | K c | c o u n t ( k c j , T i ) (Equation 19) The post is classified into the category with the maximum matching count: L i = { arg max c ∈ C c o u n t c ( i ) , i f max c ∈ C c o u n t c ( i ) > 0 , ” o t h e r w i s e ” , o t h e r w i s e . (Equation 20) (2) BERT Training and Optimization ①BERT Training After weak supervision labeling, Weibo posts receive pseudo-labels that may be noisy. To enhance semantic representation, the BERT pre-trained model is trained on this pseudo-labeled corpus to capture contextual dependencies. BERT is based on the Transformer encoder architecture, employing multi-layer self-attention to achieve bidirectional contextual understanding: A t t e n t i o n ( Q , K , V ) = s o f t max ( Q K T d k ) V (Equation 21) where Q 、 K 、 V represent query, key, and value matrices, and d k is the dimensionality scaling factor. Each input sequence is represented as the sum of token embeddings, segment embeddings, and position embeddings: E i = T i + S i + P i (Equation 22) The study employs the tokenizer of BERT to perform subword-level encoding on Weibo posts, which are then fed into the model. The model’s output layer aggregates the semantic information of the entire sentence using the special [CLS] token, and a Softmax classification layer is added on top to predict eight categories of negative sentiment causes: y = S o f t max ( W h [ C L S ] + b ) (Equation 23) Figure 4 illustrates the structure of the BERT-based negative emotion-cause classification model, which comprises an input embedding layer, the BERT encoder (consisting of multiple Transformer layers), and an output classification layer. The base model used in this study is BERT-base-Chinese, which has a moderate parameter size, comprising 12 Transformer encoder layers, 12 self-attention heads, and 768-dimensional hidden layers. Its pre-training corpus covers a large-scale Chinese text dataset. This model demonstrates stable performance in Chinese syntactic and semantic understanding and can effectively handle the colloquial and non-standard language characteristics commonly found in Weibo posts. The training process at this stage can be regarded as semantic adaptation learning under weak supervision. By training on large-scale pseudo-labeled data, the model learns the semantic distinctions among different emotional cause categories, forming relatively stable feature representations that provide a semantic foundation for the subsequent fine-tuning stage. ②BERT Fine-Tuning After weakly supervised pre-training, the model is fine-tuned on a small manually annotated dataset to refine classification boundaries and correct pseudo-label noise. Only higher-level parameters and the output layer are updated to balance efficiency and performance. This two-stage strategy of lexicon-based weak supervision + fine-tuning supervision enables the model to retain strong semantic understanding while substantially reducing the demand for labeled data. ③User-Level Aggregation After classifying each negative Weibo post, results are aggregated at the user level. For each user, the frequency of each cause category is counted and sorted in descending order, forming a cause list such as [“Social Issues,” “Appearance Anxiety,” “Health Problems,” “Other”]. To ensure the accuracy of subsequent recommendations, users whose emotional causes are exclusively classified as “Other” are filtered out. Since their posts do not explicitly match any predefined negative emotion keywords, it is difficult to determine their specific emotional triggers; recommendations for these users would be based solely on interests, which falls outside the scope of this emotion-corrective recommendation study. Removing such users ensures that the emotional classification and intervention strategies remain targeted and prevents potentially ineffective or inefficient interventions, thereby improving the overall efficiency of the recommendation model. Emotion-corrective recommendation based on weibo Users’Negative emotion causes The goal of this stage is to achieve final emotion correction based on the work of the previous modules. Drawing inspiration from the core concept of cognitive restructuring in cognitive behavioral therapy (CBT), this study designs a lightweight intervention strategy. Cognitive restructuring aims to help individuals identify and modify negative thinking patterns that lead to emotional distress. 18 Accordingly, the core task of this module is to recommend Weibo content that is related to users’negative emotion causes but presented from a more positive perspective, gently guiding users to view their difficulties from new angles, thereby stimulating positive emotions and alleviating negative moods. To achieve this, the recommendations must satisfy both content relevance and positive sentiment. Therefore, this study constructs a recommendation model integrating sentiment filtering and content filtering. In the content filtering stage, a hybrid semantic similarity method is designed, combining TF-IDF feature similarity and Sentence-BERT semantic similarity to balance the matching of emotional causes with user-level semantic similarity. The experimental workflow of this module is shown in Figure 5 . (1) Emotion Filtering The recommended content must not only relate to the user’s emotional cause but also convey positive sentiment to avoid reinforcing negativity. Therefore, a Word2Vec + BiLSTM sentiment classifier (as in Section “User Sentiment Analysis Model Based on Word2Vec and BiLSTM”) is used to predict sentiment polarity. Sentiment filtering is first performed, which ensures the positivity of recommended content while significantly reducing the scale of subsequent similarity computations. (2) Similarity Filtering and Recommendation The similarity-based filtering employed here is essentially a content-based recommendation method. Traditional content-based approaches rely solely on users’historical behaviors to construct preferences, which makes it difficult to capture users’latent emotional needs. This study extends the classical framework. As shown in Figure 6 , the original pathway “user historical behaviors → user profiling → user preference vector” is replaced by two parallel processes: “negative emotion cause keywords → user emotional cause profiling → negative emotion cause vector” and “user Weibo posts → user semantic profiling → Weibo semantic vector”. This modification allows recommendations to be guided by users’emotional causes, while still accounting for individual semantic differences. Candidate Weibo content is matched against these vectors via similarity calculation, and the results are fused to generate recommendations that more effectively achieve emotion correction. ①TF-IDF–Based Cause Relevance Calculation Firstly, TF-IDF is used to vectorize both the keywords of emotion cause lexicon and candidate Weibo posts. Cosine similarity quantifies their relatedness: S i m T F − I D F = cos ( F E C , F i ) = F E C × F i | F E C | 2 | F i | 2 (Equation 24) Where F EC 、 F i denote the emotion cause keyword vector and the candidate posts vector, respectively. ②Sentence-BERT–Based Semantic Personalization To capture deeper semantic relationships, Sentence-BERT (text2vec-base-chinese) encodes posts into dense vectors. For each emotion cause category, the semantic center of users’negative posts is computed as: v u s e r = 1 n ∑ i = 1 n E ( t i ) (Equation 25) Where E ( t i ) represents the sentence embedding of the Weibo post t i . Then, cosine similarity is calculated between the candidate post F i and this semantic center: S i m S e n t e n c e − B E R T = cos ( v u s e r , F i ) (Equation 26) ③Hybrid Similarity Fusion and Recommendation Generation The final similarity score is a weighted sum: S i m f i n a l = α · S i m T F − I D F + β · S i m S e n t e n c e − B E R T (Equation 27) Because our objective is emotion-cause correction rather than matching user interests, the weighting emphasizes cause matching. Therefore, the weight coefficients are set as α = 0.7 and β = 0.3, ensuring that the recommended results are highly aligned with the user’s negative emotional causes while moderately integrating semantic-level personalization factors to prevent the recommendations from becoming overly uniform or mechanical. After threshold filtering, the top-N posts with the highest similarity are selected as candidates for recommendation. Personalized recommendation results are subsequently generated for each user. Let the set of emotional causes for user u be R = { r 1 , r 2 , …, r m }, with corresponding occurrence counts c i , the associated weights are then defined as: w i = c i ∑ j = 1 m c j (Equation 28) Recommendation quotas for each emotional cause are assigned based on their weights, forming the user’s final recommendation set: R e c ( u ) = ∪ i = 1 m w i · T o p N ( R i ) (Equation 29) This method achieves a dual objective in the recommendation process: “emotional cause matching + personalized corrective guidance”. It not only ensures that the recommended content aligns with users’negative emotion causes but also generates differentiated, positive recommendations oriented toward emotion correction by modeling individual differences at the semantic level. Quantification and statistical analysis All computations were performed using the Python programming language, and all figures reporting model performance were generated using standard Python libraries. No statistical hypothesis tests (e.g., t-tests or ANOVA) were conducted, as the comparisons between models are based on deterministic evaluation metrics computed on fixed test sets. Model performance is evaluated using standard quantitative metrics, including accuracy, precision, recall, and F1-score for classification tasks, as well as Precision@k, NDCG@k, and MAP@k for the recommendation task, as defined in Section “Evaluation Metrics”. For the sentiment analysis module, evaluation is conducted on a held-out test set comprising 15% of the full dataset. For the emotion-cause classification module, the test set consists of n = 520 manually annotated Weibo posts, with 65 samples in each of the eight emotion-cause categories. For the recommendation module, evaluation is performed on n = 50 users, where each user is associated with a recommendation list of length 10 and corresponding relevance judgments. References 1. Olawade D.B., Wada O.Z., Odetayo A., David-Olawade A.C., Asaolu F., Eberhardt J. Enhancing mental health with Artificial Intelligence: Current trends and future prospects. J. Med. Surg. Public Health. 2024;3 doi: 10.1016/j.glmedi.2024.100099. [ DOI ] [ Google Scholar ] 2. Sun M., Gao X., Jing X., Wang W., Wang X. Public attention and sentiment towards clean heating in China: A text mining analysis based on Sina Weibo. Energy. 2025;331 doi: 10.1016/j.energy.2025.136975. [ DOI ] [ Google Scholar ] 3. Ahmad W., Khan H.U., Alarfaj F.K., Alreshoodi M. Aspect-base sentiment analysis: a comprehensive review and open research challenges. IEEE Access. 2025;13:65138–65182. doi: 10.1109/ACCESS.2025.3555744. [ DOI ] [ Google Scholar ] 4. Naz A., Khan H.U., Bukhari A., Alshemaimri B., Daud A., Ramzan M. Machine and deep learning for personality traits detection: a comprehensive survey and open research challenges. Artif. Intell. Rev. 2025;58:239. doi: 10.1007/s10462-025-11245-3. [ DOI ] [ Google Scholar ] 5. Khan H.U., Naz A., Alarfaj F.K., Almusallam N. Identifying artificial intelligence-generated content using the DistilBERT transformer and NLP techniques. Sci. Rep. 2025;15 doi: 10.1038/s41598-025-08208-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Brown D.J., Scott R., Ireland R., Harness J., Phipps D.J., Keech J.J. Rethinking social media and mental health: The role of emotion regulation difficulties. Comput. Hum. Behav. 2026;174 doi: 10.1016/j.chb.2025.108825. [ DOI ] [ Google Scholar ] 7. Lin Z., Wang Y., Zhou Y., Du F., Yang Y. MLM-EOE: Automatic depression detection via sentimental annotation and multi-expert ensemble. IEEE Trans. Affect. Comput. 2025;16:2842–2858. doi: 10.1109/TAFFC.2025.3585599. [ DOI ] [ Google Scholar ] 8. Yin F., Xia X., Pan Y., She Y., Feng X., Wu J. Sentiment mutation and negative emotion contagion dynamics in social media: A case study on the Chinese Sina Microblog. Inf. Sci. 2022;594:118–135. doi: 10.1016/j.ins.2022.02.029. [ DOI ] [ Google Scholar ] 9. Zhu P., Wang B., Tang K., Zhang H., Cui X., Wang Z. A knowledge-guided graph attention network for emotion-cause pair extraction. Knowl. Base Syst. 2024;286 doi: 10.1016/j.knosys.2023.111342. [ DOI ] [ Google Scholar ] 10. Wang Y., Li Y., Yu K., Yang J. A semantic structure-based emotion-guided model for emotion-cause pair extraction. Pattern Recogn. 2025;161 doi: 10.1016/j.patcog.2024.111296. [ DOI ] [ Google Scholar ] 11. Su X., Huang Z., Su Y., Trisedya B.D., Dou Y., Zhao Y. Hierarchical shared encoder with task-specific transformer layer selection for emotion-cause pair extraction. IEEE Trans. Affect. Comput. 2024;15:1934–1948. doi: 10.1109/TAFFC.2024.3390223. [ DOI ] [ Google Scholar ] 12. Miu A.C., Szentágotai-Tătar A., Balázsi R., Nechita D., Bunea I., Pollak S.D. Emotion regulation as mediator between childhood adversity and psychopathology: A meta-analysis. Clin. Psychol. Rev. 2022;93 doi: 10.1016/j.cpr.2022.102141. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Troy A.S., Willroth E.C., Shallcross A.J., Giuliani N.R., Gross J.J., Mauss I.B. Psychological resilience: an affect-regulation framework. Annu. Rev. Psychol. 2023;74:547–576. doi: 10.1146/annurev-psych-020122-041854. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Stellern J., Xiao K.B., Grennell E., Sanches M., Gowin J.L., Sloan M.E. Emotion regulation in substance use disorders: A systematic review and meta-analysis. Addiction. 2023;118:30–47. doi: 10.1111/add.16001. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Xu X., Wang Y., Wei X., Wang F., Zhang X. Attention-based acoustic feature fusion network for depression detection. Neurocomputing. 2024;601 doi: 10.1016/j.neucom.2024.128209. [ DOI ] [ Google Scholar ] 16. Zhu J., Zhang Z., Li Z., Hu B. Anxiety recognition based on multimodal social media data and cross-attention mechanism. Neurocomputing. 2025;646 doi: 10.1016/j.neucom.2025.130473. [ DOI ] [ Google Scholar ] 17. Beck A.T. Thinking and depression: I. Idiosyncratic content and cognitive distortions. Arch. Gen. Psychiatry. 1963;9:324–333. doi: 10.1001/archpsyc.1963.01720160014002. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Beck A.T., Rush A.J., Shaw B.F., Emery G., DeRubeis R.J., Hollon S.D. Guilford Publications; 2024. Cognitive Therapy of Depression. https://books.google.com/books?isbn=1572305827 [ Google Scholar ] 19. Onan A. Bidirectional convolutional recurrent neural network architecture with group-wise enhancement mechanism for text sentiment classification. J. King Saud Univ., Comput. Inf. Sci. 2022;34:2098–2117. doi: 10.1016/j.jksuci.2022.02.025. [ DOI ] [ Google Scholar ] 20. Talaat A.S. Sentiment analysis classification system using hybrid BERT models. J. Big Data. 2023;10:110. doi: 10.1186/s40537-023-00781-w. [ DOI ] [ Google Scholar ] 21. Ouyang J., Xuan C., Wang B., Yang Z. Aspect-based sentiment classification with aspect-specific hypergraph attention networks. Expert Syst. Appl. 2024;248 doi: 10.1016/j.eswa.2024.123412. [ DOI ] [ Google Scholar ] 22. Meng T., Shou Y., Ai W., Du J., Liu H., Li K. A multi-message passing framework based on heterogeneous graphs in conversational emotion recognition. Neurocomputing. 2024;569 doi: 10.1016/j.neucom.2023.127109. [ DOI ] [ Google Scholar ] 23. Su X., Huang Z., Zhao Y., Chen Y., Dou Y., Pan H. Recent trends in deep learning based textual emotion cause extraction. IEEE/ACM Trans. Audio Speech Lang. Process. 2023;31:2765–2786. doi: 10.1109/TASLP.2023.3254166. [ DOI ] [ Google Scholar ] 24. Lee S.Y.M., Chen Y., Li S., Huang C.-R. Emotion cause events: Corpus construction and analysis. 2010. http://www.lrec-conf.org/proceedings/lrec2010/pdf/322_Paper.pdf 25. Xia R., Ding Z. Emotion-cause pair extraction: A new task to emotion analysis in texts. 2019. 1003–1012. [ DOI ] 26. Devlin J., Chang M.-W., Lee K., Toutanova K. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 4171–4186. [ DOI ] 27. Wan B., Wu P., Yeo C.K., Li G. Emotion-cognitive reasoning integrated BERT for sentiment analysis of online public opinions on emergencies. Inf. Process. Manag. 2024;61 doi: 10.1016/j.ipm.2023.103609. [ DOI ] [ Google Scholar ] 28. Li Z., Cao S., Zhai M., Ding N., Zhang Z., Hu B. Multi-level semantic enhancement based on self-distillation BERT for Chinese named entity recognition. Neurocomputing. 2024;586 doi: 10.1016/j.neucom.2024.127637. [ DOI ] [ Google Scholar ] 29. Lv S., Lu S., Wang R., Yin L., Yin Z., A. AlQahtani S., Tian J., Zheng W. Enhancing chinese dialogue generation with word–phrase fusion embedding and sparse softmax optimization. Systems. 2024;12:516. doi: 10.3390/systems12120516. [ DOI ] [ Google Scholar ] 30. Zhou Z.-H. A brief introduction to weakly supervised learning. Natl. Sci. Rev. 2018;5:44–53. doi: 10.1093/nsr/nwx106. [ DOI ] [ Google Scholar ] 31. Lien Y.-C., Zamani H., Croft B. Generalized Weak Supervision for Neural Information Retrieval. ACM Trans. Inf. Syst. 2024;42:1–26. doi: 10.1145/3647639. [ DOI ] [ Google Scholar ] 32. Huang S.J., Liu H.Z., Niu G., Sugiyama M., Xiao J., Xie M.K. Class-distribution-aware pseudo-labeling for semi-supervised multi-label learning. Adv. Neural Inf. Process. Syst. 2023;36:25731–25747. doi: 10.52202/075280-1119. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Clifton D., Dai W., Duncan J., Liu F., Min Y., Staib L., You C., Zhou S.K. Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. Adv. Neural Inf. Process. Syst. 2023;36:9984–10021. doi: 10.52202/075280-0437. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Jing X., Li Z., Wu Z., Mu D. Enhancing Text Classification with Neural Label Embedding and Weakly-Supervised Learning. Expert Syst. Appl. 2026;298 doi: 10.1016/j.eswa.2025.129569. [ DOI ] [ Google Scholar ] 35. Wallaart O., Frasincar F., van der Knaap F. Weakly-supervised sentence-based aspect category and sentiment classification. Knowl. Base Syst. 2025;317 doi: 10.1016/j.knosys.2025.113403. [ DOI ] [ Google Scholar ] 36. Abbasi-Moud Z., Vahdat-Nejad H., Sadri J. Tourism recommendation system based on semantic clustering and sentiment analysis. Expert Syst. Appl. 2021;167 doi: 10.1016/j.eswa.2020.114324. [ DOI ] [ Google Scholar ] 37. Choudhary C., Singh I., Kumar M. SARWAS: Deep ensemble learning techniques for sentiment based recommendation system. Expert Syst. Appl. 2023;216 doi: 10.1016/j.eswa.2022.119420. [ DOI ] [ Google Scholar ] 38. Darraz N., Karabila I., El-Ansari A., Alami N., Mallahi M.E. Advancing recommendation systems with DeepMF and hybrid sentiment analysis: Deep learning and Lexicon-based integration. Expert Syst. Appl. 2025;279 doi: 10.1016/j.eswa.2025.127432. [ DOI ] [ Google Scholar ] 39. Kalimeris D., Bhagat S., Kalyanaraman S., Weinsberg U. Proc. 27th ACM SIGKDD Conf. Knowl. Discov. Data Min. Association for Computing Machinery (ACM); 2021. Preference Amplification in Recommender Systems; pp. 805–815. [ DOI ] [ Google Scholar ] 40. Liu Z., Xu W., Zhang W., Jiang Q. An emotion-based personalized music recommendation framework for emotion improvement. Inf. Process. Manag. 2023;60 doi: 10.1016/j.ipm.2022.103256. [ DOI ] [ Google Scholar ] 41. Liu X., Wang J., Xiong X., Sun H. Federated learning data protection scheme based on personalized differential privacy in psychological evaluation. Neurocomputing. 2025;611 doi: 10.1016/j.neucom.2024.128653. [ DOI ] [ Google Scholar ] 42. Rutkowska A., Szyszko M. Dictionary-based sentiment analysis of monetary policy communication: on the applicability of lexicons: A. Rutkowska. Qual. Quantity. 2024;58:5421–5444. doi: 10.1007/s11135-024-01896-9. [ DOI ] [ Google Scholar ] 43. Valkenburg P.M., Meier A., Beyens I. Social media use and its impact on adolescent mental health: An umbrella review of the evidence. Curr. Opin. Psychol. 2022;44:58–68. doi: 10.1016/j.copsyc.2021.08.017. [ DOI ] [ PubMed ] [ Google Scholar ] 44. Jost G.M., Hang S., Shaikh U., Hostinar C.E. Understanding adolescent stress during the COVID-19 pandemic. Curr. Opin. Psychol. 2023;52 doi: 10.1016/j.copsyc.2023.101646. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement • The datasets generated and analyzed during this study have been deposited at GitHub and are publicly available as of the date of publication at: GitHub: https://github.com/zj521Q/WeiboEmotionCorrection . • All original code has been deposited at GitHub and is publicly available as of the date of publication at: GitHub: https://github.com/zj521Q/WeiboEmotionCorrection . • Any additional information required to reanalyze the data reported in this paper is available from the Lead contact upon request. Articles from iScience are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (5.7 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Inherent emotional quality of human speech sounds#755799
• Neo-Classic: A Benchmark for Evaluating Linguistic-Aesthetic Reasoning in Classical Chinese Poetry#1013755
• Dehumanizing Speech#725056
• The Effect of Variability in Appearance and Context on Face Learning#286148
• An overview of computer-based natural language processing#470253