A unified transformer-BiLSTM and graph attention network framework for explainable multilingual opinion mining and relationship inference in complex social-citation networks

Transformer-Based Topic Modeling and Sentiment Analysis for Multilingual Text Data

Several new trends related to the rising volume of multilingual digital content have spurred a greater need for Natural Language Processing (NLP) systems capable of capturing the latent themes and understanding sentiment without stripping out the structure within a specific language. In this study, we introduce an integrated framework based on transformer architecture for topic modeling and sentiment analysis for Hindi and Italian social-media discourse, customer reviews and news corpus. The framework integrates language-aware representations in context, the sentiment classification system of supervised learning and the BERTopic-based topic discovery system. Hindi Processing focuses on Devanagari normalization, transliteration, mixed code and Hindi-English while Italian Processing focuses on inflectional, clitic morphology, colloquial and regionalization variation. In addition to classical baselines (Support Vector Machines (SVM), Long Short-Term Memory (LSTM) networks, and Latent Dirichlet Allocation (LDA)), MuRIL, UmBERTo, AlBERTo and LLaMAntino are studied. The performance is assessed descriptively in terms of accuracy, F1 Score, topic coherence (Cᵥ), and training behavior and out of distribution error analysis. On sentiment classification using Hindi, MuRIL achieved an accuracy of 0.89 and an F1-score of 0.88 which were better than the best reported sentiment classification in classical Hindi by 0.08 on both measures. UmBERTo's results for Italian were 0.90 accuracy and 0.89 F1-score, while LLaMAntino obtained the best reported sentiment scores of 0.91 and 0.90. For Hindi, the relative improvement over LDA was 53.7% and for Italian, it was 50.0% with respect to the topic coherence score measured by BERTopic. The findings suggest that there is clear benefit for morphologically complex text and mixed script text for using contextual embeddings and language-specific pretraining. However, there are still some limitations such as dialect-sensitive mistakes, spurious correlations, computational cost and low interpretability. Finally, given that error and fairness audits are not widely available, the study offers a more comprehensive grounding in error and fairness for the responsible deployment of multilingual topic and sentiment systems.

Feature-level enhanced syntactic-semantic graph networks via optimal transport for aspect-based sentiment analysis

Joint multi-task learning with knowledge graph embedding and semantic-enhanced encoding for teaching evaluation text analysis

To address the difficulty of accurately analyzing large-scale unstructured teaching evaluation texts, this study proposes a teaching evaluation text analysis framework integrating knowledge graph embedding, semantic-enhanced encoding, and joint multi-task learning. Based on approximately 30,000 course evaluation records, a teaching evaluation knowledge graph containing about 8,000 entity nodes and 35,000 relation edges is constructed to model heterogeneous associations among courses, teachers, evaluation aspects, and labels. A semantic-enhanced multi-channel encoding mechanism is then designed to fuse contextual text representations, graph embeddings, and structural features into a shared representation space. On this basis, a joint multi-task learning framework is developed to simultaneously perform sentiment polarity classification, evaluation aspect classification, and satisfaction regression. Experimental results on real university data show that the proposed model achieves an Accuracy of 0.915 in sentiment polarity classification and 0.886 in evaluation aspect classification, with Macro F1 improving by approximately 1.4 and 2.1 percentage points, respectively, compared with TF-IDF + SVM and text-only deep models. In addition, the deployed system maintains an average response time of about 295 ms and a QPS of approximately 200 under 200 concurrent requests, demonstrating good engineering efficiency and robustness.

From Graphs to Hypergraphs: Enhancing Aspect-Term Sentiment Analysis via Multi-Level Relational Modeling

Aspect-Term Sentiment Analysis (ATSA) aims to predict sentiment polarity for specific aspect terms, a task complicated by conflicting sentiments and limited context in short texts. Existing graph-based approaches rely on pre-defined pairwise structures to capture different linguistic views. However, this leads to two key limitations: (1) their pairwise formulation often requires multiple graphs to improve expressive capacity, and (2) their reliance on predefined parsers or heuristic graph construction limits adaptability to sentence-specific sentiment composition. We propose HyperATSA, a dynamic hypergraph framework that overcomes these limitations through a single instance-specific hypergraph constructed directly from contextual token representations. Hyperedges are dynamically induced via Hierarchical Agglomerative Clustering (HAC) over token embeddings, where an acceleration-based cutoff identifies sentence-specific semantic groupings and enables adaptive hypergraph construction. Experiments on Lap14, Rest14, and MAMS demonstrate consistent improvements over strong graph-based baselines, suggesting that hypergraph-based relational modeling generalizes effectively to short-text sentiment composition.

Graph Attention-Based Engineering Framework for Aspect-Level Sentiment-Driven Recommendation Using BiLSTM–CRF Hybrid Model

Abstract The rapid expansion of user-generated content in the form of reviews and ratings on e-commerce sites calls for sophisticated recommendation algorithms that can comprehend both the sentiments expressed in text and numeric evaluations. Existing recommendation strategies based on collaborative filtering and content-based filtering have been shown to neglect the meaning conveyed in user reviews, resulting in poor performance and interpretability. In this study, we introduce a novel recommendation algorithm that combines Aspect-Based Sentiment Analysis (ABSA) and a Graph Attention Network (GAT) to improve the recommendation quality. Initially, user reviews are processed by a Bidirectional Long Short-Term Memory (BiLSTM) model supplemented with an attention mechanism to obtain detailed sentiments toward aspects and generate an implied rating score. The difference between the implicit and explicit rating scores is then used to estimate sentiment-ratings consistency. Finally, a GAT is introduced to explore intricate relationships among users, items, and sentiment dimensions by weighting useful relations and ignoring noises through attention coefficients. The proposed approach is based on the aggregation of sentiments, scores, and embeddings. Experimental results on multiple benchmark data sets show that the proposed method for hybrid recommendations outperforms other approaches in terms of precision, and recall. Moreover, the proposed framework also allows for improved interpretability because of the aspect-level sentiment representation.

Dual-Graph Convolutional Network with Context Fusion for Chinese Sentiment Classification

Existing research on sentiment classification primarily focuses on textual contextual semantics while neglecting word co-occurrence and syntactic structure information. It also emphasizes semantic analysis of sequential structures but overlooks non-linear structural semantics. To address these limitations, we propose the DGCNCF (Dual-Graph Convolutional Network with Context Fusion) model for sentiment classification. The model employs a BiLSTM with an attention mechanism to extract textual contextual information. Subsequently, a text co-occurrence graph and a syntactic dependency graph are constructed separately. Meanwhile, contextual information is fused with node information within these graphs to obtain graph embedding representations of the text. Then, the model performs representation learning on the two graphs using GCN, thereby capturing the global word co-occurrence features and global syntactic dependency structure characteristics of the text. By integrating textual contextual information, word co-occurrence information, and syntactic dependency structure information, the model captures text features from semantic, lexical, and syntactic perspectives, thereby mitigating the limitations of relying on single-dimensional features. The validity of the model is verified on two public datasets, and the experimental results demonstrate that the model achieves effective sentiment classification performance.