Dynamic Span Interaction and Graph-Aware Memory for Entity-Level Sentiment Classification

TL;DR

SpanEIT tackles entity-level sentiment classification by dynamically aligning entity spans with sentiment spans through bidirectional attention while leveraging a span-level graph that merges syntactic and semantic relations. A memory module enforces cross-document consistency across coreferent mentions, and auxiliary supervision guides precise span and relation reasoning. Empirical results on BARU, FSAD, and IMDB show SpanEIT achieving state-of-the-art accuracy, micro-F1, and macro-F1, with robust ablations highlighting the contribution of each component. The work demonstrates strong performance across domains, offers interpretability via attention and memory visualizations, and points to practical applications in social media monitoring and customer feedback analysis.

Abstract

Entity-level sentiment classification involves identifying the sentiment polarity linked to specific entities within text. This task poses several challenges: effectively modeling the subtle and complex interactions between entities and their surrounding sentiment expressions; capturing dependencies that may span across sentences; and ensuring consistent sentiment predictions for multiple mentions of the same entity through coreference resolution. Additionally, linguistic phenomena such as negation, ambiguity, and overlapping opinions further complicate the analysis. These complexities make entity-level sentiment classification a difficult problem, especially in real-world, noisy textual data. To address these issues, we propose SpanEIT, a novel framework integrating dynamic span interaction and graph-aware memory mechanisms for enhanced entity-sentiment relational modeling. SpanEIT builds span-based representations for entities and candidate sentiment phrases, employs bidirectional attention for fine-grained interactions, and uses a graph attention network to capture syntactic and co-occurrence relations. A coreference-aware memory module ensures entity-level consistency across documents. Experiments on FSAD, BARU, and IMDB datasets show SpanEIT outperforms state-of-the-art transformer and hybrid baselines in accuracy and F1 scores. Ablation and interpretability analyses validate the effectiveness of our approach, underscoring its potential for fine-grained sentiment analysis in applications like social media monitoring and customer feedback analysis.

Figures (3)

• Figure 1: SpanEIT architecture for entity-level sentiment analysis, extracting entity and sentiment spans from text and building a combined syntactic-semantic graph processed by Graph Attention Networks to generate context-aware embeddings. Multi-head cross-attention captures interactions between entities and sentiments, while a memory bank maintains coreference-aware representations. The fused features are classified via a Multi-Layer Perceptron for fine-grained sentiment prediction.
• Figure 2: Visual analysis of the attention mechanisms in SpanEIT. (a) Entity-to-token attention demonstrates the alignment of entity representations with input tokens, while (b) entity-to-sentiment-span attention highlights the model’s focus on contextual sentiment information at the span level. These visualizations provide insights into how the model grounds entity-level sentiment predictions in both lexical and contextual features.
• Figure 3: Memory vector heatmap visualization for a sample input: "[CLS] spy index shows stable performance this [SEP] spy [SEP]". Each column represents the value of one dimension in the memory vector corresponding to the entity "spy". Warmer (red) and cooler (blue) colors indicate higher and lower activations, respectively, across the memory bank. This visualization demonstrates how SpanEIT maintains and updates entity-specific memory representations, enabling the model to retain salient contextual information relevant for entity-level sentiment inference.