Extensible Multi-Granularity Fusion Network for Aspect-based Sentiment Analysis

2024-02-12

arXiv.org

AI Summary

Existing Aspect-Based Sentiment Analysis (ABSA) methods rely on external knowledge or Graph Neural Networks (GNNs) to enhance semantic/syntactic modeling, yet suffer from high feature heterogeneity and lack of unified, extensible frameworks. Method: We propose a multi-granularity fusion framework that jointly models dependency syntax, constituent syntax, self-attention semantics, and knowledge graph features within a single architecture. We introduce novel multi-anchor triplet learning and orthogonal projection to enable cross-granularity feature co-optimization at zero computational overhead, while supporting flexible integration of new linguistic structures. The model integrates GNNs, knowledge graph embedding, and orthogonal constraint optimization. Contribution/Results: Our approach achieves significant improvements over state-of-the-art methods on SemEval-2014 Task 4 and the Twitter dataset, demonstrating both the effectiveness and generalizability of fusing heterogeneous structured information for ABSA.

Technology Category

• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Machine Learning: Multimodal Learning
• Data Mining & Knowledge Management: Linked Open Data, Knowledge Graphs & KB Completion

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding

Abstract

Aspect-based Sentiment Analysis (ABSA) evaluates sentiment expressions within a text to comprehend sentiment information. Previous studies integrated external knowledge, such as knowledge graphs, to enhance the semantic features in ABSA models. Recent research has examined the use of Graph Neural Networks (GNNs) on dependency and constituent trees for syntactic analysis. With the ongoing development of ABSA, more innovative linguistic and structural features are being incorporated (e.g. latent graph), but this also introduces complexity and confusion. As of now, a scalable framework for integrating diverse linguistic and structural features into ABSA does not exist. This paper presents the Extensible Multi-Granularity Fusion (EMGF) network, which integrates information from dependency and constituent syntactic, attention semantic , and external knowledge graphs. EMGF, equipped with multi-anchor triplet learning and orthogonal projection, efficiently harnesses the combined potential of each granularity feature and their synergistic interactions, resulting in a cumulative effect without additional computational expenses. Experimental findings on SemEval 2014 and Twitter datasets confirm EMGF's superiority over existing ABSA methods.

Problem

• Integrates multi-granularity linguistic features for aspect-based sentiment analysis.
• Reduces model complexity with a unified, extensible fusion framework.
• Enhances generalization via curriculum learning from easy to hard examples.

Innovation

• Multi-granularity fusion network integrates syntax and semantics
• Curriculum learning framework trains models from easy to hard
• Multi-anchor triplet learning fuses features without extra overhead