[Paper Review] Knowledge Graph Enhanced Aspect-Level Sentiment Analysis

Kavita Sharma, Ritu Patel+1 more|arXiv (Cornell University)|Dec 2, 2023

Sentiment Analysis and Opinion Mining4 citations

TL;DR

This paper proposes a knowledge graph-enhanced BERT model that integrates dynamic attention and a memory bank with positional encoding to improve aspect-level sentiment analysis. By leveraging external synonym knowledge and a DCGRU for sentiment characterization, the method achieves state-of-the-art performance on three benchmark datasets, demonstrating superior accuracy in detecting aspect-specific sentiment.

ABSTRACT

In this paper, we propose a novel method to enhance sentiment analysis by addressing the challenge of context-specific word meanings. It combines the advantages of a BERT model with a knowledge graph based synonym data. This synergy leverages a dynamic attention mechanism to develop a knowledge-driven state vector. For classifying sentiments linked to specific aspects, the approach constructs a memory bank integrating positional data. The data are then analyzed using a DCGRU to pinpoint sentiment characteristics related to specific aspect terms. Experiments on three widely used datasets demonstrate the superior performance of our method in sentiment classification.

Motivation & Objective

• To address the challenge of context-specific word meanings in aspect-level sentiment analysis.
• To enhance BERT-based models with external knowledge from a knowledge graph to improve sentiment classification accuracy.
• To develop a dynamic attention mechanism that generates knowledge-driven state vectors for better context representation.
• To construct a memory bank with positional encoding to preserve aspect-term location and improve sentiment characterization.
• To evaluate the proposed method on standard benchmark datasets and demonstrate its superiority over existing approaches.

Proposed method

• The method integrates a BERT model with a knowledge graph-based synonym dataset to enrich contextual embeddings.
• A dynamic attention mechanism computes attention weights using knowledge-enhanced representations to generate a knowledge-driven state vector.
• A memory bank is constructed to store aspect-specific representations, incorporating positional encoding to preserve term location in the sequence.
• A DCGRU (Deep Convolutional Gated Recurrent Unit) is applied to the memory bank to extract and refine sentiment characteristics for each aspect term.
• The final sentiment classification is performed using the refined representations from the DCGRU, enabling precise aspect-level sentiment prediction.
• The model is trained end-to-end on aspect-level sentiment datasets with standard loss functions for sequence classification.

Experimental results

Research questions

• RQ1 Can knowledge graph-enhanced embeddings improve the representation of context-specific word meanings in aspect-level sentiment analysis?
• RQ2 How does integrating dynamic attention with knowledge-driven state vectors affect sentiment classification performance?
• RQ3 To what extent does a memory bank with positional encoding improve the identification of aspect-specific sentiment?
• RQ4 Does the use of a DCGRU on the memory bank lead to better extraction of sentiment patterns compared to standard RNNs or transformers?
• RQ5 How does the proposed method compare to state-of-the-art models on standard aspect-level sentiment analysis benchmarks?

Key findings

• The proposed method achieves state-of-the-art performance on three widely used aspect-level sentiment analysis datasets.
• The integration of knowledge graph-based synonyms significantly improves the model's ability to handle polysemous words and context-dependent sentiment.
• The dynamic attention mechanism effectively focuses on relevant knowledge-enhanced features, enhancing representation quality.
• The memory bank with positional encoding improves the model's sensitivity to aspect term location, leading to more accurate sentiment prediction.
• The DCGRU component outperforms standard RNN and transformer-based modules in capturing sequential sentiment patterns for aspect terms.
• The model demonstrates consistent improvements across all three benchmark datasets, confirming its robustness and generalization capability.