Abstract
Cross-domain sentiment analysis remains a challenge due to "sentiment transfer errors," where the polarity of a word changes depending on the context or the specific aspect it describes. We present AOAKM, a novel model that leverages self-supervised graph learning to align aspect-opinion pairs. By utilizing document-level consistency and relational graph convolutional networks (R-GCN), AOAKM effectively captures domain-specific sentiment nuances. Evaluated on the Amazon-reviews benchmark, AOAKM achieves a state-of-the-art average accuracy of 93.7%, outperforming existing BERT-based and knowledge-enhanced models.
Introduction
Sentiment Analysis (SA) models often suffer from performance degradation when applied to a target domain that differs from the source training data. A primary driver of this issue is the sentiment transfer error. Many words are "sentiment-variant"; for instance, the word "fast" is positive when describing a kitchen appliance's efficiency but negative when describing how quickly batteries drain in an electronic device.
Standard models frequently overlook the relationship between an opinion word and the specific aspect it modifies. To address this, we propose the Aspect-Opinion Sentiment Alignment (AOAKM) model, which focuses on the relational nature of sentiment.
Methodology
The AOAKM framework integrates structural knowledge with contextual embeddings through three key stages:
1. Knowledge Graph Construction
Using ConceptNet and dependency parsing, the model extracts nouns (aspects), adjectives, and adverbs (opinions) from unlabeled reviews. It constructs a graph where nodes are connected by "described" relations (e.g., battery $\leftrightarrow$ fast).
2. Self-Supervised Graph Pre-training
A Relational Graph Convolutional Network (R-GCN) autoencoder is trained on two tasks:
- Relation Classification: Predicting syntactic connections between words.
- Sentiment Alignment Classification: This task assumes that aspect-opinion pairs within the same document share the same sentiment polarity. The model learns to align these pairs, allowing it to infer the sentiment of a variant word (like "fast") based on other invariant words (like "terrible" or "great") present in the same context.
3. Sentence Encoding and Classification
The structural features from the R-GCN are fused with contextual embeddings from a pre-trained BERT model. This hybrid representation is then passed through a softmax layer for final sentiment prediction.
Experimental Results
The model was evaluated using the Amazon-reviews benchmark across four domains: Books (B), DVDs (D), Electronics (E), and Kitchen appliances (K). The evaluation involved 12 cross-domain tasks.
As shown in the results, AOAKM achieved an average accuracy of 93.7%, surpassing the previous state-of-the-art model, SENTIX (92.7%). Notably, it significantly outperformed KinGDOM (84.0%), demonstrating that simply incorporating external knowledge is insufficient without proper sentiment alignment between aspect-opinion pairs.
Discussion and Conclusion
The success of AOAKM highlights that sentiment is inherently relational. By leveraging document-level consistency, the model can correct transfer errors that purely contextual models like BERT might miss. While the model currently relies on external knowledge bases like ConceptNet and assumes document-level sentiment homogeneity, it sets a new benchmark for cross-domain sentiment analysis. Future work may explore handling mixed-sentiment reviews and reducing the computational overhead of the R-GCN and BERT fusion.
Summary updated