Introduction
Aspect-Based Sentiment Analysis (ABSA) is a critical task in natural language processing that aims to identify the sentiment polarity (positive, negative, or neutral) associated with specific aspect terms within a sentence. While Graph Convolutional Networks (GCNs) have become a standard for modeling syntactic dependencies in ABSA, they are often hindered by "noise"—errors in syntactic parsing that lead the model to associate sentiment words with the wrong targets. Furthermore, standard training paradigms often fail to explicitly leverage label information during the representation learning phase.
Methodology: The C³LPGCN Framework
To address these challenges, we introduce C³LPGCN, a framework that integrates Contrastive learning, Cooperative learning, and Convolutional networks with Prompts. The architecture is designed to filter syntactic noise and align sentiment representations more closely with their ground-truth labels.
Core Components
- Mask-aware Aspect Information Filter (MAF): Rather than relying solely on raw syntactic trees, the MAF utilizes prompt tuning (e.g., "The sentiment of [aspect] is [MASK]"). By analyzing the attention patterns of a pre-trained language model (BERT) toward the [MASK] token, the MAF identifies sentiment-relevant nodes in the GCN, effectively filtering out syntactic noise.
- Prompt-based Contrastive Learning (PCL): Unlike traditional contrastive learning that compares arbitrary sentence pairs, PCL constructs explicit "label samples." It generates positive samples using the true sentiment label and negative samples using incorrect labels, forcing the model to learn discriminative features tied directly to sentiment classes.
- Cooperative Learning: This module ensures consistency between the original input and its "true label" counterpart. By minimizing the KL divergence between their internal representations and output distributions, the model learns a more robust mapping from text to sentiment.
Experimental Results
We evaluated C³LPGCN on three benchmark datasets: Laptop, Restaurant, and Twitter. The model consistently outperformed state-of-the-art (SOTA) syntactic and contrastive baselines.
Key Findings
- Superior Performance: On the Laptop dataset, C³LPGCN achieved an accuracy of 82.75%, surpassing previous SOTA models like APARN-BERT and DualGCN-BERT. Similar gains were observed in the Restaurant and Twitter datasets.
- Ablation Insights: Removing the MAF, PCL, or Cooperative Learning modules resulted in a measurable performance drop. Specifically, removing the MAF led to the most significant decline in the Restaurant dataset, highlighting the importance of filtering syntactic noise.
- Prompts as Semantic Anchors: The use of prompts provides a semantic "spotlight" that guides the GCN. This prevents the model from being misled by the informal or ungrammatical structures often found in social media and review data.
Conclusion and Future Work
C³LPGCN demonstrates that integrating prompt-based learning with syntactic GCNs significantly improves ABSA performance by reducing noise and explicitly incorporating label signals. While the current model relies on manually constructed prompt templates and incurs higher computational costs due to sample generation, future research will focus on automating template discovery through continuous prompts and optimizing the training process to reduce GPU memory overhead.
Summary updated