Learn from Syntax: Improving Pair-wise Aspect and Opinion Terms Extractionwith Rich Syntactic Knowledge

Shengqiong Wu, Hao Fei, Yafeng Ren, ji Donghong

TL;DR

This paper introduces a novel framework for jointly extracting aspect and opinion terms (PAOTE) by deeply integrating rich syntactic information. Their approach uses a "syntax fusion encoder" that combines dependency labels and POS tags, along with high-order and syntactic-aware scoring mechanisms, to significantly improve performance on this challenging task.

What Problem Does This Paper Address?

Fine-grained aspect-based sentiment analysis (ABSA) aims to understand detailed opinions expressed about specific aspects of a product or service. A key sub-task is Pair-wise Aspect and Opinion Terms Extraction (PAOTE), which involves identifying and linking aspect terms (e.g., "screen") with their corresponding opinion terms (e.g., "bright"). Existing methods often struggle with the nuances of language, such as long-range dependencies and overlapping terms. This paper addresses the limitations of prior work by proposing a more sophisticated way to leverage syntactic information, which is crucial for understanding the relationships between words in a sentence, and by introducing higher-order scoring to better handle complex term interactions.

Methods

The proposed model, named SynFue, enhances PAOTE through several key innovations:

  • Syntax Fusion Encoder (SynFue): This module is designed to capture rich syntactic and linguistic features. It comprises:
  • Label-Aware Graph Convolutional Network (LAGCN): This network models dependency edges, their labels (e.g., 'nsubj', 'conj'), and Part-of-Speech (POS) tags in a unified manner. This is an improvement over previous methods that only considered dependency edges.
  • Local-Attention Module: This module specifically encodes POS tags to improve the detection of term boundaries, which is important for identifying multi-word terms.
  • High-Order Aspect-Opinion Term Pairing: Instead of just considering pairs of terms, the model employs:
  • Biaffine Scoring: A standard method for measuring pairwise relationships.
  • Triaffine Scoring: A novel addition that models triadic relations, allowing for a better understanding of overlapping term structures.
  • Syntactic-Aware Scoring: The representations learned by LAGCN are re-harnessed to perform scoring that is explicitly aware of the syntactic relationships, further guiding the pairing process.

Datasets

The paper evaluates its model on four benchmark datasets commonly used for PAOTE:

  • 14lap
  • 14res
  • 15res
  • 16res

The authors also provide statistics on the number of sentences, aspect terms, opinion terms, and overlapping pairs for each dataset, noting that overlapping pairs constitute a significant portion (around 24.42% in their data).

Key Insights and Ideas

  • Syntactic Labels Matter: The paper strongly emphasizes that not just the structure of dependency trees (edges) but also the labels on those edges (e.g., 'nsubj' vs. 'conj') provide critical clues for aspect-opinion pairing. This insight is directly incorporated into their LAGCN.
  • POS Tags Aid Boundary Detection: POS tags are shown to be valuable for identifying the precise boundaries of aspect and opinion terms, especially for multi-word phrases.
  • Higher-Order Interactions are Crucial: The introduction of Triaffine scoring highlights the importance of considering interactions beyond simple pairwise relationships, particularly in scenarios with overlapping aspect-opinion terms.
  • Syntactic Information Enhances Pairing: Re-using the learned syntactic representations for scoring directly improves the model's ability to correctly pair aspect and opinion terms.
  • Two Layers are Often Optimal: Experiments suggest that a two-layer SynFue encoder provides the best balance, implying that deeper syntax propagation can lead to redundancy and overfitting.

Key Results

The proposed model consistently outperforms state-of-the-art baselines across all four benchmark datasets, achieving an average F1 score of 72.75%. Ablation studies reveal that:

  • Removing BERT leads to significant performance drops, underscoring its importance.
  • Ablating dependency labels, using a vanilla GCN (without labels), or removing the local attention for POS tags all result in performance degradation, validating the contribution of each component.
  • The syntactic-aware scoring mechanism is crucial for pairing performance.
  • The model demonstrates strong capabilities in both aspect term extraction (ATE) and opinion term extraction (OTE) as sub-tasks.

Limitations and Future Work

  • Overlapping Pairs: While the Triaffine scorer is introduced to handle overlapping pairs, the paper notes that the effectiveness of this component might be limited by the relatively fewer overlapping pairs in some datasets.
  • Layer Number: The optimal number of layers for the SynFue encoder was found to be two, suggesting that excessive layers can lead to information redundancy and overfitting.
  • Future Work: The paper implicitly suggests further exploration into more advanced methods for handling complex syntactic structures and potentially larger datasets with more overlapping pairs to fully leverage the benefits of higher-order scoring. The authors also mention that their model provides explainable predictions, hinting at future work in interpretability.

Summary updated

Abstract

In this paper, we propose to enhance the pair-wise aspect and opinion terms extraction (PAOTE) task by incorporating rich syntactic knowledge. We first build a syntax fusion encoder for encoding syntactic features, including a label-aware graph convolutional network (LAGCN) for modeling the dependency edges and labels, as well as the POS tags unifiedly, and a local-attention module encoding POS tags for better term boundary detection. During pairing, we then adopt Biaffine and Triaffine scoring for high-order aspect-opinion term pairing, in the meantime re-harnessing the syntax-enriched representations in LAGCN for syntactic-aware scoring. Experimental results on four benchmark datasets demonstrate that our model outperforms current state-of-the-art baselines, meanwhile yielding explainable predictions with syntactic knowledge.
BibTeX
@inproceedings{wu2021learn,
  author = {Shengqiong Wu and Hao Fei and Yafeng Ren and ji Donghong},
  title = {{Learn from Syntax: Improving Pair-wise Aspect and Opinion Terms Extractionwith Rich Syntactic Knowledge}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/545},
}
Figure