Explicit Interaction Network for Aspect Sentiment Triplet Extraction

Peiyi Wang, Lianzhe Huang, Tianyu Liu, Damai Dai, Runxin Xu, Houfeng Wang, Baobao Chang, Zhifang SuiarXiv 2021 · OpenReview

Abstract

Aspect Sentiment Triplet Extraction (ASTE) aims to recognize targets, their sentiment polarities and opinions explaining the sentiment from a sentence. ASTE could be naturally divided into 3 atom subtasks, namely target detection, opinion detection and sentiment classification. We argue that the proper subtask combination, compositional feature extraction for target-opinion pairs, and interaction between subtasks would be the key to success. Prior work, however, may fail on ‘one-to-many’ or ‘many-to-one’ situations or derive non-existent sentiment triplets due to defective subtask formulation, sub-optimal feature representation or the lack of subtask interaction. In this paper, we divide ASTE into target-opinion joint detection and sentiment classification subtasks, which is in line with human cognition, and correspondingly utilize sequence encoder and table encoder to handle them. Table encoder extracts sentiment at token-pair level, so that the compositional feature between targets and opinions can be easily captured. To establish explicit interaction between subtasks, we utilize the table representation to guide the sequence encoding, and inject the sequence features back into the table encoder. Experiments show that our model outperforms state-of-the-art methods on six popular ASTE datasets.

Introduction

In fine-grained sentiment analysis, the goal is to extract triplets consisting of an aspect target, a sentiment polarity, and an opinion term. For instance, in the triplet (sofa, POSITIVE, nice), "sofa" is the target, "POSITIVE" is the sentiment, and "nice" is the opinion. This task, known as Aspect Sentiment Triplet Extraction (ASTE), is challenging because real-world reviews often contain complex mappings.

Examples of complex ASTE scenarios: one-to-many (S1) and many-to-one (S2).

As shown in Figure 1, a single target can be associated with multiple opinions (one-to-many), or a single opinion can describe multiple targets (many-to-one). Previous models often struggled with these scenarios because they relied on token-level features or lacked sufficient interaction between the subtasks of target/opinion detection and sentiment classification.

Methodology: The Explicit Interaction Network (EIN)

The proposed Explicit Interaction Network (EIN) addresses these limitations by mimicking human cognitive processes. It divides ASTE into two main subtasks: target-opinion joint detection and sentiment classification.

Overview of the Explicit Interaction Network (EIN) architecture and the interaction between Sequence and Table Encoders.

Dual-Encoder Architecture

The model utilizes two specialized encoders that operate in parallel and interact at every layer:

  1. Sequence Encoder: This component performs target and opinion detection using standard BIO sequence labeling. It focuses on identifying the boundaries of relevant terms.
  2. Table Encoder: This component treats sentiment classification as a relation extraction task on a 2D grid. Each cell $(m, n)$ in the table represents the relationship between token $m$ and token $n$. This allows the model to capture "compositional features"—how a specific pair of words (e.g., "low" and "price") interacts to produce a sentiment.

Explicit Interaction Mechanisms

The core innovation is the bidirectional feedback loop between these encoders:

  • Table Guided Attention (TGA): High-level sentiment features from the table encoder are used to guide the sequence encoder's attention mechanism, helping it focus on tokens that are likely part of a triplet.
  • Sequence Feature Injection (SFI): Features from the sequence encoder (identifying targets and opinions) are injected back into the table encoder. This provides the table encoder with structural information about term boundaries, refining its sentiment predictions.

Experimental Results

The EIN model was evaluated on the ASTE-DATA-V1 and V2 benchmarks, which include laptop and restaurant reviews. It achieved state-of-the-art (SOTA) performance across all datasets.

  • Superior Accuracy: On the ASTE-DATA-V2 14Rest dataset, EIN reached an F1 score of 71.13, a significant improvement over previous top-performing models like GTS (68.22).
  • Handling Complexity: The model's strength is most evident in sentences with multiple triplets. While other models' performance degrades as the number of triplets increases, EIN maintains high accuracy due to its token-pair level table encoding.
  • Efficiency: EIN demonstrates strong performance even with simpler base encoders like GloVe, often outperforming BERT-based competitors that lack explicit subtask interaction.

Conclusion

The Explicit Interaction Network represents a significant step forward in ASTE by moving beyond simple multi-task learning. By establishing explicit, bidirectional communication between target detection and sentiment classification, the model effectively captures the relational nature of sentiment. Future work could involve extending this interaction framework to other complex information extraction tasks, such as joint entity and relation extraction.

Summary updated

Abstract

Aspect Sentiment Triplet Extraction (ASTE) aims to recognize targets, their sentiment polarities and opinions explaining the sentiment from a sentence. ASTE could be naturally divided into 3 atom subtasks, namely target detection, opinion detection and sentiment classification. We argue that the proper subtask combination, compositional feature extraction for target-opinion pairs, and interaction between subtasks would be the key to success. Prior work, however, may fail on `one-to-many' or `many-to-one' situations or derive non-existent sentiment triplets due to defective subtask formulation, sub-optimal feature representation or the lack of subtask interaction. In this paper, we divide ASTE into target-opinion joint detection and sentiment classification subtasks, which is in line with human cognition, and correspondingly utilize sequence encoder and table encoder to handle them. Table encoder extracts sentiment at token-pair level, so that the compositional feature between targets and opinions can be easily captured. To establish explicit interaction between subtasks, we utilize the table representation to guide the sequence encoding, and inject the sequence features back into the table encoder. Experiments show that our model outperforms state-of-the-art methods on six popular ASTE datasets.
BibTeX
@misc{wang2021explicit,
  author = {Peiyi Wang and Lianzhe Huang and Tianyu Liu and Damai Dai and Runxin Xu and Houfeng Wang and Baobao Chang and Zhifang Sui},
  title = {{Explicit Interaction Network for Aspect Sentiment Triplet Extraction}},
  year = {2021},
  eprint = {2106.11148},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.11148},
}
Figure