[Paper Review] Explicit Interaction Network for Aspect Sentiment Triplet Extraction

Peiyi Wang, Tianyu Liu+4 more|arXiv (Cornell University)|Jun 21, 2021

Sentiment Analysis and Opinion Mining43 references4 citations

TL;DR

This paper proposes an Explicit Interaction Network for Aspect Sentiment Triplet Extraction (ASTE) that splits the task into target-opinion joint detection and sentiment classification, using a sequence encoder and a table encoder to capture compositional features at the token-pair level. Bidirectional explicit interaction between modules improves performance, achieving state-of-the-art results on six ASTE datasets, especially in challenging 'many-to-one' scenarios with multiple triplets.

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.

Motivation & Objective

• To address the limitations of prior ASTE methods in handling 'many-to-one' and 'one-to-many' sentiment triplet scenarios.
• To improve subtask interaction by replacing implicit, shared encoder-based communication with explicit, bidirectional interaction between target-opinion detection and sentiment classification.
• To reduce erroneous predictions of non-existent triplets by modeling compositional features at the token-pair level instead of relying on token-level features.
• To align the model's subtask formulation with human cognitive processes—evaluating sentiment based on both target and opinion together.
• To achieve state-of-the-art performance on six benchmark ASTE datasets, particularly in high-difficulty, multi-triplet sentences.

Proposed method

• The model divides ASTE into two components: target-opinion joint detection using a sequence encoder, and sentiment classification using a table encoder.
• The table encoder generates sentiment features directly at the token-pair level, capturing compositional features between targets and opinions.
• Explicit bidirectional interaction is established via table-guided attention, where table representation guides sequence encoding, and sequence features are injected back into the table encoder.
• The method uses a sequence encoder (e.g., BERT) for contextual encoding, and a table encoder that processes all target-opinion pairs in parallel to predict sentiment polarities.
• The interaction mechanism allows each module to selectively use useful information from the other, reducing interference and improving robustness.
• The model is trained end-to-end with cross-entropy loss on triplet prediction, with joint optimization of detection and sentiment classification.

Experimental results

Research questions

• RQ1 How does explicit, bidirectional interaction between subtasks improve ASTE performance compared to implicit or unidirectional interaction?
• RQ2 Can modeling sentiment at the token-pair level reduce false positive predictions of non-existent triplets?
• RQ3 Does a subtask combination that aligns with human cognition—evaluating sentiment based on target-opinion pairs—lead to better performance on complex sentences with multiple triplets?
• RQ4 How does the model perform in 'many-to-one' scenarios where one target has multiple sentiment polarities?
• RQ5 To what extent does the table encoder’s compositional feature extraction outperform token-level feature modeling in handling ambiguous or overlapping sentiment expressions?

Key findings

• The proposed model achieves state-of-the-art performance on six popular ASTE datasets, with absolute F1 improvements of +3.72, +4.90, +7.93, and +8.73 over baselines on 15Lap, 14Rest, 15Rest, and 16Rest respectively.
• On the 14Lap dataset, the model achieves an F1 score of 71.13, significantly outperforming the next-best method JET_o (68.73) and GTS (63.83).
• The model shows a marked advantage in sentences with multiple triplets, achieving a 58.21% ratio of such sentences on 14Rest, where performance gain is most pronounced.
• In a case study with five golden triplets, the model correctly extracts all five, while JET_o misses two and both JET_o and GTS generate two false positives due to strong sentiment words.
• The model’s performance gain increases with the number of triplets per sentence, as shown in Figure 8, indicating superior robustness in complex, high-difficulty cases.
• The ratio of sentences with multiple triplets (R_t>1) correlates strongly with performance gain, with the largest improvement observed on 14Rest (58.21% of sentences have multiple triplets), where the model gains +8.73 F1 points.