[Paper Review] First Target and Opinion then Polarity: Enhancing Target-opinion Correlation for Aspect Sentiment Triplet Extraction

Lianzhe Huang, Peiyi Wang+6 more|arXiv (Cornell University)|Feb 17, 2021

Sentiment Analysis and Opinion Mining26 references28 citations

TL;DR

The paper proposes a two-stage ASTE framework that first extracts targets and opinions, then uses perceivable target-opinion pairs with restricted attention to jointly determine triplet polarity, achieving state-of-the-art results on multiple ASTE datasets.

ABSTRACT

Aspect Sentiment Triplet Extraction (ASTE) aims to extract triplets from a sentence, including target entities, associated sentiment polarities, and opinion spans which rationalize the polarities. Existing methods are short on building correlation between target-opinion pairs, and neglect the mutual interference among different sentiment triplets. To address these issues, we utilize a two-stage framework to enhance the correlation between targets and opinions: at stage one, we extract targets and opinions through sequence tagging; then we append a group of artificial tags named Perceivable Pair, which indicate the span of a specific target-opinion tuple, to the input sentence to obtain closer correlated target-opinion pair representation. Meanwhile, we reduce the negative interference between triplets by restricting tokens' attention field. Finally, the polarity is identified according to the representation of the Perceivable Pair. We conduct experiments on four datasets, and the experimental results show the effectiveness of our model.

Motivation & Objective

• Motivate ASTE by addressing weak target-opinion correlation in prior methods.
• Propose a two-stage framework: (1) extract targets and opinions; (2) judge polarity using target-opinion-tuple aware representations.
• Enforce stronger target-opinion coupling via perceivable pairs appended to input and restricted attention to reduce inter-triplet interference.
• Demonstrate effectiveness on ASTE-DATA-V2 across four datasets and show advantages in complex many-to-one/multi-triplet scenarios.

Proposed method

• Stage 1 uses a BERT-based tagging model to jointly extract targets and opinions from the sentence.
• Stage 2 enumerates all possible target-opinion pairs as perceivable pairs and appends them to the input to form a pair-aware representation.
• A compound computation scheme concatenates all perceivable pairs into a single sequence to enable efficient encoding.
• Perceivable pairs share position embeddings to indicate span boundaries and use distinct segment IDs to separate original input from the pair tokens.
• A restricted attention field confines each perceivable pair’s attention to its own context, reducing interference from other pairs.
• A BERT-based encoder produces pair representations by concatenating T-B/T-E and O-B/O-E to form r_ij, which is then used for 4-class polarity/match prediction (POS, NEU, NEG, O).

Experimental results

Research questions

• RQ1 Can a two-stage extraction-pairing approach improve ASTE performance over end-to-end or tagging-only methods?
• RQ2 Do perceivable pairs and restricted attention effectively strengthen target-opinion coupling and reduce cross-triplet interference?
• RQ3 How does the proposed method perform on complex scenarios like many-to-one or multiple triplets per sentence?
• RQ4 Is the approach scalable with standard encoders (e.g., BERT) on ASTE-DATA-V2 datasets?

Key findings

• The method achieves state-of-the-art results on 14Lap and 16Rest datasets and is close to state-of-the-art on 14Rest, while outperforming baselines on average across four datasets.
• Ablation shows perceivable pairs, their start/end tags, and restricted attention are critical for performance gains.
• The approach better handles many-to-one and multi-triplet cases, with larger performance gaps in complex sentences.
• Compared to OTE-MTL, the proposed method better captures target-opinion correlations by evaluating all possible pairs instead of relying on single representations.
• Qualitative analysis demonstrates that perceivable pairs help focus attention on the truly informative tokens for polarity prediction.