Source-linked AI summary
Learning Span-Level Interactions for Aspect Sentiment Triplet Extraction
Lu Xu, Yew Ken Chia, Lidong Bing
TL;DR
ASTE extracts aspect targets, opinions, and sentiment relations, but existing approaches rely on word-level interactions and pipeline methods can suffer from broken triplet interactions and error propagation. Span-ASTE models target–opinion span interactions, uses ATE and OTE supervision for dual-channel pruning, and significantly improves ASTE, ATE, and OTE performance, especially for multi-word triplets.
Problem
ASTE must extract aspect targets, their opinions, and sentiment relations, while existing approaches rely on word-level interactions and pipeline methods can suffer from broken triplet interactions and error propagation.
Method
Span-ASTE explicitly models interactions between whole target and opinion spans and uses ATE and OTE supervision in dual-channel span pruning to reduce enumerated candidates.
Results
Span-ASTE significantly outperforms previous methods on ASTE, ATE, and OTE across four benchmark datasets with BiLSTM and BERT encoders; with BiLSTM and GloVe, it exceeds the best pipeline model by 15.62, 8.93, 5.24, and 10.16 F1 points.
Takeaways & Limitations
The span-level approach yields more significant improvements on triplets with multi-word targets or opinions, while a substantial performance gap remains between single-word and multi-word triplets.
Takeaways & Limitations
A significant performance gap remains between single-word and multi-word triplets, leaving multi-word triplet extraction as a potential area for future work.
Abstract
from arXiv · showhide
Aspect Sentiment Triplet Extraction (ASTE) is the most recent subtask of ABSA which outputs triplets of an aspect target, its associated sentiment, and the corresponding opinion term. Recent models perform the triplet extraction in an end-to-end manner but heavily rely on the interactions between each target word and opinion word. Thereby, they cannot perform well on targets and opinions which contain multiple words. Our proposed span-level approach explicitly considers the interaction between the whole spans of targets and opinions when predicting their sentiment relation. Thus, it can make predictions with the semantics of whole spans, ensuring better sentiment consistency. To ease the high computational cost caused by span enumeration, we propose a dual-channel span pruning strategy by incorporating supervision from the Aspect Term Extraction (ATE) and Opinion Term Extraction (OTE) tasks. This strategy not only improves computational efficiency but also distinguishes the opinion and target spans more properly. Our framework simultaneously achieves strong performance for the ASTE as well as ATE and OTE tasks. In particular, our analysis shows that our span-level approach achieves more significant improvements over the baselines on triplets with multi-word targets or opinions.
1 Introduction
ASTE extends aspect-based sentiment analysis by extracting aspect targets, opinion terms, and their sentiment relations as triplets. Span-ASTE addresses limitations of prior pipelines and word-level end-to-end methods by modeling target–opinion span interactions and using ATE/OTE supervision for pruning.
- ASTE forms a complete sentiment representation from an aspect target, its corresponding opinion term, and the expressed sentiment.
- Earlier pipeline methods separate term extraction from relation prediction, breaking triplet interactions and risking error propagation.
- Prior end-to-end methods rely on independent word-to-word interactions, although roughly one-third of benchmark triplets contain multi-word targets or opinions.
- Span-ASTE directly captures interactions between whole target and opinion spans when predicting their sentiment relation.
- Span-ASTE significantly improves ASTE performance, especially for triplets containing multi-word targets or opinions, and outperforms previous methods on ASTE, ATE, and OTE across four benchmark datasets.
- The dual-channel pruning strategy uses explicit ATE and OTE supervision to reduce span-enumeration cost and better pair valid target and opinion candidates.
2 Span-based ASTE
Span-ASTE represents enumerated target and opinion spans, prunes them into separate candidate pools using ATE and OTE supervision, and predicts sentiment relations for paired candidates. This design addresses the computational cost of span pairing while preserving target–opinion type distinctions.
- Task formulation: ASTE extracts triplets consisting of a target span, an opinion span, and their sentiment relation.
- Sentence encoding: Span-ASTE constructs span representations from contextualized token representations, span endpoints, and a trainable width feature.BiLSTM and BERT provide alternative contextualized token representations; span representations may also use max- or mean-pooling.
- Mention module: The mention module predicts Target, Opinion, or Invalid for each enumerated span using a feed-forward network.ATE and OTE supervise the dual-channel pruning strategy before target and opinion spans are paired.
- Span pruning: For a sentence of length n, enumerating spans requires O(n2) candidates and pairing target and opinion spans can require O(n4) interactions.These costs make considering every span pair computationally impractical.
- Span pruning: Dual-channel pruning creates separate target and opinion pools instead of mixing mention types in one pool.The selected candidate count for each pool is proportional to sentence length through the threshold hyper-parameter z.
- Triplet module: Each target candidate is coupled with each opinion candidate, and a classifier predicts Positive, Negative, Neutral, or Invalid sentiment relations.The pair representation includes a trainable feature based on the distance between the target and opinion spans.
3 Experiment
Span-ASTE is evaluated on four refined ASTE benchmark datasets using BiLSTM and BERT encoders, with comparisons against pipeline and end-to-end baselines. It consistently achieves stronger ASTE results and also improves ATE and OTE performance without additional retraining or tuning.
- Datasets: The evaluation uses four ASTE datasets derived from SemEval, covering three restaurant datasets and one laptop dataset.The datasets were refined to add missing triplets and remove conflicting sentiments.
- ASTE Results: Table 2 compares Precision, Recall, and F1 on four datasets for Span-ASTE and previous models.Results are reported with both BiLSTM and BERT sentence encoders.
- ASTE Results: 15.62, 8.93, 5.24, and 10.16 F1 points: Span-ASTE exceeds the best pipeline model with a BiLSTM encoder using GloVe.The comparison is reported on the four datasets in order.
- ASTE Results: 4.35, 5.02, 3.12, and 2.33 F1 points: Span-ASTE exceeds the previous best GTS results with the BERT encoder.The gains are reported on the four datasets in order.
- Additional Experiments: Without additional retraining or tuning, Span-ASTE significantly improves F1 over GTS on both ATE and OTE across the four datasets.The model directly addresses both extraction tasks using its existing framework.
4 Analysis
Analyses show that Span-ASTE is especially beneficial for triplets containing multi-word targets or opinions, while dual-channel pruning improves efficiency and span distinction. Ablations further support retaining positional information and boundary-based span representations.
- 4.1 Comparison of Single-word and Multi-word Spans: Multi-word triplets account for one-third or even half of datasets, and their F1 results drop by more than 10 points for both GTS and Span-ASTE.The drop exceeds 20 points for Rest 14.
- 4.1 Comparison of Single-word and Multi-word Spans: Span-ASTE achieves more significant F1 improvements over GTS on multi-word triplets than on single-word triplets.Recall improves more than precision, and span enumeration benefits recall of multi-word spans.
- 4.1 Comparison of Single-word and Multi-word Spans: Performance is lower when a triplet contains a multi-word opinion term, a trend attributed to imbalanced data distribution.
- 4.2 Pruning Efficiency: 4 times more pairs and computational cost: SC-Adjusted must double z to match the dual-channel strategy’s candidate coverage.Under matched efficiency, Dual-Channel has a significant F1 advantage, especially at lower z; the reported model uses z = 0.5.
- 4.3 Qualitative Analysis: Qualitative examples show Span-ASTE accurately extracts multi-word target-opinion pairs and determines their overall sentiment where GTS makes polarity or extraction errors.GTS’s word-level decoding can produce ties or incomplete multi-word targets.
- 4.4 Ablation Study: Removing span width and distance embeddings reduces ASTE performance, indicating that positional information remains useful.Pooling-based span representations also perform worse, suggesting information loss for distinguishing valid and invalid spans.
5 Related Work
Related work situates ASTE within ABSA and distinguishes it from established aspect sentiment classification, aspect term extraction, and opinion term extraction tasks. Prior ASTE methods include pipelines, joint frameworks, and end-to-end structured or word-pair approaches.
- Joint ABSA Frameworks: Prior work includes frameworks jointly solving multiple ABSA subtasks, including relation-aware collaborative learning and aspect-category-oriented extraction.
- ASTE: ASTE extracts and forms an aspect term, its associated sentiment, and the corresponding opinion term into a triplet.
6 Conclusions
Span-ASTE learns interactions between whole target and opinion spans and uses ATE/OTE supervision for dual-channel pruning. Its strong ASTE performance is mostly attributable to improvements on multi-word triplets, while a substantial single-word versus multi-word gap remains.
- Span-ASTE models interactions between target spans and opinion spans rather than only word-to-word interactions.The approach explicitly learns span-level interactions for ASTE.
- ATE and OTE supervision guides dual-channel pruning to reduce enumerated candidates and improve valid target-opinion pairing efficiency.The pruning strategy also supports distinguishing target and opinion spans.
- Span-ASTE significantly outperforms previous methods on ASTE, ATE, and OTE tasks.The conclusion reports strong performance across all three tasks.
- The ASTE gains are mostly attributable to improvements on multi-word triplets.The authors identify this as the main source of the strong ASTE performance.
- A significant performance gap remains between single-word and multi-word triplets, leaving the latter as a potential area for future work.The conclusion identifies this unresolved difference as a future-work direction.
A Additional Experimental Settings
The experiments use either GloVe with a BiLSTM encoder or uncased BERT_base, with specified optimization, embedding, batching, and sequence-length settings.
- BERT-based model runtimes are 157, 115, 87, and 111 sec/epoch on Rest 14, Lap 14, Rest 15, and Rest 16, respectively.The reported total parameter counts are 2.24M with GloVe and 110M with BERT base.
- Span width and distance embeddings have 20 and 128 dimensions, with bucketed inputs ranging from 0 through 64+.The bucket boundaries are [0, 1, 2, 3, 4, 5-7, 8-15, 16-31, 32-63, 64+].
- Parameters are updated after each sentence, producing a batch size of 1, and input sequences are restricted to a maximum of 512 tokens.
B Additional Data Statistics
The additional data-statistics section reports the numbers of target and opinion terms across the four datasets.
- Table 9 reports the numbers of target terms and opinion terms in the four datasets.
C Dev Results
The development-results section presents the model’s results on the development datasets.
- Table 10 reports the model’s results on the development datasets.
D Additional Comparisons
Additional comparisons indicate that Span-ASTE generally outperforms previous end-to-end approaches across four datasets, while comparisons with RACL and JET require care because their task settings or evaluation procedures differ.
- Direct comparison with JET is unavailable because JET cannot directly solve ATE and OTE unless evaluation uses triplet results.
- Span-ASTE generally outperforms the previous two end-to-end approaches on the four datasets.
- RACL comparisons are not fully fair because RACL uses target-term sentiment polarity, whereas Span-ASTE and GTS use target-opinion pairing relations.
- Table 9 reports the numbers of target and opinion terms, while Table 10 reports results on development datasets.
- Table 8 reports additional ATE and OTE test comparisons among RACL, GTS, and Span-ASTE on datasets released by Xu et al. (2020b).