Source-linked AI summary
Bidirectional Machine Reading Comprehension for Aspect Sentiment Triplet Extraction
Shaowei Chen, Yu Wang, Jie Liu, Yuelin Wang
TL;DR
ASTE requires coordinated extraction of opinion entities, their relations, and sentiments, but existing approaches do not fully capture these associations. BMRC reformulates ASTE as multi-turn MRC with bidirectional query sequences, and experiments report state-of-the-art performance across four benchmark datasets.
Problem
ASTE must jointly identify aspects, opinion expressions, their relations, and sentiments, whose associations and complex mappings challenge existing approaches.
Method
BMRC reformulates ASTE as multi-turn MRC using non-restrictive, restrictive, and sentiment-classification queries in aspect-first and opinion-first directions.
Results
BMRC achieves state-of-the-art performance on four benchmark datasets.
Takeaways & Limitations
The framework unifies opinion-entity extraction, relation detection, and sentiment classification while allowing either entity type to trigger pair extraction.
Abstract
from arXiv · showhide
Aspect sentiment triplet extraction (ASTE), which aims to identify aspects from review sentences along with their corresponding opinion expressions and sentiments, is an emerging task in fine-grained opinion mining. Since ASTE consists of multiple subtasks, including opinion entity extraction, relation detection, and sentiment classification, it is critical and challenging to appropriately capture and utilize the associations among them. In this paper, we transform ASTE task into a multi-turn machine reading comprehension (MTMRC) task and propose a bidirectional MRC (BMRC) framework to address this challenge. Specifically, we devise three types of queries, including non-restrictive extraction queries, restrictive extraction queries and sentiment classification queries, to build the associations among different subtasks. Furthermore, considering that an aspect sentiment triplet can derive from either an aspect or an opinion expression, we design a bidirectional MRC structure. One direction sequentially recognizes aspects, opinion expressions, and sentiments to obtain triplets, while the other direction identifies opinion expressions first, then aspects, and at last sentiments. By making the two directions complement each other, our framework can identify triplets more comprehensively. To verify the effectiveness of our approach, we conduct extensive experiments on four benchmark datasets. The experimental results demonstrate that BMRC achieves state-of-the-art performances.
Introduction
ASTE unifies aspect, opinion-expression, and sentiment identification, but their associations and complex relations make joint extraction challenging. BMRC addresses this by casting ASTE as multi-turn MRC with bidirectional query sequences.
- Existing methods generally solve fine-grained opinion-mining subtasks individually or in pairs, leaving complete triplet identification unresolved.
- ASTE identifies aspects, corresponding opinion expressions, and sentiments as triplets such as (food, delicious, positive).
- Aspect–opinion relations can be one-to-many, many-to-one, overlapped, or embedded, while multiple sentiments may depend on those relations.
- The paper formalizes ASTE as a multi-turn MRC task, using prior answers in later queries to learn associations among subtasks.
- BMRC uses non-restrictive, restrictive, and sentiment-classification queries in two directions, beginning with either aspects or opinion expressions.
- Experiments on four benchmark datasets report state-of-the-art performance for BMRC.
Related Work
Prior work studies fine-grained opinion-mining subtasks, joint extraction, and machine reading comprehension architectures. BMRC extends MRC to ASTE with bidirectional processing intended to identify triplets more comprehensively.
- Fine-grained opinion mining includes aspect term extraction, opinion term extraction, and aspect-level sentiment classification.
- Earlier studies increasingly combine subtasks through aspect–polarity, aspect–opinion, aspect-category–sentiment, and aspect–opinion-pair extraction.
- MRC answers queries from contexts using architectures such as BiDAF and QANet to model query–context interaction.
- MRC has been applied to named entity recognition, entity relation extraction, and summarization, motivating its use for ASTE.
- Unlike existing MRC applications, BMRC adds bidirectional processing so either an aspect or opinion expression can initiate triplet extraction.
Problem Formulation
ASTE identifies aspect–opinion–sentiment triplets from review sentences, representing the task as collections of triplets with sentiment assigned to each aspect. Its multi-turn formulation uses extraction and classification queries to connect subtasks.
- Task Definition: ASTE represents each review sentence as a collection of triplets containing an aspect, opinion expression, and sentiment.The task uses T = {(a_i, o_i, s_i)} for a sentence X.
- Query Structure: The formalization constructs three query types: non-restrictive extraction, restrictive extraction, and sentiment classification.
- Query Structure: In the first turn, non-restrictive queries extract entities from the review sentence to trigger aspect–opinion pairs.
- Query Structure: Restrictive queries then identify corresponding aspects or opinion expressions based on previously extracted entities, followed by sentiment classification.
- Task Definition: Each aspect receives a sentiment label from Positive, Negative, or Neutral.
Methodology
BMRC extracts aspect–opinion pairs in both directions and then predicts aspect-oriented sentiment, using query dependencies to jointly connect extraction, relation detection, and classification.
- Bidirectional Framework: BMRC uses A→O and O→A directions so either an aspect or an opinion expression can trigger extraction of the corresponding entity.The framework fuses answers from both directions to form triplets during inference.
- A→O Direction: In the A→O direction, non-restrictive queries extract aspects and restrictive queries identify their corresponding opinion expressions.
- O→A Direction: In the O→A direction, non-restrictive queries extract opinion expressions and restrictive queries identify the aspects they describe.
- Sentiment Classification: Sentiment classification queries predict each aspect’s polarity using the aspect and its corresponding opinion expressions.The query is designed to predict positive, negative, or neutral sentiment for each aspect.
- Joint Learning: The query design jointly learns dependencies among opinion entity extraction, relation detection, and sentiment classification.The framework states that aspect and opinion semantics are considered during sentiment prediction.
Encoding Layer
BMRC encodes each query together with its review sentence using BERT, combining token, position, and segment information before Transformer encoding.
- Input Construction: The encoder concatenates each query with the review sentence to form a combined input beginning with [CLS].
- Input Construction: Each token’s initial representation sums word, position, and segment embeddings.
- BERT Encoding: BERT encodes the resulting representation sequence through stacked Transformer blocks.
Answer Prediction
BMRC predicts extraction answers with token-level start and end classifiers, while sentiment queries use a classifier over the [CLS] representation.
- Extraction Prediction: Non-restrictive and restrictive extraction queries can return multiple opinion entities from the review sentence.
- Extraction Prediction: Two binary classifiers predict whether each token is an answer-span start or end position.
- Extraction Prediction: The start and end classifiers use hidden token representations with separate parameter matrices W_s and W_e.The cited formulation defines both matrices as model parameters over hidden representations.
- Sentiment Prediction: Sentiment classification uses the hidden representation of [CLS] and a three-class classifier to predict y_S.The classifier predicts sentiment for the aspect–opinion context supplied by the query.
Joint Learning
BMRC jointly learns ASTE subtasks by combining losses from non-restrictive extraction, restrictive extraction, and sentiment classification queries. The resulting objective is optimized with gradient descent using AdamW in this paper.
- Loss design: Non-restrictive extraction queries in both directions use cross-entropy loss against the gold distributions.The predicted distributions are compared with gold distributions for optimization.
- Loss design: Restrictive extraction queries in both directions are trained with a corresponding loss to learn entity-relation dependencies.
- Loss design: Sentiment classification queries use cross-entropy loss, and all query losses are combined into the model’s overall objective.
- Optimization: The combined optimization problem can be solved by gradient descent; BMRC uses AdamW for optimization.
Inference
During inference, BMRC fuses query answers from aspect-to-opinion and opinion-to-aspect directions to construct aspect-opinion pairs and sentiment triplets. A threshold filters asymmetric pairs before sentiment classification completes the triplets.
- Aspect-to-opinion inference: In the aspect-to-opinion direction, non-restrictive queries predict aspects, while restrictive queries identify their corresponding opinion expressions.
- Bidirectional inference: The opinion-to-aspect direction independently identifies aspect-opinion pairs in reverse order, complementing the aspect-to-opinion direction.
- Pair fusion: BMRC retains the intersection of directional pair sets and adds difference-set pairs whose joint probability exceeds threshold δ.The pair probability is defined differently depending on whether the pair comes from the aspect-to-opinion or opinion-to-aspect set.
- Pair fusion: Each opinion entity’s probability is computed by multiplying the probabilities of its start and end positions.
- Triplet construction: Sentiment classification queries predict the sentiment for each extracted aspect, producing triplets from the fused aspect-opinion pairs and sentiments.
Experiments
BMRC is evaluated on four benchmark datasets using multiple ASTE-related subtasks and repeated test runs. It achieves state-of-the-art performance, with especially large gains in pair and triplet extraction, and remains competitive on datasets retaining one-to-many opinion relations.
- Experimental setup: The experiments use four benchmark datasets from the SemEval ABSA Challenges, with annotations and splits derived from prior work.
- Experimental setup: Performance is evaluated with Precision, Recall, and F1-score across aspect-sentiment, opinion, pair, and triplet extraction subtasks.Testing results are averaged over five runs with different random seeds, selecting the best development-set checkpoint each time.
- Main results: BMRC achieves state-of-the-art performance on all datasets, with average F1 gains of 5.14% for aspect-opinion pair extraction and 9.58% for triplet extraction.Improvements on aspect-sentiment co-extraction and opinion term extraction are described as slight.
- Main results: The results indicate that pipeline extraction causes severe error accumulation, whereas BMRC fuses and simplifies aspect, opinion, and relation tasks.The reported F1 improvement is attributed mainly to increased precision, indicating more reliable predictions than the baselines.
- Additional evaluation: On AFOE datasets retaining cases where one opinion expression corresponds to multiple aspects, BMRC and GTS significantly outperform the IMN+IOG pipeline.Compared with GTS, BMRC still achieves competitive performance.
Ablation Study
The ablations show that bidirectional extraction, restrictive queries, relation-aware sentiment classification, and BERT each contribute to BMRC’s performance. Bidirectional extraction reduces directional bias, while relation information strengthens sentiment co-extraction.
- Restrictive Extraction Query: Restrictive extraction queries are evaluated for capturing dependencies between opinion entity extraction and relation detection.The ablation removes the restrictive query and separates relation detection into a later MRC stage.
- Bidirectional Structure: Bidirectional MRC improves aspect-opinion pair and triplet extraction by allowing either aspects or opinions to initiate pair detection.Removing either direction harms extraction of the corresponding entity and subsequently weakens pair and triplet extraction.
- Relation-Aware Sentiment Classification: Relation-aware joint learning significantly improves aspect term and sentiment co-extraction compared with separating relation detection and sentiment classification.Aspect term extraction is not seriously harmed by removing relation detection, but aspect-and-sentiment co-extraction is weakened.
- Effect of BERT: BMRC without BERT surpasses TSF by an average of 8.15% F1-score on triplet extraction.This comparison shows that the MRC framework achieves state-of-the-art performance without BERT.
- Effect of BERT: BERT adds 7.69% F1-score, while the bidirectional MRC structure and three query types add 10.4% over the BERT-based ablation model.The analyses report a larger improvement from the proposed contributions than from BERT alone.
Conclusion
The paper formalizes ASTE as a multi-turn reading-comprehension task and proposes BMRC with specialized queries and bidirectional extraction. The framework integrates entity extraction, relation detection, and sentiment classification, achieving state-of-the-art performance empirically.
- Conclusion: BMRC formalizes ASTE as MTMRC and uses non-restrictive, restrictive, and sentiment classification queries to connect its subtasks.The framework jointly incorporates opinion entity extraction, relation detection, and sentiment classification.
- Conclusion: Bidirectional MRC lets either an aspect or an opinion expression trigger an aspect-opinion pair, complementing the two extraction directions.This design follows the possibility that triplets can originate from either entity type.
- Conclusion: Joint sentiment classification incorporates aspect-opinion relations to further promote sentiment prediction.The sentiment classification query is used together with joint learning.
- Conclusion: BMRC achieves state-of-the-art performance in the empirical study.