Source-linked AI summary
Knowing What, How and Why: A Near Complete Solution for Aspect-based Sentiment Analysis
Haiyun Peng, Lu Xu, Lidong Bing, Fei Huang, Wei Lu, Luo Si
TL;DR
Existing ABSA methods do not extract aspects, sentiment, and the opinions causing it together, leaving the full sentiment picture incomplete. This paper introduces ASTE and a two-stage framework that predicts these elements and pairs them, achieving benchmark performance and outperforming adapted strong baselines.
Problem
Existing ABSA methods handle individual or paired subtasks but do not provide one-shot extraction of aspects, their sentiments, and the opinions causing them.
Method
The two-stage framework jointly predicts aspect-sentiment labels and opinion terms, then pairs candidate aspects with the opinions they modify.
Results
The model outperforms strong baselines across stage-one extraction and achieves a steady advantage in final triplet extraction.
Takeaways & Limitations
ASTE provides a benchmarked framework for extracting aspect-sentiment-opinion triplets in one shot.
Takeaways & Limitations
Pairing multiple opinions with one aspect can produce false-positive triplets during the second stage.
Abstract
from arXiv · showhide
Target-based sentiment analysis or aspect-based sentiment analysis (ABSA) refers to addressing various sentiment analysis tasks at a fine-grained level, which includes but is not limited to aspect extraction, aspect sentiment classification, and opinion extraction. There exist many solvers of the above individual subtasks or a combination of two subtasks, and they can work together to tell a complete story, i.e. the discussed aspect, the sentiment on it, and the cause of the sentiment. However, no previous ABSA research tried to provide a complete solution in one shot. In this paper, we introduce a new subtask under ABSA, named aspect sentiment triplet extraction (ASTE). Particularly, a solver of this task needs to extract triplets (What, How, Why) from the inputs, which show WHAT the targeted aspects are, HOW their sentiment polarities are and WHY they have such polarities (i.e. opinion reasons). For instance, one triplet from "Waiters are very friendly and the pasta is simply average" could be ('Waiters', positive, 'friendly'). We propose a two-stage framework to address this task. The first stage predicts what, how and why in a unified model, and then the second stage pairs up the predicted what (how) and why from the first stage to output triplets. In the experiments, our framework has set a benchmark performance in this novel triplet extraction task. Meanwhile, it outperforms a few strong baselines adapted from state-of-the-art related methods.
Introduction
The introduction argues that existing individual and compound ABSA tasks do not fully explain sentiment because they omit the opinion reason. It therefore motivates aspect sentiment triplet extraction and presents a two-stage approach that extracts aspects, polarities, and opinions before pairing them.
- Task definition: Aspect sentiment triplet extraction targets the three ABSA goals: the explicit aspect term, its sentiment polarity, and the opinion term expressing the reason.The aspect term names the discussed target, while the opinion term is the evaluative comment or phrase.
- Motivation: Existing compound ABSA tasks remain insufficient because identifying sentiment toward an aspect does not reveal the opinion term causing that sentiment.The introduction uses “waiters” and “friendly” to illustrate why sentiment classification alone cannot provide a complete sentiment picture.
- Two-stage framework: The first stage jointly extracts potential aspect terms with sentiment and potential opinion terms using two label sequences.Aspect extraction and sentiment classification use a unified tagging system, while opinion extraction uses a BIO-like tagging system.
- Two-stage framework: The second stage pairs extracted aspects and opinion expressions, using distance embeddings because word distance is informative when sentences contain multiple aspects and opinions.The framework first obtains aspect–sentiment candidates and opinion candidates, then models their correspondence.
Proposed Framework · Problem Formulation
The ASTE task extracts sentiment triplets describing targets, their polarities, and the opinions causing those polarities. Its formulation uses two stages: sequence labeling followed by candidate-pair identification.
- Problem Formulation: ASTE extracts triplets consisting of aspects or targets, their sentiment polarities, and the opinions causing those sentiments.These dimensions correspond to What, How, and Why.
- Proposed Framework: The framework formulates ASTE in two stages, as illustrated by the proposed two-stage model.Stage one performs labeling, while stage two produces final triplets from candidate pairs.
- Problem Formulation: Stage one contains unified-tag and opinion-tag sequence-labeling subtasks.The unified tags locate aspects and encode their sentiment, while the opinion tags identify opinion expressions.
- Problem Formulation: Opinion expressions are paired with the targets or aspects they modify in a many-to-many setting.This assumption permits one opinion to modify multiple targets and one target to associate with multiple opinions.
- Problem Formulation: The unified tag schema uses BIOES-style polarity labels for positive, negative, and neutral aspects, together with O.The passage lists B-, I-, E-, and S- labels for each polarity category.
- Problem Formulation: The opinion tag schema uses B, I, E, and S labels, together with O, to represent opinion-expression spans.The predicted opinion sequence contains one tag per input position.
- Proposed Framework: Stage two constructs a candidate pool by coupling every extracted aspect with every extracted opinion expression.For n aspects and m opinion expressions, the pool contains all listed (T_i, O_j) combinations.
- Proposed Framework: Stage two identifies legitimate candidate pairs and outputs them as the final results, with sentiment polarity embedded in each aspect representation.The candidate pairs are formed from aspects and opinions labeled in the same sentence.
Model Overview
The framework uses two stages: a unified first stage predicts target and opinion labels, while a second stage pairs predicted aspects and opinion expressions. Its architecture combines boundary-guided BLSTMs with dependency-based graph convolution to support target and opinion extraction.
- Stage One: Stage one jointly predicts target labels YT S and opinion labels YOPT in a unified model.Its target side adapts Li et al. (2019a)’s unified tag schema and shares a target-guidance component with opinion prediction.
- Stage One: Two stacked BLSTMs predict target boundaries and use boundary guidance and target guidance to provide signals for target extraction.The lower BLSTMT performs auxiliary BIO boundary prediction for the upper BLSTM and guidance components.
- Stage One: A GCN models mutual target–opinion influence through dependency relations before sending this signal to target-guidance and opinion-prediction modules.The target-guidance component concatenates target boundary information with the GCN output to leverage target information for opinion extraction.
- Stage Two: Stage two generates all possible aspect–opinion pairs from stage-one predictions and applies position embeddings based on target–opinion distance.Non-target and non-opinion terms share a zero position embedding, followed by a BLSTM encoder and concatenation of aspect and opinion hidden states.
Stage One
Stage One jointly predicts target boundaries, sentiment-appended unified tags, and opinion terms through interacting sequence, graph, guidance, consistency, and fusion mechanisms. It is trained by minimizing the summed cross-entropy losses of its four output signals.
- Unified tag prediction: A BLSTM target tagger predicts aspect boundary labels, while a second BLSTM accumulates sentiment information for unified aspect-sentiment tagging.The sentiment tag set appends positive, negative, or neutral polarity to boundary tags.
- Unified tag prediction: The Sentiment Consistency module uses a gate to inherit features from the previous time step and reduce drastic sentiment-label changes across multi-term aspects.This addresses contradictory term-level sentiment predictions within an aspect.
- Unified tag prediction: Boundary Guidance constrains boundary-to-sentiment transitions with a probability transformation matrix, while opinion representations reinforce unified tag prediction.The opinion representation hOPT is integrated into unified prediction, unlike in Li et al. (2019a).
- Opinion extraction: A GCN captures syntactic dependencies, and Target Guidance combines target-boundary information with these representations to guide opinion-term extraction.The resulting BLSTMOPT output both predicts opinions and assists unified tag prediction through Boundary Guidance.
- Training: Stage One minimizes the total objective J(θ) = LT + LT S + LT G + LOPT using stochastic gradient descent and cross-entropy losses for four output signals.The task indicators are T, TS, TG, and OPT.
Stage Two
Stage two enumerates all possible aspect–opinion pairs from stage one and classifies each candidate as valid or invalid. It uses positional information and contextual BLSTM representations to support pair classification.
- Candidate Pair Classification: Stage two enumerates every possible aspect–opinion pair and classifies whether each candidate pair is valid.Given n aspects and m opinion expressions, it forms n×m candidate pairs.
- Position Encoding: The model encodes aspect–opinion positional relations using their absolute word-length distance as relative position information.Position embeddings are created by treating the distance as a position index for aspects and opinions.
- Contextual Encoding: It concatenates pretrained GloVe embeddings with trainable position embeddings, then uses a BLSTM to encode sentence context for aspects and opinions.For each aspect and opinion expression, the model averages the BLSTM hidden states corresponding to its sentence term indices.
Experiments
Experiments use merged SemEval-derived data and evaluate the framework in two stages: unified aspect/sentiment and opinion extraction, followed by triplet or pair prediction. The model generally outperforms strong baselines, while ablations and examples expose component effects and pairing limitations.
- Experimental setup: The datasets originate from SemEval Challenges, merge same-sentence samples with differing target and opinion annotations, and use a randomly selected 20% of training data for validation.Each sample contains the original sentence, unified aspect/target tags, and opinion tags; overlapping targets and opinions are corrected.
- Stage one: Stage one compares unified aspect extraction and sentiment classification against RINANTE, CMLA, and Li-Unified, alongside opinion-term extraction baselines.The evaluation follows the framework’s two-stage design, with separate comparisons for aspect/opinion co-extraction and opinion extraction.
- Stage one: Our model outperforms RINANTE, CMLA, and Li-Unified on unified aspect extraction and sentiment classification across all datasets.Li-unified-R is nevertheless better than Li-unified on all datasets and surpasses the full model on 14res and 14lap.
- Stage one: The core model achieves the best F score among existing baselines for opinion-term extraction, while the our–TG variant leads all baselines in the laptop domain.Li-unified-R remains competitive, exceeding the model on 14lap but generally trailing it on restaurant datasets.
- Stage two: Stage two sends candidate triplets to a binary classifier trained on ground-truth aspect-opinion pairs and evaluates both triplet and pair performance.Because no identical-setting baseline exists, the stage-two model is stacked on the best stage-one baselines to form pipeline models.
- Ablation tests: Removing BLSTMOPT usually reduces performance, whereas removing TG lowers unified-tag results but can improve opinion extraction on 14res and 14lap.The authors attribute the mixed TG effect to potentially weak mutual relations between aspects and opinion terms.
Conclusions
The paper introduces aspect sentiment triplet extraction, a one-shot task that jointly identifies aspects, their sentiment polarities, and opinion reasons. Its two-stage framework generates mutually informed candidates and then pairs the correct aspects with opinion terms.
- Task and framework: The proposed task jointly extracts aspects, sentiment polarities, and opinion terms to answer what, how, and why in one shot.It couples aspect extraction, aspect-term sentiment classification, and opinion-term extraction.
- Task and framework: The first stage generates candidate aspects with sentiment polarities and candidate opinion terms using their mutual influence.
- Task and framework: The second stage pairs the correct aspects and opinion terms to produce sentiment triplets.