Source-linked AI summary
A Multi-task Learning Framework for Opinion Triplet Extraction
Chen Zhang, Qiuchi Li, Dawei Song, Benyou Wang
TL;DR
ABSA approaches either extract aspect-sentiment pairs without opinion terms or co-extract aspects and opinions without their sentiment dependencies. OTE-MTL jointly extracts aspects and opinions while parsing their sentiment dependencies, and consistently outperforms state-of-the-art baselines on the evaluated datasets.
Problem
Aspect-sentiment pair extraction omits opinion terms, while aspect-opinion co-extraction does not determine sentiment dependencies needed to form meaningful pairs.
Source: §1 · pp. 1–2Method
OTE-MTL jointly extracts aspect and opinion terms and predicts word-level sentiment dependencies with a biaffine scorer for triplet decoding.
Source: §1 · p. 2Results
OTE-MTL consistently outperforms all state-of-the-art baselines on the evaluated datasets with and without OOTs.
Source: §1 · p. 2, §4.1 · p. 6Takeaways & Limitations
The triplet perspective incorporates opinions and aspect-opinion sentiment dependencies, which the authors present as supporting sentiment determination and model interpretability.
Source: §1 · p. 1The authors identify needs for more robust aspect and opinion taggers, a more flexible evaluation metric, and a stronger triplet interaction mechanism.
Source: §6 · pp. 8–9
Abstract
from arXiv · showhide
The state-of-the-art Aspect-based Sentiment Analysis (ABSA) approaches are mainly based on either detecting aspect terms and their corresponding sentiment polarities, or co-extracting aspect and opinion terms. However, the extraction of aspect-sentiment pairs lacks opinion terms as a reference, while co-extraction of aspect and opinion terms would not lead to meaningful pairs without determining their sentiment dependencies. To address the issue, we present a novel view of ABSA as an opinion triplet extraction task, and propose a multi-task learning framework to jointly extract aspect terms and opinion terms, and simultaneously parses sentiment dependencies between them with a biaffine scorer. At inference phase, the extraction of triplets is facilitated by a triplet decoding method based on the above outputs. We evaluate the proposed framework on four SemEval benchmarks for ASBA. The results demonstrate that our approach significantly outperforms a range of strong baselines and state-of-the-art approaches.
1 Introduction
The introduction reframes ABSA as extracting aspect–opinion–sentiment triplets, addressing limitations of pair extraction and aspect–opinion co-extraction. It presents OTE-MTL and reports evaluation on four SemEval benchmarks.
1 Introduction
- Aspect-sentiment pairs omit opinion terms, while aspect-opinion co-extraction omits sentiment dependencies needed to form meaningful pairs.The introduction also notes that unified aspect-sentiment tags inadequately model aspect-sentiment interaction and overlook triplet interactions in overlapping cases.Source: §1 · pp. 1–2
- Opinion triplet extraction integrates aspect-sentiment pair extraction with aspect-opinion co-extraction, adding opinions as sentiment references and dependencies that can support interpretability.Figure 1 contrasts the extraction patterns; the proposed perspective treats the approaches as complementary.Source: §1 · p. 1, Figure 1 · p. 1
- Peng et al. (2019) first extracts unified aspect-sentiment and opinion tags, then pairs them with a classifier, but this formulation leaves interaction modeling limited.The introduction identifies unified tagging and disregard for opinion interactions as issues; overlapping triplets may share an aspect or opinion and tend to share sentiment.Source: §1 · p. 2
- OTE-MTL jointly detects aspects, opinions, and sentiment dependencies, using separate sequence-tagging heads and a biaffine scorer for word-level dependencies.The introduction describes joint training as regularizing the dependency parser with aspect- and opinion-aware constraints.Source: §1 · p. 2
- OTE-MTL consistently outperforms state-of-the-art baselines on all evaluated datasets, and a case study shows better handling of overlapping cases.The evaluation covers four SemEval ABSA benchmark collections.Source: §1 · p. 2
2 Proposed Framework
OTE-MTL predicts aspect and opinion spans alongside word-level sentiment dependencies, then decodes those outputs into triplets. Its trainable prediction stage combines separate tagging tasks with dependency parsing.
2.1 Problem Formulation
- Each output triplet contains an aspect span, an opinion span, and a sentiment, and the model maps each sentence to a set of such triplets.Spans are represented by their start and end positions.Source: §2.1 · pp. 2–3
2.2 The OTE-MTL Framework
- OTE-MTL uses a trainable prediction stage for aspect tagging, opinion tagging, and word-level sentiment dependency parsing, followed by heuristic triplet decoding.The prediction stage builds on sentence encoding, and the decoder uses the detected aspects, opinions, and dependencies.Source: §2.2 · p. 3
2.3 Sentence Encoding
- A bidirectional LSTM encodes pretrained word embeddings into contextualized word representations by concatenating forward and backward states.The passages specify GloVe embeddings and explain that the two directional states are concatenated.Source: §2.3 · p. 3
2.4 Aspect and Opinion Representation
- Dimension-reducing nonlinear layers form aspect- and opinion-specific features, while separately parameterized representations are prepared for sentiment parsing.The transformations are intended to strip irrelevant information and reduce overfitting risk in subsequent computations.Source: §2.4 · pp. 3–4
2.5 Multi-task Architecture
- Separate B-I-O taggers identify aspects and opinions; dependency parsing classifies word pairs, focusing on the final words of aspect and opinion spans.The dependency labels are NEU, NEG, POS, and NO-DEP; restricting parsing to span-final word pairs reduces learning redundancy.Source: §2.5 · p. 4
- A biaffine scorer assigns word-pair dependency scores, and the framework jointly trains tagging and parsing with an L2-regularized objective.The overall objective balances tagging and dependency parsing with a trade-off term.Source: §2.5 · pp. 4–5
2.6 Triplet Decoding
- For decoding, sentiment dependencies serve as pivots for reverse-order traversal of aspect and opinion tags to assemble triplets.In the example, dependency (6, 1, POS) yields aspect span (4, 6), opinion span (1, 1), and positive sentiment.Source: §2.6 · p. 5
3 Experimental Setup
The experiments use four SemEval datasets with opinion annotations, evaluate exact aspect–opinion–sentiment triplets, and compare pipeline baselines and OTE-MTL variants.
3.1 Datasets and Evaluation Metrics
- The evaluation covers REST14, REST15, REST16, and LAPTOP14, adding opinion annotations from Wang et al. (2017) and Peng et al. (2019) to datasets originally annotated for aspects and sentiments.Overlapping cases are retained and account for 24.2% of triplets on average across the four datasets.Source: §3.1 · p. 5
- A triplet counts as a true positive only when its aspect, opinion, and sentiment all exactly match the gold standard.Results are averaged over 10 random-initialization runs, and paired t-tests assess significance.Source: §3.1 · p. 5
3.2 Implementation Details
- Training uses patience of five epochs to stop early when validation performance no longer improves.The supplied setup also specifies a batch size of 32.Source: §3.2 · pp. 5–6
3.3 Baselines and Variants
- The pipeline baselines include systems adapted from aspect-sentiment pair extraction and aspect-opinion co-extraction, alongside Pipeline.CMLA-MTL and HAST-MTL extend co-extraction systems into multi-task learning.Source: §3.3 · p. 6
- OTE-MTL variants test tag-embedding input, a concatenation-based scorer, unified aspect-sentiment tags, and a collapsed tag set that prevents aspect–opinion overlap.The unified variant uses a binary pair classifier, while the collapsed tag set combines aspect and opinion tagging in one module.Source: §3.3 · p. 6
4 Results and Analysis
OTE-MTL outperforms the state-of-the-art baselines across datasets with and without OOTs. Variant comparisons and case studies examine its components and performance on overlapping triplets, while error analysis identifies persistent false positives and false negatives.
4.1 Quantitative Evaluation
- OTE-MTL consistently outperforms all state-of-the-art baselines on every dataset, both with and without OOTs.Table 2 reports quantitative evaluation results as percentages.Source: §4.1 · p. 6, Table 2 · p. 7
- Results are generally better on datasets without OOTs, except LAPTOP14; OTE-MTL also surpasses its variants, slightly exceeding OTE-MTL-Inter and beating other variants by larger margins.The authors suggest that datasets without OOTs are easier and that triplet interaction modelling can partly address overlap challenges.Source: §4.1 · p. 6
- Variant analysis attributes poorer performance to approaches that concatenate representations without modelling priors, couple aspect and sentiment extraction, or collapse aspect and opinion representations.The authors suggest that collapsing both representations into one space may limit expressive capacity.Source: §4.1 · pp. 6–7
4.2 Qualitative Evaluation
- In the overlapping second case, OTE-MTL-Unified performs less well; in a long-range case with both aspect and opinion overlap, OTE-MTL does better but still misses correct predictions.Both models work well on the first, non-overlapping case.Source: §4.2 · p. 7, Table 3 · p. 8
- False positives are often triggered by one incorrect triplet element, while overlapping cases account for almost half of false negatives.The authors suggest more robust span detection and note that exact match may penalize harmless minor span discrepancies.Source: §4.2 · p. 7, Figure 4 · p. 8
5 Related Work
Related ABSA work either extracts aspect-sentiment pairs or co-extracts aspects and opinions, while other formulations assume an aspect is already available. This paper casts the first two approaches as triplet extraction and relates its joint subtasks to table-filling methods for triplet extraction tasks.
5.1 Aspect-based Sentiment Analysis
- ABSA approaches include extracting aspect-sentiment pairs, co-extracting aspects and opinions, and predicting sentiment for a given aspect, which the authors call inflexible for practical use.The paper unifies pair extraction and aspect-opinion co-extraction as a triplet extraction problem.Source: §5.1 · p. 8
5.2 Triplet Extraction-based Task
- For triplet extraction, the paper draws on table-filling methods for joint entity and relation extraction and decomposes the task into three subtasks solved jointly with multi-task learning.Word-level sentiment dependency parsing is described as a table-filling problem.Source: §5.2 · p. 8
6 Conclusions and Future Work
The paper presents opinion triplet extraction for ABSA and a multi-task framework addressing limitations of existing approaches. Experiments support the framework’s effectiveness, while future work targets more robust taggers, flexible evaluation, and richer triplet interaction.
6 Conclusions and Future Work
- The framework is reported effective against a wide range of strong baselines, and variant comparisons signify the necessity of its core components.Source: §6 · p. 8
- The proposed framework uses joint training, decoupled aspect and sentiment prediction, and regularization among correlated tasks to address limitations of existing approaches.Source: §6 · p. 8
- The authors identify more robust aspect and opinion taggers, a more flexible triplet evaluation metric, and a stronger triplet interaction mechanism as future research directions.Source: §6 · pp. 8–9