Source-linked AI summary
A Joint Training Dual-MRC Framework for Aspect Based Sentiment Analysis
Yue Mao, Yi Shen, Chao Yu, Longjun Cai
TL;DR
ABSA lacks a unified end-to-end solution covering aspect extraction, opinion extraction, and sentiment classification, especially for triple extraction. The paper addresses this gap with jointly trained dual-MRC models and reports results that outperform existing baselines and state-of-the-art methods across evaluated tasks. Its scope includes a conditional-independence assumption and a noted difficulty when sentiment polarity may correspond to multiple opinion terms.
Problem
Previous ABSA approaches did not solve all subtasks in a unified end-to-end framework, while triple extraction methods could require two stages and three separate models.
Method
The paper converts triple extraction into two MRC problems and jointly trains two BERT-MRC models with parameter sharing: one for AE and one for AOE and SC.
Results
The proposed framework outperforms all compared baselines across experiments covering multiple ABSA subtasks and benchmark datasets.
Takeaways & Limitations
A single jointly trained dual-MRC framework can handle all ABSA subtasks in one model.
Takeaways & Limitations
The analysis notes that sentiment polarity may point to multiple opinion terms, including terms not paired with the given aspect.
Abstract
from arXiv · showhide
Aspect based sentiment analysis (ABSA) involves three fundamental subtasks: aspect term extraction, opinion term extraction, and aspect-level sentiment classification. Early works only focused on solving one of these subtasks individually. Some recent work focused on solving a combination of two subtasks, e.g., extracting aspect terms along with sentiment polarities or extracting the aspect and opinion terms pair-wisely. More recently, the triple extraction task has been proposed, i.e., extracting the (aspect term, opinion term, sentiment polarity) triples from a sentence. However, previous approaches fail to solve all subtasks in a unified end-to-end framework. In this paper, we propose a complete solution for ABSA. We construct two machine reading comprehension (MRC) problems and solve all subtasks by joint training two BERT-MRC models with parameters sharing. We conduct experiments on these subtasks, and results on several benchmark datasets demonstrate the effectiveness of our proposed framework, which significantly outperforms existing state-of-the-art methods.
Introduction
ABSA comprises multiple subtasks, culminating in extraction of aspect–opinion–sentiment triples. The paper proposes a dual-MRC framework that jointly handles these subtasks in one model and reports strong benchmark performance.
- ABSA includes aspect term extraction, opinion term extraction, and aspect-level sentiment classification.
- Triple extraction targets aspect terms, their corresponding opinion terms, and sentiment polarities, but prior work used computationally inefficient multi-stage modeling.
- The framework converts triple extraction into two MRC problems: the left handles AE, while the right handles AOE and SC.
- The dual-MRC framework jointly trains three objectives and uses a single model to handle all ABSA subtasks.
- Experiments compare the framework across ABSA tasks and report performance that outperforms state-of-the-art methods.
Related Work
Prior ABSA research developed task-specific or partially joint methods, with recurring issues including pipeline error propagation, inefficient tagging, and unpaired extracted terms. Triple extraction introduced a broader target but retained a multi-stage design.
- SC: SC research models the relationship between an aspect and its context using architectures including CNNs, RNNs, attention networks, and memory networks.
- AE: AE identifies all aspect terms in a sentence and has been formulated as sequence labeling or sequence-to-sequence learning.
- AESC: AESC methods jointly extract aspects and sentiment, but pipeline systems propagate errors and unified tagging approaches can be inefficient or sentiment-inconsistent.
- OE: OE can assist AE or SC, but independently extracted aspect and opinion terms are not paired and therefore cannot provide an aspect's corresponding polarity cause.
- AOE and Pair: AOE extracts aspect–opinion pairs with aspects given in advance, whereas Pair extraction identifies span types and their relationships from scratch.
- Triple: Triple extraction combines aspect terms, corresponding opinion terms, and sentiment polarity, but the prior method used two stages and three separate models.
Proposed Framework
The framework converts triple extraction into two MRC problems and jointly trains BERT-based components to cover aspect extraction, opinion extraction, and sentiment classification. At inference, the left MRC supplies aspect candidates to repeated right-MRC queries that produce opinion terms and sentiment polarities.
- Joint Training: Triple extraction is decomposed into AE, AOE, and SC, with a left MRC handling AE and a right MRC handling AOE and SC.The joint likelihood combines the three subtasks with weights α, β, and γ.
- Dual-MRC Architecture: Both MRC parts use BERT to encode context; the left extracts aspect spans, while the right extracts opinion spans for a given aspect and classifies sentiment.The left and right components use separate start/end-position outputs, and the right component additionally produces sentiment polarity.
- Joint Training: The model minimizes a weighted joint loss combining aspect extraction, sentiment classification, and opinion-aspect extraction objectives.The loss is J = α · JAE + β · JSC + γ · JAOE, with α, β, and γ controlling objective contributions.
- Dataset Conversion: Original triple annotations are converted into two MRC inputs: one query returns all aspects, and aspect-specific queries return opinions and sentiment.The left MRC uses the sentence as context; the right MRC is constructed for each extracted or known aspect.
- Inference Process: Training uses ground-truth aspects to construct right-MRC inputs, whereas inference pipelines predicted aspects from the left MRC into right-MRC queries.The inference algorithm loops over aspect candidates and adds each returned opinion and polarity to the extracted triple set.
- Inference Process: The framework’s outputs support the individual AE, AOE, and SC tasks, while AESC and pair extraction use combinations of the MRC outputs.The paper directs readers to Table 1 for the task-specific output combinations.
Experiments
Experiments evaluate the dual-MRC framework across three ABSA research lines with different datasets, annotations, tasks, and baselines. The framework generally outperforms prior methods, with gains in sentiment classification, aspect–opinion matching, and pair/triple extraction, while joint learning supports multiple tasks in one model.
- Experimental Setup: Three research lines use distinct datasets, annotations, subtasks, baselines, and experimental settings, with matched settings maintained for each comparison.The datasets derive from Wang et al. (2017), Fan et al. (2019), and Peng et al. (2020).
- Main Results: On Wang et al. (2017), the framework reaches first or second place for AE, SC, and AESC, with especially strong gains in SC and AESC.Its AE results are close to SPAN-BERT, while MRC substantially improves SC and AESC relative to the comparison methods.
- Main Results: On Fan et al. (2019), most AOE results significantly outperform prior baselines, although 16res is slightly below LOTN.The strongest relative performance occurs on lap14, which the authors associate with domain differences between laptop and restaurant comments.
- Main Results: On Peng et al. (2020), the framework significantly outperforms baselines, especially in precision for pair extraction and in Pair and Triple tasks.The authors attribute the extraction gains to combining MRC with span-based extraction, which reduces the choices per token from 7^n to 4^n.
- Joint Learning Analysis: Joint learning changes F1 scores little overall but is more efficient and handles more tasks with one model.For AESC, jointly extracting opinion terms produces a small improvement in aspect-level sentiment classification.
- Case Study: In case studies, the MRC approach improves aspect–opinion matching, detects complete entity boundaries, and produces better sentiment classification in the reported examples.It avoids matching “tuna” with “too dry” and detects “log on” rather than only “log.”
Conclusions
The proposed joint-training dual-MRC framework handles all ABSA subtasks in one shot and outperforms all compared baselines.
- The framework jointly handles all ABSA subtasks in one shot using two MRC components.The left MRC extracts aspect terms; the right performs aspect-oriented opinion extraction and sentiment classification.
- The original dataset is converted and fed into the dual-MRC framework for joint training.
- Experiments across three research lines compare the framework with different ABSA subtasks and baselines.
- The proposed framework outperforms all compared baselines.