Source-linked AI summary
Open-Domain Targeted Sentiment Analysis via Span-Based Extraction and Classification
Minghao Hu, Yuxing Peng, Zhen Huang, Dongsheng Li, Yiwei Lv
TL;DR
Open-domain targeted sentiment analysis must identify opinion targets and their polarities, while sequence tagging suffers from a large search space and sentiment inconsistency. The paper proposes a span-based extract-then-classify framework with pipeline, joint, and collapsed variants; it consistently outperforms sequence-tagging baselines, and the pipeline variant performs best.
Problem
Targeted sentiment analysis requires detecting opinion targets and their polarities, but sequence tagging introduces a huge search space and can produce inconsistent word-level sentiment predictions.
Method
The framework uses a BERT encoder, extracts multiple candidate targets from start and end span positions, and classifies each candidate using its summarized span representation.
Results
The span-based methods consistently outperform sequence-tagging baselines on three benchmark datasets, with SPAN-pipeline performing best among the span-based variants.
Takeaways & Limitations
Span-level polarity classification provides the main performance improvement, while the extractor is more suitable for long sentences and the pipeline model surpasses joint and collapsed models.
Abstract
from arXiv · showhide
Open-domain targeted sentiment analysis aims to detect opinion targets along with their sentiment polarities from a sentence. Prior work typically formulates this task as a sequence tagging problem. However, such formulation suffers from problems such as huge search space and sentiment inconsistency. To address these problems, we propose a span-based extract-then-classify framework, where multiple opinion targets are directly extracted from the sentence under the supervision of target span boundaries, and corresponding polarities are then classified using their span representations. We further investigate three approaches under this framework, namely the pipeline, joint, and collapsed models. Experiments on three benchmark datasets show that our approach consistently outperforms the sequence tagging baseline. Moreover, we find that the pipeline model achieves the best performance compared with the other two models.
1 Introduction
Open-domain targeted sentiment analysis identifies opinion targets and their sentiment polarities, but sequence tagging can create a large search space and inconsistent word-level polarity predictions. The proposed span-based extract-then-classify framework addresses these issues and outperforms sequence-tagging baselines, with the pipeline variant performing best among the compared span-based models.
- 1 Introduction: The task detects opinion targets and predicts the sentiment polarity associated with each target.Unlike sentence-level sentiment analysis, it must identify entities such as “Windows 7” and “Vista” before assigning sentiment classes.
- 1 Introduction: Prior approaches formulate target extraction and polarity prediction as sequence tagging with target and polarity tags.Joint variants use separate target and polarity tags or collapsed tags that directly encode targeted sentiment boundaries.
- 1 Introduction: Sequence tagging faces a huge extractive search space and may assign different polarities to words within the same multi-word target.Word-level polarity tagging also ignores the semantics of the entire opinion target, so consistency is not guaranteed.
- 1 Introduction: The proposed span-based scheme extracts multiple targets from their span boundaries and classifies polarity using summarized span representations.Its search space grows linearly with sentence length, while span-level classification considers all target words before prediction.
- 1 Introduction: Experiments on three benchmark datasets show consistent gains over sequence-tagging baselines, with the pipeline model best among pipeline, joint, and collapsed variants.The study compares tagging-based and span-based models and evaluates all three span-based approaches.
2 Related Work
Related work treats targeted sentiment analysis as target extraction followed by polarity classification, using sequence tagging and neural models. This paper instead frames target identification as span-level extract-then-classify, drawing on span extraction in machine comprehension and question answering.
- 2 Related Work: Targeted sentiment analysis requires detecting entities and their associated sentiments in open-domain text.The task is distinguished from sentence-level sentiment analysis by its entity-specific sentiment requirement.
- 2 Related Work: Traditional target extraction methods widely use Conditional Random Fields, while later work applies CNNs and RNNs.These approaches primarily address identifying entities from the input sentence.
- 2 Related Work: The proposed approach differs by formulating target identification as a span-level extract-then-classify process rather than word detection.Its span-based labeling scheme is related to extractive question answering, where continuous text spans are selected as answers.
3 Extract-then-Classify Framework
The framework extracts multiple opinion-target spans using boundary prediction and then classifies each target’s polarity from a summarized span representation. It also compares pipeline, joint, and collapsed variants, using heuristic decoding to reduce redundant multi-span predictions.
- Framework Overview: The framework uses BERT contextual representations, a multi-target extractor, and a polarity classifier under an extract-then-classify design.The extractor proposes candidate targets, while the classifier predicts sentiment for each extracted span.
- Multi-Target Extractor: Target candidates are generated by predicting span start and end positions rather than assigning sequence tags to each token.The extractor obtains start and end scores and probability distributions from contextual token representations.
- Multi-Target Extractor: Heuristic multi-span decoding selects high-scoring valid spans, penalizes span length, and removes overlapping candidates until the output limit is reached.Candidates must satisfy end-position and score-threshold constraints; the procedure addresses redundant spans among top-K predictions.
- Polarity Classifier: The polarity classifier summarizes contextual vectors over each target span with attention and predicts its sentiment probability using feed-forward transformations.During inference, the highest-probability sentiment class is selected for each candidate span.
- Model Variants: The pipeline, joint, and collapsed models differ in how extraction and polarity prediction share representations, objectives, and label spaces.The pipeline trains separate models, the joint model shares a BERT backbone, and the collapsed model predicts span boundaries separately for each sentiment class.
4 Experiments
Experiments on three benchmark datasets compare span-based and sequence-tagging approaches for targeted sentiment analysis. Span-based models achieve stronger overall results, with the pipeline variant performing best, while analyses identify sentence length and target length as important factors.
- 4.1 Setup: Experiments use LAPTOP, REST, and TWITTER, evaluating exact target-and-polarity matches with P, R, F1, and polarity ACC.BERTLARGE is used for both TAG and SPAN models to make the comparison fair.
- 4.3 Main Results: SPAN methods outperform TAG baselines, with the best span-based method gaining 1.55%, 0.94%, and 3.43% absolute on the three datasets.The TAG baselines already outperform the previous best UNIFIED approach.
- 4.3 Main Results: SPAN-pipeline performs best among span-based models, whereas SPAN-collapsed drops severely across all datasets.The results suggest only a weak connection between target extraction and polarity classification, and that merging polarity labels into target spans is ineffective.
- 4.4 Analysis on Target Extraction: For target extraction, BIO tagging outperforms the extractor on LAPTOP and REST, where short sentences limit the tagger’s search space.In REST, 98% of sentences contain fewer than 40 words.
- 4.4 Analysis on Target Extraction: When sentence length exceeds 40, the extractor surpasses the tagger by 16.1 F1 on LAPTOP and 1.0 F1 on REST.The extractor is more robust for long sentences because its search space increases linearly with sentence length.
- 4.5 Analysis on Polarity Classification: Span-level polarity classification gains 9.97%, 8.15%, and 15.4% over the CRF-based tagger across the three datasets.Using the entire target span representation also makes accuracy decline less as target length increases, alleviating sentiment inconsistency.
5 Conclusion
The paper proposes a span-based extract-then-classify framework for targeted sentiment analysis and reports consistent advantages over sequence tagging and alternative span-based designs.
- The framework uses a multi-target extractor to propose candidate targets and a span-level polarity classifier based on summarized span representations.This design replaces word-level sentiment tagging with target-span extraction followed by polarity prediction.
- The proposed approach firmly outperforms sequence tagging baselines and previous state-of-the-art methods on three benchmark datasets.
- The pipeline model consistently surpasses both the joint and collapsed models under the span-based framework.