CACKT: Sentiment-Conditioned Boundary Learning and Candidate Transfer for Targeted Sentiment Analysis
Abstract This paper focuses on Targeted Sentiment Analysis (TSA), aiming to detect specific aspects from text and predict their corresponding sentiment polarities. Precise aspect boundary detection is pivotal to TSA, as erroneous candidate spans inevitably compromise downstream sentiment classification. Nevertheless, prevalent span-based approaches largely overlook sentiment signals during boundary generation, yielding boundary drift and aspect--sentiment misalignment. To alleviate this problem, we present CACKT, a candidate-optimized TSA framework. CACKT incorporates a Sentiment-Conditioned Boundary Extraction (SCBE) module that learns prototypical representations of sentiment polarities, allowing sentiment features to proactively steer the identification of span start and end positions. Building upon SCBE, a confidence-aware candidate transfer mechanism dynamically weights candidates by their reliability, suppressing noise from low-confidence spans. Furthermore, boundary-aware hard negative contrastive learning constructs informative negatives---including boundary shifts, competing aspects, and mismatched sentiments---to sharpen the model's discrimination against plausible errors. Extensive experiments show that CACKT refines candidate boundaries and delivers superior overall TSA performance.
Authors
- Junfeng Shen (ORCID: https://orcid.org/0000-0003-0049-6563)
- Jinghao Yan (ORCID: https://orcid.org/0009-0001-0199-8525)
- Yutong Liu
Institutions
- Hubei University (CN)
- Zhejiang University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11289899/v1
- Primary Topic
- Sentiment Analysis and Opinion Mining
- Type
- preprint