Hard Example Mining-Driven Label Alignment for LLMs in Aspect Sentiment Triplet Extraction

ACL ARR 2026 January Submission3201 Authors

Keywords: Aspect Sentiment Triplet Extraction, Large Language Models, Label Alignment, Hard Example Mining

Abstract: In recent years, large language models (LLMs) have achieved remarkable success across various natural language processing tasks. However, recent studies indicate that LLMs still underperform smaller supervised fine-tuned models in the fine-grained Aspect Sentiment Triplet Extraction (ASTE) task. To investigate the reasons behind this phenomenon, we conducted an in-depth analysis of cases where LLM predictions failed. Our findings reveal that the primary cause for this suboptimal performance lies in the inconsistency between the internal extraction behaviors of LLMs and task-specific annotation standards. Motivated by this insight, we introduce HEMLA, a Hard Example Mining-driven Label Alignment framework. Specifically, we use LLM-generated responses as prompts to train a lightweight alignment model, while dynamically deciding whether to include each response in the training process based on a hard example mining strategy. Extensive experimental results demonstrate that our method consistently outperforms existing state-of-the-art approaches and offers a new paradigm for adapting LLMs to downstream tasks without fine-tuning the underlying LLMs.

Paper Type: Long

Research Area: Information Extraction and Retrieval

Research Area Keywords: named entity recognition and relation extraction, safety and alignment

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 3201