Extract Then Adjust: A Two-Stage Approach for Automatic Term Extraction

Automatic Term Extraction (ATE) is a fundamental natural language processing task that extracts relevant terms from domain-specific texts. Existing transformer-based approaches have indeed achieved impressive improvement. However, we observe that even state-of-the-art (SOTA) extractors suffer from boundary errors, which are distinguished by incorrect start or end positions of a candidate term. The minor differences between candidate terms and ground-truth leads to a noticeable performance decline. To alleviate the boundary errors, we propose a two-stage extraction approach. First, we design a span-based extractor to provide high-quality candidate terms. Subsequently, we adjust the boundaries of these candidate terms to enhance performance. Experiment results show that our approach effectively identifies and corrects boundary errors in candidate terms, thereby exceeding the performance of previous state-of-the-art models.