T-Extractor: A Hybrid Unsupervised Approach for Term and Named Entity Extraction Using Rules, Statistical, and Semantic Methods

Abstract

Automatic term extraction is a key technology for optimizing natural language processing tasks such as machine translation, sentiment analysis, knowledge graph construction, and/or ontology population. This study presents the T-Extractor approach for unsupervised term extraction. The research goal is to develop an efficient method that does not require labeled data, and to analyze its applicability on scientific texts. T-Extractor combines rule-based, statistical, and semantic analysis, treating unigram and phrase extraction as two subtasks. Part-of-speech templates are used in the candidate selection phase, while a filter based on raw and rectified frequencies refines phrase boundaries. TopicScore is applied for final term filtering, improving extraction precision. Additionally, simple rules help identify abbreviations and named entities, improving recall. T-Extractor was tested on the ACTER (three languages, four domains) and ACL RD-TEC 2.0 datasets. In English, the best result was achieved in the equi domain, with an F1- measure of 48.5%, precision of 41.6%, and recall of 58.2%. On the ACTER dataset, the approach outperformed existing unsupervised methods and performed better than the supervised GPT-3.5-Turbo and BERT models in the corp and wind domains. Specifically, in the corp domain, T-Extractor's F1- measure approached that of the HAMLET model, lagging by 3.7%. In addition, the method showed results comparable to those of promtATE and TALN-LS2N.