Term-Unigram Extractions Using Embedding-Based Filtering

With the accelerating growth of textual data volumes, there is a need to develop methods of automatic natural language processing. One of the key tasks in this direction is term extraction, which finds application in machine learning, information retrieval and knowledge graph construction.There are supervised and unsupervised methods for term extraction. Supervised approaches show high performance but require the availability of labeled data for training. Unsupervised methods, on the other hand, are universal and not tied to a specific subject domain but are generally inferior to supervised solutions. In addition, natural language processing tasks are prone to data noise due to the presence of common words that make it difficult to extract informative terms.This paper presents the UNI-TEX (Unigram-Term Extraction) method for automatic and unsupervised extraction of unigram terms. The aim of the research is to develop an unsupervised method for unigram term extraction and to evaluate its performance in different subject domains and multiple languages. The main steps of the methodology include unigram extraction using particle tags (POS tags), pre-filtering and semantic filtering based on stop words. In addition, the proposed approach partially addresses the problem of data noise.The UNI-TEX method was tested on the ACTER corpus including four domains and three languages. The extracted unigrams were compared with a reference list of single-word terms, and the proposed method showed an average F1-score of about 46% for English. This approach can be applied to identify informative and domain-specific words as well as to clean the text data from common lexemes.