Methodological aspects of semantic relationship extraction for automatic thesaurus generation

The paper is devoted to analysis of methods for automatic generation of a specialized thesaurus. The main algorithm of generation consists of three stages: selection and preprocessing of a text corpus, recognition of thesaurus terms, and extraction of relations among terms. Our work is focused on exploring methods for semantic relation extraction. We developed a test bench that allow to test well-known algorithms for extraction of synonyms and hypernyms. These algorithms are based on different relation extraction techniques: lexico-syntactic patterns, morpho-syntactic rules, measurement of term information quantity, general-purpose thesaurus WordNet, and Levenstein distance. For analysis of the result thesaurus we proposed a complex assessment that includes the following metrics: precision of extracted terms, precision and recall of hierarchical and synonym relations, and characteristics of the thesaurus graph (the number of extracted terms and semantic relationships of different types, the number of connected components, and the number of vertices in the largest component). The proposed set of metrics allows to evaluate the quality of the thesaurus as a whole, reveal some drawbacks of standard relation extraction methods, and create more efficient hybrid methods that can generate thesauri with better characteristics than thesauri generated by using separate methods. In order to illustrate this fact, one of such hybrid methods is considered in the paper. It combines the best standard algorithms for hypernym and synonym extraction and generates a specialized medical thesaurus. The hybrid method leaves the thesaurus quality on the same level and finds more relations between terms than well-known algorithms.