Coherence Evaluation Method Based on Analyses of Phrases Graph

Abstract

Coherence evaluation falls into a category of natural language processing tasks. The state-of-the-art methods utilize different machine learning techniques to estimate the coherence value of texts. In this paper, the graph based method of coherence evaluation based on the extraction of phrases and coreference resolution has been suggested. State-of-the-art graph-based methods of coherence evaluation have been analyzed. The approach based on the analysis of the set of phrases has been suggested. The extraction of phrases from a text has been implemented according to the features of Ukrainian language with the usage of a syntactic parser. The usage of a bipartite digraph to estimate the lexical connectivity of sentences has been proposed. The effectiveness of the method on the Ukrainian-language corpus has been examined using document discrimination and insertion tasks. The metrics retrieved have been compared with the calculated metrics of other analytical and graph-based methods. The results obtained may indicate that the method suggested can be used to perform the coherence evaluation of Ukrainian texts. The method can be adapted to features of other languages by the changing of several additional components.