Using Semantic Relatedness Measures in Coreference Resolution for Russian

Abstract
The paper is devoted to the role of semantic information (in the form of semantic relatedness measures) in coreference resolution for the Russian language. It describes a series of experiments in calculating metrics of semantic relatedness based on Russian material and evaluating the possibility of using them in systems of natural language processing, as well as the performance of such systems. The goal of the first stage of experiments was to find out, which semantic relatedness measures better correspond to coreference relations between referential expressions. For this purpose, several metrics calculated from different parameters were chosen and evaluated on the test set, derived from the Russian coreference corpus RuCor. Semantic data for the metrics was obtained from two sources: Russian Wikipedia and RuThes thesaurus. The results showed that while RuThes provided more reliable data for common nouns, Wikipedia data correlated better with named entities. Based on the obtained results, metrics that corresponded to coreference relations the most were chosen to be implemented during the next stage of experiments. For the second stage of experiments a machine-learning based coreference resolution system that could use semantic relatedness measures as features was developed, based on the decision trees classification algorithm. Four versions of the system were tested: without any features derived from semantic information, with features derived from only one of the sources, and with features derived from both sources. Tests were performed on the subset of RuCor corpus that already included gold standard mark-up as the base for evaluation. The tests showed noticeable improvement for the version that was using semantic information from both data sources. The experiments made demonstrate the increase of quality of coreference resolution with the implementation of features based on semantic information. The results obtained are comparable to or exceed the ones described in similar papers on the topic of Russian coreference resolution.

Keywords
natural language processing, coreference resolution, semantic relatedness measures, machine learning, Russian language