Contribution of Semantic Information to Coreference Resolution

Main goals of this work was to 1) evaluate the degree of influence of semantic information not contained in the original text on the quality of performance of automated coreference analyzers, and 2) to offer a way to integrate the information of such kind into existing algorithms that don’t originally take it into account. The first goal was achieved by conducting an experiment in comparing the performance quality of an algorithm for resolving anaphoric and coreferential relations with and without taking semantic information into account. The necessity of modifying the algorithm allowed to achieve the second goal as well. To achieve the aforementioned goals, several midline tasks had to be completed. Firstly, a theoretical approach, according to which the experiment needed to be conducted, was defined. Then a corpus of texts was gathered and marked-up according to the approach chosen that became the base for further method quality evaluation and modification of features used. Finally, because no available systems for coreference analysis of Russian exist, as it became clear during the research, the necessity arose to develop a system for this task and integrate into it the acquired information. The results we achieved allow to say with confidence that semantically oriented systems of coreferential analysis have an advantage in performance quality. It also became clear that including semantic information in the set of features used for analysis allows for an increase in performance even for a not so sophisticated algorithm.