DeepMistake: Which Senses are Hard to Distinguish for a Word-in-Context Model
In this paper, we describe our solution of the Lexical Semantic Change Detection (LSCD) problem. It is based on a WordinContext (WiC) model detecting whether two occurrences of a particular word carry the same meaning. We propose and compare several WiC architectures and training schemes, and also different ways to convert WiC predictions into final word scores estimating the degree of semantic change. We participated in the RuShiftEval LSCD competition for the Russian language, where our model achieved 2nd best result during the competition. During postevaluation experiments we improved the WiC model and managed to outperform the best system. An important part of this paper is detailed error analysis where we study the discrepancies between WiC predictions and human annotations and their effect on the LSCD results.