Lexical Complexity Evaluation based on Context for Russian Language

Abstract

The task of identifying complex words within a context usually referred to as Complex Word Identification (CWI) or Lexical Complexity Prediction (LCP), is a vital component in Lexical Simplification pipelines. Correctness of complexity estimation depends on presented features, i.e. hand-crafted features, word embeddings, and presence of surrounding context, as well as on exploited rules or models, i.e. manually designed filtering, classic machine learning models, recurrent neural networks, and Transformer-based models. To our knowledge, the majority of existing works in CWI and LCP areas are devoted to investigating properties of English words and texts, accompanied by studies of German, Spanish, French and Hindu languages with little to no attention to Russian. In this paper, we present a study on lexical complexity estimation for the Russian language, by investigating the following topics: how well do morphological, semantic, and syntactic properties of a word represent its complexity; does a surrounding context significantly affect the accuracy of complexity estimation. We provide a brief description of the dataset of lexical complexity in context based on the Russian Synodal Bible and expand it by presenting a dataset of morphological, semantic, and syntactic features for annotated words. Additionally, we present linear regression and RuBERT models as baselines for lexical complexity estimation respectively.