Character n-gram-based word embeddings for morphological analysis of texts

The paper presents modifications of fastText word embedding model based solely on n-grams, for morphological analysis of texts. fastText is a library for classifying texts and teaching vector representations. The representation of each word is calculated as the sum of its individual vector and the vectors of its symbolic n-grams. fastText stores and uses a separate vector for the whole word, but in extra-vocabular cases there is no such vector, which leads to a deterioration in the quality of the resulting word vector. In addition, as a result of storing vectors for whole words, fastText models usually require a lot of memory for storage and processing. This becomes especially problematic for morphologically rich languages, given the large number of word forms. Unlike the original fastText model, the proposed modifications only pretrain and use vectors for the character n-grams of a word, eliminating the reliance on word-level vectors and at the same time helping to significantly reduce the number of parameters in the model. Two approaches are used to extract information from a word: internal character n-grams and suffixes. Proposed models are tested in the task of morphological analysis and lemmatization of the Russian language, using SynTagRus corpus, and demonstrate results comparable to the original fastText.