Tonal spaces of vector language models

Objectives. Tonality as a positive-negative mood is a key parameter of text analysis and generation algorithms. Existing machine learning methods encode tonality in computationally extensive and uninterpretable ways which hampers the development of the corresponding applications. The work aims to solve this problem for Russian language. Methods. Pre-trained vector language models are used to encode words as vectors in multidimensional spaces. In these spaces, the tonality corresponds to a specific direction which optimally discriminates the positive and negative prototypes. The tonality of word is then determined by its projection onto this direction. Adding a tonal vector to a key word defines a one-dimensional subspace, containing its positive and negative associations. Results. The algorithm is tested on GloVe and FastText machine language models, encoding individual Russian words and morphemes with vectors in 300-dimensional space. Commonly used verbs and nouns served as key words. The average reliability of the found tonal associations estimates as 80 %. Conclusion. The results indicate the applicability of pre-trained vector language models for fast and interpretable working with tonal information. The developed approach is applicable for the tasks of aspect-based sentiment analysis, as well as for the machine generation of object-oriented texts with a required tonality. Generalization of the tonal axis to the triple of Osgood's semantic factors allows expanding the method to a full range of affectively-semantic information.