Kramov A. Design and application of neural networks for the creation of the methods of the coherence estimation of Ukrainian-language texts

The paper is devoted to the investigation of the designing and applying of different neural networks for the creation of the methods of the coherence estimation of Ukrainian-language texts. The coherence estimation of a document falls into the category of natural language processing tasks. The generation of an output coherence estimation value involves the analysis of the thematic integrity of all text’s parts basing on the correspondence of the document’s content to the background knowledge of a reader and the different types of the consistency of text’s spans. The consideration of the listed features of a text during the estimation of its coherence leads to the applying of the correspondent coherence estimation methods for the solving of the tasks that belong to different areas (medical diagnostics, search engines, etc.). It should be mentioned that the analysis of the coherence evaluation of the documents of Ukrainian and other Slavic languages is still at the initial stage. It is advisable to perform the experimental verification of the effectiveness of state-of-the-art methods on Ukrainian-language corpora; moreover, an important task consists in the creation of new models for the detailed analysis of the formation process of the coherence estimation value of Ukrainian-language texts. The lack of the fixed text’s structure formed by a person according to its thoughts and the complexity of the unambiguous assessment of the coherence of a document (subjectivity of the reader's perception of the text based on their own impressions and background knowledge) complicate the evaluation of the mentioned metric of the thematic integrity of a text by utilizing a pre-defined set of instructions. The listed issues of the coherence estimation of a text cause the expediency of the applying of different machine learning techniques for the solving of assigned tasks. Taking into account the increase of devices’ computational power and the development of the parallelized and distributed computing, the state-of-the-art methods of the coherence evaluation of a text are based on the designing of the multilayer neural network of different architecture. The applying of deep learning models allows the achieving of accuracy increase according to an assigned task by means of the generalization of the different features of the texts of an input corpus. However, taking into account the complexity of the unambiguous definition of the criteria of a coherent document, besides the achievement of the desired accuracy of correspondent methods it is advisable to find out the reasons for the retrieving of an output result. An important task is to investigate the formation of the features of an input text and the connection types between its elements that are analyzed during the evaluation of the output coherence value of a Ukrainian-language document. In the paper, the analysis of current coherence estimation methods based on machine learning techniques has been performed. The principle of work of Entity Grid, Entity Graph, and an RST-based method that consists in the analysis of the regularity of syntactic and discourse roles change within adjacent sentences and a whole document at all has been considered in details. The key disadvantages of the usage of the mentioned methods are the neglecting of other text’s elements, the semantic properties of components; the dependency of the accuracy of the solving of assigned tasks on external instruments. The analysis of the methods of the local and global coherence estimation of a document based on the usage of different neural networks for the representation of text’s elements and the evaluation of its coherence at the level of the semantic consistency of sentences has been performed. The appropriateness of the utilizing of recurrent and convolutional layers for the designing of a neural network model for the estimation of both local and global coherence of a document has been shown. Basing on the performed comparative analysis of methods, the effectiveness of the applying of neural network models for the solving of the assigned tasks of the coherence estimation of a text has been demonstrated.