Nilufar Abdurakhmonova

The article describes a new model of a universal linguistic knowledge graph for Turkic languages as a basis for creating ontolinguistic resources. In the model, world knowledge is formalized as an ontological core that integrates taxonomic and situational ontologies (by analogy with WordNet and FrameNet) and is closely interrelated with linguistic data from Turkic languages. The relevance of the work is determined by the low-resource nature of most Turkic languages and the loss of effectiveness of neural network approaches in processing agglutinative languages with rich morphology. The goal of the project is to develop a new model of a universal linguistic knowledge graph for Turkic languages. The objectives are: integrating world knowledge and deep linguistic information at all levels of the language system; constructing the graph according to the principle of a semiotic vertical; ensuring the morphemocentricity of the model; and integrating it with international language standards. The foundational method is a pragmatically-oriented approach that focuses the research domain on the structural and functional features of Turkic languages. The knowledge graph is constructed according to the principle of a semiotic vertical that unites five levels – phonological, morphological, syntactic, semantic, and pragmatic – as well as three inter-level bridges (morphonological, morphosyntactic, and morphosemantic). A key feature of the model is its morphemocentricity: the morpheme acts as the basic linguistic unit, connected with units of all language levels of the model, which fundamentally distinguishes this approach from word-centric models for fusional languages. The model is integrated with international standards: IPA, Universal Dependencies, UDS, UMR, FrameNet, WordNet, and DiAML, each of which is used to describe linguistic elements and relations at its own level. The model has already found practical application in several software products, such as the linguistic portal “Turkic Morpheme” and the new version of the Tatar corpus “Tugan Tel”. Prospects include improving the interpretability of LLMs, improving the quality of machine translation, and unifying annotations of electronic corpora.

This paper addresses the problem of classifying taxpayer requests (complaints and suggestions) received by regulatory authorities at two levels: (1) determining the text type (complaint/suggestion) and (2) assigning the complaint to one of five pre-defined categories (Tax Calculation Error, IT Problems, Benefits/Deductions, Procedural Issues, Other). The study proposed two approaches: a rule-oriented (dictionary) method and a combination of machine learning algorithms. The rule-oriented technique is based on pre-compiled dictionaries of keywords and phrases, allowing for quick identification of requests based on a set of “trigger” lexemes. However, testing on real data showed that this technique in some cases (especially with ambiguous text) demonstrates reduced accuracy and completeness. To improve the results, a system was developed in which Naïve Bayes solves the problem of binary classification “offer/complaint”, and the support vector machine classifies any text (complaint or offer) into one of five topics. Training was carried out on a corpus of 2000 texts marked by experts. The testing results confirmed that machine-learning algorithms are superior to dictionary methods: Naïve Bayes gives F1 up to 88-90% when determining “offer/complaint”, and Support vector machine distributes complaints into categories with an accuracy of 90-92%. The obtained results and comparative analysis show that the combination of Naïve Bayes and Support Vector Machine allows achieving high efficiency in processing various requests, while maintaining clear operating logic and interpretability at the binary classifier stage.

In this article, the authors developed a two-stage method for analyzing user comments from social networks. At the first stage, the authors cluster comments by hashtags using K-Means, where they obtained 18 clusters at the output. From these classers, eight supercategories of comments were formed, which were used to train the classification model. The multilingual neural network model was used as a model, which was further trained on the formed corpus mentioned above. Experiments have shown that using hashtags as “weak” labels can significantly reduce the labor costs of manual labeling and at the same time achieve fairly high classification rates. On a separate test sample of 2000 comments, Precision and Recall reached average values of 0.81 and 0.78, respectively, with a final F1-measure of 0.79. The article also analyzes typical errors (false positive and false negative) related to the intersection of topics, polysemy of terms, and the specifics of the informal language of social networks. The proposed approach can be useful for automatic monitoring of public opinion, identifying trends, and solving other applied problems of content analysis.

In this study, the authors presented an algorithm for automatic sentiment analysis in school essays written in the Uzbek language. The algorithm is implemented on the basis of a on a convolutional neural network architecture, designed to classify text using TensorFlow and Keras. Authors created a training dataset consisting of almost 5000 sentences and phrases, most commonly used in everyday communication. The text data underwent preprocessing, including punctuation removal and conversion to lowercase, before being transformed into numerical representations using an embedding layer that was trained simultaneously with the model. Besides, the authors tested the effectiveness of the model, where the evaluation was carried out using such metric as precision. As a result of testing, the precision reached 88 for sentiment analysis in 50 essays, which consist of 811 sentences overall. Moreover, the authors conducted a comparative analysis of existing works and proposed further options for the development of the algorithm.

Allusions play a significant role in literary and cultural works, serving as a tool for conveying deep meanings embedded by authors. In the Uzbek language, as in many other languages, allusions are commonly used to create more complex narratives, which allows us to connect the work with broader cultural or historical traditions. However, identifying these allusions, and even more so interpreting them, presents significant challenges for readers. In recent years, it might be seen the active development of algorithms and different kind of solutions for various text analysis tasks. Among these tasks, identifying metaphors, idioms, and allusions can be singled out as one of the most difficult tasks, due to the need for a deeper understanding of the text. Moreover, interpreting allusions requires interdisciplinary knowledge in the field of literature, cultural studies, linguistics, and more. This article discusses the development of an algorithm for the automatic detection of allusions in the Uzbek language, aimed at helping researchers and readers fully understand the meaning conveyed by authors. In addition, the proposed solution can help linguists and those who study the Uzbek language. The algorithm uses two dictionaries, where first consist of 10,000 tagged sentences, while second contains over 80 thousand word forms. At the same time, the algorithm was tested and as a result of a number of experiments, the efficiency of the work was revealed, where the accuracy of the algorithm reached range from 82,6% to 100%.

This study presents the development of a tool for identifying named entities in Uzbek legal texts. It should be noted, that besides of detecting named entities, the authors developed an algorithm, which is able to standardize word forms by replacing the detected dialect words (Karluk, Kypchak and Oghuz) with their formal forms. This will help to fix popular grammatical mistakes among native speakers from different regions of the Uzbekistan. The proposed hybrid approach combines the traditional approach, which is used in the preprocessing (standardization of word forms), where a dictionary with more than 10 thousand marked words is actively used. At the same time, a custom language model is used to work with detecting named entities, which was trained on 2000 legal sentences. The testing results showed quite high indicators, in particular, the language model detected named entities with an accuracy of 90%, and the recall reached 94%. Moreover, the algorithm used to standardize dialect word forms showed even higher rates, ranging from 90% to 100 ^% depending on the dialect.