SYSTEMATIZATION AND COMPARATIVE REVIEW OF TOPIC MODELING METRICS

This paper presents a systematic and comparative review of metrics used to assess the quality of topic models in natural language processing (NLP) tasks. Both traditional metrics, such as perplexity and topic coherence, and modern approaches, including word embedding tags (WETC) and graph methods, are considered. The advantages and limitations of each metric, as well as their applicability in various practical scenarios, are analyzed. Particular attention is paid to the criteria for choosing metrics depending on the modeling goal and the features of the text corpus. The work can be useful for researchers and developers seeking to improve the interpretability and accuracy of topic models.