MODERN APPROACHES TO MULTICLASS INTENT CLASSIFICATION BASED ON PRE-TRAINED TRANSFORMERS

Subject of Research. The paper considers modern approaches to the multiclass intention classification problem. The user intention is the incoming user requests when interacting with voice assistants and chatbots. The algorithm is meant for determination what class the call belongs to. Modern technologies such as transfer learning and transformers can improve significantly the multiclass classification results. Method. This study uses a comparative model analysis technique. In turn, each model is inlined into a common pipeline for data preparing and clearing, and the model training but with regard to its specific requirements. The following models applied in real projects have been selected for comparison: Logistic Regression + TF-IDF, Logistic Regression + FastText, LSTM + FastText, Conv1D + FastText, BERT, and XLM. The sequence of models corresponds to their historical origin, but in practice these models are used without regard to the time period of their creation but depending on the effectiveness of the problem being solved. Main Results. The effectiveness of the multiclass classification models on real data is studied. Comparison results of modern practical approaches are described. In particular, XLM confirms the superiority of transformers over other approaches. An assumption is made considering the reason why the transformers show such a gap. The advantages and disadvantages of modern approaches are described. Practical Relevance. From a practical point of view, the results of this study can be used for projects that require automatic classification of intentions, as part of a complex system (voice assistant, chatbot or other system), as well as an independent system. The pipeline designed during the study can be applied for comparison and selection of the most effective model for specific data sets, both in scientific research and production.