Identifying the intentions of a user communicating with a bot

The challenge of the classification of the intents of a user during the interaction with a chat-bot has been considered. The comparative analysis of the existing intent classification dataset and corresponding methods based on machine learning and deep learning techniques has been performed. Moreove...

Full description

Saved in:
Bibliographic Details
Date:2023
Main Authors: Теленик, С. Ф., Погорілий, С. Д., Крамов, А. А., Вовк, Є. А.
Format: Article
Language:Ukrainian
Published: Інститут проблем реєстрації інформації НАН України 2023
Subjects:
Online Access:http://drsp.ipri.kiev.ua/article/view/300436
Tags: Add Tag
No Tags, Be the first to tag this record!
Journal Title:Data Recording, Storage & Processing

Institution

Data Recording, Storage & Processing
Description
Summary:The challenge of the classification of the intents of a user during the interaction with a chat-bot has been considered. The comparative analysis of the existing intent classification dataset and corresponding methods based on machine learning and deep learning techniques has been performed. Moreover, the task of the detection of intents with the performing of the zero-shot classification using NLI models has been defined. Different intent classification approaches, NLI models, and hypothesis templates for the detection of a user’s intent have been suggested. The experimental verification of the effectiveness of the suggested approaches and models for the zero-shot classification on the corpora of different domains has been performed. In addition, the analysis of the results of the mentioned experimental configurations and existing intent classification methods based on pre-trained and large language models (GPT-3.5) has been implemented. The results obtained may indicate the advisability of the usage of different NLI models and their hypothesis templates in order to increase the accuracy of the zero-shot intent classification on the corpora of different domains. However, the highest accuracy metrics that were obtained by the NLI models don’t exceed the corresponding values of pre-trained models and GPT-3.5 which allows drawing a conclusion about the advisability of the conducting of further research devoted to the increase of the zero-shot intent classification with the usage of NLI models or other models within the natural language processing area. Tabl.: 4. Fig.: 1. Refs: 50 titles.