Intent Classification based on Deep Learning Language Model in Turkish Dialog Systems

摘要

Dialog systems are information retrieval tools that utilize artificial intelligence algorithms to facilitate humanmachine interactions in conversational or written form, deployed in a number of applications, such as conversational search and virtual assistants. Intent Classification, being the first step in dialog systems, models the task of determining the topic of the search or the command to be executed by the virtual assistant as a classification task. In this study, we use an automatic translation tool to adapt the existing English intent classification datasets for Turkish. Using the adapted datasets that belong to various application domains, we fine-tune Turkish, English and multilingual variations of the deep-learning-based BERT model for intent detection, which achieves competitive performance in various Natural Language Processing tasks. We employ Support Vector Machine with bag-of-words modeling and TF-IDF term weighting as a baseline. Experiment results show that deep-learning-based models outperform bag-of-words model and achieve state-of-theart results for Turkish intent classification.