Slot Filling for Voice Assistants

Abstract

Voice Assistant applications are increasing in popularity and getting deployment in industrial and daily life tasks. In Voice Assistant and Dialog Systems applications, the task of Slot Filling constitutes a major design problem. The voice assistant, tasked with executing a designated job, needs to extract the necessary parameters from the spoken query input by the user. In this study, we use Turkish and English datasets to attack Slot Filling problem with Machine Learning and Deep Learning algorithms. We perform Slot Filling with Conditional Random Fields and compare its performance with recently popularized BERT-like Transformer-based pre-trained language models. We make use of Precision, Recall and F1 Score metrics in model evaluation. Experimental results show that Slot Filling performance of Transformer-based pre-trained language models exceeds the performance of conventional Conditional Random Fields. Furthermore, we observe that multilingual and cross-lingual pretrained language models outperform the models that are pretrained only on the target language. It is expected that the deployed methods and obtained results would contribute to the Dialog Systems and Voice Assistant technologies.