Goal-oriented spoken dialogue systems based on natural language understanding

Abstract

In the post-pandemic era, the online education market embraces a boom with productive investment and increasing opportunities. This project aims at developing a speech dialogue system based on an NLU model that can understand task-oriented natural language texts to provide education-related services. It is believed that the system will assist in reducing the communication costs between students and teachers in the online teaching environment, ease the shift of students from physical to virtual classrooms, and alleviate the pressure on the teaching staff.

The past few years have witnessed a significant breakthrough in Natural Language Processing with advanced Deep Neural Network techniques; therefore, this project explores neural-network-based natural language processing methods for information extraction and task identification. A novel SOTA joint training approach for intent classification and slot filling is adopted in the development of the model. In order to incisively understand the model, great efforts are paid to experiments for different components of this model, including the alternation of BERT variants, the adoption of CRF, the optimal value of hyperparameters such as the number of training epochs, and the value of slot loss coefficient. A custom intent and slot filling dataset with goal-oriented corpus in the specific application scenarios of online teaching platforms are generated for model training due to the lack of such ready-made education-focused data resources.

The implementation of the front-end application is designed to cross various platforms with the support of Flutter, from web pages to mobile devices, which is innovative compared with most of the on-shelf virtual assistants that are dependent on operating systems such as Apple Siri, Microsoft Cortana, or Huawei Xiao Ai. The back-end system, based on Python Flask library, is integrated with the mentioned model trained on the custom dataset to provide support for intent identification and critical information capture.