A Voice Interactive 3-Phase Hybrid Question Answering System
The technology of Artificial Intelligence (AI) makes it possible for machines to learn from experience through data and perform tasks in a human-like way. Deep learning is a subfield of AI, which is a machine learning method that has been successfully applied to a variety of tasks. In Natural Language Processing (NLP) domain, question answering systems can help humans to get the precise information they want quickly instead of searching online in web contents. In this work, we integrated a voice and text chat interaction system with a 3-Phase hybrid QA system to provide a knowledge service. Our 3-Phase QA consists of a Phase-1 rule-based Question Answering (QA) System, a Phase-2 semantic question similarity-based QA system, and a Phase-3 information retrieval-based QA system. We started by investigating the existing voice and text chat-based QA systems and designed the Phase-1 system. For voice system, we studied different speech recognition APIs which can be used for Text-To-Speech and Speech-To-Text. For Phase-1 system, we investigated the functionality of IBM Watson Assistant Chatbot Service and used it to build our rule-based dialog flow. After that, we investigated and explored different deep learning approaches for the Phase-2 and Phase-3 system. Our proposed Similarity Enhanced Concatenation-based Network (SECoN) on the Phase-2 system achieved 89.5\% accuracy and 89.4\% F1 score on question duplication detection task. In the Phase-3 system, we explored different data preprocessing strategies and BERT-style pre-trained models and validated the performance on reading comprehension dataset SQuAD2.0, and achieved 75.4\% Exact Match (EM) score and 78.8\% F1 score by using ALBERT-base pre-trained model.