Date of Submission
Summer 7-26-2021
Degree Type
Thesis
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
Committee Chair/First Advisor
Dr. Jing (Selena) He
Track
Big Data
NLP
Chair
Dr. Jing (Selena) He
Committee Member
Dr. Meng Han
Committee Member
Dr. Chao Mei
Abstract
Understanding on customer service with comprehensive information has become attracted in recent years due to its importance to business and consumers. Traditional method, which collect questionnaires in paper-format from consumers, is considered to inefficient and time-consuming. As Natural Language Processing (NLP) technologies developing, sentiment analysis and emotion detection has been demonstrated to understand customers’ satisfaction effectively. However, these popular methods only devote the polarity or emotional expression of products or service, they have limitations on exploring relevant knowledge as side information in specific domain. Therefore, a specific knowledge graph can be utilized to construct a question and answering system on customer service. In this thesis, we propose a knowledge graph based method named Custom Understanding and Responding Knowledge Graph (KG) Question and Answering system (CurKG-QA) on language understanding for customer service. Our method utilize two-way trigger including simple similarity match and hierarchical multi-label classification on hierarchical knowledge to effective answer user’s input question in human language. In addition, we explore a new model named Hierar-BERT-RCNN to recognize and classify vague question in hierarchical multi-label classification step. This model outperforms over hierarchical baseline models (BERT, BERT-CNN, BERT-DPCNN) on DuEE dataset on average 0.83% higher in main level and 9.49% higher in sub-level, and it achieves 96.51% accuracy in main level classification and 95.58% accuracy in sub-level classification. Also, the results show that simple similarity match of our CurKG-QA performs well on hierarchical air-service dataset even input question has jump-level or poor format.