A Training Strategy of Lecture Video-based Dataset for Chatbot Development in Civil Engineering Education

Le résumé fourni par la source

In the field of civil engineering, there is a growing demand to bridge the gap between academia and industry by equipping students with advanced digital technologies and fostering thinking creatively. A chatbot has emerged as a potential solution to alleviate the challenges of civil engineering higher education by supporting students' autonomous learning. Although the recent increase in available lecture videos has made it possible to build a domain-specific knowledge base, it remains unclear how to enhance the Question Answering (QA) performance on lecture video dataset that exhibits spoken language using limited train datasets. This study aims to investigate the potential of lecture video-based QA datasets and propose a training strategy by evaluating the impact of linguistic features, dataset quantity, and train order on QA performance. The experimental results show that the lecture video-based dataset has the enough potential to be used alone, but when its size is small, it is recommended to train the large-scale benchmark dataset first, even if the linguistic features are different.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.