Feasibility Study of a BERT-based Question Answering Chatbot for Information Retrieval from Construction Specifications
Jin-Gee Kim, Sehwan Chung, Seonghyeon Moon, Seokho Chi
Abstract
Checking construction specification in every construction phase is critical to ensure proper construction quality and to avoid contractual problems. However, manual review is inefficient, expensive, and error-prone. There have been efforts to automatically review specifications, but these studies are limited in their practical applicability. As a solution, the use of retrieval-based user interface (as known as a chatbot) can extract specific information from construction specifications as a user wants. For the development of an information retrieval chatbot for construction specifications, this paper tested the application feasibility of a question answering methodology using Bidirectional Encoder Representations from Transformers (BERT). By taking advantages of the pre-trained BERT, user-wanted information was successfully extracted from construction specifications. With this approach, variety of questions can be responded flexibly without time-consuming manual tasks such as labeling.