Large Language Models for Investigating Co-Occurrences in Multi-Domain Indoor Environmental Quality

Department of Civil, Environmental, & Architectural Engineering (CEAE), Worcester Polytechnic Institutes, Worcester, MA 01609, USA

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Abstract. Building occupants are continuously exposed to multiple indoor environmental stimuli, including thermal, visual, acoustic, and indoor air quality (IAQ) related factors. Prior research typically examines these multiple stimuli separately. However, our understanding of how problems in different indoor environmental quality (IEQ) domains frequently co-occur and the extent to which they lead to occupant dissatisfaction remains very limited. This limitation stems from the labor-intensive, costly, and time-consuming nature of collecting observational data for a substantial number of buildings and their occupants. In this study, we introduce a state of-the-art large language modeling (LLM) approach to text-mine social media data from 12 major U.S. cities and identify multi-domain IEQ issues. Our advanced domain-adaptive LLM was developed based on the large-sized BART model and fine-tuned using a zero-shot natural language inference (NLI) corpus, complemented by a domain-specific vocabulary dictionary to enhance the integration of domain-specific knowledge. From 4.2 million collected reviews, our model identified over 230,000 occupant comments related to IEQ, among which 27,733 reviews discussed more than one IEQ topic. Analyzing the multi-domain IEQ reviews revealed that noise pollution is a dominant concern in the U.S., most frequently co-occurring with IAQ and, to a lesser extent, thermal discomfort. We also discovered that cities such as Los Angeles, Mexico City, Chicago, New York, and Miami are more prone to encounter multidimensional IEQ issues. These cities particularly show divergent sentiment scores across different IEQ domains, with one domain receiving extremely positive evaluations while others notably negative. Our findings highlight the importance of addressing co-occurring IEQ factors affecting occupant behavior and well-being. By applying LLMs to large-scale occupant reviews, we offer a new way to better understand what drives occupant satisfaction in the built environment.