[1Yin-A] Poster 1

[1Yin-A-08] Near-Miss Case Recommendation for Hazard Prediction Support at Construction Sites Using Unsupervised Contrastive Learning

〇Sumiya Yuta¹, Naoyuki Echizenya¹, Yoshiaki Oida¹ (1. Fujitsu Limited)

Keywords:

Near-miss reports,Recommender systems,Unsupervised contrastive learning,Re-ranking,Natural Language Processing

In construction sites, near-miss reporting activities are conducted where workers report experiences of perceiving danger even when no accident occurred, accumulating valuable knowledge for disaster prevention. However, it is not easy for individual workers to retrieve and relate the accumulated data to their own work contexts. This study proposes a recommendation system that suggests relevant near-miss incidents based on work backgrounds. Both work backgrounds and near-miss experiences consist of free-form text, and no supervised labels indicating relevance between them exist. We construct a recommendation model that embeds work backgrounds and incidents into a shared representation space through unsupervised contrastive learning, using document similarity of response pairs from the same respondent as soft labels. However, when recommendations rely solely on contrastive learning similarity, generic incidents such as "almost tripped" tend to appear at the top regardless of work context. To address this, we introduce a reranking method considering similarity between work backgrounds, prioritizing incidents relevant to the input context. Human evaluation using over 100,000 real records showed that reranking improved relevance accuracy from 80% to 93%, confirming the effectiveness of the proposed method.