Original paper title: Towards Safer Operations: An Expert-involved Dataset of High-Pressure Gas Incidents for Preventing Future Failures

We have published a paper by our “IncidentAI” research team, which develops trouble search solutions in the manufacturing industry using large-scale language models (LLMs), titled “Toward Safer Operations: Knowledge to Prevent Accidents.” Utilizing High Pressure Gas Accident Data'' was held at the top international conference in the field of natural language processing, ``EMNLP 2023We would like to inform you that we have been selected for the Industrial Track category of ``.

EMNLP (Empirical Methods in Natural Language Processing) is known as the most difficult international conference in the field of natural language processing (NLP), attended by many top companies, researchers, and developers from around the world. The recently selected "Industry Track" category is a session to present innovative applications applying NLP in the manufacturing field.This year, in particular, papers on practical applications of large-scale language models, which are rapidly developing, will be submitted. It attracted a lot of attention and was encouraged.

We will continue to strengthen our research and development to effectively utilize domain knowledge (specific knowledge and knowledge possessed by industries and companies) and propose practical AI solutions in business situations.

■Research overview: Proving that datasets utilizing domain knowledge are useful for accident prevention
Until now, corpora (large-scale natural language databases) that are generally easy to obtain in AI development have limited the use of AI for safety prediction in the manufacturing field to information sufficient for practical use that requires specialized knowledge. It was considered difficult because it was not organized. In this paper, the dataset that Cinnamon AI's research team built together with companies and experts proves useful for analyzing high-pressure gas plant accident reports. This dataset is based on actual accident report data*, and it has been found that the dataset constructed by experts can be used as a practical system in the field of safety prediction to prevent accidents.

Our research team constructed a dataset using three tasks: named entity recognition (NER), cause and effect extraction (Cause & Effect), and information retrieval (IR). By utilizing these datasets, it is possible to consistently structure, analyze, and utilize accident information written in natural sentences (sentences similar to spoken language). This paper proves that the scope of NLP technology can be applied not only to analysis of accident reporting and accident prevention, but also to applications that require a variety of specialized knowledge.

*Collected from publicly available reports on high pressure gas accidents published in 2022 by the High Pressure Gas Safety Association
*The dataset presented in the paper isThis pageIt is published in