LOTTE INNOVATE Paper Accepted for EMNLP, an International Conference on Natural Language Processing

[Financial News] LOTTE INNOVATE had a paper accepted to the Industry Track of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026). The paper presents natural language processing research on restoring technical terms damaged by speech recognition (STT) errors in corporate meeting environments during meeting summarization.
According to LOTTE INNOVATE on the 29th, EMNLP is considered one of the world's leading international conferences in natural language processing, along with ACL and NAACL. In particular, the Industry Track addresses challenges and solutions arising from the design, development, deployment, and analysis of natural language processing and speech technologies used in real-world industrial settings. Through this work, LOTTE INNOVATE demonstrated its expertise in artificial intelligence (AI) by researching AI technologies optimized for corporate environments and connecting them to improvements in actual service quality.
The accepted paper is titled “A Study of the Conditions for Restoring Technical Terms Corrupted by Speech Recognition Errors Using a User-Customized Glossary During Meeting Summarization and of Summarization Quality Evaluation.” It examines how a glossary tailored to the user or meeting context can be used to restore domain-specific terms corrupted by speech recognition errors during meeting summarization.
Corporate meetings frequently include technical terms specific to particular industries and jobs, as well as proper nouns closely tied to work, such as project names, product names, organization names, and system names. If these terms are transcribed incorrectly during speech recognition, key information can be distorted or omitted in meeting minutes and summaries.
The LOTTE INNOVATE research team studied an approach that provides a glossary tailored to the meeting and work context in the summarization prompt, without separately retraining the STT model it was already training and operating for meeting recognition. In experiments using glossaries built from annotations of technical terms for each meeting, glossaries tailored to the meeting context were more effective at restoring technical terms than broad domain or integrated glossaries. Performance declined as more low-relevance terms were included, confirming the importance of carefully building glossaries around meeting-related terms rather than making them overly broad.
The findings could become a key competitive advantage for corporate AI meeting summarization services, including AI Meeting Minutes, one of the features of LOTTE INNOVATE’s AI platform, iMember Work. By using a glossary narrowed to the scope of each meeting and restoring technical terms damaged by speech recognition errors with accurate spellings in the keyword section of summaries, the approach is expected to improve the accuracy of technical terms in summary results and help differentiate the iMember Work AI Meeting Minutes service.
Ki-hwan Kim, the paper’s first author and a member of LOTTE INNOVATE’s AI Technology Team, said, "This study examined what role terminology information reflecting the context of each meeting can play in supplementing damage to technical terms caused by speech recognition errors."
The study was conducted by members of LOTTE INNOVATE’s AI Technology Team, including Ki-hwan Kim, Hyeong-jun Lim, Yun-seo Jung, Hee-yong Park, Chang-hyun Jung, Seung-hee Ma, Joohye Park, and Jung-hwan Kim.


[email protected] Ji-an Yeon Reporter