Sogang University AI research team's paper accepted at top NLP conference ACL 2026
Translated from Korean, summarized and contextualized by DistantNews.
At a glance
- A research team from Sogang University's Artificial Intelligence department has had its paper accepted for an oral presentation at ACL 2026, a top natural language processing conference.
- The paper introduces 'Omni-Embed-Audio (OEA),' a model leveraging multimodal large language models for robust audio-text retrieval, addressing limitations of existing search methods.
- The research achieved strong performance, particularly in understanding complex user queries and distinguishing similar-sounding audio, placing it in the top 15% of submissions.
A research team from Sogang University's Artificial Intelligence department has achieved a significant milestone, with their paper accepted for an oral presentation at the prestigious ACL 2026 conference, the world's leading academic gathering for natural language processing. The research, led by Professor Jang Doo-sung, focuses on enhancing audio-text retrieval using advanced AI models.
The paper, titled 'Omni-Embed-Audio: Leveraging Multimodal LLMs for Robust Audio-Text Retrieval,' proposes a novel approach that encodes text and audio into a shared embedding space within a multimodal large language model. This method aims to overcome the limitations of current audio-text search systems, which are often evaluated based on detailed text descriptions rather than how users actually search. As user queries become more complex with the rise of LLMs, the team developed a new evaluation benchmark, User-Intent Queries (UIQ), incorporating five types of search behaviors: questions, commands, keywords, paraphrases, and negative queries.
The proposed Omni-Embed-Audio (OEA) model demonstrated impressive results. While achieving performance comparable to existing top models in traditional text-to-audio retrieval, it significantly improved text-to-text search by 22%. Notably, OEA excelled in distinguishing between audio clips that sound similar but have different meanings, a task known as 'hard negative' retrieval, showing a marked advantage. This success underscores the potential of using LLMs to create encoders with superior semantic understanding capabilities for complex queries.
Selected from over 12,000 submissions, the paper's acceptance for an oral presentation places it among the top 15% of contributions to ACL 2026. The research team, including PhD student Yoo Hae-jun, master's student Lee In-seong, and undergraduate Shin Yong-seop, will present their findings in San Diego on July 6th. The article notes that this information was provided by Sogang University and may not reflect Hankyoreh's views.
Originally published by Hankyoreh in Korean. Translated, summarized, and contextualized by our editorial team with added local perspective. Read our editorial standards.