Two papers from NTT have been accepted at ACL 2026, the flagship international conference in Natural Language Processing

At the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), to be held in San Diego, USA, from July 2 to 7, NTT Laboratories had 2 papers accepted for the main conference. ACL highlights advanced research in artificial intelligence for human language, including large language models, language understanding, information extraction, and machine translation. ACL 2026 is highly selective, with an acceptance rate of 19% for the main conference (from 12,145 submissions).

Abbreviated names of the laboratories:
CS: NTT Communication Science Laboratories

■NeoAMT: Neologism-Aware Agentic Machine Translation with Reinforcement Learning

Zhongtao Miao (University of Tokyo), Kaiyan Zhao (University of Tokyo), Masaaki Nagata (CS), Yoshimasa Tsuruoka (University of Tokyo)

We developed a mechanism in which, during the translation of sentences containing neologisms, a large language model (LLM) consults an external dictionary and uses the results to generate the output when its internal knowledge is insufficient. We also created benchmark data to evaluate the accuracy of this mechanism. Our research demonstrates that agent-based machine translation, in which an LLM autonomously uses external information to assess context, cultural background, and specialized terminology, is feasible and can produce high-quality, non-literal translations.

■One Single Hub Text Breaks CLIP: Identifying Vulnerabilities in Cross-Modal Encoders via Hubness

Hiroyuki Deguchi (CS), Katsuki Chousa (CS), Yusuke Sakai (Nara Institute of Science and Technology)

Cross-modal embeddings, which enable information retrieval and quality evaluation across different modalities such as images and text, play an important role as a foundational technology for AI systems that process multi-modal data. In this study, we focus on “hubness,” a known vulnerability in cross-modal embeddings, and reveal issues in existing quality evaluation and information retrieval models. These findings are expected to contribute to improving the reliability of information retrieval and AI-generated text evaluation, mitigating vulnerabilities, and designing safer AI systems that avoid unintended behaviors.