In May 2025,one paper from the Natural Language Processing Group was accepted by ICML 2025. The full name of ICML 2025 is The Forty-Second International Conference on Machine Learning. ICML,along with ICLR and NeurIPS,is one of the three major conferences in machine learning and artificial intelligence,holding significant influence and attention in related fields. ICML 2025 will be held from July 13 to July 19,2025,in Vancouver,Canada.
The accepted paper is briefly introduced as follows:
Overcoming Non-monotonicity in Transducer-based Streaming Generation (Zhengrui Ma,Yang Feng，Min Zhang)
Abstract:Streaming generation models have broad applications across various domains,with the Transducer architecture being particularly popular in industry. However,its input-synchronous decoding mechanism faces challenges when handling tasks with non-monotonic alignment properties. To address this issue,this study proposes a learnable monotonic attention mechanism that integrates the decoding and input streams of the Transducer. Our method leverages the forward-backward algorithm to infer the posterior probabilities of alignment between predictor states and input timestamps. During training,it estimates the contextual representation of monotonic attention based on these probabilities,avoiding the exhaustive search over an exponentially large alignment space. Extensive experimental results demonstrate that our proposed MonoAttn-Transducer significantly improves performance in handling non-monotonic alignment in streaming generation,providing a robust solution for complex streaming generation tasks.
