Non-Monotonic Attention-based Read/Write Policy Learning for Simultaneous Translation

2025-03-28

AI Summary

Simultaneous machine translation (SiMT) faces a fundamental trade-off between low latency and high translation quality. To address this, we propose a dynamic read/write policy learning framework based on non-monotonic attention. Our method leverages a high-quality offline pre-trained seq2seq model and employs source–target alignment points—derived from word-level alignments—as weak supervision signals to train a lightweight binary classifier that learns an adjustable triggering decision boundary. Crucially, this is the first approach to explicitly incorporate alignment information as weak supervision for policy learning, enabling continuous, inference-time control over the latency–quality trade-off. Experiments on multiple streaming benchmarks demonstrate that our method significantly outperforms strong baselines: BLEU scores closely approach those of offline translation models, while maintaining millisecond-level latency. This effectively alleviates the inherent quality–latency tension in SiMT.

Technology Category

• Search and Optimization: Learning to Search
• Machine Learning: Active Learning
• Natural Language Processing: Learning & Optimization for NLP

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems

Abstract

Simultaneous or streaming machine translation generates translation while reading the input stream. These systems face a quality/latency trade-off, aiming to achieve high translation quality similar to non-streaming models with minimal latency. We propose an approach that efficiently manages this trade-off. By enhancing a pretrained non-streaming model, which was trained with a seq2seq mechanism and represents the upper bound in quality, we convert it into a streaming model by utilizing the alignment between source and target tokens. This alignment is used to learn a read/write decision boundary for reliable translation generation with minimal input. During training, the model learns the decision boundary through a read/write policy module, employing supervised learning on the alignment points (pseudo labels). The read/write policy module, a small binary classification unit, can control the quality/latency trade-off during inference. Experimental results show that our model outperforms several strong baselines and narrows the gap with the non-streaming baseline model.

Problem

• Balancing translation quality and latency in simultaneous machine translation
• Converting pretrained non-streaming models into efficient streaming models
• Learning read/write policies using alignment-based pseudo labels

Innovation

• Enhances pretrained model with alignment
• Learns read/write policy via supervision
• Binary module controls quality/latency trade-off