Learning Adaptive Segmentation Policy for End-to-End Simultaneous Translation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7862 - 7874 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Learning Adaptive Segmentation Policy for End-to-End Simultaneous Translation Ruiqing Zhang, Zhongjun He∗, Hua Wu and Haifeng Wang Baidu Inc. No. 10, Shangdi 10th Street, Beijing, 100085, China {zhangruiqing01, hezhongjun, wu_hua, wanghaifeng}@baidu.com Abstract End-to-end simultaneous speech-to-text trans- lation aims to directly perform translation from streaming source speech to target text with high translation quality and low latency. A typical simultaneous translation (ST) sys- tem consists of a speech translation model and a policy module, which determines when to wait and when to translate. Thus the policy is crucial to balance translation qual- ity and latency. Conventional methods usu- ally adopt fixed policies, e.g. segmenting the source speech with a fixed length and gener- ating translation. However, this method ig- nores contextual information and suffers from low translation quality. This paper proposes an adaptive segmentation policy for end-to- end ST. Inspired by human interpreters, the policy learns to segment the source streaming speech into meaningful units by considering both acoustic features and translation history, maintaining consistency between the segmen- tation and translation. Experimental results on English-German and Chinese-English show that our method achieves a good accuracy- latency trade-off over recently proposed state- of-the-art methods. 1 Introduction Recent years have witnessed extensive studies and rapid progress of Simultaneous translation (ST). It aims to perform translation from source speech into the target language with high quality and low latency and is widely used in many scenarios, such as international conferences, press releases, etc. Generally, the research of ST falls into two cat- egories: the cascaded method, and the end-to-end ∗ Corresponding author. 1In German, each singular noun is assigned a gender, either masculine, feminine, or neuter, which determines whether the definite article (like “The” in English) preceding the noun is “Der”, “Die” or “Das”. Therefore, translating “The” hastily without receiving the following noun may cause mistransla- tion. (b) Word-based policy ist Hund Word Boundary Detector 1 0 1 1 0 (c) Adaptive segment based on adaptive policy Meaningful Unit Boundary Detector 0 1 1 1 1 Die hat weg Die Hund The dog has gone missing Der ist (a) Fixed-length policy Translation Translation Translation weg. </s> Hund </s> . </s> weg . Figure 1: An En-De example illustrates three segmen- tation policies for end-to-end simultaneous translation. The yellow line indicates the sampling frequency of the source speech. (a) Fixed-length policy generates a target word for each interval. (b) Word-based pol- icy detects word boundaries and generates target word once a source word is detected (green triangle). (c) Our method detects meaningful units and generates transla- tion. Both the Fixed-length policy and Word-based pol- icy incorrectly generate “Die” without consideration of following context1. The last blue blocks denote the translation after receiving the complete speech. Mis- translated words are annotated in red. method. The cascaded method consists of an au- tomatic speech recognition (ASR) model which transcribes …

Content truncated to 500 words