Syntax-based Simultaneous Translation through Prediction of Unseen Syntactic Constituents

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 198–207, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Syntax-based Simultaneous Translation through Prediction of Unseen Syntactic Constituents Yusuke Oda Graham Neubig Sakriani Sakti Tomoki Toda Satoshi Nakamura Graduate School of Information Science Nara Institute of Science and Technology Takayamacho, Ikoma, Nara 630-0192, Japan {oda.yusuke.on9, neubig, ssakti, tomoki, s-nakamura}@is.naist.jp Abstract Simultaneous translation is a method to reduce the latency of communication through machine translation (MT) by di- viding the input into short segments be- fore performing translation. However, short segments pose problems for syntax- based translation methods, as it is diffi- cult to generate accurate parse trees for sub-sentential segments. In this paper, we perform the first experiments applying syntax-based SMT to simultaneous trans- lation, and propose two methods to pre- vent degradations in accuracy: a method to predict unseen syntactic constituents that help generate complete parse trees, and a method that waits for more input when the current utterance is not enough to gener- ate a fluent translation. Experiments on English-Japanese translation show that the proposed methods allow for improvements in accuracy, particularly with regards to word order of the target sentences. 1 Introduction Speech translation is an application of machine translation (MT) that converts utterances from the speaker’s language into the listener’s language. One of the most identifying features of speech translation is the fact that it must be performed in real time while the speaker is speaking, and thus it is necessary to split a constant stream of words into translatable segments before start- ing the translation process. Traditionally, speech translation assumes that each segment corresponds to a sentence, and thus performs sentence bound- ary detection before translation (Matusov et al., 2006). However, full sentences can be long, par- ticularly in formal speech such as lectures, and if translation does not start until explicit ends of Figure 1: Simultaneous translation where the source sentence is segmented after “I think” and translated according to (a) the standard method, (b) Grissom II et al. (2014)’s method of final verb prediction, and (c) our method of predicting syn- tactic constituents. sentences, listeners may be forced to wait a con- siderable time until receiving the result of trans- lation. For example, when the speaker continues to talk for 10 seconds, listeners must wait at least 10 seconds to obtain the result of translation. This is the major factor limiting simultaneity in tradi- tional speech translation systems. Simultaneous translation (Section 2) avoids this problem by starting to translate before observing the whole sentence, as shown in Figure 1 (a). However, as translation starts before the whole sentence is observed, translation units are often not syntactically or semantically complete, and the performance may suffer accordingly. The dele- terious effect of this missing information is less worrying in largely monotonic language pairs (e.g. English-French), but cannot be discounted in syn- tactically distant language pairs (e.g. English- Japanese) that often …

Content truncated to 500 words