A Comparison of Different Punctuation Prediction Approaches in a Translation Context

A Comparison of Different Punctuation Prediction Approaches in a Translation Context Vincent Vandeghinste CCL – KU Leuven vincent@ccl.kuleuven.be Joris Pelemans Apple Inc. jpelemans@apple.com Lyan Verwimp ESAT-PSI – KU Leuven lyan.verwimp@esat.kuleuven.be Patrick Wambacq ESAT-PSI – KU Leuven patrick.wambacq@esat.kuleuven.be Abstract We test a series of techniques to pre- dict punctuation and its effect on ma- chine translation (MT) quality. Sev- eral techniques for punctuation prediction are compared: language modeling tech- niques, such as n-grams and long short- term memories (LSTM), sequence labeling LSTMs (unidirectional and bidirectional), and monolingual phrase-based, hierarchi- cal and neural MT. For actual translation, phrase-based, hierarchical and neural MT are investigated. We observe that for punc- tuation prediction, phrase-based statistical MT and neural MT reach similar results, and are best used as a preprocessing step which is followed by neural MT to perform the actual translation. Implicit punctuation insertion by a dedicated neural MT system, trained on unpunctuated source and punc- tuated target, yields similar results. 1 Introduction In speech translation, the first step often consists of automatic speech recognition (ASR). Most ASR systems output an unsegmented stream of words, apart from some form of acoustic segmentation which splits a transcript into so-called utterances. Translating this stream of words, using off-the- shelf MT, results in a lower translation quality compared to translating punctuated input, as MT systems are usually trained on properly punctuated and segmented source and target text. End-to-end speech translation systems, that do not suffer from c⃝ 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. this problem, have recently achieved high-quality results too (Weiss et al., 2017), but these models require infrastructure (in terms of GPUs and train- ing time) that is not available to everyone. We compare several techniques and approaches for punctuation prediction in a translation context, starting from an input that already contains the cor- rect sentence boundaries. All techniques and ap- proaches are trained on the same dataset, allowing us to fully attribute different results to the specific techniques and approaches used. Thus, the main contribution of this paper is not introducing new methods for punctuation prediction, but a thorough comparison of methods previously used, since ex- tensive comparisons are often lacking in related work. We compare three families of approaches for punctuation prediction: (1) language modeling, (2) sequence modeling, and (3) monolingual MT. These approaches are combined in three differ- ent architectures resulting in translated and punc- tuated output: (1) Preprocessing adds punctua- tion before translating with a normal MT system, trained on punctuated source and punctuated target data; (2) Implicit insertion adds punctuation dur- ing MT, which is trained on unpunctuated source and punctuated target data; and (3) Postprocess- ing adds punctuation after MT, which is trained on unpunctuated source and unpunctuated target data. Figure 1 shows these different strategies, together with the baseline strategy, in which the unpunc- tuated data is translated by a regular MT system trained on punctuated source and target data. 2 Related work In this …

Content truncated to 500 words