Domain-independent Punctuation and Segmentation Insertion

Domain-independent Punctuation and Segmentation Insertion Eunah Cho⇤, Jan Niehues, Alex Waibel Institute for Anthropomatics and Robotics, Karlsruhe Institute of Technology, Karlsruhe, Germany firstname.lastname@kit.edu Abstract Punctuation and segmentation is crucial in spoken lan- guage translation, as it has a strong impact to translation per- formance. However, the impact of rare or unknown words in the performance of punctuation and segmentation insertion has not been thoroughly studied. In this work, we simulate various degrees of domain-match in testing scenario and in- vestigate their impact to the punctuation insertion task. We explore three rare word generalizing schemes using part-of-speech (POS) tokens. Experiments show that gener- alizing rare and unknown words greatly improves the punc- tuation insertion performance, reaching up to 8.8 points of improvement in F-score when applied to the out-of-domain test scenario. We show that this improvement in punctuation quality has a positive impact on a following machine transla- tion (MT) performance, improving it by 2 BLEU points. 1. Introduction Punctuation and segmentation for automatic speech recogni- tion (ASR) output is crucial in order to provide a better read- ability of the transcript as well as for a better performance in a subsequent application, such as machine translation (MT). Current state-of-the-art ASR systems often do not generate any or reliable punctuation marks. Thus, there has been an extensive amount of study on this issue. A widely used method for punctuation and segmentation insertion utilizes language model with prosody features [1] due to its low latency. On the other hand, translation model- inspired systems [2, 3, 4] show an outstanding performance, both in accuracy of punctuation marks and improving the fol- lowing MT performance. Using such monolingual transla- tion systems, a non-punctuated source language is translated into a punctuated source language. Recently a neural ma- chine translation (NMT)-based model [5] is shown to have a better performance, maintaining low latency in a real-time application. While the monolingual translation system for punctua- tion insertion has been thoroughly investigated for its per- formance in subsequent applications, such as MT, and for ⇤Now in Amazon: eunahch@amazon.com real-time scenario constraints, such as latency and compact- ness of the model, domain-dependency of the model and its potential impact have been left under-explored. In this paper, we investigate the domain-dependency of punctuation and segmentation insertion task and suggest that a generalization scheme over domain-specific words can greatly improve the performance. In this scheme, rare and unknown words are represented in their part-of-speech (POS) tokens for generalization. In order to analyze the problem, we consider three sce- narios where a different amount of matching in-domain data is available for training. In the first scenario, test data and training data are from the same resource. Therefore they share a same genre. In the second scenario, we consider a case where only a small amount of in-domain data is avail- able for training the punctuation insertion model. In the last scenario, no matching in-domain data is available for train- ing. Detailed data description for each scenario will be given in Section 5. …

Content truncated to 500 words