Cross-Lingual Language Modeling with Syntactic Reordering for Low-Resource Speech Recognition

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 766–776, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Cross-Lingual Language Modeling with Syntactic Reordering for Low-Resource Speech Recognition Ping Xu and Pascale Fung Human Language Technology Center Department of Electronic and Computer Engineering The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong xuping@ust.hk, pascale@ece.ust.hk Abstract This paper proposes cross-lingual language modeling for transcribing source resource- poor languages and translating them into tar- get resource-rich languages if necessary. Our focus is to improve the speech recognition performance of low-resource languages by leveraging the language model statistics from resource-rich languages. The most challeng- ing work of cross-lingual language modeling is to solve the syntactic discrepancies between the source and target languages. We therefore propose syntactic reordering for cross-lingual language modeling, and present a first result that compares inversion transduction grammar (ITG) reordering constraints to IBM and lo- cal constraints in an integrated speech tran- scription and translation system. Evaluations on resource-poor Cantonese speech transcrip- tion and Cantonese to resource-rich Mandarin translation tasks show that our proposed ap- proach improves the system performance sig- nificantly, up to 3.4% relative WER reduction in Cantonese transcription and 13.3% relative bilingual evaluation understudy (BLEU) score improvement in Mandarin transcription com- pared with the system without reordering. 1 Introduction Statistical language modeling techniques have achieved remarkable success in speech and language processing (Clarkson and Rosenfeld, 1997; Stolcke, 2002). However, this success largely depends on the availability of a large amount of suitable text data in a language. Without sufficient text data for training, it is very difficult to build a practical and usable sta- tistical language model. Therefore, most of the ad- vances have been reported in so called resource-rich language such as English, Mandarin and Japanese, after creating linguistic resources of these languages at considerable cost. Today there are more than 6000 living languages spoken in the world (Gordon et al., 2005), and most of them have little transcribed texts and are considered as resource-poor languages (Nakov and Ng, 2009). Many of these languages are actually spoken by a huge number of speakers (e.g. some Chinese and Indian languages), and thus there is still a great demand to build speech and language processing systems for these languages. Owing to data scarcity, most often an interpo- lation (Bellegarda, 2004) of language models be- tween a resource-poor language and a resource-rich language is used in most low-resource ASR sys- tems. Some researchers have proposed transform- ing resource-rich language models to resource-poor language models by word-level transduction, either in a context-independent or context-dependent man- ner (Hori et al., 2003; Akita and Kawahara, 2006; Jensson et al., 2009; Neubig et al., 2010). In (Jens- son et al., 2009), a simple dictionary based context- independent transduction from a resource-rich lan- guage to a resource-poor language is exploited to improve speech recognition of the resource-poor language. In (Hori et al., 2003; Akita and Kawahara, 2006; …

Content truncated to 500 words