Web Augmentation of Language Models for Continuous Speech Recognition of SMS Text Messages

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 157–165, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Web augmentation of language models for continuous speech recognition of SMS text messages Mathias Creutz1, Sami Virpioja1,2 and Anna Kovaleva1 1Nokia Research Center, Helsinki, Finland 2Adaptive Informatics Research Centre, Helsinki University of Technology, Espoo, Finland mathias.creutz@nokia.com, sami.virpioja@tkk.fi, annakov@gmx.de Abstract In this paper, we present an efficient query selection algorithm for the retrieval of web text data to augment a statistical language model (LM). The number of retrieved rel- evant documents is optimized with respect to the number of queries submitted. The querying scheme is applied in the do- main of SMS text messages. Continuous speech recognition experiments are con- ducted on three languages: English, Span- ish, and French. The web data is utilized for augmenting in-domain LMs in general and for adapting the LMs to a user-specific vocabulary. Word error rate reductions of up to 6.6 % (in LM augmentation) and 26.0 % (in LM adaptation) are obtained in setups, where the size of the web mixture LM is limited to the size of the baseline in-domain LM. 1 Introduction An automatic speech recognition (ASR) system consists of acoustic models of speech sounds and of a statistical language model (LM). The LM learns the probabilities of word sequences from text corpora available for training. The perfor- mance of the model depends on the amount and style of the text. The more text there is, the better the model is, in general. It is also important that the model be trained on text that matches the style of language used in the ASR application. Well matching, in-domain, text may be both difficult and expensive to obtain in the large quantities that are needed. A popular solution is to utilize the World Wide Web as a source of additional text for LM train- ing. A small in-domain set is used as seed data, and more data of the same kind is retrieved from the web. A decade ago, Berger and Miller (1998) proposed a just-in-time LM that updated the cur- rent LM by retrieving data from the web using re- cent recognition hypotheses as queries submitted to a search engine. Perplexity reductions of up to 10 % were reported.1 Many other works have fol- lowed. Zhu and Rosenfeld (2001) retrieved page and phrase counts from the web in order to update the probabilities of infrequent trigrams that occur in N-best lists. Word error rate (WER) reductions of about 3 % were obtained on TREC-7 data. In more recent work, the focus has turned to the collection of text rather than n-gram statistics based on page counts. More effort has been put into the selection of query strings. Bulyko et al. (2003; 2007) first extend their baseline vocabulary with words from a small in-domain training cor- pus. They then use n-grams with these new words in their web queries in order to retrieve text of a …

Content truncated to 500 words