G2P Conversion of Proper Names Using Word Origin Information

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 367–371, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics G2P Conversion of Proper Names Using Word Origin Information Sonjia Waxmonsky and Sravana Reddy Department of Computer Science The University of Chicago Chicago, IL 60637 {wax, sravana}@cs.uchicago.edu Abstract Motivated by the fact that the pronuncia- tion of a name may be influenced by its language of origin, we present methods to improve pronunciation prediction of proper names using word origin information. We train grapheme-to-phoneme (G2P) models on language-specific data sets and interpolate the outputs. We perform experiments on US sur- names, a data set where word origin variation occurs naturally. Our methods can be used with any G2P algorithm that outputs poste- rior probabilities of phoneme sequences for a given word. 1 Introduction Speakers can often associate proper names with their language of origin, even when the words have not been seen before. For example, many English speak- ers will recognize that Makowski and Masiello are Polish and Italian respectively, without prior knowl- edge of either name. Such recognition is important for language processing tasks since the pronuncia- tions of out-of-vocabulary (OOV) words may de- pend on the language of origin. For example, as noted by Llitj´os (2001), ‘sch’ is likely to be pro- nounced as /sh/ for German-origin names (Schoe- nenberg) and /sk/ for Italian-origin words (Schi- avone). In this work, we apply word origin recognition to grapheme-to-phoneme (G2P) conversion, the task of predicting the phonemic representation of a word given its written form. We specifically study G2P conversion for personal surnames, a domain where OOVs are common and expected. Our goal is to show how word origin information can be used to train language-specific G2P models, and how output from these models can be combined to improve prediction of the best pronunciation of a name. We deal with data sparsity in rare language classes by re-weighting the output of the language- specific and language-independent models. 2 Previous Work Llitj´os (2001) applies word origin information to pronunciation modeling for speech synthesis. Here, a CART decision tree system is presented for G2P conversion that maps letters to phonemes using local context. Experiments use a data set of US surnames that naturally draws from a diverse set of origin lan- guages, and show that the inclusion of word origin features in the model improves pronunciation accu- racy. We use similar data, as described in §4.1. Some works on lexical modeling for speech recognition also make use of word origin. Here, the focus is on expanding the vocabulary of an ASR system rather than choosing a single best pronunci- ation. Maison et al. (2003) train language-specific G2P models for eight languages and output pronun- ciations to augment a baseline lexicon. This aug- mented lexicon outperforms a handcrafted lexicon in ASR experiments; error reduction is highest for foreign names spoken by native speakers of the ori- gin language. Cremelie and ten Bosch (2001) carry out …

Content truncated to 500 words