Pair Language Models for Deriving Alternative Pronunciations and Spellings from Pronunciation Dictionaries

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1584–1589, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Pair Language Models for Deriving Alternative Pronunciations and Spellings from Pronunciation Dictionaries Russell Beckley Oregon Health and Science University beckleyr@ohsu.com Brian Roark Google Inc. roarkbr@gmail.com Abstract Pronunciation dictionaries provide a readily available parallel corpus for learning to trans- duce between character strings and phoneme strings or vice versa. Translation models can be used to derive character-level paraphrases on either side of this transduction, allowing for the automatic derivation of alternative pro- nunciations or spellings. We examine finite- state and SMT-based methods for these related tasks, and demonstrate that the tasks have different characteristics – finding alternative spellings is harder than alternative pronunci- ations and benefits from round-trip algorithms when the other does not. We also show that we can increase accuracy by modeling sylla- ble stress. 1 Introduction Robust processing of speech and language requires dealing with variation in language production, ei- ther in terms of pronunciation in the spoken domain or spelling in the written domain. Predicting the intended words of an acoustic or textual sequence is an important recognition task, often required for downstream processing such as spoken language un- derstanding or knowledge extraction. Informal text genres, such as those found in social media, share some characteristics with speech; in fact such text is often informed by pronunciation variation. For ex- ample, consider the following tweet: He aint gotta question my loyalty, cuz he knw wen sh!t get real. Ill be right here! where several tokens (e.g. “cuz”, “wen”) represent spelling alternations related to pronunciation. Work in text normalization and spelling correction – e.g., Toutanova and Moore (2002); Li and Liu (2012) – has included pronunciation information to improve recognition of the intended word, via grapheme to phoneme (g2p) conversion modeling derived from pronunciation dictionaries. Pronunciation dictionaries provide natural par- allel corpora, with strings of characters paired to strings of phones. Thus, standard lexicons have been used in recent years with machine transla- tion systems such as Moses (Koehn et al., 2007), to train g2p systems (Laurent et al., 2009; Gerosa and Federico, 2009). Further, other algorithms us- ing such dictionaries also use translation phrase tables, but not for translation tasks. For exam- ple, data-driven paraphrasing methods (Bannard and Callison-Burch, 2005) use translation phrase-tables as a “pivot” to learn sets of phrases which trans- lated to the same target phrase. In a similar manner, with a pronunciation dictionary instead of a phrse- table, pivoting can be used to learn alternative pro- nunciations (Karanasou and Lamel, 2010), i.e., di- rect phoneme-to-phoneme (p2p) “translation” sys- tems that yield alternative pronunciations. Alterna- tively, round-trip translation could be used, e.g., to map from letter strings to phone strings in one step, then from the resulting phone strings to letter strings in a second step, as the means to find alternative spellings (Li and Liu, 2012). In this study, we explore dictionary-derived mod- els to find either alternative …

Content truncated to 500 words