An MDL-based approach to extracting subword units for grapheme-to-phoneme conversion

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 713–716, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics An MDL-based approach to extracting subword units for grapheme-to-phoneme conversion Sravana Reddy Department of Computer Science The University of Chicago Chicago, IL 60637 sravana@cs.uchicago.edu John Goldsmith Departments of Linguistics and Computer Science The University of Chicago Chicago, IL 60637 goldsmith@uchicago.edu Abstract We address a key problem in grapheme-to- phoneme conversion: the ambiguity in map- ping grapheme units to phonemes. Rather than using single letters and phonemes as units, we propose learning chunks, or subwords, to re- duce ambiguity. This can be interpreted as learning a lexicon of subwords that has min- imum description length. We implement an algorithm to build such a lexicon, as well as a simple decoder that uses these subwords. 1 Introduction A system for converting written words to their pro- nunciations is an important component of speech- related applications, especially in large vocabulary tasks. This problem, commonly termed “grapheme- to-phoneme conversion”, or g2p, is non-trivial for several written languages, including English, since a given letter (grapheme) may represent one of several possible phonemes, depending on the context. Be- cause the length of the context varies throughout the dictionary, fixed-length contexts may overfit some words, or inaccurately model others. We approach this problem by treating g2p as a function from contiguous sequences of graphemes, which we call ‘grapheme subwords’, to sequences of phonemes (‘phoneme subwords’), so that there is minimal ambiguity in finding the phoneme subword that corresponds to a given grapheme subword. That is, we seek to minimize both these quantities: 1. The conditional entropy of the phoneme sub- words given a grapheme subword. This di- rectly tackles the problem of ambiguity – a per- fectly unambiguous phoneme subword condi- tional distribution would have entropy = 0. 2. The entropy of the grapheme subwords. This prevents the model from getting arbitrarily complex. As a toy example, consider the following word- pronunciation1 pairs: time T AY M sting S T IH NG negation N AH G EY SH AH N There are at least 5 graphemes whose correspond- ing phoneme distribution is ambiguous (‘i’, ‘e’, ‘t’, ‘n’, ‘g’). In the segmentation below, every grapheme subword corresponds to only one phoneme subword: t + ime T + AY M s + t + ing S + T + IH NG neg + a + tion N AH G + EY + SH AH N 2 Related Work Many grapheme-to-phoneme algorithms rely on something resembling subwords; these are mainly used to account for sequences of letters representing a single phoneme (‘ph’ for F), or vice versa (‘x’ for K S). Some of the early works that create one-to- one alignments between a word and its pronuncia- tion address these cases by allowing a letter to map to one phoneme, a null phoneme, or 2-3 phonemes. Jiampojamarn and Kondrak (2009) use expectation-maximization (EM) to learn many- to-many alignments between words and pro- nunciations, …

Content truncated to 500 words