Automatic Syllabification with Structured SVMs for Letter-to-Phoneme Conversion

Proceedings of ACL-08: HLT, pages 568–576, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Automatic Syllabification with Structured SVMs for Letter-To-Phoneme Conversion Susan Bartlett† Grzegorz Kondrak† Colin Cherry‡ †Department of Computing Science ‡Microsoft Research University of Alberta One Microsoft Way Edmonton, AB, T6G 2E8, Canada Redmond, WA, 98052 {susan,kondrak}@cs.ualberta.ca colinc@microsoft.com Abstract We present the first English syllabification system to improve the accuracy of letter-to- phoneme conversion. We propose a novel dis- criminative approach to automatic syllabifica- tion based on structured SVMs. In comparison with a state-of-the-art syllabification system, we reduce the syllabification word error rate for English by 33%. Our approach also per- forms well on other languages, comparing fa- vorably with published results on German and Dutch. 1 Introduction Pronouncing an unfamiliar word is a task that is of- ten accomplished by breaking the word down into smaller components. Even small children learn- ing to read are taught to pronounce a word by “sounding out” its parts. Thus, it is not surprising that Letter-to-Phoneme (L2P) systems, which con- vert orthographic forms of words into sequences of phonemes, can benefit from subdividing the input word into smaller parts, such as syllables or mor- phemes. Marchand and Damper (2007) report that incorporating oracle syllable boundary information improves the accuracy of their L2P system, but they fail to emulate that result with any of their automatic syllabification methods. Demberg et al. (2007), on the other hand, find that morphological segmenta- tion boosts L2P performance in German, but not in English. To our knowledge, no previous English orthographic syllabification system has been able to actually improve performance on the larger L2P problem. In this paper, we focus on the task of automatic orthographic syllabification, with the explicit goal of improving L2P accuracy. A syllable is a subdi- vision of a word, typically consisting of a vowel, called the nucleus, and the consonants preceding and following the vowel, called the onset and the coda, respectively. Although in the strict linguistic sense syllables are phonological rather than orthographic entities, our L2P objective constrains the input to or- thographic forms. Syllabification of phonemic rep- resentation is in fact an easier task, which we plan to address in a separate publication. Orthographic syllabification is sometimes re- ferred to as hyphenation. Many dictionaries pro- vide hyphenation information for orthographic word forms. These hyphenation schemes are related to, and influenced by, phonemic syllabification. They serve two purposes: to indicate where words may be broken for end-of-line divisions, and to assist the dictionary reader with correct pronunciation (Gove, 1993). Although these purposes are not always con- sistent with our objective, we show that we can im- prove L2P conversion by taking advantage of the available hyphenation data. In addition, automatic hyphenation is a legitimate task by itself, which could be utilized in word editors or in synthesizing new trade names from several concepts. We present a discriminative approach to ortho- graphic syllabification. We formulate syllabifica- tion as a tagging problem, and learn a discriminative tagger from labeled data …

Content truncated to 500 words