Reducing the Annotation Effort for Letter-to-Phoneme Conversion

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 127–135, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Reducing the Annotation Effort for Letter-to-Phoneme Conversion Kenneth Dwyer and Grzegorz Kondrak Department of Computing Science University of Alberta Edmonton, AB, Canada, T6G 2E8 {dwyer,kondrak}@cs.ualberta.ca Abstract Letter-to-phoneme (L2P) conversion is the process of producing a correct phoneme sequence for a word, given its letters. It is often desirable to reduce the quantity of training data — and hence human anno- tation — that is needed to train an L2P classifier for a new language. In this pa- per, we confront the challenge of building an accurate L2P classifier with a minimal amount of training data by combining sev- eral diverse techniques: context ordering, letter clustering, active learning, and pho- netic L2P alignment. Experiments on six languages show up to 75% reduction in an- notation effort. 1 Introduction The task of letter-to-phoneme (L2P) conversion is to produce a correct sequence of phonemes, given the letters that comprise a word. An ac- curate L2P converter is an important component of a text-to-speech system. In general, a lookup table does not suffice for L2P conversion, since out-of-vocabulary words (e.g., proper names) are inevitably encountered. This motivates the need for classification techniques that can predict the phonemes for an unseen word. Numerous studies have contributed to the de- velopment of increasingly accurate L2P sys- tems (Black et al., 1998; Kienappel and Kneser, 2001; Bisani and Ney, 2002; Demberg et al., 2007; Jiampojamarn et al., 2008). A common assump- tion made in these works is that ample amounts of labelled data are available for training a classifier. Yet, in practice, this is the case for only a small number of languages. In order to train an L2P clas- sifier for a new language, we must first annotate words in that language with their correct phoneme sequences. As annotation is expensive, we would like to minimize the amount of effort that is re- quired to build an adequate training set. The ob- jective of this work is not necessarily to achieve state-of-the-art performance when presented with large amounts of training data, but to outperform other approaches when training data is limited. This paper proposes a system for training an ac- curate L2P classifier while requiring as few an- notated words as possible. We employ decision trees as our supervised learning method because of their transparency and flexibility. We incorporate context ordering into a decision tree learner that guides its tree-growing procedure towards gener- ating more intuitive rules. A clustering over letters serves as a back-off model in cases where individ- ual letter counts are unreliable. An active learning technique is employed to request the phonemes (labels) for the words that are expected to be the most informative. Finally, we apply a novel L2P alignment technique based on phonetic similarity, which results in impressive gains in accuracy with- out relying on any training data. Our empirical evaluation on several L2P …

Content truncated to 500 words