Learning Pronunciation Dictionaries: Language Complexity and Word Selection Strategies

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 232–239, New York, June 2006. c⃝2006 Association for Computational Linguistics Learning Pronunciation Dictionaries Language Complexity and Word Selection Strategies John Kominek Alan W Black Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {jkominek,awb}@cs.cmu.edu Abstract The speed with which pronunciation dictio- naries can be bootstrapped depends on the ef- ficiency of learning algorithms and on the or- dering of words presented to the user. This pa- per presents an active-learning word selection strategy that is mindful of human limitations. Learning rates approach that of an oracle sys- tem that knows the final LTS rule set. 1 Introduction The construction of speech-to-speech translation systems is difficult, complex, and prohibitively ex- pensive for all but handful of major languages. De- veloping systems for new languages is a highly skilled job requiring considerable effort, as is the process of training people to acquire the necessary technical knowledge. Ideally, a native speaker of a (minor) language – with the right tools – should be able to develop a speech system with little or no technical knowl- edge of speech recognition, machine translation, dialog management, or speech synthesis. Rapid de- velopment of machine translation, for example, is the goal of (Lavie et al., 2003). Similarly, com- bined development of speech recognition and speech synthesis is the stated goal of (Engelbrecht and Schultz, 2005). Here we concentrate on lexicon creation for synthesis and recognition tasks, with the affiliated problem of letter-to-sound rule inference. Two central questions of dictionary building are: what letter-to-sound rule representation lends itself well to incremental learning? – and which words should be presented to the user, in what order? In this paper we investigate various approaches to the word ordering problem, including an active learning algorithm. An “active learner” is a class of machine learning algorithms that choose the or- der in which it is exposed to training examples (Auer, 2000). This is valuable when there isn't a pre-existing set of training data and when the cost of acquiring such data is high. When humans are adding dictionary entries the time and accuracy de- pends on the selected word (short words are easier than long; familiar are easier than unfamiliar), and on how quickly the learner's error rate drops (long words are more informative than short). Also, mindful that no answer key exists for new lan- guages – and that humans easily become impatient – we would like to know when a language's letter to sound rule system is, say, 90% complete. This turns out to be surprising elusive to pin down. The next section outlines our working assump- tions and issues we seek to address. Section 3 de- scribes our LTS learning framework, an elabora- tion of (Davel and Barnard, 2003). The learning behavior on multiple test languages is documented in Section 4, followed in Section 5 by a compari- son of several word selection strategies. 2 Assumptions and Issues In designing language technology …

Content truncated to 500 words