Leveraging supplemental representations for sequential transduction

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 396–406, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Leveraging supplemental representations for sequential transduction Aditya Bhargava Department of Computer Science University of Toronto Toronto, ON, Canada, M5S 3G4 aditya@cs.toronto.edu Grzegorz Kondrak Department of Computing Science University of Alberta Edmonton, AB, Canada, T6G 2E8 kondrak@cs.ualberta.ca Abstract Sequential transduction tasks, such as grapheme-to-phoneme conversion and ma- chine transliteration, are usually addressed by inducing models from sets of input-output pairs. Supplemental representations offer valu- able additional information, but incorporating that information is not straightforward. We apply a unified reranking approach to both grapheme-to-phoneme conversion and ma- chine transliteration demonstrating substantial accuracy improvements by utilizing heteroge- neous transliterations and transcriptions of the input word. We describe several experiments that involve a variety of supplemental data and two state-of-the-art transduction systems, yielding error rate reductions ranging from 12% to 43%. We further apply our approach to system combination, with error rate reductions between 4% and 9%. 1 Introduction Words exist independently of writing, as abstract enti- ties shared among the speakers of a language. Those abstract entities have various representations, which in turn may have different realizations. Orthographic forms, phonetic transcriptions, alternative transliter- ations, and even sound-wave spectrograms are all related by referring to the same abstract word and they all convey information about its pronunciation. Figure 1 shows various representations of the word Dickens. The primary (canonical) orthographic rep- resentation of the word corresponds to the language to which the word belongs. The secondary ortho- graphic representations in different writing scripts are transliterations of the word, which exhibit phono- orthography Dickens �डक� स ディケンズ Диккенс Ντίκενς ⁞ transliterations /dɪkɪnz/ dIkInz D IH K AH N Z dIk@nz d I k x n z ⁞ transcriptions MTL G2P BTL TTS SR P2G Figure 1: Several NLP tasks involve conversion between various word representations. The tasks on which we focus are shown in black. logical adaptation to the target language. The vari- ous phonetic transcriptions consist of sequences of phonemes representing the pronunciation of the word. Transcription schemes may differ in the number and coverage of various phonemes, as well as the choice of the underlying speech variety. The spoken pro- nunciation (represented by the waveform) presents a common latent influence on the representations. Several well-known NLP tasks involve matching, alignment, and conversion between different word representations. Grapheme-to-phoneme conversion (G2P) aims at generating a transcription of a word from its orthographic representation. The reverse task is phoneme-to-grapheme conversion (P2G). Ma- chine transliteration (MTL) deals with conversion be- tween orthographic representations in different writ- ing scripts; forward transliteration proceeds from the primary representation to secondary representations, while the reverse task is called back-transliteration (BTL). The conversion between a sound and an or- thography is the goal of text-to-speech synthesis 396 (TTS) and speech recognition (SR), where transcrip- tions are often used as intermediate forms. Although both MTL and G2P take orthographic representations as input, the two tasks …

Content truncated to 500 words