An Unsupervised Probability Model for Speech-to-Translation Alignment of Low-Resource Languages

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1255–1263, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics An Unsupervised Probability Model for Speech-to-Translation Alignment of Low-Resource Languages Antonios Anastasopoulos and David Chiang University of Notre Dame {aanastas, dchiang}@nd.edu Long Duong University of Melbourne lduong@student.unimelb.edu.au Abstract For many low-resource languages, spoken lan- guage resources are more likely to be an- notated with translations than with transcrip- tions. Translated speech data is potentially valuable for documenting endangered lan- guages or for training speech translation sys- tems. A first step towards making use of such data would be to automatically align spoken words with their translations. We present a model that combines Dyer et al.’s reparam- eterization of IBM Model 2 (fast_align) and k-means clustering using Dynamic Time Warping as a distance measure. The two com- ponents are trained jointly using expectation- maximization. In an extremely low-resource scenario, our model performs significantly better than both a neural model and a strong baseline. 1 Introduction For many low-resource languages, speech data is easier to obtain than textual data. And because speech transcription is a costly and slow process, speech is more likely to be annotated with transla- tions than with transcriptions. This translated speech is a potentially valuable source of information – for example, for documenting endangered languages or for training speech translation systems. In language documentation, data is usable only if it is interpretable. To make a collection of speech data usable for future studies of the language, some- thing resembling interlinear glossed text (transcrip- tion, morphological analysis, word glosses, free translation) would be needed at minimum. New technologies are being developed to facilitate col- lection of translations (Bird et al., 2014), and there already exist recent examples of parallel speech collection efforts focused on endangered languages (Blachon et al., 2016; Adda et al., 2016). As for the other annotation layers, one might hope that a first pass could be done automatically. A first step to- wards this goal would be to automatically align spo- ken words with their translations, capturing informa- tion similar to that captured by word glosses. In machine translation, statistical models have tra- ditionally required alignments between the source and target languages as the first step of training. Therefore, producing alignments between speech and text would be a natural first step towards MT systems operating directly on speech. We present a model that, in order to learn such alignments, adapts and combines two components: Dyer et al.’s reparameterization of IBM Model 2 (Dyer et al., 2013), more commonly known as fast_align, and k-means clustering using Dy- namic Time Warping (Berndt and Clifford, 1994) as a distance measure. The two components are trained jointly using expectation-maximization. We experiment on two language pairs. One is Spanish-English, using the CALLHOME and Fisher corpora. The other is Griko-Italian; Griko is an endangered language for which we created (and make freely available)1 gold-standard translations and word alignments (Lekakou et al., 2013). In all cases, our model outperforms …

Content truncated to 500 words