Collecting Bilingual Audio in Remote Indigenous Communities

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1015–1024, Dublin, Ireland, August 23-29 2014. Collecting Bilingual Audio in Remote Indigenous Communities Steven Bird Dept of Computing and Information Systems, University of Melbourne; Linguistic Data Consortium, University of Pennsylvania Lauren Gawne Department of Linguistics and Multilingual Studies, Nanyang Technological University, Singapore Katie Gelbart School of Oriental and African Studies, University of London Isaac McAlister Department of Languages, Literatures, and Cultures, University of Massachusetts, Amherst Abstract Most of the world’s languages are under-resourced, and most under-resourced languages lack a writing system and literary tradition. As these languages fall out of use, we lose important sources of data that contribute to our understanding of human language. The first, urgent step is to collect and orally translate a large quantity of spoken language. This can be digitally archived and later transcribed, annotated, and subjected to the full range of speech and language process- ing tasks, at any time in future. We have been investigating a mobile application for recording and translating unwritten languages. We visited indigenous communities in Brazil and Nepal and taught people to use smartphones for recording spoken language and for orally interpreting it into the national language, and collected bilingual phrase-aligned speech recordings. In spite of sev- eral technical and social issues, we found that the technology enabled an effective workflow for speech data collection. Based on this experience, we argue that the use of special-purpose soft- ware on smartphones is an effective and scalable method for large-scale collection of bilingual audio, and ultimately bilingual text, for languages spoken in remote indigenous communities. 1 Introduction Past the top one to three hundred economically significant languages, there are few prospects for re- sourcing the production of annotated corpora. Advances in natural language processing have relied on such corpora – including treebanks and wordnets – though they are expensive to produce and depend on substantial prior scholarship on the language. An alternative is to collect bilingual aligned text, relating a low-resource language to a high-resource language, and then infer lexical and syntactic information from the high-resource language via alignments (Abney and Bird, 2010; Baldwin et al., 2010; Palmer et al., 2010; Das and Petrov, 2011). This approach only works for written languages. Over half the world’s languages lack a literary tradition. In some cases they have a writing system, but it is not in regular use and so these languages remain effectively unwritten. Collecting data for unwritten languages necessarily involves speech. w1 w2 w3 w4 w5 w6 w7 w8 w9 w10 w11w12 f1f2f3f4f5f6f7f8 f9f10f11f12f13f14 f15f16f17f18f19f20f21 f22f23f24f25f26f27 f1f2 f5f6 f7f8 w3 w2 f9f10f11 f15f16 f22f23 f12f13f14 w1 w6 w5 f17f18f19 f20f21 w8 w7 w9 w10 f24f25 w12 w11 w'1 w'2 w'3 w'4 w'5 w'6 w'7 w'8 w'8 w'10 w'11 w'12 f26f27 f3f4 w4 Figure 1: The Vision: phrase-aligned bilingual audio from an unwritten language to a language of wider communication, along with extracted acoustic features and crowdsourced transcription (left); interlinear glossed text with word segmentation, word-level glosses, and …

Content truncated to 500 words