Large-Scale Text Collection for Unwritten Languages

International Joint Conference on Natural Language Processing, pages 1134–1138, Nagoya, Japan, 14-18 October 2013. Large-scale text collection for unwritten languages Florian R. Hanke and Steven Bird Department of Computing and Information Systems, University of Melbourne florian.hanke@gmail.com, sbird@unimelb.edu.au Abstract Existing methods for collecting texts from endangered languages are not creating the quantity of data that is needed for cor- pus studies and natural language process- ing tasks. This is because the process of transcribing and translating from audio recordings is too onerous. A more effec- tive method, we argue, is to involve lo- cal speakers in the field location, using an audio-only translation interface that is portable and easy to use. We present en- couraging early results of an experimental investigation of the efficiency of creating translations using this method, and report on the quality of the resulting content. 1 Introduction Language documentation aims to “provide a com- prehensive record of the linguistic practices char- acteristic of a given speech community” (Himmel- mann, 1998). In a typical language documenta- tion workflow, a linguistic event is recorded, then metadata is added concerning participants, loca- tion, language, and so forth. Later, the record- ing is transcribed, glossed at the word or mor- pheme level, and then a translation is provided. Not all of these activities occur in the field: usually recording, metadata capture, and some transcrip- tion work take precedence over word-level glosses and phrasal translations (Thieberger, 2011). Woodbury (2007) argues that for an archive en- try to be analysable for a future linguist, multi- ple kinds of translations are needed, for example audio recordings of UN-style simultaneous oral translations, or sentence by sentence translations. The typical workflow of documentary linguistics does not produce the amount of data required for large-scale corpus-based analysis of the language once it is no longer spoken (Abney and Bird, 2010). In addition, the typical workflow necessi- tates the creation of transcriptions before any an- notations can be made, for example in ELAN,1 a popular software tool for linguistic annotation (Berez, 2007). We propose to add a new path to this workflow (see Figure 1) to facilitate crowdsourcing of trans- lations, whether by local (typically village-based) speakers or by geographically distributed speak- ers from the diaspora (Reiman, 2010; Bird, 2010; Zaidan and Callison-Burch, 2011). To this end, we have developed mobile phone software with easy- to-use interfaces for collecting oral translations. Figure 1: Current and proposed workflows for early oral annotation. In this paper, we focus on a single activity, oral translation, via one of the mentioned inter- faces, a “phonecall” interface. We have developed software that runs on mobile phones, and which has been successfully deployed in off-grid indige- nous village settings. It can be used by linguists and native speakers to rapidly collect a substantial amount of high-quality, time-aligned bilingual au- dio. We report on an experimental investigation of the efficiency of creating translations, and the quality of the translated content. 1http://tla.mpi.nl/tools/tla-tools/elan/ 1134 2 Oral translation using Aikuma Oral translation is the process of listening to …

Content truncated to 500 words