Parallel Corpora in Mboshi (Bantu C25, Congo-Brazzaville)

Parallel Corpora in Mboshi (Bantu C25, Congo-Brazzaville) Annie Rialland1, Martine Adda-Decker 1, Guy-Noël Kouarata1, Gilles Adda2, Laurent Besacier3, Lori Lamel2, Elodie Gauthier3, Pierre Godard2, Jamison Cooper-Leavitt2 1 LPP, CNRS-Paris 3/Sorbonne Nouvelle, France, 2 LIMSI, CNRS, Université Paris-Saclay, France, 3 Laboratoire d’Informatique de Grenoble (LIG)/GETALP group, France {annie.rialland, martine.adda-decker}@univ-paris3.fr, guy_kouarata@yahoo.com, {gilles.adda, lamel, pierre.godard, cooperleavitt}@limsi.fr {laurent.besacier, elodie.gauthier}@univ-grenoble-alpes.fr Abstract This article presents multimodal and parallel data collections in Mboshi, as part of the French-German BULB project. It aims at supporting documentation and providing digital resources for less resourced languages with the help of speech and language-based technology. The data collection specifications thus have to meet both field linguists' and computer scientists' requirements, which are large corpora for the latter and linguistically dense data for the former. Beyond speech, the collection comprises pictures and videos documenting social practices, agriculture, wildlife and plants. Visual supports aimed at encouraging people to comment on objects which are meaningful in their daily lives. Speech recordings are composed of the original speech in Mboshi, a respoken version and a translated version to French. These three speech streams remain time-aligned thanks to LIG-AIKUMA, which adds new features to a previous AIKUMA application. The speech corpus includes read material (5k sentences, Bible), verb conjugations and a large part of spontaneous speech (conversations, picture descriptions) resulting in over 50 hours of Mboshi speech, of which 20 hours are already respoken and orally translated to French. These parallel oral data are intended for linguistic documentation (tonology, phonology...) and automatic processing (corpus annotation, alignment between Mboshi speech and French translations). Keywords: fieldwork, ASR, Bantu, Mboshi, French, speech, parallel corpus, pictures 1. Introduction According to UNESCO, 43% of the 6000 estimated languages of the World are endangered. Even languages spoken by more than a million people can be threatened due to lack of transmission from one generation to the next one. Archiving languages (as well as the knowledge associated to them) is currently an emergency and an overwhelming task. The challenge is even greater when the languages are under-resourced, often lacking writing system, written texts and translated corpora. The BULB (Breaking the Unwritten Language Barrier) project aims at providing tools to language documentation and description for unwritten languages (or languages with scarce textual material) with the help of language-based technologies (in section 2). The data collection specifications thus have to meet both field linguists' and computer scientists' requirements, which are large corpora for the latter and linguistically dense data for the former. By linguistically dense, we mean calibrated material to speed up the grammar development, typically conjugations and sentence lists as proposed by Bouquiaux and Thomas (1976) among others. The first step of the methodology of the BULB project involves the collection of parallel corpora (speech, respeaking and oral translation). They could be recorded with an application developed within the BULB project: LIG-AIKUMA, a user friendly device for linguists and speakers of the language community (in section 3). In this article, we will focus on the Mboshi language (in section 4), on the parallel corpora …

Content truncated to 500 words