Analyse des performances de modèles de langage sub-lexicale pour des langues peu-dotées à morphologie riche (Performance analysis of sub-word language modeling for under-resourced languages with rich morphology: case study on Swahili and Amharic) [in French]

Performance analysis of sub-word language modeling for under-resourced languages with rich morphology : case study on Swahili and Amharic This paper investigates the impact on ASR performance of sub-word units for two underresourced african languages with rich morphology (Amharic and Swahili). Two subword units are considered : syllable and morpheme, the latter being obtained in a supervised or unsupervised way. The important issue of word reconstruction from the syllable (or morpheme) ASR output is also discussed. For both languages, best results are reached with morphemes got from unsupervised approach. It leads to very significant WER reduction for Amharic ASR for which LM training data is very small (2.3M words) and it also slightly reduces WER over a Word-LM baseline for Swahili ASR (28M words for LM training). A detailed analysis of the OOV word reconstruction is also presented ; it is shown that a high percentage (up to 75% for Amharic) of OOV words can be recovered with morph-based language model and appropriate reconstruction method. MOTS-CLÉS : Modèle de langage, Morphème, Hors vocabulaire, Langues peu-dotées.