Lexical and Language Modeling of Diacritics and Morphemes in Arabic Automatic Speech Recognition

[摘要] Arabic is a morphologically rich language which rarely displays diacritics. These two features of the language pose challenges when building Automatic Speech Recognition (ASR) systems. Morphological complexity leads to many possible combinations of stems and affixes to form words, and produces texts with high Out Of Vocabulary (OOV) rates. In addition, texts rarely display diacritics which informs the reader about short vowels, geminates, and nunnations (word ending /n/). A lack of diacritics means that 30% of textual information is missing, causing ambiguities in lexical and language modeling when attempting to model pronunciations, and the context of a particular pronunciation. Intuitively, from an English centric view, the phrase th;;wrtr wrt n thwrt with ;;morphological decomposition;; is realized as, th wrtr wrt n th wrt. Including ;;diacritics;; produces, the writer wrote in the writ. Thus our investigations in this thesis are twofold. Firstly, we show the benefits and interactions between modeling all classes of diacritics (short vowels, geminates, nunnations) in the lexicon. On a Modern Standard Arabic (MSA) corpus of broadcast news, this provides a 1.9% absolute improvement in Word Error Rate (WER) (p < 0.001). We also extend this graphemic lexicon with pronunciation rules, yielding a significant improvement over a lexicon that does not explicitly nodel diacritics. This results in a of 2.4% absolute improvement in WER (p < 0.001). Secondly, we show the benefits of language modeling at the morphemic level with diacritics, over the commonly available, word-based, nondiacratized text. This yields an absolute WER improvement of 1.0% (p < 0.001).

[发布机构] Massachusetts Institute of Technology