Capturing Word-level Dependencies in Morpheme-based Language Modeling

Morphologically rich languages suffer from data sparsity and out-of-vocabulary words problems. As a result, researchers use morphemes (sub-words) as units in language modeling instead of full-word forms. The use of morphemes in language modeling, however, might lead to a loss of word level dependency since a word can be segmented into 3 or more morphemes and the scope of the morpheme n-gram might be limited to a single word. In this paper we propose the use of roots to capture word-level dependencies in Amharic language modeling. Our experiment shows that root-based language models are better than the word based and other factored language models when compared on the basis of the probability they assign for the test set. However, no benefit has been obtained (in terms of word recognition accuracy) as a result of using root-based language models in a speech recognition task.