Corpus Generation to Develop Amharic Morphological Segmenter

Morphological segmenter is an important component in Amharic natural language processing systems. Despite this fact, Amharic lacks large amount of morphologically segmented corpus. Large amount of corpus is often a requirement to develop neural network-based language technologies. This paper presents an alternative method to generate large amount of morph-segmented corpus for Amharic language. First, a relatively small (138,400 words) morphologically annotated Amharic seed-corpus is manually prepared. The annotation enables to identify prefixes, stem, and suffixes of a given word. Second, a supervised approach is used to create a conditional random field-based seed-model (on the seed-corpus). Applying the seed-model (an unsupervised technique on a large unsegmented raw Amharic words) for prediction, a large corpus size (3,777,283) of segmented words are automatically generated. Third, the newly generated corpus is used to train an Amharic morphological segmenter (based on a supervised neural sequence-to-sequence (seq2seq) approach using character embeddings). Using the seq2seq method, an F-score of 98.65% was measured. Results show an agreement with previous efforts for Arabic language. The work presented here has profound implications for future studies of Ethiopian language technologies and may one day help solve the problem of the digital-divide between resource-rich and under-resourced languages.