Improving Coverage of an Inuktitut Morphological Analyzer Using a Segmental Recurrent Neural Network
Jeffrey C. Micher · 2017
Languages such as Inuktitut are particularly challenging for natural language processing because of polysynthesis, abundance of grammatical features represented via morphology, morphophonemics, dialect variation, and noisy data.We make use of an existing morphological analyzer, the Uqailaut analyzer, and a dataset, the Nunavut Hansards, and experiment with improving the analyzer via bootstrapping of a segmental recurrent neural network onto it.We present results of the accuracy of this approach which works better for a coarse-grained analysis than a fine-grained analysis.We also report on accuracy of just the "closed-class" suffix parts of the Inuktitut words, which are better than the overall accuracy on the full words.