Multi-dialect neural machine translation and dialectometry

We present a multi-dialect neural machinetranslation (NMT) model tailored to Japanese.While the surface forms of Japanese dialectsdiffer from those of standard Japanese, mostof the dialects share fundamental properties such as word order, and some also usemany of the same phonetic correspondencerules. To take advantage of these properties,we integrate multilingual, syllable-level, andfixed-order translation techniques into a general NMT model. Our experimental resultsdemonstrate that this model can outperform abaseline dialect translation model. In addition,we show that visualizing the dialect embeddings learned by the model can facilitate geographical and typological analyses of dialects.