Character-level Neural Architectures for Jointly Predicting Word Alignments and Word-internal Structure in Morphologically Complex Languages

Abstract

Through a word alignment task between English and Turkish, this project investigates ways to more effectively approach morphologically complex languages in the field of \ac{NLP}. Our models create an inductive bias to focus on word-internal structure, by taking character-level input and jointly predicting alignment, lemmas and morphological tags. Current versions of the model are able to exploit the lemma distribution so that the predicted alignment distribution improves in quality, while possible improvements to the morphological tag side of the architecture are identified. Furthermore, different methods of encoding character-level input are explored, suggesting that modern neural architectures might benefit from using multiple types of encoders in conjunction. Finally, the benefit of moving away from word-level input data towards the character level is further supported.