Experiments on Deep Morphological Inflection

Abstract

Morphological inflection (MI) is the task of generating a target word form based on a source word and a set of target morphological tags. We present different language-agnostic systems for MI and report results on datasets of three different sizes: low, medium and high. All these systems are deep neural network based. We implement and describe a lower baseline, and show that our systems improve on this baseline, as well as meet the state-of-art. One significant contribution through this work is studying the different neural architectures that perform best on different dataset sizes as well as on different languages. Another contribution is exploring the use of phonological features of the language in addition to characters, as well as pre-training of word embeddings. We also implement a hybrid system which combines rules learnt from string alignments along with deep learning. The significance of our work lies in the fact that the systems presented can be used for any language (we present results on 52 languages we experimented with) and in our analysis of how linguistic properties of each language has a strong bearing on the design of the neural architecture used for that language. © 2023, Springer Nature Switzerland AG.