MorphIRe · Morphologically-informed representations for natural language processing

The morphological structure of a word plays an important role in determining its function and meaning, yet it is often disregarded by current machine learning models aimed at natural language processing (NLP). State-of-the-art NLP models typically rely on word-level or character-level representations. This arguably works well for English, the dominant language in NLP research, since it is morphologically simple, but poses a challenge for morphologically-rich languages like Basque, Estonian, or Kurdish. As a consequence, the current state of the art is biased against these languages, preventing us from building better NLP technology for them.The MorphIRe project aims to learn morphologically-informed representations for NLP. It proposes to explore the fine-grained morphological analysis of word forms in order to learn representations that are grounded in morphemes, the smallest grammatical unit of language. Using these representations as input to NLP models is expected to improve their performance particularly for morphologically-rich languages. To this end, MorphIRe will make use of deep learning with neural network architectures both to learn the representations and to apply them to state-of-the-art models for a variety of NLP tasks, such as language modelling and dependency parsing.The impact of MorphIRe is twofold: 1) Learning input representations that can be used in a variety of models encourages reusability of the results and promises that improvements will carry over to future NLP research. 2) Through improving the state of the art on morphologically-rich languages, speakers of these languages will ultimately benefit from better NLP technology. This way, MorphIRe has the potential for making both a scientific and a societal impact.