Modelling Verbal Morphology in Nen

Modelling Verbal Morphology in Nen Saliha Murado˘gluΩΦ Nicholas EvansΩΦ Ekaterina Vylomovaµ ΩThe Australian National University (ANU) µThe University of Melbourne ΦARC Centre of Excellence for the Dynamics of Language (CoEDL) saliha.muradgolu@anu.edu.au, nicholas.evans@anu.edu.au, ekaterina.vylomova@unimelb.edu.au Abstract Nen verbal morphology is remarkably com- plex; a transitive verb can take up to 1, 740 unique forms. The combined effect of having a large combinatoric space and a low-resource setting amplifies the need for NLP tools. Nen morphology utilises distributed exponence – a non-trivial means of mapping form to mean- ing. In this paper, we attempt to model Nen verbal morphology using state-of-the-art ma- chine learning models for morphological rein- flection. We explore and categorise the types of errors these systems generate. Our results show sensitivity to training data composition; different distributions of verb type yield differ- ent accuracies (patterning with E-complexity). We also demonstrate the types of patterns that can be inferred from the training data through the case study of syncretism. 1 Introduction A long-standing research direction in NLP targets the development of robust language technology applicable across the wide variety of the world’s languages. Unfortunately, the vast majority of ma- chine learning models are being developed for a small fraction of nearly 7,000 languages in the world, such as English, German, French, or Chi- nese. With introduction of highly multilingual corpora such as UniversalDependencies (Nivre et al., 2016) and UniMorph (Sylak-Glassman et al., 2015; Kirov et al., 2018) the situation started to change. For instance, SIGMORPHON organized a number of shared tasks on morphological reinflec- tion starting from 10 languages in 2016 (Cotterell et al., 2016) and up to 90 languages in 2020 (Vylo- mova et al., 2020). In 2020, languages were sam- pled from various typologically diverse families: Indo-European, Oto-Manguean, Tungusic, Turkic, Niger-Congo, Bantu, and others. Still, just one lan- guage, namely, Murrinh-patha, an Australian Abo- riginal language (Mansfield, 2019), represented the whole linguistic variety of the Oceania region. In this paper, we aim at filling the gap by exploring Nen, a Papuan language spoken by approximately 400 people in Papua New Guinea. Nen is known for its rich verbal morphology, with a transitive verb in- flecting for up to 1, 740 feature combinations. Dis- tributed exponence, the phenomenon which gives rise to this large paradigm size, provides insight into modelling complex mappings between surface forms and feature bundles. We conduct a series of experiments on morpho- logical reinflection task recently introduced under the umbrella of SIGMORPHON (Cotterell et al., 2016, 2018). We train several state-of-the-art ma- chine learning models for verbal inflection in Nen and provide an extensive error analysis. We inves- tigate the relationship between the distribution of verb type (inflection classes) in the data and per- formance. Finally, we show that the system learns properties of the data that are not explicitly given, but may be inferred. The rest of the paper is organized as follows: In Section 2, we give a brief overview of related work. Section 3 provides an overview of Nen verbal mor- phology, Section …

Content truncated to 500 words