Bootstrapping Techniques for Polysynthetic Morphological Analysis

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6652–6661 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6652 Bootstrapping Techniques for Polysynthetic Morphological Analysis William Lane Charles Darwin University william.lane@cdu.edu.au Steven Bird Charles Darwin University steven.bird@cdu.edu.au Abstract Polysynthetic languages have exceptionally large and sparse vocabularies, thanks to the number of morpheme slots and combinations in a word. This complexity, together with a general scarcity of written data, poses a challenge to the development of natural language technologies. To address this challenge, we offer linguistically-informed approaches for bootstrapping a neural morphological analyzer, and demonstrate its application to Kunwinjku, a polysynthetic Australian language. We generate data from a finite state transducer to train an encoder- decoder model. We improve the model by “hallucinating” missing linguistic structure into the training data, and by resampling from a Zipf distribution to simulate a more natural distribution of morphemes. The best model accounts for all instances of reduplication in the test set and achieves an accuracy of 94.7% overall, a 10 percentage point improvement over the FST baseline. This process demonstrates the feasibility of bootstrapping a neural morph analyzer from minimal resources. 1 Introduction Polysynthesis represents the high point of morpho- logical complexity. For example, in Kunwinjku, a language of northern Australia (ISO gup), the word ngarriwokyibidbidbuni contains six morphs: (1) ngarri- 1pl.excl- wok- word- yi- COM- bid- REDUP- bidbu- go.up- ni PI ‘We were talking as we climbed up’ Example (1) illustrates common features of polysynthesis: fusion, incorporation, and reduplication. Fusion combines multiple grammatical functions into a single morph, leading to large morph classes, and challenging the item-and-arrangement leanings of finite state morphology. Incorporation presents a modelling challenge because rule-based methods are unable to enumerate an open class, and machine learning methods need to learn how to recognize the boundary between contiguous large or open morph classes. Reduplication is also a challenge because it copies and prepends a portion of the verb root to itself, requiring a nonlinear or multi-step process. Tackling these phenomena using finite state transducers (FSTs) involves a combination of technical devices whose details depend on subtleties of the morphological analysis (cf. Arppe et al., 2017). There remains a need for more investigation of polysynthetic languages to deepen our understanding of the interplay between the options on the computational side, and the most parsimonious treatment on the linguistic side. Morphological complexity leads to data spar- sity, as the combinatorial possibilities multiply with each morpheme slot: most morphologically complex words will be rare. Furthermore, many morphologically complex languages are also endan- gered, making it difficult to collect large corpora. Thus, polysynthetic languages challenge existing ways of building tools and applications for the communities that speak these languages. In this work we investigate Kunwinjku, spoken by about 2,000 people in West Arnhem in the far north of Australia. Members of the community have expressed interest in using technology to sup- port language learning and literacy development. Thus, we face the challenge of developing useful language technologies on …

Content truncated to 500 words