Modeling Morphological Typology for Unsupervised Learning of Language Morphology

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6672–6681 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6672 Modeling Morphological Typology for Unsupervised Learning of Language Morphology Hongzhi Xu1,3, Jordan Kodner2, Mitch Marcus1, Charles Yang2 1CIS Department, University of Pennsylvania, Philadelphia, USA 2Linguistics Department, University of Pennsylvania, Philadelphia, USA 3ICSA Institute, Shanghai International Studies University, Shanghai, China hongz.xu@gmail.com, jkodner@sas.upenn.edu mitch@cis.upenn.edu, charles@ling.upenn.edu Abstract This paper describes a language-independent model for fully unsupervised morphological analysis that exploits a universal framework leveraging morphological typology. By model- ing morphological processes including suffixa- tion, prefixation, infixation, and full and partial reduplication with constrained stem change rules, our system effectively constrains the search space and offers a wide coverage in terms of morphological typology. The system is tested on nine typologically and genetically diverse languages, and shows superior perfor- mance over leading systems. We also investi- gate the effect of an oracle that provides only a handful of bits per language to signal morpho- logical type. 1 Introduction Morphological analysis aims to identify languages’ word-internal structures. Early approaches to the computational analysis of morphology modeled the structure of each language with hand-built rules, (e.g. Sproat, 1992). Such systems require a signif- icant amount of work from domain experts, and while they tend to be very accurate, they also suffer from low coverage. Supervised and semi- supervised machine learning approaches require expert input and will suffer from out-of-vocabulary problems. This paper focuses primarily on fully unsupervised morphological learning, which offers the most flexibility and can be deployed for new languages with no data annotation. Concatenation-based morphological learning systems aim to identify morphemes or morpheme boundaries within words (Virpioja et al., 2013; Goldwater and Johnson, 2004; Creutz and Lagus, 2005, 2007; Lignos, 2010; Poon et al., 2009; Sny- der and Barzilay, 2008). The Morpho-Challenge tasks1 provide a set of morphologically annotated 1http://morpho.aalto.fi/events/morphochallenge/ data for testing concatenation. However, systems designed directly for identifying morpheme bound- aries are limited in that non-linear structures such as infixation cannot be well captured. Another approach exploits morphological rela- tions between word pairs. Related words form mor- phological chains through processes of derivation. There are many such processes including affixation at the edges or middle of a word, reduplication, stem transformations, and so on. Of these, only edge-affixation is available to concatenation-based models, so leveraging derivation directly allows for wider cross-linguistic coverage (Schone and Juraf- sky, 2001; Narasimhan et al., 2015; Soricut and Och, 2015; Luo et al., 2017; Xu et al., 2018). A more holistic line of work builds learning on the concept of morphological paradigms (Parkes et al., 1998; Goldsmith, 2001; Chan, 2006; Xu et al., 2018). Paradigms can be defined as sets of morphological processes applicable to homoge- neous groups of words. For example, the paradigm (NULL, -er, -est, -ly) in English can be applied to ad- jectives (e.g., high, higher, highest, highly), while (NULL, -ing, -ed, -s, -er) is defined over verbs (e.g, walk, walking, walked, walks, walker). Paradigms have several merits. First, …

Content truncated to 500 words