Fortification of Neural Morphological Segmentation Models for Polysynthetic Minimal-Resource Languages

Proceedings of NAACL-HLT 2018, pages 47–57 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Fortification of Neural Morphological Segmentation Models for Polysynthetic Minimal-Resource Languages Katharina Kann∗ Center for Information and Language Processing LMU Munich, Germany kann@cis.lmu.de Manuel Mager∗ Instituto de Investigaciones en Matem´aticas Aplicadas y en Sistemas Universidad Nacional Aut´onoma de M´exico mmager@turing.iimas.unam.mx Ivan Meza-Ruiz Instituto de Investigaciones en Matem´aticas Aplicadas y en Sistemas Universidad Nacional Aut´onoma de M´exico ivanvladimir@turing.iimas.unam.mx Hinrich Sch¨utze Center for Information and Language Processing LMU Munich, Germany inquiries@cislmu.org Abstract Morphological segmentation for polysynthetic languages is challenging, because a word may consist of many individual morphemes and training data can be extremely scarce. Since neural sequence-to-sequence (seq2seq) mod- els define the state of the art for morpho- logical segmentation in high-resource settings and for (mostly) European languages, we first show that they also obtain competitive perfor- mance for Mexican polysynthetic languages in minimal-resource settings. We then propose two novel multi-task training approaches— one with, one without need for external un- labeled resources—, and two corresponding data augmentation methods, improving over the neural baseline for all languages. Finally, we explore cross-lingual transfer as a third way to fortify our neural model and show that we can train one single multi-lingual model for related languages while maintaining compara- ble or even improved performance, thus reduc- ing the amount of parameters by close to 75%. We provide our morphological segmentation datasets for Mexicanero, Nahuatl, Wixarika and Yorem Nokki for future research. 1 Introduction Due to the advent of computing technologies to indigenous communities all over the world, natural language processing (NLP) applications ∗*The first two authors contributed equally. for languages with limited computer-readable textual data are getting increasingly important. This contrasts with current research, which fo- cuses strongly on approaches which require large amounts of training data, e.g., deep neural net- works. Those are not trivially applicable to minimal-resource settings with less than 1, 000 available training examples. We aim at closing this gap for morphological surface segmentation, the task of splitting a word into the surface forms of its smallest meaning-bearing units, its morphemes. Recovering morphemes provides information about unknown words and is thus especially im- portant for polysynthetic languages with a high morpheme-to-word ratio and a consequently large overall number of words. To illustrate how seg- mentation helps understanding unknown multiple- morpheme words, consider an example in this pa- per’s language of writing: even if the word uncon- ditionally did not appear in a given training corpus, its meaning could still be derived from a combina- tion of its morphs un, condition, al and ly. Due to its importance for down-stream tasks (Creutz et al., 2007; Dyer et al., 2008), segmenta- tion has been tackled in many different ways, con- sidering unsupervised (Creutz and Lagus, 2002), supervised (Ruokolainen et al., 2013) and semi- supervised settings (Ruokolainen et al., 2014). Here, we add three new questions to this line of re- search: (i) Are data-hungry neural network models 47 applicable to segmentation of polysynthetic …

Content truncated to 500 words