Tackling the Low-resource Challenge for Canonical Segmentation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5237–5250, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 5237 Tackling the Low-resource Challenge for Canonical Segmentation Manuel Mager1, ¨Ozlem C¸ etino˘glu1 and Katharina Kann2 1Institute for Natural Language Processing, University of Stuttgart, Germany 2University of Colorado Boulder, USA {manuel.mager, ozlem}@ims.uni-stuttgart.de, katharina.kann@colorado.edu Abstract Canonical morphological segmentation con- sists of dividing words into their standardized morphemes. Here, we are interested in ap- proaches for the task when training data is limited. We compare model performance in a simulated low-resource setting for the high- resource languages German, English, and In- donesian to experiments on new datasets for the truly low-resource languages Popoluca and Tepehua. We explore two new models for the task, borrowing from the closely related area of morphological generation: an LSTM pointer-generator and a sequence-to-sequence model with hard monotonic attention trained with imitation learning. We find that, in the low-resource setting, the novel approaches out- perform existing ones on all languages by up to 11.4% accuracy. However, while ac- curacy in emulated low-resource scenarios is over 50% for all languages, for the truly low- resource languages Popoluca and Tepehua, our best model only obtains 37.4% and 28.4% ac- curacy, respectively. Thus, we conclude that canonical segmentation is still a challenging task for low-resource languages. 1 Introduction Morphological segmentation denotes the task of dividing words into their constituting morphemes, i.e., their smallest meaning-bearing units, and has been studied extensively in natural language pro- cessing (NLP) (Ruokolainen et al., 2016). The most common form of segmentation consists of separating morphemes at the surface level. How- ever, this is not always well suited: in fusional languages, morphemes are merged during word formation and, thereby, change their surface forms. Thus, in this paper, we tackle the task of canoni- cal segmentation (Cotterell et al., 2016b), which consists of segmenting a word while restoring the original forms of its morphemes. Considering, e.g., profitably profitable-ly künstlich kunst-lich penyusup pen-susup čyuʔmuʔk y-tuʔmuʔk šwiilakał iš-wiila-kan-łi Tepehua Surface form Canonical segmentation Popoluca Indonesian German English Figure 1: Canonical segmentation examples for all lan- guages in our experiments. the English word collision, its surface segmenta- tion is collis+ion, while its canonical segmentation is collide+ion. Figure 1 provides examples for all five languages we experiment on. Neural models have shown to perform well on this task when large amounts of training data are available (Kann et al., 2016; Ruzsics and Samardzic, 2017). Nevertheless, datasets with mor- phological annotations are difficult to obtain, since they require expert annotators. Furthermore, many languages with complex morphology are spoken by a limited number of people or are listed as en- dangered languages (Mager et al., 2018), which re- duces the possible annotator pool even more. How- ever, morphological segmentation is important for downstream tasks like machine translation (Con- forti et al., 2018; Vania and Lopez, 2017), depen- dency parsing (Seeker and C¸ etino˘glu, 2015; Vania et al., 2018), or semantic role labeling (Sahin and Steedman, 2018). Moreover, high performance on these tasks …

Content truncated to 500 words