Minimal Supervision for Morphological Inflection

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2078–2088 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2078 Minimal Supervision for Morphological Inflection Omer Goldman Bar Ilan University omer.goldman@gmail.com Reut Tsarfaty Bar Ilan University reut.tsarfaty@biu.ac.il Abstract Neural models for the various flavours of mor- phological reinflection tasks have proven to be extremely accurate given ample labeled data, yet labeled data may be slow and costly to ob- tain. In this work we aim to overcome this an- notation bottleneck by bootstrapping labeled data from a seed as small as five labeled in- flection tables, accompanied by a large bulk of unlabeled text. Our bootstrapping method ex- ploits the orthographic and semantic regulari- ties in morphological systems in a two-phased setup, where word tagging based on analogies is followed by word pairing based on distances. Our experiments with the Paradigm Cell Fill- ing Problem over eight typologically different languages show that in languages with rela- tively simple morphology, orthographic reg- ularities on their own allow inflection mod- els to achieve respectable accuracy. Com- bined orthographic and semantic regularities alleviate difficulties with particularly complex morpho-phonological systems. We further show that our bootstrapping methods substan- tially outperform hallucination-based methods commonly used for overcoming the annotation bottleneck in morphological reinflection tasks. 1 Introduction The introduction of neural models into natural lan- guage processing in the last decade has led to huge improvements in all supervised generation tasks, including morphological inflection.1 In particular, previous works (Cotterell et al., 2017; Silfverberg and Hulden, 2018) have achieved near-perfect per- formance over the Paradigm Cell Filling Problem (PCFP) (Ackerman et al., 2009), wherein models are required to provide any form in an inflection 1In recent years the term reinflection has surfaced as a reference to morphological inflection done not necessarily from the lemma. In this paper we will refer to both inflection and reinflection as "inflection", and specify whenever we refer to inflection done exclusively from the lemma. table, given a few forms of the same lexeme.2 Two lines of recent work made progress towards less supervision, in different fashions. The first sim- ply provided scenarios with smaller training sets — for example, in SIGMORPHON’s shared tasks (Cotterell et al., 2017, 2018). The second research avenue aims to discover the paradigmatic structure of an unknown language given a large bulk of un- labeled data, either alone (Soricut and Och, 2015; Elsner et al., 2019), accompanied by a list of all relevant forms in the vocabulary (Erdmann et al., 2020), or by a list of lemmas (Jin et al., 2020). The problem with the first kind of attempts is that given the neural nature of the most success- ful models, their performance on limited supervi- sion is capped, and data augmentation is likely to help only if the initial data is diverse enough. As for the second scenario of no supervision at all, it is somewhat pessimistic and unrealistic. Even if much labeled data for a language does not exist for a low-resourced language, …

Content truncated to 500 words