A Cautious Generalization Goes a Long Way: Learning Morphophonological Rules

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1793–1805 July 9-14, 2023 ©2023 Association for Computational Linguistics A Cautious Generalization Goes a Long Way: Learning Morphophonological Rules Salam Khalifa†‡, Sarah Payne†‡, Jordan Kodner†‡, Ellen Broselow†, and Owen Rambow†‡ †Department of Linguistics, and ‡Institute for Advanced Computational Science (IACS) Stony Brook University {first.last}@stonybrook.edu Abstract Explicit linguistic knowledge, encoded by re- sources such as rule-based morphological an- alyzers, continues to prove useful in down- stream NLP tasks, especially for low-resource languages and dialects. Rules are an impor- tant asset in descriptive linguistic grammars. However, creating such resources is usually ex- pensive and non-trivial, especially for spoken varieties with no written standard. In this work, we present a novel approach for automatically learning morphophonological rules of Arabic from a corpus. Motivated by classic cognitive models for rule learning, rules are generalized cautiously. Rules that are memorized for in- dividual items are only allowed to generalize to unseen forms if they are sufficiently reliable in the training data. The learned rules are fur- ther examined to ensure that they capture true linguistic phenomena described by domain ex- perts. We also investigate the learnability of rules in low-resource settings across different experimental setups and dialects 1 Introduction Discovering patterns and generalizing them is the core concept behind learning in the vast major- ity of NLP models throughout time regardless of how they are learned or represented. Tasks such as morphological (re)inflection and grapheme-to- phoneme conversion have direct parallels with lan- guage learning in humans, and there is often a de- sire to compare the performance of modern systems (especially deep neural networks) to that in humans due to the relatively salient patterns in the transfor- mations that the learners (machine or human) learn. Representing such transformations with explicit rules would further enhance the efforts on language acquisition modeling and reduce the gap between NLP and domain experts such as linguists and cog- nitive scientists. Additionally, in low-resource set- tings in NLP, rule-based resources continue to with- stand the test of time when it comes to downstream kitaab+ha kaatib+ha kaatib+iin+ha Egyptian kitabha katibha katbinha Sudanese kitaaba kaatiba kaatbinna Hijazi kitaabaha kaatibha kaatbiinaha Emirati kitaabha kaatbinha kaatbiinha her book he is/I'm writing it they/we are writing it ������ ������ �������� Table 1: Different realizations of three words across four dialects. The dialects share the same underlying repre- sentation of the words. Changes in the realized forms are highlighted as follows: shortened vowels are bolded, epenthetic phones are underlined, deleted phones are not shown, and finally, realizations faithful to the under- lying representations (i.e., no change) are italicized. tasks; however, creating such resources is a tedious task and often labor-intensive. Moreover, neural networks are opaque and require additional efforts to extract human-interpretable patterns from them. Therefore, there is a crucial need for rule-learning systems that produce well-generalizable rules and are able to learn rules from a small amount of data. In this paper, we present a theory-backed rule- learning …

Content truncated to 500 words