A Graph Auto-encoder Model of Derivational Morphology

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1127–1138 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1127 A Graph Auto-encoder Model of Derivational Morphology Valentin Hofmann*‡, Hinrich Sch¨utze‡, Janet B. Pierrehumbert†* *Faculty of Linguistics, University of Oxford †Department of Engineering Science, University of Oxford ‡Center for Information and Language Processing, LMU Munich valentin.hofmann@ling-phil.ox.ac.uk Abstract There has been little work on modeling the morphological well-formedness (MWF) of derivatives, a problem judged to be complex and difficult in linguistics (Bauer, 2019). We present a graph auto-encoder that learns em- beddings capturing information about the com- patibility of affixes and stems in derivation. The auto-encoder models MWF in English sur- prisingly well by combining syntactic and se- mantic information with associative informa- tion from the mental lexicon. 1 Introduction A central goal of morphology is, as famously put by Aronoff (1976), “to tell us what sort of new words a speaker can form.” This defini- tion is tightly intertwined with the notion of mor- phological well-formedness (MWF). While non- existing morphologically well-formed words such as pro$computer$ism conform to the mor- phological patterns of a language and could be formed, non-existing morphologically ill-formed words such as pro$and$ism violate the patterns and are deemed impossible (Allen, 1979). More recent research has shown that MWF is a gradient rather than binary property: non-existing words that conform to the morphological patterns of a language differ in how likely they are to be actually created by speakers (Pierrehumbert, 2012). This is particularly true in the case of derivational morphology, which is not obligatory and often serves communicative needs (Bauer, 2019). As a result, the degree of MWF of a non-existing deriva- tive is influenced by a multitude of factors and judged to be hard to predict (Bauer, 2001). In NLP, the lack of reliable ways to estimate the MWF of derivatives poses a bottleneck for genera- tive models, particularly in languages exhibiting a rich derivational morphology; e.g., while inflected read listen kiss reader listener kisser readable listenable kissable (a) Mental lexicon (L) read listen kiss $able $er (b) Two-mode projection (B) Figure 1: Derivatives in the mental lexicon L (a) and the derivational graph (DG), their derivational projec- tion B (b). Predicting whether a word is part of a derivational abstraction corresponds to predicting a sin- gle edge in the DG (dotted lines). forms can be translated by generating morpholog- ically corresponding forms in the target language (Minkov et al., 2007), generating derivatives is still a major challenge for machine translation systems (Sreelekha and Bhattacharyya, 2018). Similar prob- lems exist in the area of automatic language gener- ation (Gatt and Krahmer, 2018). This study takes a first step towards computation- ally modeling the MWF of English derivatives. We present a derivational graph auto-encoder (DGA) that combines semantic and syntactic information with associative information from the mental lexi- con, achieving very good results on MWF predic- tion and performing on par with a character-based LSTM at a fraction of the number …

Content truncated to 500 words