Constructing a Derivational Morphology Resource with Transformer Morpheme Segmentation

Constructing a Derivational Morphology Resource with Transformer Morpheme Segmentation Łukasz Knigawka Warsaw University of Technology Faculty of Electrical Engineering ul. Koszykowa 75 00-662 Warszawa, Poland lukasz.knigawka.stud@pw.edu.pl Abstract This paper describes a framework for the creation of new derivational morphology databases for a selected set of productive af- fixes in English. The sample resource ob- tained comprises almost 120k English words with morpheme segmentations generated by Transformer. The model and the database have been compared against other existing solutions. Moreover, this study offers an overview of po- tentially problematic cases encountered during the process of automatic word segmentation. 1 Introduction Derivational morphology studies the formation of new words (lexemes) ”rather than forms of a single word (cf. inflection)” (Bauer, 2004). The most common way of deriving new English words is af- fixation, which involves combining potential bases with affixes so that a new, morphologically complex word can be built. In the present study, two kinds of affixation are considered: suffixation (suffixes are the affixes placed after a base) and prefixation (prefixes precede a base). Affixes, as well as the bases, can be subsumed under morphemes, which are the smallest meaningful morphological units of a language (Hockett, 1958). Morphological seg- mentation divides words into morphemes, hence automatic morpheme segmentation employs com- putational methods of morpheme boundary iden- tification. The main focus of this paper is canoni- cal segmentation, first introduced in Cotterell et al. (2016b). It analyses a word as a sequence of canon- ical morphemes representing the underlying forms of morphemes, which may differ from their ortho- graphic representations. For example, the canoni- cal segmentation of the word funniest is fun-y-est. In principle, canonical morphological segmenta- tion constitutes a useful, though insufficient, tool for the analysis of morphologically complex words. In this work, methods of automatic morpheme seg- mentation are reviewed with the aim to create new morphological resources. Initially, a machine learn- ing model is trained to perform canonical morpho- logical segmentation. Subsequently, English words consisting of more than one morpheme are selected for further analysis. All the model input words are potentially affixed, i.e. they contain one of the affixes (prefixes and suffixes) under review. This study also investigates how the trained segmenta- tion model would deal with problematic morpho- logical cases. 2 Related Work Several recent studies have focused on automatic morphological segmentation. The log-linear model proposed in Cotterell et al. (2016b) is to learn to segment and restore orthographic changes jointly. In Kann et al. (2016), a character-level model con- sisting of five encoder-decoders is introduced and has become the new state-of-the-art. Convolutional neural networks have been applied in the process of morphological segmentation of Russian words in Sorokin and Kravtsova (2018). A discrimina- tive joint model for canonical segmentation, with a context-free grammar backbone, has been intro- duced in Cotterell et al. (2016a). After applying it to a subset of the English portion of the CELEX data (Baayen et al., 1996), an annotated treebank consisting of over 7k English words was released. Importantly, Mager et al. …

Content truncated to 500 words