MORSE: Semantic-ally Drive-n MORpheme SEgment-er

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 552–561 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1051 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 552–561 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1051 MORSE: Semantic-ally Drive-n MORpheme SEgment-er Tarek Sakakini University of Illinois Urbana, IL 61820 sakakini@illinois.edu Suma Bhat University of Illinois Urbana, IL 61820 spbhat2@illinois.edu Pramod Viswanath University of Illinois Urbana, IL 61820 pramodv@illinois.edu Abstract In this paper we present a novel frame- work for morpheme segmentation which uses the morpho-syntactic regularities pre- served by word representations, in addi- tion to orthographic features, to segment words into morphemes. This framework is the first to consider vocabulary-wide syntactico-semantic information for this task. We also analyze the deficiencies of available benchmarking datasets and intro- duce our own dataset that was created on the basis of compositionality. We validate our algorithm across different datasets and languages and present new state-of-the-art results. 1 Introduction Morpheme segmentation is a core natural lan- guage processing (NLP) task used as an inte- gral component in related-fields such as infor- mation retrieval (IR) (Zieman and Bleich, 1997; Kurimo et al., 2007), automatic speech recog- nition (ASR) (Bilmes and Kirchhoff, 2003; Ku- rimo et al., 2006), and machine translation (MT) (Lee, 2004; Virpioja et al., 2007). Most previ- ous works have relied solely on orthographic fea- tures (Harris, 1970; Goldsmith, 2000; Creutz and Lagus, 2002, 2005, 2007), neglecting the under- lying semantic information. This has led to an over-segmentation of words because a change of the surface form pattern is a necessary but insuffi- cient indication of a morphological change. For example, the surface form of “freshman”, hints that it should be segmented to “fresh-man”, al- though “freshman” does not describe semantically the compositional meaning of “fresh” and “man”. To compensate for this lack of semantic knowl- edge, previous works (Schone and Jurafsky, 2000; Baroni et al., 2002; Narasimhan et al., 2015) have incorporated semantic knowledge lo- cally by checking the semantic relatedness of possibly morphologically related pair of words. Narasimhan et al. (2015) check for semantic relat- edness using cosine similarity in word represen- tations (Mikolov et al., 2013a; Pennington et al., 2014). A limitation of such an approach is the inherent “sample noise” in specific word repre- sentations (exacerbated in the case of rare words). Moreover, limitation to local comparison enforces modeling morphological relations via semantic re- latedness, although it has been shown that differ- ence vectors model morphological relations more accurately (Mikolov et al., 2013b). To address this issue, we introduce a new framework (MORSE), the first to bring semantics into morpheme seg- mentation both on a local and a vocabulary-wide level. That is, when checking for the morpho- logical relation between two words, we not only check for the semantic relatedness of the pair at hand (local), but also check if the difference vec- tors of pairs showing similar orthographic change …

Content truncated to 500 words