A Comparative Study of Embedding Models in Predicting the Compositionality of Multiword Expressions

A Comparative Study of Embedding Models in Predicting the Compositionality of Multiword Expressions Navnita Nandakumar Bahar Salehi Timothy Baldwin School of Computing and Information Systems The University of Melbourne Victoria 3010, Australia nnandakumar@student.unimelb.edu.au {salehi.b,tbaldwin}@unimelb.edu.au Abstract In this paper, we perform a compara- tive evaluation of off-the-shelf embedding models over the task of compositional- ity prediction of multiword expressions (“MWEs”). Our experimental results sug- gest that character- and document-level models do capture some aspects of MWE compositionality and are effective at mod- elling varying levels of compositionality, but ultimately are not as effective as a sim- ple word2vec baseline. However they have the advantage over word-level models that they do not require token-level identifica- tion of MWEs in the training corpus. 1 Introduction In recent years, the study of the semantic id- iomaticity of multiword expressions (“MWEs”: Baldwin and Kim (2010)) has focused on com- positionality prediction, a regression task involv- ing the mapping of an MWE onto a continuous scale, representing its compositionality either as a whole or for each of its component words (Reddy et al., 2011; Ramisch et al., 2016; Cordeiro et al., to appear). In the case of couch potato “an idler who spends much time on a couch (usually watch- ing television)”, e.g., on a scale of [0, 1] the over- all compositionality may be judged to be 0.3, and the compositionality of couch and potato as 0.8 and 0.1, respectively. The main motivation for the study of compositionality is to better understand the semantic of the compound and the semantic relationships between the component words of the MWEs, which has applications in various infor- mation retrieval and natural language processing tasks (Venkatapathy and Joshi, 2006; Acosta et al., 2011; Salehi et al., 2015b). Separately, there has been burgeoning interest in learning distributed representations of words and their meanings, starting out with word em- beddings (Mikolov et al., 2013; Pennington et al., 2014) and now also involving the study of character- and document-level models (Baroni et al., 2014; Le and Mikolov, 2014; Bojanowski et al., 2017; Conneau et al., 2017). This work has been applied in part to predicting the composition- ality of MWEs (Salehi et al., 2015a; Hakimi Parizi and Cook, 2018), work that this paper builds on directly, in performing a comparative study of the performance of a range of off-the-shelf representa- tion learning methods over the task of MWE com- positionality prediction. Our contributions are as follows: (1) we show that, despite their effectiveness over a range of other tasks, recent off-the-shelf character- and document-level embedding learning methods are inferior to simple word2vec at modelling MWE compositionality; and (2) we demonstrate the util- ity of using paraphrase data in addition to simple lemmas in predicting MWE compositionality. 2 Related work The current state-of-the-art in compositionality prediction involves the use of word embeddings (Salehi et al., 2015a). The vector representa- tions of each component word (e.g. couch and potato) and the overall MWE (e.g. couch potato) are taken as a proxy for their respective …

Content truncated to 500 words