Abstract Meaning Representation for MWE: A study of the mapping of aspectuality based on Mandarin light verb jiayi

Abstract Meaning Representation for MWE: A study of the mapping of aspectuality based on Mandarin light verb jiayi Lu Lu Nianwen Xue Hong Kong Polytechnic University Brandeis University lu-cbs.lu@polyu.edu.hk xuen@brandeis.edu Chu-Ren Huang Hong Kong Polytechnic University churen.huang@polyu.edu.hk Abstract Multiword expression (MWE) refers to various types of linguistic units that are made up of more than one word. Light verb constructions (LVCs), as one of the least explored areas in MWEs, have idiosyncratic features that are difficult to capture in computational linguistics. In this paper, we addressed the aspectual differences between LVCs and their corresponding regular verb constructions with corpus data. Specifically, the jiayi- LVC in Mandarin Chinese was investigated as a case study, where idiosyncratic aspectual information in the LVC was proposed. This feature was not yet previously represented in abstract meaning representation (AMR), in which LVCs and its regular verb counterparts shared the same AMR. Given the semantic difference in jiayi-LVCs, we expand AMR by introducing aspect as the root node, while maintaining the core predicative component in meaning representation. 1 Introduction Light verb constructions (LVCs) are a crucial type of multiword expression (MWE) that can be found across different languages (Butt 2010). One of the prominent semantic features in LVCs (e.g. have a bath in English and jinxing yanjiu ‘carry out a study’ in Mandarin Chinese) is that the light verb (LV; i.e. have and jinxing) bears little semantic content and the actions are largely described by the eventive nominals (i.e. bath and yanjiu ‘research’), encoding the same semantic information as their regular verb counterparts. Since the meaning of LVCs differ from usual predicative structures or the direct aggregation of its semantic components, LVCs, as one of the least explored areas of MWEs in computational linguistics, pose a number of challenges in computational grammar, such as automatic word alignment, annotation and semantic representation. In the model of Abstract Meaning Representation (AMR), it is often assumed that the LVCs and its corresponding regular verb construction (RVC) share the same representation (Banarescu et al. 2012; Flanigan et al. 2014, Bu et al. 2016). AMR is a semantic framework addressing the predicate-argument relation of the whole sentence (to be more fully described in Section 4). However, corpus data suggest that LVCs and RVCs have slightly different semantic meaning. In Urdu, for example, LVs play a central role in the meaning and morphosyntactic choices of the whole construction in (1). Although the LVs par ‘fall’ and ḍaal ‘put’ both occur with ciikh ‘scream’ in Urdu, the LV in (1a), which involves an involuntary action, is preceded by an unmarked nominative subject, whereas the LV ḍaal ‘put’ in (1b), denoting a conscious control over the action, requires the marked ergative case on the subject argument. Based on this observation, Butt (1995) argues that it is the particular LV ḍaal ‘put’ that contributes to the meaning of conscious choices of a given action, which further assigns a marked ergative case to the subject in (3b). In this analysis, the so-called ‘LV’ is not completely …

Content truncated to 500 words