Automatic Classification of Verb-Direction Constructions In Mandarin Chinese

Lin, Tao

Graduate Thesis Or Dissertation

Automatic Classification of Verb-Direction Constructions In Mandarin Chinese Public Deposited

Analytics

Download PDF

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/jw827c71w

Abstract

Processing Multi-Word Expressions (MWEs) presents a challenge for Natural Language Processing (NLP) systems (Sag et al. 2002). In Mandarin Chinese, there are different kinds of MWEs, such as compounding constructions and serial verb constructions containing multiple predicates (Chao 1968). In this project, I will focus on parsing the semantics of a family of constructions called Verb-Direction Constructions (VDCs) in Chinese. Similar to English Verb-Particle Constructions, VDCs include a verb of precondition followed by a directional verb (e.g., na chu (lit ‘take exit’) ‘take out’). VDC functions include Self-Motion, Caused-Motion, Aspect, Discourse-Connective, and Evidential, among others (Liu et al. 1998).

Achieving native speakers’ interpretation of a language in machine learning systems can support different applications. Inspired by the framework of Sign-Based Construction Grammar (Sag 2012, Michaelis 2009, 2013) as well as Conceptual Metaphor Theory (Lakoff & Johnson 1980, 1998), I conducted three classification tasks. In the first task, I designed a VDC taxonomy, which categorizes distinct functions of VDCs, such as event structures (both causative and non-causative) that involve the movement of an entity through space to a final location. Two versions of the taxonomy were developed and learned. The annotation guideline was mainly based on an analysis of Frame Semantics (Baker, Fillmore, & Lowe 1998) for different VDC events. In the second task, VDCs were annotated as metaphoric and literal expressions, and metaphor detection was performed. The third task makes preliminary steps aimed at detecting the coerced use of VDCs, in which VDCs alter the canonical argument structures of verbs (Goldberg 1995, 1999).

This research makes two primary contributions. First, it establishes linguistic analyses of the VDC properties in question, including taxonomies of event types, metaphorical mappings, and coercion, most of which directly support the VDC classification tasks. Second, based on the linguistically motivated categories, it develops an automated method for semantic classification of VDC constructions, surpassing the scope of classification resources previously devised within Chinese NLP (Xue et al. 2000; Xue & Palmer 2009; Huang et al. 2010; Lu & Wang 2017). The system developed potentially supports other NLP applications, such as machine translation, event detection, metaphor processing, and word sense disambiguation.

Creator