Enriching and Evaluating Meaning Representations

Yuxin Ji

Highest Honor in Quantitative Theory and Methods


Abstract

Meaning representations receive increasing attention in the field of computational linguistics in recent years. Works include developing frameworks to represent the meaning of a sentence and exploring schemes to extract document-level interpretations. Abstract Meaning Representation (AMR) is a semantic graph framework which fails to adequately represent a number of important semantic features, including number (singular and plural), definiteness, quantifiers, and intensional contexts. Several proposals have been made to improve the representational adequacy of AMR by enriching its graph structure. However, these modifications are rarely implemented on existing AMR corpora due to the labor costs associated with manual annotation. In addition to sentence-level, there are attempts to extend such representations to the document-level, one of which is on coreference resolution. In this paper, I develop an automated annotation tool which algorithmically enriches AMR graphs to better represent number, (in)definite articles, quantificational determiners, and intensional arguments. I compare the automatically produced annotations to gold-standard manual annotations and show that the automatic annotator achieves impressive results, even matching those of human annotators for certain tasks. (1) Through implementing the enriched structure to the large AMR 3.0 corpus and train models using the enriched graphs, I attested the feasibility of my proposals for enrichment. Additionally, I develop an annotation scheme for document-level coreference (2) and conduct a comparison study for the text type effects across news, fables, and a novel Reddit data. The experiment results indicate the need to develop schemes adjusted for each text type due to their distinct characteristics in language use and content.

Department / School

Quantitative Theory and Methods (Informatics) / Emory University

Degree / Year

BS / Spring 2022

Committee

Jinho D. Choi, Computer Science and QTM, Emory University (Chair)
Marjorie Pak, Linguistics, Emory University
Li Xiong, Computer Science, Emory University

Links

Anthology | Paper | Presentation

Yuxin Ji (top-left), Jinho Choi (top-right), Gregor Williamson (bottom-left), Marjorie Pak (bottom-middle), Li Xiong (bottom-right)