Year: 2023
Type: conference-paper
Abstract: In this paper, we examine the concept of coreference in natural language text, and the challenge of identifying when two or more narrative entities should be resolved as anaphorically bound, and hence viewed as semantically identical or related.To help answer this question, we propose a coreference scale (Scalar Anaphora) for determining the degree of similarity between an anaphoric expression and its antecedent in narratives.We create a corpus of pairs of such anaphors and antecedents and annotate the relations between them based on the newly defined scale.Our data shows that the ratio of human annotators' agreement score aligns with the scale of coreference.We also present the baseline results of predicting the scales using recent T5 and GPT-4 models, which suggests that predicting such finegrained scales is still a challenging task for large language models.We will make the code and the data publicly available.
Authors: Bingyang Ye, Jingxuan Tu, James Pustejovsky
Institution: Brandeis University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4389519132