| dc.description.abstract | Event coreference resolution aims to determine and cluster event mentions that refer to the same real-world event. It is a relatively less studied natural language processing (NLP) task despite being crucial for various NLP applications such as topic detection and tracking, question answering, and summarization.
A typical event coreference resolution system relies on scoring similarity between two event mentions in a document followed by clustering. However, event coreference chains are sparsely distributed and only certain key events that connect other peripheral events in a document are repeated to organize content and produce a coherent story. This makes manually labeling many event coreference relations very time-consuming. Furthermore, event mentions tend to appear in diverse contexts and few are accompanied by a full set of their arguments. The three challenges, the distributional sparsity of coreferential event mentions, the absence of abundant human-annotated event coreference data, and the high diversity of contexts containing coreferential event mentions, make it hard to build effective event coreference resolution systems.
The primary goal of this dissertation is to develop a holistic approach that can successfully model document-level content structures to overcome the problems arising due to the sparse distribution of event coreference chains. To that end, we first study the discourse-level significance of an event that has many coreferential mentions in a document and devise a heuristics-based approach that captures several specific distributional patterns of coreferential event mentions. Inspired by the empirical improvement of the heuristics-based approach, we propose a new task of news discourse profiling, grounded in the news discourse theories, to identify document-level content structures and present a systematic method to incorporate them into an event coreference resolution system. Besides outperforming the heuristics-based model, the news discourse profiling-based system is capable of explaining the nature of correlations between coreferential event mentions and content structures. Consequently, we leverage the correlations between news discourse profiling and event coreference relations and define several rules to automatically collect event pairs from unlabeled news documents. Through both manual validation and empirical evaluations, we show that news discourse profiling additionally enables us to overcome the annotational sparsity.
Overall, this dissertation contributes to the current literature on event coreference resolution by adopting news discourse structure-centric approaches that are orthogonal to supervised feature-based pairwise classifiers. News discourse structure, when incorporated through explicit constraints or used to automatically acquire data from unlabeled news documents, adds to the performance of pairwise event coreference classifiers. I hope that the work done in this dissertation potentially inspires new work on analyzing and modeling discourse structure theories to improve event coreference resolution across text genres and languages. | en |
