Weakly-supervised Learning Approaches for Event Knowledge Acquisition and Event Detection

Abstract

Capabilities of detecting events and recognizing temporal, subevent, or causality relations among events can facilitate many applications in natural language understanding. However, supervised learning approaches that previous research mainly uses have two problems. First, due to the limited size of annotated data, supervised systems cannot sufficiently capture diverse contexts to distill universal event knowledge. Second, under certain application circumstances such as event recognition during emergent natural disasters, it is infeasible to spend days or weeks to annotate enough data to train a system. My research aims to use weakly-supervised learning to address these problems and to achieve automatic event knowledge acquisition and event recognition. In this dissertation, I first introduce three weakly-supervised learning approaches that have been shown effective in acquiring event relational knowledge. Firstly, I explore the observation that regular event pairs show a consistent temporal relation despite of their various contexts, and these rich contexts can be used to train a contextual temporal relation classifier to further recognize new temporal relation knowledge. Secondly, inspired by the double temporality characteristic of narrative texts, I propose a weakly supervised approach that identifies 287k narrative paragraphs using narratology principles and then extract rich temporal event knowledge from identified narratives. Lastly, I develop a subevent knowledge acquisition approach by exploiting two observations that 1) subevents are temporally contained by the parent event and 2) the definitions of the parent event can be used to guide the identification of subevents. I collect rich weak supervision to train a contextual BERT classifier and apply the classifier to identify new subevent knowledge. Recognizing texts that describe specific categories of events is also challenging due to language ambiguity and diverse descriptions of events. So I also propose a novel method to rapidly build a fine-grained event recognition system on social media texts for disaster management. My method creates high-quality weak supervision based on clustering-assisted word sense disambiguation and enriches tweet message representations using preceding context tweets and reply tweets in building event recognition classifiers.