Event Type Induction and Annotation using Unsupervised and Synthetically Supervised Methods

Event extraction is a key task in information extraction and involves the extraction of event mentions, event arguments, and event roles from a corpus of raw text. The creation of event extraction datasets usually involves manual annotation by domain experts. This process is expensive and prone to human biases. In recent years, several weakly supervised methods have been proposed to automatically obtain the event ontology from raw text data. Event ontology creation consists of event type induction and naming the induced event clusters.  A key challenge in the event induction task is tackling polysemous verbs and disambiguating verb senses. Prior works in this task have used verb sense dictionaries and summary datasets, which are hard to obtain and have been created using expensive manual annotation. To alleviate these concerns, we introduce CLUE, a contrastive learning based event induction and annotation framework. The framework leverages unsupervised event clustering and synthetic supervision from LLMs to induce and name event clusters. The unsupervised learning paradigm is first applied in the form of a contrastive K-Means clustering to induce event clusters from unstructured text. The induced event clusters are then named using synthetic supervision from an LLM. The event names and consequently the event clusters are further consolidated and refined using a similarity based metric. The framework outperforms existing baselines in this task. Our solution aims to reduce the cost of event annotation and also to remove the biases involved in event dataset creation.

Committee: Qi Li (major professor), Liyi Li and Yang Li