Project description

Event extraction methods are used to encapsulate n-ary relationships between a number of concepts. Recently, neural network-based methods have been proposed to capture events, their arguments and context (Nguyen and Nguyen, 2019; Trieu et al., 2020) at sentence level. These event extraction models include embeddings from language models, representations from transformers (Devlin et al., 2019), etc. but are mostly sentence based and as a result event mentions are not captured within and across documents. To improve the performance of event extraction systems in downstream applications, such as question-answering, text summarisation, and contradiction detection of events, we need to detect the connection between event mentions beyond sentence level (Lu and Ng, 2018). The task is defined as event coreference, in which we determine if two or more event mentions refer to the same event.

This task is challenging since event mentions and their arguments may be expressed in different ways, e.g., in noun phrases or verb phrases in different sentences and different documents. Previous approaches for document-level event coreference utilised similar embeddings for coreferential events (Kenyon-Dean et al., 2018) or a joint model to learn both entity coreference and event coreference (Lee at al., 2012; Barhom et al., 2019). However, they did not have a document clustering step to limit the number of coreferent candidates. Moreover, their models focus only on the similarity between events within their respective contexts (i.e., mentions refer to the same event), but may not take into account their general semantic level (i.e., mentions that are paraphases).

Therefore, in this project, we propose to utilise paraphrase detection (Meged et al., 2020; Zeng et al., 2020) to capture the similarities between event mentions as well as event arguments. Following (Zeng et al., 2020), we plan to tackle the document clustering step, in which we will use hierarchical agglomerative clustering instead of k-means.