Event and Entity Coreference using Trees to Encode Uncertainty in Joint Decisions

Abstract

Coreference decisions among event mentions and among co-occurring entity mentions are highly interdependent, thus motivating joint inference. Capturing the uncertainty over each variable can be crucial for inference among multiple dependent variables. Previous work on joint coreference employs heuristic approaches, lacking well-defined objectives, and lacking modeling of uncertainty on each side of the joint problem. We present a new approach of joint coreference, including (1) a formal cost function inspired by Dasgupta's cost for hierarchical clustering, and (2) a representation for uncertainty of clustering of event and entity mentions, again based on a hierarchical structure. We describe an alternating optimization method for inference that when clustering event mentions, considers the uncertainty of the clustering of entity mentions and viceversa. We show that our proposed joint model provides empirical advantages over state-ofthe-art independent and joint models.

What happens before and after: Multi-Event Commonsense in Event Coreference Resolution

“…First, it is a simple model following the standard approach presented in Figure 1. Later approaches rely on hierarchical representations (Yadav et al, 2021a) or discourse coherence theory (Held et al, 2021). Second, it is based on RoBERTa and is more efficient and less memory consuming than the succeeding CDLM model (Caciularu et al, 2021) that is based on the much larger Longformer model (Beltagy et al, 2020).…”

“…More recently, Yadav et al (2021a) built on Cattan et al (2021a) by proposing a hierarchical approach to representing uncertainty of clustering event and entity mentions. The state-of-the-art models for cross document coreference are Caciularu et al (2021), which models cross-text relationships by using larger context windows, and Held et al (2021), which applies discourse coherence theory to coreference.…”

“…However, using gold mentions is common practice among many coreference systems where the focus is on improving the pairwise scorer (e.g. Barhom et al, 2019;Yadav et al, 2021a). For a fair comparison, we report the baseline performance by re-running Cattan et al (2021a) using gold mentions similar to the baseline used in Yadav et al (2021b).…”

“…and Baldwin, 1998), MUC (Vilain et al, 1995), and CEAF e (Luo, 2005). We follow the evaluation setup used in recent work (Cattan et al, 2021a;Yadav et al, 2021a;Held et al, 2021;Cattan et al, 2021b) and evaluate all our models at the topic level. That is, each metric is computed for each topic separately and averaged across all topics.…”