Joint Inference for Event Coreference Resolution

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3264–3275, Osaka, Japan, December 11-17 2016. Joint Inference for Event Coreference Resolution Jing Lu1 and Deepak Venugopal2 and Vibhav Gogate1 and Vincent Ng1 1Department of Computer Science, The University of Texas at Dallas, Richardson, TX 75080 2Department of Computer Science, The University of Memphis, Memphis, TN 38152 jxl125430@utdallas.edu, dvngopal@memphis.edu, {vgogate,vince}@hlt.utdallas.edu Abstract Event coreference resolution is a challenging problem since it relies on several components of the information extraction pipeline that typically yield noisy outputs. We hypothesize that exploiting the inter-dependencies between these components can significantly improve the performance of an event coreference resolver, and subsequently propose a novel joint inference based event coreference resolver using Markov Logic Networks (MLNs). However, the rich features that are important for this task are typically very hard to explicitly encode as MLN formulas since they significantly increase the size of the MLN, thereby making joint inference and learning infeasible. To address this problem, we propose a novel solution where we implicitly encode rich features into our model by augmenting the MLN distribution with low dimensional unit clauses. Our approach achieves state-of-the-art results on two standard evaluation corpora. 1 Introduction Within-document event coreference resolution is the task of determining which event mentions in a text refer to the same real-world event. Event coreference is arguably more challenging and less studied than entity coreference. The challenge stems in part from the fact that an event coreference resolver typically lies towards the end of the standard information extraction (IE) pipeline, assuming as input the noisy outputs of its upstream components. Specifically, a standard event coreference resolver takes as input the extracted event triggers, their arguments, and the entity coreference information, and aggregates this information through rules to resolve coreferent event mentions. Each component of this pipeline can introduce errors that naturally propagate to the event coreference resolver, thereby significantly affecting its performance. Further, the aforementioned pipeline architecture also fails to exploit inter-dependencies between the various components that can provide valuable insights to the resolver. In light of these weaknesses, we propose a novel approach to within-document event coreference resolution based on Markov Logic Networks (MLNs) (Domingos and Lowd, 2009). In our approach, we jointly perform four key tasks in the IE pipeline: trigger identification and subtyping, argument identification and role determination, entity coreference resolution, and event coreference resolution. To our knowledge, this is the first attempt to design an MLN for event coreference resolution. MLNs are particularly well-suited for modeling joint inference tasks in natural language processing (NLP) due to the inherent relational structure and uncertainty typically associated with challenging NLP problems. A major obstacle to the successful application of MLNs to NLP tasks is the high computational complexity of probabilistic inference and learning algorithms. The MLNs used in NLP are so large that even linear time inference algorithms are computationally infeasible. For instance, the rich sets of features that are typically used to solve the four tasks in the IE pipeline for …

Content truncated to 500 words