Joint Event and Temporal Relation Extraction with Shared Representations and Structured Prediction

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 434–444, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 434 Joint Event and Temporal Relation Extraction with Shared Representations and Structured Prediction Rujun Han,1,2 Qiang Ning,3 Nanyun Peng1,2 1Department of Computer Science, University of Southern California 2Information Sciences Institute, University of Southern California 3Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign rujunhan@isi.edu, qning2@illinois.edu, npeng@isi.edu Abstract We propose a joint event and temporal rela- tion extraction model with shared representa- tion learning and structured prediction. The proposed method has two advantages over ex- isting work. First, it improves event represen- tation by allowing the event and relation mod- ules to share the same contextualized embed- dings and neural representation learner. Sec- ond, it avoids error propagation in the conven- tional pipeline systems by leveraging struc- tured inference and learning methods to assign both the event labels and the temporal rela- tion labels jointly. Experiments show that the proposed method can improve both event ex- traction and temporal relation extraction over state-of-the-art systems, with the end-to-end F1 improved by 10% and 6.8% on two bench- mark datasets respectively. 1 Introduction The extraction of temporal relations among events is an important natural language understanding (NLU) task that can benefit many downstream tasks such as question answering, information re- trieval, and narrative generation. The task can be modeled as building a graph for a given text, whose nodes represent events and edges are la- beled with temporal relations correspondingly. Figure 1a illustrates such a graph for the text shown therein. The nodes assassination, slaugh- tered, rampage, war, and Hutu are the candi- date events, and different types of edges spec- ify different temporal relations between them: as- sassination is BEFORE rampage, rampage IN- CLUDES slaughtered, and the relation between slaughtered and war is VAGUE. Since “Hutu” is actually not an event, a system is expected to an- notate the relations between “Hutu” and all other nodes in the graph as NONE (i.e., no relation). As far as we know, all existing systems treat this task as a pipeline of two separate subtasks, (a) Temporal Relation Graph (b) Pipeline Model (c) Structured Joint Model Figure 1: An illustration of event and relation models in our proposed joint framework. (a) is a (partial) graph of the output of the relation extraction model. “Hutu” is not an event and hence all relations including it should be annotated as NONE. (b) and (c) are comparisons be- tween a pipeline model and our joint model. i.e., event extraction and temporal relation classi- fication, and they also assume that gold events are given when training the relation classifier (Ver- hagen et al., 2007, 2010; UzZaman et al., 2013; Chambers et al., 2014; Ning et al., 2017; Meng and Rumshisky, 2018). Specifically, they built end-to- end systems that extract events first and then pre- dict temporal relations between them (Fig. 1b). In …

Content truncated to 500 words