Constructing Information Networks Using One Single Model

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1846–1851, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Constructing Information Networks Using One Single Model Qi Li† Heng Ji† Yu Hong†‡ Sujian Li¶ †Computer Science Department, Rensselaer Polytechnic Institute, USA ‡School of Computer Science and Technology, Soochow University, China ¶Key Laboratory of Computational Linguistics, Peking University, MOE, China †{liq7,hongy2,jih}@rpi.edu, ¶lisujian@pku.edu.cn Abstract In this paper, we propose a new frame- work that unifies the output of three infor- mation extraction (IE) tasks - entity men- tions, relations and events as an informa- tion network representation, and extracts all of them using one single joint model based on structured prediction. This novel formulation allows different parts of the information network fully interact with each other. For example, many rela- tions can now be considered as the re- sultant states of events. Our approach achieves substantial improvements over traditional pipelined approaches, and sig- nificantly advances state-of-the-art end-to- end event argument extraction. 1 Introduction Information extraction (IE) aims to discover entity mentions, relations and events from unstructured texts, and these three subtasks are closely inter- dependent: entity mentions are core components of relations and events, and the extraction of rela- tions and events can help to accurately recognize entity mentions. In addition, the theory of eventu- alities (D¨olling, 2011) suggested that relations can be viewed as states that events start from and result in. Therefore, it is intuitive but challenging to ex- tract all of them simultaneously in a single model. Some recent research attempted to jointly model multiple IE subtasks (e.g., (Roth and Yih, 2007; Riedel and McCallum, 2011; Yang and Cardie, 2013; Riedel et al., 2009; Singh et al., 2013; Li et al., 2013; Li and Ji, 2014)). For example, Roth and Yih (2007) conducted joint inference over entity mentions and relations; Our previous work jointly extracted event triggers and arguments (Li et al., 2013), and entity mentions and relations (Li and Ji, 2014). However, a single model that can ex- tract all of them has never been studied so far. Asif Mohammed Hanif detonated explosives in Tel Aviv Attack Person Weapon Geopolitical Entity Place Instrument Attacker Agent-Artifact Physical x1 x2 x3 x4 x5 x6 x7 x8 x: y: Figure 1: Information Network Representation. Information nodes are denoted by rectangles. Ar- rows represent information arcs. For the first time, we uniformly represent the IE output from each sentence as an information net- work, where entity mentions and event triggers are nodes, relations and event-argument links are arcs. We apply a structured perceptron framework with a segment-based beam-search algorithm to con- struct the information networks (Collins, 2002; Li et al., 2013; Li and Ji, 2014). In addition to the per- ceptron update, we also apply k-best MIRA (Mc- Donald et al., 2005), which refines the perceptron update in three aspects: it is flexible in using var- ious loss functions, it is a large-margin approach, and it can use mulitple candidate structures to tune feature weights. In …

Content truncated to 500 words