Incremental Joint Extraction of Entity Mentions and Relations

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 402–412, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Incremental Joint Extraction of Entity Mentions and Relations Qi Li Heng Ji Computer Science Department Rensselaer Polytechnic Institute Troy, NY 12180, USA {liq7,jih}@rpi.edu Abstract We present an incremental joint frame- work to simultaneously extract entity men- tions and relations using structured per- ceptron with efficient beam-search. segment-based decoder based on the ideaA of semi-Markov chain is adopted to the new framework as opposed to traditional token-based tagging. In addition, by virtue of the inexact search, we developed a num- ber of new and effective global features as soft constraints to capture the inter- dependency among entity mentions and relations. Experiments on Automatic Con- tent Extraction (ACE)1 corpora demon- strate that our joint model significantly outperforms a strong pipelined baseline, which attains better performance than the best-reported end-to-end system. 1 Introduction The goal of end-to-end entity mention and re- lation extraction is to discover relational struc- tures of entity mentions from unstructured texts. This problem has been artificially broken down into several components such as entity mention boundary identification, entity type classification and relation extraction. Although adopting such a pipelined approach would make a system com- paratively easy to assemble, it has some limita- tions: First, it prohibits the interactions between components. Errors in the upstream components are propagated to the downstream components without any feedback. Second, it over-simplifies the problem as multiple local classification steps without modeling long-distance and cross-task de- pendencies. By contrast, we re-formulate this task as a structured prediction problem to reveal the linguistic and logical properties of the hidden 1http://www.itl.nist.gov/iad/mig//tests/ace structures. For example, in Figure 1, the output structure of each sentence can be interpreted as a graph in which entity mentions are nodes and re- lations are directed arcs with relation types. By jointly predicting the structures, we aim to address the aforementioned limitations by capturing: (i) The interactions between two tasks. For exam- ple, in Figure 1a, although it may be difficult for a mention extractor to predict “1,400” as a Per- son (PER) mention, the context word “employs” between “tire maker” and “1,400” strongly in- dicates an Employment-Organization (EMP-ORG) relation which must involve a PER mention. (ii) The global features of the hidden structure. Var- ious entity mentions and relations share linguis- tic and logical constraints. For example, we can use the triangle feature in Figure 1b to en- sure that the relations between “forces”, and each of the entity mentions “Somalia/GPE”, “Haiti/GPE” and “Kosovo/GPE”, are of the same type (Physical (PHYS), in this case). Following the above intuitions, we introduce a joint framework based on structured percep- tron (Collins, 2002; Collins and Roark, 2004) with beam-search to extract entity mentions and rela- tions simultaneously. With the benefit of inexact search, we are also able to use arbitrary global features with low cost. The underlying learning algorithm has been successfully applied to some other Natural Language Processing (NLP) …

Content truncated to 500 words