Noisy Or-based model for Relation Extraction using Distant Supervision

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1937–1941, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Noisy Or-based model for Relation Extraction using Distant Supervision Ajay Nagesh1,2,3 1IITB-Monash Research Academy ajaynagesh@cse.iitb.ac.in Gholamreza Haffari 2Faculty of IT, Monash University gholamreza.haffari@monash.edu Ganesh Ramakrishnan 3Dept. of CSE, IIT Bombay ganesh@cse.iitb.ac.in Abstract Distant supervision, a paradigm of rela- tion extraction where training data is cre- ated by aligning facts in a database with a large unannotated corpus, is an attractive approach for training relation extractors. Various models are proposed in recent lit- erature to align the facts in the database to their mentions in the corpus. In this paper, we discuss and critically analyse a popular alignment strategy called the “at least one” heuristic. We provide a sim- ple, yet effective relaxation to this strat- egy. We formulate the inference proce- dures in training as integer linear program- ming (ILP) problems and implement the relaxation to the “at least one ” heuris- tic via a soft constraint in this formulation. Empirically, we demonstrate that this sim- ple strategy leads to a better performance under certain settings over the existing ap- proaches. 1 Introduction Although supervised approaches to relation ex- traction (GuoDong et al., 2005; Surdeanu and Cia- ramita, 2007) achieve very high accuracies, they do not scale as they are data intensive and the cost of creating annotated data is quite high. To alle- viate this problem, Mintz et al. (2009) proposed relation extraction in the paradigm of distant su- pervision. In this approach, given a database of facts (e.g. Freebase1) and an unannotated docu- ment collection, the goal is to heuristically align the facts in the database to the sentences in the corpus which contain the entities mentioned in the fact. This is done to create weakly labeled train- ing data to train a classifier for relation extraction. The underlying assumption is that all mentions of 1www.freebase.com an entity pair2 (i.e. sentences containing the en- tity pair) in the corpus express the same relation as stated in the database. The above assumption is a weak one and is often violated in natural language text. For in- stance, the entity pair, (Barack Obama, United States) participate in more than one relation: citizenOf, presidentOf, bornIn and every men- tion expresses either one of these fixed set of rela- tions or none of them. Consequently, a number of models have been proposed in literature to provide better heuristics for the mapping between the entity pair in the database and its mentions in the sentences of the corpus. Riedel et al. (2010) tightens the assump- tion of distant supervision in the following man- ner: “Given a pair of entities and their mentions in sentences from a corpus, at least one of the men- tions express the relation given in the database”. In other words, it models the problem as that of multi-instance (mentions) single-label (relation) learning. Following this, Hoffmann et al. (2011) and Surdeanu et al. (2012) propose …

Content truncated to 500 words