Exploring Fine-grained Entity Type Constraints for Distantly Supervised Relation Extraction

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2107–2116, Dublin, Ireland, August 23-29 2014. Exploring Fine-grained Entity Type Constraints for Distantly Supervised Relation Extraction Yang Liu Kang Liu Liheng Xu Jun Zhao National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences Zhongguancun East Road #95, Beijing 100190, China {yang.liu, kliu, lhxu, jzhao}@nlpr.ia.ac.cn Abstract Distantly supervised relation extraction, which can automatically generate training data by align- ing facts in the existing knowledge bases to text, has gained much attention. Previous work used conjunction features with coarse entity types consisting of only four types to train their model- s. Entity types are important indicators for a specific relation, for example, if the types of two entities are “PERSON” and “FILM” respectively, then there is more likely a “DirectorOf” rela- tion between the two entities. However, the coarse entity types are not sufficient to capture the constraints of a relation between entities. In this paper, we propose a novel method to explore fine-grained entity type constraints, and we study a series of methods to integrate the constraints with the relation extracting model. Experimental results show that our methods achieve bet- ter precision/recall curves in sentential extraction with smoother curves in aggregated extraction which mean more stable models. 1 Introduction Relation Extraction is the task of extracting semantic relations between a pair of entities from sentences containing them. It can potentially benefit many applications, such as knowledge base construction, question answering (Ravichandran and Hovy, 2002), textual entailment (Szpektor et al., 2005), etc. Tra- ditional supervised approaches for relation extraction (Zhou et al., 2005)(Zhou et al., 2007) need to manually label training data, which is expensive and limits the ability to scale up. Due to the shortcom- ing of supervised approaches mentioned above, recently, a more promising approach named distantly supervised relation extraction (or distant supervision for relation extraction) (Mintz et al., 2009) has be- come popular. Instead of manual labeling, it automatically generates training data by aligning facts in existing knowledge bases to text. However, the paradigm of distant supervision also causes new problems of noisy training data both in positive training instances and negative training instances. To overcome the false positive problem caused by the distant supervision assumption, researches in (Riedel et al., 2010)(Hoffmann et al., 2011)(Sur- deanu et al., 2012) proposed multi-instance models to model noisy positive training data, where they assumed that at least one sentence in those containing an entity pair is truly positive. Takamatsu et al. (Takamatsu et al., 2012) claimed that the at-least-one assumption in multi-instance models would fail when there was only one sentence containing both entities. They proposed a method to learn and filter noisy pattern features from training instances to overcome the false positive problem. Researchers (Xu et al., 2013)(Zhang et al., 2013)(Ritter and Etzioni, 2013) tried to address the problem of false negative training data caused by the incomplete knowledge base. Xu el al. (Xu et al., 2013) used the pseudo- relevance feedback method trying …

Content truncated to 500 words