Citation Context

...rom Wikipedia. A few strategies for seed selection for distant supervision have already been investigated: at-least-one models [10,21,18,26,13], hierarchical topic models [1,19], pattern correlations =-=[22]-=-, and an information retrieval approach [25]. At-least-one models [10,21,18,26,13] are based on the idea that “if two entities participate in a relation, at least one sentence that mentions these two ...

Citation Context

...ong assumption and replaced all sentences with at least one sentence. Hoffmann et al. (2011) pointed out that many entity pairs have more than one relation. They extend11It is the abbreviation for Distant supervision for Relation extraction with Matrix Completion ed the multi-instance learning framework (Riedel et al., 2010) to the multi-label circumstance. Surdeanu et al. (2012) proposed a novel approach to multi-instance multi-label learning for relation extraction, which jointly modeled all the sentences in texts and all labels in knowledge bases for a given entity pair. Other literatures (Takamatsu et al., 2012; Min et al., 2013; Zhang et al., 2013; Xu et al., 2013) addressed more specific issues, like how to construct the negative class in learning or how to adopt more information, such as name entity tags, to improve the performance. Our work is more relevant to Riedel et al.’s (2013) which considered the task as a matrix factorization problem. Their approach is composed of several models, such as PCA (Collins et al., 2001) and collaborative filtering (Koren, 2008). However, they did not concern about the data noise brought by the basic assumption of distant supervision. 3 Model We apply a new tec...

Citation Context

...mann et al., 2011; Surdeanu et al., 2012; Riedel et al., 2010; Yao et al., 2010; Min et al., 2013), 22 hierarchical topic models (Alfonseca et al., 2012; Roth and Klakow, 2013), pattern correlations (=-=Takamatsu et al., 2012-=-), and an information retrieval approach (Xu et al., 2013). At-least-one models (Hoffmann et al., 2011; Surdeanu et al., 2012; Riedel et al., 2010; Yao et al., 2010; Min et al., 2013) are based on the...

Citation Context

...election: A few strategies for seed selection for distant supervision have already been investigated: at-least-one models [13][29][23][33][17], hierarchical topic models [1][25], pattern correlations =-=[30]-=-, and an information retrieval approach [32]. At-least-one models assume that “if two entities particpate in a relation, at least one sentence that mentions these two entities might express that relat...

Citation Context

...Extraction (Open IE) (Etzioni et al., 2008; Banko et al., 2008; Mesquita et al., 2013) is a domain-independent information extraction paradigm to extract relation tuples from collected corpus (Shinyama and Sekine, 2006) and Web (Etzioni et al., 2008; Banko et al., 2008). Open IE systems are initialized with a few domain-independent extraction patterns. To create labeled data, the texts are dependency-parsed, and the domain-independent patterns on the parses form the basis for extractions. Recently, to reduce 808 labeling effort for relation extraction, distant supervision (Mintz et al., 2009; Takamatsu et al., 2012; Min et al., 2013; Xu et al., 2013) has been proposed. This is an unsupervised approach that exploits textual features in large unlabeled corpora. In contrast to Open IE, we tune the relation patterns for a domain of interest, using labeled relation instances in source and target domains and unlabeled instances in the target domain. Our work is also different from the multischema matching in database integration (Doan et al., 2003). Multi-schema matching finds relations between columns of schemas, which have the same semantic. In addition, current weighted schema matching methods do not addre...

Citation Context

...ctors without using manually labeled data (Mintz et al., 2009; Jiang, 2009; Chan and Roth, 2010; Wang et al., 2011; Riedel et al., 2010; Ji et al., 2011; Hoffmann et al., 2011; Surdeanu et al., 2012; =-=Takamatsu et al., 2012-=-; Min et al., 2013). Various relation extraction approaches have been adapted to the medical domain, most of which focus on designing heuristic rules targeted for diagnosis and integrating the medical...

Citation Context

...is a skilful basketball player w ho w as born in Shanghai, China. Yao Ming announced his retirement f rom basketball in a press conference in Shanghai. ... Fig. 1. Noise in training data by distant supervision. The first sentence is the correct labeling and the second one is incorrect. To overcome the problem of noisy labeled data in DS, work in [13][9][14] attempted to model the noisy data with multi-instance learning methods. They assume that at least one of the sentences containing both e1 and e2 expresses r(e1, e2). However, this at-least-one assumption can fail, for that Takamatsu et al. [16] showed 91.7% of entity pairs only have one labeled sentence in Wikipedia articles which do not fit for the multi-instance learning assumption. Moreover, they used binary features in their model. This setting will enforce some frequent indistinctive features. For example, labeled sentence “Life of Pie, by Ang Lee, can be said to be the best...” for r(e1, e2) = “DirectorOf(Ang Lee, Life Of Pi)” has one lexical feature “e2 by e1” where e1 and e2 are placeholders for two entities. However, it can also be found in the “AuthorOf ” relation like the sentence “One Hundred Years of Solitude, by Gabrie...

Citation Context

...]) and Multi-Instance MultiLabel (MIML) learning ([59]) further improve it to allow a pair to have multiple relations. Their models allow different mentions for a pair to express different relations. =-=[63]-=- view the problem differently and propose a method that estimates the probability of each pattern showing each relation, based on the automatically labeled dataset. Their algorithm removes false posit...