Robust Domain Adaptation for Relation Extraction via Clustering Consistency

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 807–817, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Robust Domain Adaptation for Relation Extraction via Clustering Consistency Minh Luan Nguyen†∗, Ivor W. Tsang‡, Kian Ming A. Chai§, and Hai Leong Chieu§ †Institute for Infocomm Research, Singapore ‡Centre for Quantum Computation & Intelligent Systems, University of Technology, Sydney, Australia §DSO National Laboratories, Singapore mlnguyen@i2r.a-star.edu.sg, ivor.tsang@gmail.com {ckianmin,chaileon}@dso.org.sg Abstract We propose a two-phase framework to adapt existing relation extraction classi- fiers to extract relations for new target do- mains. We address two challenges: neg- ative transfer when knowledge in source domains is used without considering the differences in relation distributions; and lack of adequate labeled samples for rarer relations in the new domain, due to a small labeled data set and imbalance rela- tion distributions. Our framework lever- ages on both labeled and unlabeled data in the target domain. First, we determine the relevance of each source domain to the target domain for each relation type, using the consistency between the clus- tering given by the target domain labels and the clustering given by the predic- tors trained for the source domain. To overcome the lack of labeled samples for rarer relations, these clusterings operate on both the labeled and unlabeled data in the target domain. Second, we trade-off between using relevance-weighted source- domain predictors and the labeled target data. Again, to overcome the imbalance distribution, the source-domain predictors operate on the unlabeled target data. Our method outperforms numerous baselines and a weakly-supervised relation extrac- tion method on ACE 2004 and YAGO. 1 Introduction The World Wide Web contains information on real-world entities, such as persons, locations and ∗The work is done while Nguyen was a research staff in Nanyang Technological University, Singapore. organizations, which are interconnected by vari- ous semantic relations. Detecting these relations between two entities is important for many tasks on the Web, such as information retrieval (Salton and McGill, 1986) and information extraction for question answering (Etzioni et al., 2008). Recent work on relation extraction has demonstrated that supervised machine learning coupled with intelli- gent feature engineering can provide state-of-the- art performance (Jiang and Zhai, 2007b). How- ever, most supervised learning algorithms require adequate labeled data for every relation type to be extracted. Due to the large number of relations among entities, it may be costly to annotate a large enough set of training data to cover each relation type adequately in every new domain of interest. Instead, it can be more cost-effective to adapt an existing relation extraction system to the new do- main using a small set of labeled data. This paper considers relation adaptation, where a relation ex- traction system trained on many source domains is adapted to a new target domain. There are at least three challenges when adapt- ing a relation extraction system to a new domain. First, the same semantic relation between two en- tities can be expressed using different lexical or syntactic patterns. For example, …

Content truncated to 500 words