Joint Learning of Entity Linking Constraints Using a Markov-Logic Network

Computational Linguistics and Chinese Language Processing Vol. 19, No. 1, March 2014, pp. 11-32 11  The Association for Computational Linguistics and Chinese Language Processing Joint Learning of Entity Linking Constraints Using a Markov-Logic Network Hong-Jie Dai, Richard Tzong-Han Tsai, and Wen-Lian Hsu Abstract Entity linking (EL) is the task of linking a textual named entity mention to a knowledge base entry. Traditional approaches have addressed the problem by dividing the task into separate stages: entity recognition/classification, entity filtering, and entity mapping, in which different constraints are used to improve the system’s performance. Nevertheless, these constraints are executed separately and cannot be used interactively. In this paper, we propose an integrated solution to the task based on a Markov logic network (MLN). We show how the stage decision can be formulated and combined in an MLN. We conducted experiments on the biomedical EL task, gene mention linking (GML), and compared our model’s performance with those of two other GML approaches. Our experimental results provide the first comprehensive GML evaluations from three different perspectives: article-wide precision/recall/F-measure (PRF), instance-based PRF, and question answering accuracy. This paper also provides formal definitions of all of the above EL tasks. Experimental results show that our method outperforms the baseline and state-of-the-art systems under all three evaluation schemes. Keywords: Entity Linking, Entity Disambiguation, Markov Logic Network, Gene Normalization 1. Introduction Developing a system that can identify entities, such as personal names and gene or disease mentions, and that can classify the relations between them is useful for several applications in natural language processing and knowledge acquisition. There are several possible uses for  Graduate Institute of BioMedical Informatics, Taipei Medical University E-mail: hjdai@tmu.edu.tw  Dep. of Computer Science & Information Engineering, National Central University E-mail: thtsai@csie.ncu.edu.tw  Institute of Information Science, Academia Sinica E-mail: hsu@iis.sinica.edu.tw 12 Hong-Jie Dai et al. such a system in different fields, e.g., improving document retrieval for specific entities, relation extraction, and attribute assignment (e.g., gene ontology annotations). In these applications, recognized entities must be linked to unique database entries. McNamee and Dang (2009b) named the task of matching a textual entity mention to a knowledge base (KB) entry Entity Linking (EL). In Figure 1, we provide a biomedical abstract to illustrate this task. The abstract discusses the relationship of the gene “CD59” to other lymphocyte antigens. TITLE: Structure of the CD59-encoding gene: further evidence of a relationship to murine lymphocyte antigen Ly-6 protein ABSTRACT: The gene for CD59 [membrane inhibitor of reactive lysis (MIRL), protectin], a phosphatidylinositol-linked surface glycoprotein that regulates the formation of the polymeric C9 complex of complement and that is deficient on the abnormal hematopoietic cells of patients with paroxysmal nocturnal hemoglobinuria, consists of four exons spanning 20 kilobases. … PMID [1381503] Figure 1. An example of entity linking. After EL, the gene mention “CD59” in the first sentence must be linked to ID 966 in the Entrez Gene database of PubMed. In the first sentence, the authors also listed other designations of the gene, including “membrane inhibitor …

Content truncated to 500 words