N-ary Biographical Relation Extraction using Shortest Path Dependencies

N-ary Biographical Relation Extraction using Shortest Path Dependencies Gitansh Khirbat Jianzhong Qi Rui Zhang Department of Computing and Information Systems The University of Melbourne Australia gkhirbat@student.unimelb.edu.au {jianzhong.qi, rui.zhang}@unimelb.edu.au Abstract Modern question answering and summa- rizing systems have motivated the need for complex n-ary relation extraction sys- tems where the number of related entities (n) can be more than two. Shortest path dependency kernels have been proven to be effective in extracting binary relations. In this work, we propose a method that employs shortest path dependency based rules to extract complex n-ary relations without decomposing a sentence into con- stituent binary relations. With an aim of extracting biographical entities and re- lations from manually annotated datasets of Australian researchers and department seminar mails, we train an information ex- traction system which first extracts enti- ties using conditional random fields and then employs the shortest path dependency based rules along with semantic and syn- tactic features to extract n-ary affiliation relations using support vector machine. Cross validation of this method on the two datasets provides evidence that it outper- forms the state-of-the-art n-ary relation ex- traction system by a margin of 8% F-score. 1 Introduction Information extraction (IE) is the process of ex- tracting factual information from unstructured and semi-structured data and storing it in a structured queryable format. Two important components of an IE system are entity extraction and relation ex- traction. These components are sequential and to- gether form the backbone of a classic IE system. Entity extraction systems have achieved a high ac- curacy in identifying certain entities such as men- tion of people, places and organizations (Finkel et al., 2005). However, such named entity recogni- tion (NER) systems are domain-dependent and do not scale up well to generalize across all entities. Relation extraction systems utilize the identi- fied entities to extract relations among them. Past two decades have witnessed a significant advance- ment in extracting binary domain-dependent re- lations (Kambhatla, 2004), (Zhao and Grishman, 2005) and (Bunescu and Mooney, 2005a). How- ever, modern question answering and summariz- ing systems have triggered an interest in captur- ing detailed information in a structured and se- mantically coherent fashion, thus motivating the need for complex n-ary relation extraction systems (where the number of entities, n ≥ 2). Some no- table n-ary relation extraction systems are (Mc- Donald et al., 2005) and (Li et al., 2015). Mc- Donald et al. (2005) factorized complex n-ary re- lation into binary relations, representing them in a graph and tried to reconstruct the complex relation by making tuples from selected maximal cliques in the graph. While they obtained reasonable preci- sion and recall using a maximum entropy binary classifier on a corpus of 447 selected abstracts from MEDLINE, they have not explored the con- stituency and dependency parse features which have been proven to be efficient in relation extrac- tion. Li et al. (2015) make use of lexical semantics to train a model based on distant-supervision for n- ary relation extraction. However, the applicability of this …

Content truncated to 500 words