Mining Inter-Entity Semantic Relations Using Improved Transductive Learning

Mining Inter-Entity Semantic Relations Using Improved Transductive Learning Zhu Zhang School of Information and Department of EECS, University of Michigan, Ann Arbor, MI 48105, U.S.A Abstract. This paper studies the problem of mining relational data hidden in natural language text. In particular, it approaches the relation classification problem with the strategy of transductive learning. Dif- ferent algorithms are presented and empirically evaluated on the ACE corpus. We show that transductive learners exploiting various lexical and syntactic features can achieve promising classification performance. More importantly, transductive learning performance can be significantly im- proved by using an induced similarity function. 1 Introduction The world today is full of various information sources, with different ways of representing the same information. One common problem that arises in the data management community is that data existing in one format may be needed in a different format for another purpose. An instance of this general problem is that relational data don’t always exist in the form of relational tables; lots of them are hidden in natural language text. For example, (author, book) pairs can be instantiated as “. . . Shakespeare’s famous work Hamlet . . . ” or “. . . A Brief History of Time was written by Stephen Hawking . . . ” in text. On the other hand, within the information retrieval and natural language processing community, Information Extraction (IE) systems are understood as techniques for automatically extracting information from text, specifically, iden- tifying relevant information (usually of pre-defined types) from text documents in a certain domain. Once extracted, the information can be used for purposes such as database population and text indexing. While significant progress has been made in IE research, stimulated in particular by the Message Understand- ing Conferences (MUC) 1 and the recent ACE (Automatic Content Extraction) program 2 organized by the LDC (Linguistic Data Consortium), it is generally agreed that many barriers exist to the wider use of IE technologies due to the difficulties in adapting systems to new applications and domains. Keeping track of dynamic information sources (e.g., web pages) is challenging as well. To address these challenges, there has been a recent trend shift in the research community from knowledge-based approaches to machine learning techniques. Moreover, due to the cost related to acquiring large amount of labeled training 1 http://www.itl.nist.gov/iaui/894.02/related projects/muc/ 2 http://www.ldc.upenn.edu/Projects/ACE/ R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 402–413, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 Mining Inter-Entity Semantic Relations 403 data, researchers have been looking at various learning algorithms exploiting cheaply available unlabeled data (usually in much larger amounts), which aim at minimizing the need for labeled data while still achieving comparable results. According to the scope of ACE, current IE research has three main objectives: Entity Detection and Tracking (EDT), Relation Detection and Characterization (RDC), and Event Detection and Characterization (EDC). This study focuses on the second subproblem, RDC. In particular, the goal is to automatically classify binary relations between entities, i.e., to decide in which relational table to …

Content truncated to 500 words