A convex relaxation for weakly supervised relation extraction

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1580–1590, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics A convex relaxation for weakly supervised relation extraction ´Edouard Grave EECS Department University of California, Berkeley grave@berkeley.edu Abstract A promising approach to relation extrac- tion, called weak or distant supervision, exploits an existing database of facts as training data, by aligning it to an unla- beled collection of text documents. Using this approach, the task of relation extrac- tion can easily be scaled to hundreds of different relationships. However, distant supervision leads to a challenging multi- ple instance, multiple label learning prob- lem. Most of the proposed solutions to this problem are based on non-convex formu- lations, and are thus prone to local min- ima. In this article, we propose a new approach to the problem of weakly su- pervised relation extraction, based on dis- criminative clustering and leading to a convex formulation. We demonstrate that our approach outperforms state-of-the-art methods on the challenging dataset intro- duced by Riedel et al. (2010). 1 Introduction Information extraction refers to the broad task of automatically extracting structured information from unstructured documents. An example is the extraction of named entities and the relations be- tween those entities from natural language texts. In the age of the world wide web and big data, information extraction is quickly becoming perva- sive. For example, in 2013, more than 130, 000 scientific articles were published about cancer. Keeping track with that quantity of information is almost impossible, and it is thus of utmost im- portance to transform the knowledge contained in this massive amount of documents into structured databases. Traditional approaches to information extrac- tion relies on supervised learning, yielding high Knowledge base r e1 e2 BornIn Lichtenstein New York City DiedIn Lichtenstein New York City Sentences Latent labels Roy Lichtenstein was born in New York City, into an upper- middle-class family. BornIn In 1961, Leo Castelli started displaying Lichtenstein’s work at his gallery in New York. None Lichtenstein died of pneumonia in 1997 in New York City. DiedIn Figure 1: An example of a knowledge database comprising two facts and training sentences ob- tained by aligning this database to unlabeled text. precision and recall results (Zelenko et al., 2003). Unfortunately, these approaches need large amount of labeled data, and thus do not scale well to the great number of different types of fact found on the Web or in scientific articles. A promising approach, called distant or weak supervision, is to exploit an existing database of facts as training data, by aligning it to an unlabeled collection of text documents (Craven and Kumlien, 1999). In this article, we are interested in weakly super- vised extraction of binary relations. A challenge pertaining to weak supervision is that the obtained training data is noisy and ambiguous (Riedel et al., 2010). Let us start with an example: if the fact Attended(Turing, King′s College) exists in the knowledge database and we observe the sen- tence …

Content truncated to 500 words