<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2007">
  <Title>Semi-supervised Relation Extraction with Label Propagation</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Relation extraction is the task of finding relationships between two entities from text. For the task, many machine learning methods have been proposed, including supervised methods (Miller et al., 2000; Zelenko et al., 2002; Culotta and Soresen, 2004; Kambhatla, 2004; Zhou et al., 2005), semi-supervised methods (Brin, 1998; Agichtein and Gravano, 2000; Zhang, 2004), and unsupervised method (Hasegawa et al., 2004).</Paragraph>
    <Paragraph position="1"> Supervised relation extraction achieves good performance, but it requires a large amount of manually labeled relation instances. Unsupervised methods do not need the definition of relation types and manually labeled data, but it is difficult to evaluate the clustering result since there is no relation type label for each instance in clusters. Therefore, semi-supervised learning has received attention, which can minimize corpus annotation requirement.</Paragraph>
    <Paragraph position="2"> Current works on semi-supervised resolution for relation extraction task mostly use the bootstrapping algorithm, which is based on a local consistency assumption: examples close to labeled examples within the same class will have the same labels. Such methods ignore considering the similarity between unlabeled examples and do not perform classification from a global consistency viewpoint, which may fail to exploit appropriate manifold structure in data when training data is limited. The objective of this paper is to present a label propagation based semi-supervised learning algorithm (LP algorithm) (Zhu and Ghahramani, 2002) for Relation Extraction task. This algorithm works by representing labeled and unlabeled examples as vertices in a connected graph, then propagating the label information from any vertex to nearby vertices through weighted edges iteratively, finally inferring the labels of unlabeled examples after the propagation process converges. Through the label propagation process, our method can make the best of the information of labeled and unlabeled examples to realize a global consistency assumption: similar examples should have similar labels. In other words, the labels of unlabeled examples are determined by considering not only the similarity between labeled and unlabeled examples, but also the similarity between unlabeled examples.</Paragraph>
  </Section>
class="xml-element"></Paper>