Attribute Relation Extraction from Template-inconsistent Semi-structured Text by Leveraging Site-level Knowledge

International Joint Conference on Natural Language Processing, pages 1097–1101, Nagoya, Japan, 14-18 October 2013. Attribute Relation Extraction from Template-inconsistent Semi-structured Text by Leveraging Site-level Knowledge Yang Liu, Fang Liu, Siwei Lai, Kang Liu, Guangyou Zhou, Jun Zhao National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences {yang.liu,fliu,swlai,kliu,gyzhou,jzhao}@nlpr.ia.ac.cn Abstract A variety of methods have been proposed for attribute-value extraction from semi- structured text with consistent templates (strict semi-text). However, when the tem- plates in semi-structured text are inconsis- tent (weak semi-text), these methods will work poorly. To overcome the template- inconsistent problem, in this paper, we proposed a novel method to leverage site- level knowledge for attribute-value extrac- tion. First, we use a graph-based random walk model to acquire site-level knowl- edge. Then we utilize such knowledge to identify weak semi-text in each page and extract attribute-value pairs. The experi- ments show that, comparing to the base- line method which does not utilize site- level knowledge, our method can improve the extraction performance significantly. 1 Introduction Among types of relations, attributes (e.g. nation- ality, date of birth) have emerged as one of the most popular types (Alfonseca et al., 2010), as they capture properties of respective objects (or instances) (e.g. Kobe Bryant). Generally, an at- tribute relation consists of an object, an attribute and its associated value (e.g. Kobe Bryant - date of birth - August 23, 1978, where “August 23, 1978” is the value of “date of birth”). In this paper, we call such a relation an object-attribute-value (OAV) tuple. Many methods have been proposed to extract attributes from semi-structured text (Ca- farella et al., 2008)(Venetis et al., 2011)(Crescenzi et al., 2001)(Arasu and Garcia-Molina, 2003) and unstructured text like webpages and Web search query logs (Reisinger and Pas¸ca, 2009)(Pas¸ca et al., 2010)(Pasca and Van Durme, 2007). Semi- structured text (strict semi-text) often has dis- tinctive HTML tags and consistent templates like HTML tables (eg: Wikipedia infoboxes). How- ever, a lot of user-generated semi-structured text with weak structures exist, where their templates generating records are inconsistent and the HTML tags in these templates are less distinctive. In this paper, we focus on the issue of extracting attribute- value (AV) pairs from semi-structured text with in- consistent templates (weak semi-text). In previous work, Yoshinaga and Torisawa (Yoshinaga and Torisawa, 2007) extracted AV pairs of given objects from semi-structured text. They induced templates via a set of attributes ob- tained beforehand and used the templates to ex- tract AV pairs. There are two constraints of their method. First, it heavily depends on the initial set of attributes. However, the quality and coverage of the initial set of attributes is hard to control. Sec- ond, they hold the assumption that attributes in the same block of semi-structured text are generated with the same template which weak semi-text does not satisfy. Their method mainly concentrated on extraction from semi-structured text with consis- tent templates (strict semi-text). When facing the weak semi-text with inconsistent templates, it will fail to obtain satisfactory results. To …

Content truncated to 500 words