Generalized Expectation Criteria for Bootstrapping Extractors using Record-Text Alignment

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 131–140, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Generalized Expectation Criteria for Bootstrapping Extractors using Record-Text Alignment Kedar Bellare Dept. of Computer Science University of Massachusetts Amherst, MA 01003 kedarb@cs.umass.edu Andrew McCallum Dept. of Computer Science University of Massachusetts Amherst, MA 01003 mccallum@cs.umass.edu Abstract Traditionally, machine learning ap- proaches for information extraction require human annotated data that can be costly and time-consuming to produce. However, in many cases, there already exists a database (DB) with schema related to the desired output, and records related to the expected input text. We present a conditional random field (CRF) that aligns tokens of a given DB record and its realization in text. The CRF model is trained using only the available DB and unlabeled text with generalized expecta- tion criteria. An annotation of the text induced from inferred alignments is used to train an information extractor. We eval- uate our method on a citation extraction task in which alignments between DBLP database records and citation texts are used to train an extractor. Experimental results demonstrate an error reduction of 35% over a previous state-of-the-art method that uses heuristic alignments. 1 Introduction A substantial portion of information on the Web consists of unstructured and semi-structured text. Information extraction (IE) systems segment and label such text to populate a structured database that can then be queried and mined efficiently. In this paper, we mainly deal with information extraction from text fragments that closely resem- ble structured records. Examples of such texts include citation strings in research papers, con- tact addresses on person homepages and apart- ment listings in classified ads. Pattern match- ing and rule-based approaches for IE (Brin, 1998; Agichtein and Gravano, 2000; Etzioni et al., 2005) that only use specific patterns, and delimiter and font-based cues for segmentation are prone to fail- ure on such data because these cues are gen- erally not broadly reliable. Statistical machine learning methods such as hidden Markov models (HMMs) (Rabiner, 1989; Seymore et al., 1999; Freitag and McCallum, 1999) and conditional ran- dom fields (CRFs) (Lafferty et al., 2001; Peng and McCallum, 2004; Sarawagi and Cohen, 2005) have become popular approaches to address the text extraction problem. However, these methods require labeled training data, such as annotated text, which is often scarce and expensive to pro- duce. In many cases, however, there already exists a database with schema related to the desired out- put, and records that are imperfectly rendered in the available unlabeled text. This database can serve as a source of significant supervised guid- ance to machine learning methods. Previous work on using databases to train information extrac- tors has taken one of three simpler approaches. In the first, a separate language model is trained on each column of the database and these mod- els are then used to segment and label a given text sequence (Agichtein and Ganti, 2004; Can- isius and Sporleder, 2007). However, this ap- proach does not …

Content truncated to 500 words