D. Lange et al., CIKM 2010

Citation

Dustin Lange, Christoph Böhm, Felix Naumann. 2010. Extracting structured information from Wikipedia articles to populate infoboxes. In CIKM '10 Proceedings of the 19th ACM international conference on Information and knowledge management.

Online version

ACM Digital Library

Summary

This is a paper introducing iPopulator system, which automatically populates infoboxes of Wikipedia articles by extracting attribute values from the article's text. (Also known as Infobox completion problem.

System Architecture and Applied Methods

iPopulator's extraction workflow contains four steps:

Structure Analysis

For each attribute of the infobox template, the system analyzes its values given in the training articles' infoboxes to determine a structure that represents the attribute's syntactical characteristics. For example, the infobox_company attribute number_of_employees might be 12,500 (2003), which means that 12,500 people were employed in the year 2003. The authors employed a heuristic method to find out a regular expression to get values from infobox values.

Training Data Creation

This step is similar to the Matcher' in the Hoffmann et al., ACL 2010 paper, which matches the infobox attributes to correspond Wikipedia article. Because the attribute values usually do not occur verbatim in texts. The paper proposes some ideas to address the issue.

• Article Paragraph Filtering: Only the first few paragraphs will be chosen, ignore others;
• Fuzzy Matching: Some simple heuristic is applied to discover fuzzy matches of attribute values, and numbers are processed separately from all other strings;
• Labeling Value Parts: All attribute values are divided into several parts according to the corresponding attribute value structure. Each part of the value structure is labeled separately. To retain the identity of the value part that is being labeled, a number is assigned to each structure part and used as the actual label.

Value Extractor Creation

The labeled training data are used to generate extractors for as many attributes as possible. In this paper, authors manually design feature function $f$ for each attribute (seems like heuristically, not specified in the paper). And then employed Conditional Random Fields (CRFs) to generate attribute value extractors. These extractors are automatically evaluated, so that ineffective extractors can be discarded ($precision<0.75$).

Attribute Value Extraction

The extractors can then be applied to all articles to fnd missing attribute values for existing infoboxes.

The process of the steps mentioned above can be demostracted in the following figure:

Experimental Result

Overall Performance

Evaluating on 1,727 distinct infobox template attributes, the authors claims they achieve Precision $P\geq 0.8$ for 1,521 attributes and $P\geq 0.9$ for 1,127 attributes. The overall average measures is given precision with 0.91, Recall 0.66, giving a $F1=0.73$.

Comparison with related work

By comparing with related work Kylin and successor K2, the iPopulator system out perform Kylin's and K2's results, result is summarized by the table below:

Related papers

The experiment result is compared with Kylin Information Extraction system.

Similar problem of infobox population is also discussed in Wu and Weld ACL 2010's Open Information Extraction system and Hoffmann et al., ACL 2010's LUCHS system.

Result generated by iPopulator is also used the knowledge base of DBpedia.