Tables are widely used in Wikipedia articles to display relational information { they are inherently concise and information rich. However, aside from info-boxe s, there are no automatic methods to exploit the integrated content of these tables. We thus present DRETa: a tool that uses DBpedia as a reference knowledge-base to extract RDF triples from generic Wikipedia tables.
Large amounts of data on the Web are presented in tables [
Instead of interpreting generic Web tables, we have rather been focussing on (partially) interpreting the tables embedded in Wikipedia (henceforth \Wikitables"). In particular, we have created DRETa: a prototype for (semi)automatically performing a best-e ort extraction of RDF triples from Wikitables. Though many of the challenges remain the same, focussing on Wikitables has a number of distinct advantages over the more general Web table scenario: (1) Wikitable cells often contain links to Wikipedia articles that disambiguate the entities being talked about; (2) Wikitables contain a high ratio of rich encyclopaedic knowledge; (3) the context of a Wikitable can be mapped to the article in which it appears; (4) existing RDF knowledge-bases, that o er partial exports of Wikipedia content, can be used for reference and for mining legacy entity URIs and predicates. To maximise the precision of the triples extracted from Wikitables, DRETa exploits these unique advantages insofar as possible.
The DRETa system works by using a suitable reference
knowledgebase|such as DBpedia [
We sketch the extraction process by way of a real-world example. Figure 1 presents a Wikitable abridged from the \Manchester United F.C." Wikipedia article, containing relations between players, their shirt number, country and position. There are also relations between players and the entity described by the article (their current club is Manchester United F.C.).
Aside from the No. columns, the cells of the table contain hyperlinks to other articles in Wikipedia, including countries, football positions, and individual players. For example, the ags link to articles for the country; GK links to the article for Goalkeeper (associated football). These links provide unambiguous referents to Wikipedia entities, which can in turn be mapped directly to DBpedia entities and descriptions. Currently, DRETa focuses on the extraction of relations between cells containing wiki-links and does not consider plain-string values. For example, it would not try to extract player numbers from the table.
Previous works on extracting RDF from such tables (e.g., [
Candidate triples that we extract are further associated with a number of features to help classify them as correct/incorrect, associating each triple with a con dence score. Details of the features are out-of-scope, but, for example, we hypothesise that the more rows a given relation holds for across entities in two xed columns in the reference KB, the higher the likelihood that that relationship exists on all such rows. Other features, such as a match between the label of the candidate relation and a column header, can further strengthen con dence in the match. Using a selection of 750 random triples labelled by three judges, we employed o ine various machine learning methods (SVM, Nave Bayes, Bagging Decision Trees, Random Forest, Logistic) to train a range of binary classi ers that are then made available to the DRETa system for classifying (in)correct triples, and ultimately for ranking and ltering candidate triples at runtime. 3
The current DRETa prototype works with DBpedia 3.8 and Wikipedia article names (for auto-completion) as last updated in May 2013. The user submits a Wikipedia article title (with the help of auto-complete) and optionally selects a classi er. The extraction process is then as follows: i) the selected article is downloaded and cached in memory for future queries; ii) all the tables are extracted, repaired (as applicable), and ltered; iii) for each table, mappings from wiki-links to KB entities are executed; iv) candidate relations for pairs of resources are collected from the reference KB and candidate triples proposed; v) the selected classi er is run to rank triples by con dence, vi) each candidate triple is tested against the KB to determine if it is a pre-existing triple or not.
Following our motivating example, Figure 2 presents DRETa's results for the Wikipedia article \Manchester United F.C.". After ca. 11 secs. we get 457 RDF triples extracted from tables contained in that article as a result. Some of these triples already exist in DBpedia (rows with an DBpedia icon visible). Others are novel: from the top-10 extracted triples sorted by con dence, for example, we extract triples for the birth-places of the footballers Rafael da Silva (Brazil) and Phil Jones (England), which were not previously known to DBpedia. 4
In this paper, we have presented DRETa: a prototype system for extracting RDF triples from Wikipedia tables. The process maps entities in table cells to entities in a reference knowledge-base and then looks for potential relations that hold between entities on the same row across two given columns, or that hold between entities in a single column and the article entity. Triples are associated with a set of features that help classify correct/incorrect triples. A selection of machine learning methods are used o ine to train classi ers, where these classi ers can then be used to rank the con dence of triples. The prototype is available online at http://deri-srvgal36.nuigalway.ie:8080/wikitables-demo-0.1.0/.
We are currently investigating methods to perform a high-quality \bulk" tripli cation of all tables in English Wikipedia. We have used the architecture of DRETa to extract 22 million triples from over one million tables (all tables) in English Wikipedia. We have already estimated a 52% precision measure for the raw candidate triples extracted by our methods and our next steps are to evaluate the extent to which machine learning techniques and di erent classi ers improve this baseline precision by ltering incorrect triples at various thresholds, and we will identify a gold standard to be able to estimate recall. We also wish generalize our methods with YAGO2 and Freebase as reference knowledge-bases, towards a more ambitious generalisation that employs entity-recognition tools (instead of wiki-links) for processing generic Web tables.