DBpedia Abstracts: A Large-Scale, Open, Multilingual NLP Training Corpus

DBpedia Abstracts: A Large-Scale, Open, Multilingual NLP Training Corpus 1Martin Brümmer, 1,2Milan Dojchinovski, 1Sebastian Hellmann 1AKSW, InfAI at the University of Leipzig, Germany 2Web Intelligence Research Group, FIT, Czech Technical University in Prague {bruemmer,dojchinovski,hellmann}@informatik.uni-leipzig.de Abstract The ever increasing importance of machine learning in Natural Language Processing is accompanied by an equally increasing need in large-scale training and evaluation corpora. Due to its size, its openness and relative quality, the Wikipedia has already been a source of such data, but on a limited scale. This paper introduces the DBpedia Abstract Corpus, a large-scale, open corpus of annotated Wikipedia texts in six languages, featuring over 11 million texts and over 97 million entity links. The properties of the Wikipedia texts are being described, as well as the corpus creation process, its format and interesting use-cases, like Named Entity Linking training and evaluation. Keywords: training, dbpedia, corpus, named entity recognition, named entity linking, nlp 1. Introduction Wikipedia is the most important and comprehensive source of open, encyclopedic knowledge. The English Wikipedia alone features over 5 million articles, representing a vast source of openly licensed natural language text. Besides the Wikipedia web site, Wikipedia data is avail- able via a RESTful API as well as complete XML dumps. However, API access is officially limited to one request per second, prohibiting a web scraping approach to acquire the data. Even after downloading the XML dump files, con- suming and processing them is a complicated task, further hindering large-scale data extraction. To alleviate these issues, the DBpedia project (Auer et al., 2008) has been extracting, mapping, converting and publishing Wikipedia data since 2007. The main focus of the DBpedia extrac- tion lies in mapping of info boxes, templates and other eas- ily identified structured data found in Wikipedia articles to properties of an ontology. Article texts and the data they may contain are not focused in the extraction process, al- though they are the largest part of most articles in terms of time spent on writing, informational content and size. Only the text of the first introductory section up until the first heading of the articles is extracted and contained in the DBpedia, and called abstract. Links inside the articles are only extracted as an unordered bag, showing only an un- specified relation between the article that contains the link and the articles that are being linked, but not where in the text the linked article was mentioned or what its surface form was. But because the links are set by the contributors to Wikipedia themselves, they represent entities in the text that were intellectually disambiguated by URL. This prop- erty makes extracting the abstracts including the links and their exact position in the text an interesting opportunity to create a corpus usable for, among other cases, NER and NEL algorithm training and evaluation. This paper describes the creation of a large-scale, multi- lingual Wikipedia abstract corpus that contains Wikipedia abstract texts, entity mentions occuring in the text, their po- sition, surface form as well as …

Content truncated to 500 words