Linking, Searching, and Visualizing Entities in Wikipedia

Linking, Searching, and Visualizing Entities in Wikipedia Marcus Klang, Pierre Nugues Lund University, Department of Computer science, Lund, Sweden marcus.klang@cs.lth.se, pierre.nugues@cs.lth.se Abstract In this paper, we describe a new system to extract, index, search, and visualize entities in Wikipedia. To carry out the entity extraction, we designed a high-performance, multilingual, entity linker and we used a document model to store the resulting linguistic annotations. The entity linker, HEDWIG, extracts the mentions from text using a string matching engine and links them to entities with a combination of statistical rules and PageRank. The document model, Docforia (Klang and Nugues, 2017), consists of layers, where each layer is a sequence of ranges describing a specific annotation, here the entities. We evaluated HEDWIG with the TAC 2016 data and protocol (Ji and Nothman, 2016) and we reached the CEAFm scores of 70.0 on English, on 64.4 on Chinese, and 66.5 on Spanish. We applied the entity linker to the whole collection of English and Swedish articles of Wikipedia and we used Lucene to index the layers and a search module to interactively retrieve all the concordances of an entity in Wikipedia. The user can select and visualize the concordances in the articles or paragraphs. Contrary to classic text indexing, this system does not use strings to identify the entities but unique identifiers from Wikidata. A demonstration of the entity search and visualization will be available for English at this address http://vilde.cs.lth.se:9001/en-hedwig/ and for Swedish at: http://vilde.cs.lth.se:9001/sv-hedwig/. Keywords: named entity recognition, entity linker, wikipedia 1. Introduction Wikipedia has become a popular NLP resource used in many projects such as text categorization (Wang et al., 2009), information extraction, question answering (Fer- rucci, 2012), or translation (Smith et al., 2010). In addition to its size and diversity, Wikipedia, through its links, also enables to create a graph that associates concepts, entities, and their mentions in text. Wu and Weld (2010), for in- stance, used the “wikilinks”, the Wikipedia hyperlinks, to collect the mentions of an entity and build sets of synonyms for an open information extraction system. However, according to the edition rules of Wikipedia, only the first mention of an entity should be linked in an article. An automatic wikification is then necessary to associate the subsequent mentions with an entity (Mihalcea and Csomai, 2007). In addition, searching entities using names in the form of strings can be tricky as names are sometimes am- biguous and entities may have more than one name. Find- ing all the occurrences of an organization like the United Nations would require five or more queries as they can be mentioned not only as the United Nations, but also as: UN, U.N., United Nations Organization, UNO, etc. In this paper, we describe a novel multilingual system to process, index, search, and visualize all the mentions of an entity in Wikipedia. This system consists of an entity linker, HEDWIG, that extracts the mentions from text using a named-entity recognition engine and links them to entities with a combination of statistical rules …

Content truncated to 500 words