Filter and Match Approach to Pair-wise Web URI Linking

ALTA Shared-task 2016: Filter and Match Approach to Pair-wise Web URI Linking S. Shivashankar Department of CIS, University of Melbourne shivashankar@student.unimelb.edu.au Yitong Li Department of CIS, University of Melbourne yitongl4@student.unimelb.edu.au Afshin Rahimi Department of CIS, University of Melbourne arahimi@student.unimelb.edu.au Abstract This paper describes the method and re- sults of our approach, submitted as team ‘NLPCruise’ at ALTA shared task 2016. The goal of the shared task is to predict whether two given web Uniform Resource Identifiers (URIs) correspond to the same entity or not. Retrieving the URI con- tent in addition to the dataset provided, we built a two stage filter and match technique that utilises search engine scores, seman- tic similarity and machine translation fea- tures. Our model achieved an F1 score of 0.85714 on the public test-set and ranked second finally on the private leaderboard. 1 Introduction In general, establishing a mapping from entities in a knowledge base to URI end-points is a use- ful task both to collate information about entities and to disambiguate them. Typically, semantic sources such as Wikipedia, DBPedia are used as end-points. Although they provide a rich con- text for entities, they do not achieve sufficient recall over different domains and entities. On the other hand, domain-specific sources such as DBLP, IMDb or MusicBrainz cover only a sin- gle target domain (e.g. movies, music) well but aren’t useful in other domains. To benefit from both general purpose and domain-specific knowl- edge bases, one could use WebKB (URIs) as end- points. In-order to do that one must infer its ex- istence on the web. For every entity endpoint we discover, we may recover thousands of entity men- tions via inlinks. While the effectiveness of inlink- driven entity disambiguation is known for a single KB setting, this can be extended to leverage in- links across a collection of automatically discov- ered web KBs (Chisholm et al., 2016). Thus in this task, we classify whether a pair of URIs cor- respond to the same entity or not. Similar tasks were addressed in other shared tasks/challenges, such as Web People Search task (WePS 1), defined as the problem of organizing web search results for a given person name. At a more generic level, TREC relevance feedback track2 had multiple tasks related to relevance clas- sification for a set of documents given a query, which we can treat as similar problems to ours when formulated as pair-wise binary document classification. toThe goal of ALTA shared task 2016 is determine whether two URLs refer to the same underlying entity or not. For ex- ample, http://www.nytimes.com/ topic/person/barack-obama https://twitter.com/BarackObamaand refer to the same person, but https: //twitter.com/realDonaldTrump and https://www.instagram.com/ ivankatrump refer to different entities. 2 Data preparation The original data includes URLs of entities which link to the HTML pages, titles of the HTML pages, and snippets or brief descriptions of pages. Ti- tles and snippets were obtained using BING search API, where the input query was the named entity. The top two URLs from BING search are the URL pairs …

Content truncated to 500 words