IRST-BP: Web People Search Using Name Entities

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 195–198, Prague, June 2007. c⃝2007 Association for Computational Linguistics IRST-BP: Web People Search Using Name Entities Octavian Popescu FBK-irst, Trento (Italy) popescu@itc.it Bernardo Magnini FBK-irst, Trento (Italy) magnini@itc.it Abstract In this paper we describe a person clus- tering system for web pages and report the results we have obtained on the test set of the Semeval 2007 Web Person Search task. Deciding which particular person a name refers to within a text document depends mainly on the capac- ity to extract the relevant information out of texts when it is present. We con- sider “relevant” here to stand primarily for two properties: (1) uniqueness and (2) appropriateness. In order to address both (1) and (2) our method gives pri- mary importance to Name Entities (NEs), defined according to the ACE specifications. The common nouns not referring to entities are considered fur- ther as coreference clues only if they are found within already coreferred docu- ments. 1 Introduction Names are ambiguous items (Artiles, Gonzalo and Sekine 2007). As reported on an experiment carried out on an Italian news corpus (Magnini et all 2006) within a 4 consecutive days from a local newspaper the perplexity is 56% and 14% for first and last name respectively. Deciding which particular person a name refers to within a text document depends mainly on the capacity to extract the relevant information out of texts when it is present1. We consider “relevant” here to stand primarily for two properties: (1) uniqueness and (2) appropriateness. A feature is unique as long as it appears only with one per- son. Consider a cluster of web pages that charac- terizes only one person. Many of the N-grams in this cluster are unique compared to other cluster. Yet the uniqueness may come simply from the sparseness. Appropriateness is the property of an N-gram to characterize that person. Uniqueness may be assured by ontological properties (for example, “There is a unique president of a republic at a definite moment of time”, “Alberta University is in Canada). How- ever, the range of ontological information we are able to handle is quite restricted and we are not able to realize the coreference solely relying on them. Uniqueness may be assured by estimating a very unlike probability of the occurrence of certain N-grams for different persons (as, for example, “Dekang Lin professor Alberta Canada Google”). Appropriateness is a difficult issue because of two reasons: (a) it is a dynamic feature (b) it is hard to be localized and extracted from text. The greatest help comes from the name of the page, when it happens to be a suggestive name such as “homepage”, “CV”, “resume” or “about”. Gene- 1 It is very difficult to evaluate whether the informa- tion allowing the coreference of two instances of a (same) name is present in a web page or news. A crude estimation on our news corpus for the names occurring between 6-20 times, which represent 8% of …

Content truncated to 500 words