Robust Similarity Measures for Named Entities Matching

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 593–600 Manchester, August 2008 Robust Similarity Measures for Named Entities Matching Erwan Moreau1 Institut T´el´ecom ParisTech & LTCI CNRS erwan.moreau@enst.fr Franc¸ois Yvon Univ. Paris Sud & LIMSI CNRS yvon@limsi.fr Olivier Capp´e Institut T´el´ecom ParisTech & LTCI CNRS cappe@enst.fr Abstract Matching coreferent named entities with- out prior knowledge requires good similar- ity measures. Soft-TFIDF is a fine-grained measure which performs well in this task. We propose to enhance this kind of met- rics, through a generic model in which measures may be mixed, and show experi- mentally the relevance of this approach. 1 Introduction In this paper, we study the problem of matching coreferent named entities (NE in short) in text col- lections, focusing primarily on orthographic vari- ations in nominal groups (we do not handle the case of pronominal references). Identifying textual variations in entities is useful in many text min- ing and/or information retrieval tasks (see for ex- ample (Pouliquen et al., 2006)). As described in the literature (e.g. (Christen, 2006)), textual dif- ferences between entities are due to various rea- sons: typographical errors, names written in dif- ferent ways (with/without first name/title, etc.), abbreviations, lack of precision in organization names, transliterations, etc. For example, one wants “Mr. Rumyantsev” to match with “Alexan- der Rumyanstev” but not with “Mr. Ryabev”. Here we do not address the related problem of dis- ambiguation2 (e.g. knowing whether a given oc- currence of “George Bush” refers to the 41st or 43rd president of the USA), because it is techni- cally very different from the matching problem. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. 1Now at LIPN - Univ. Paris 13 & UMR CNRS 7030. 2Which is essential in the Web People Search task. There are different ways to tackle the problem of NE matching: the first and certainly most reli- able one consists in studying the specific features of the data, and then use any available tool to de- sign a specialized method for the matching task. This approach will generally take advantage of language-specific (e.g. in (Freeman et al., 2006)) and domain-specific knowledge, of any external resources (e.g. database, names dictionaries, etc.), and of any information about the entities to pro- cess, e.g. their type (person name, organization, etc.), or internal structure (e.g. in (Prager et al., 2007)). In such an in-depth approach, supervised learning is helpful: it has been used for example in a database context3 in (Bilenko et al., 2003), but this approach requires labeled data which is usu- ally costly. All those data specific appproaches would necessitate some sort of human expertise. The second approach is the robust one: we propose here to try to match any kind of NE, extracted from “real world” (potentially noisy) sources, without any kind of prior knowledge4. One looks for coreferent NE, whatever their type, source, language5 or quality6. Such robust simi- larity methods may be useful for a lot of …

Content truncated to 500 words