Disambiguating Conjunctions in Named Entities

Proceedings of the Australasian Language Technology Workshop 2005, pages 7–14, Sydney, Australia, December 2005. Disambiguating Conjunctions in Named Entities Pawe�l MAZUR Institute of Applied Informatics Wroc�law University of Technology Wyb. Wyspia´nskiego 27 50-370 Wroc�law, Poland Pawel.Mazur@pwr.wroc.pl Robert DALE Centre for Language Technology Macquarie University NSW 2109, Sydney, Australia Robert.Dale@mq.edu.au Abstract The recognition of named entities is now a well- developed area, with a range of symbolic and machine learning techniques that deliver high accuracy identification and categorisation of a variety of entity types. However, there are still some named entity phenomena that present problems for existing techniques; in particular, relatively little work has explored the disam- biguation of conjunctions appearing in candi- date named entity strings. We demonstrate that there are in fact four distinct uses of con- junctions in the context of named entities; we present the results of some experiments using machine-learned classifiers to disambiguate the different uses of the conjunction, with 81.73% of test examples being correctly classified. We pro- vide some discussion and analysis of the prob- lem of conjunction in named entities, and we show that there are some cases which are am- biguous even for humans. 1 Introduction Initially developed as a component task in infor- mation extraction (see, for example, (Grishman and Sundheim, 1996)), named entity recogni- tion, whereby entities such as people, organi- zations and geographic locations are identified and tracked in texts, has become an important part of other natural language processing appli- cations such as question answering, text sum- marisation and machine translation. Named entity recognition consists of both identifying those strings in a text that corre- spond to named entities, and then classifying each such named entity string as being of a spe- cific type, with typical categories being Com- pany, Person and Location. Sometimes an ad- ditional step is introduced for coreference res- olution, which establishes whether two named entities in a given document refer to the same (either physical or abstract) object; so, for ex- ample, we might determine that the named en- tity International Business Machines Corpora- tion has the same real world referent as both IBM and Big Blue. We can go further and de- termine whether named entities in separate doc- uments (hence, cross-document coreference res- olution) refer to the same entities (see (Bagga, 2004)), although this is a much less explored area. All of the above assumes that identifying in- dividual named entities in text is a relatively straightforward and well-defined task. How- ever, although there are reported high perfor- mance figures for named entity identification and classification in general, there are some cat- egories of named entities that remain problem- atic. We will refer to those strings in a text that may correspond to named entities as can- didate named entity strings; one type of candidate named entity string that is problem- atic, from a surface linguistic perspective, is the category that contains conjunctions. Consider the string Australia and New Zealand Banking Group Limited: in the absence of an appropri- ate domain …

Content truncated to 500 words