Year: 2008
Type: article
Abstract: Entity Cruncher (EC) is an experimental Webas-corpus approach to NER (Named Entity Recognition). The task is to identify and classify names of persons, locations, and organisations, and possibly even more detailed classes. EC looks for frequencies of names in prespecified lexical contexts (thought to be excellent determiners of the class and identity of the candidate word) on the web, using the MSN Live API (msdn2.microsoft.com). The document frequencies of the candidate words in the prespecified contexts are the basis for deciding the classification of each Named Entity candidate. The system was tested on the evaluation data used for Dutch in the CoNLL (Conference for Computational Natural Language Learning) 2002 competition. Among other things, Dutch has a significant number of surnames that are names of places. Similarly, many organizations are named after persons (such as Heineken). This makes the task slightly different from using English, where the common names for places are more separated from names of people. EC got an F-score of 63.20 using little more than 30 lexical context patterns and a tiny selection of less than 400 finite state lexical phrase patterns. The results obtained with this rather simplistic, small scale, approach give grounds for optimism regarding both the use of corpus lookup for categorisation of names and words, and for the use of the web as a high coverage backdown alternative. 1
Authors: Audun Rømcke, Christer Johansson
Institution: University of Bergen
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1892653308