Using Corpus-derived Name Lists for Named Entity Recognition

Using Corpus-derived Name Lists for Named Entity Recognition Mark Stevenson and Robert Gaizauskas Department of Computer Science, University of Sheffield Regent Court, 211 Portobello Street, Sheffield S1 4DP United Kingdom {marks, robertg}~dcs, shef. ac.uk Abstract This paper describes experiments to establish the performance of a named entity recognition system which builds categorized lists of names from manu- ally annotated training data. Names in text are then identified using only these lists. This approach does not perform as well as state-of-the-art named en- tity recognition systems. However, we then show that by using simple filtering techniques for improv- ing the automatically acquired lists, substantial per- formance benefits can be achieved, with resulting F- measure scores of 87% on a standard test set. These results provide a baseline against which the con- tribution of more sophisticated supervised learning techniques for NE recognition should be measured. 1 Introduction Named entity (NE) recognition is the process of identifying and categorising names in text. Systems which have attempted the NE task have, in general, made use of lists of common names to provide clues. Name lists provide an extremely efficient way of re- cognising names, as the only processing required is to match the name pattern in the list against the text and no expensive advanced processing such as full text parsing is required. However, name lists are a naive method for recognising names. McDonald (1996) defines internal and external evidence in the NE task. The first is found within the name string itself, while the second is gathered from its context. For example, in the sentence "President Washington chopped the tree" the word "President" is clear ex- ternal evidence that "Washington" denotes a person. In this case internal evidence from the name cannot conclusively tell us whether "Washington" is a per- son or a location ("Washington, DC"). A NE sys- tem based solely on lists of names makes use of only internal evidence and examples such as this demon- strate the limitations of this knowledge source. Despite these limitations, many NE systems use extensive lists of names. Krupke and Hausman (1998) made extensive use of name lists in their sys- tem. They found that reducing their size by more than 90% had little effect on performance, conversely adding just 42 entries led to improved results. This implies that the quality of list entries is a more im- portant factor in their effectiveness than the total number of entries. Mikheev et al. (1999) experi- mented with different types of lists in an NE system entered for MUC7 (MUC, 1998). They concluded that small lists of carefully selected names are as effective as more complete lists, a result consistent with Krupke and Hausman. However, both studies altered name lists within a larger NE system and it is difficult to tell whether the consistency of perform- ance is due to the changes in lists or extra, external, evidence being used to balance against the loss of internal evidence. In this paper a NE system which uses only …

Content truncated to 500 words