|Title||Comparison of character-level and part of speech features for name recognition in biomedical texts|
|Publication Type||Journal Article|
|Year of Publication||2004|
|Authors||Collier, N, Takeuchi, K|
|Journal||Journal of Biomedical Informatics|
The immense volume of data which is now available from experiments in molecular biology has led to an explosion in reported results most of which are available only in unstructured text format. For this reason there has been great interest in the task of text mining to aid in fact extraction, document screening, citation analysis, and linkage with large gene and gene-product databases. In particular there has been an intensive investigation into the named entity (NE) task as a core technology in all of these tasks which has been driven by the availability of high volume training sets such as the GENIA v3.02 corpus. Despite such large training sets accuracy for biology NE has proven to be consistently far below the high levels of performance in the news domain where F scores above 90 are commonly reported which can be considered near to human performance. We argue that it is crucial that more rigorous analysis of the factors that contribute to the model's performance be applied to discover where the underlying limitations are and what our future research direction should be. Our investigation in this paper reports on variations of two widely used feature types, part of speech (POS) tags and character-level orthographic features, and makes a comparison of how these variations influence performance. We base our experiments on a proven state-of-the-art model, support vector machines using a high quality subset of 100 annotated MEDLINE abstracts. Experiments reveal that the best performing features are orthographic features with F score of 72.6. Although the Brill tagger trained in-domain on the GENIA v3.02p POS corpus gives the best overall performance of any POS tagger, at an F score of 68.6, this is still significantly below the orthographic features. In combination these two features types appear to interfere with each other and degrade performance slightly to an F score of 72.3.