22.5.1 Biological Named Entity Recognition

Information extraction tasks in the biological realm are characterized by a much wider range of relevant types of entities than the PERSON, ORGANIZATION, and LOCATION semantic classes that characterize work that is focused on news-style texts. Fig. 22.31 and the following example illustrate just a small subset of the variety of semantic classes of named entities that have been the target of NER systems in the biomedical domain.

[TISSUE Plasma] [GP BNP] concentrations were higher in both the [POPULATION judo] and [POPULATION marathon groups] than in [POPULATION controls], and positively correlated with [ANAT LV] mass as well as with deceleration time.

Nearly all of the techniques described in Sec. 22.1 have been applied to the biomedical NER problem, with a particular focus on the problem of recognizing gene/protein names. This task is particularly difficult due to the wide range of forms that gene names can take: white, insulin, BRCA1, ether a go-go, and breast cancer associated 1 are all the names of genes. The choice of algorithm for gene name recognition seems to be less important than the choice of features; typical feature sets include word-shape and contextual features, as discussed earlier; ad-

ditionally, knowledge-based features, such as using the count of Google hits for a sequence like BRCA1 gene to decide whether or not a token of the string BRCA1 is a reference to a gene or not, are sometimes incorporated into statistical systems.

Surprisingly, the use of huge publicly available lists of gene names has not generally contributed to the performance of a gene/protein NER system (Yeh et al., 2005), and in fact may actually degrade it (Jr. et al., 2006). It is not uncommon for gene names to be many tokens long (e.g. breast cancer associated 1). Gene name length has a demonstrable effect on NER system performance (Kinoshita et al., 2005; Yeh et al., 2005), and any technique for correctly finding the boundaries of multi-token names seems to increase performance. Use of the abbreviation-definition-detection algorithm (Schwartz and Hearst, 2003) is common for this purpose, since many such names appear as abbreviation or symbol definitions at some point in a publication. Base noun group chunkers can also be useful in this regard, as can a surprisingly small number of heuristic rules (Kinoshita et al., 2005).