Table 5.

Impact of different context types on human gene mention normalization

Context type	Precision	Recall	F measure
Baseline: NER only	9.7	91.1	17.5
NER + GeneRifs	50.8	78.3	61.6
NER + GO terms	46.3	81.2	59.0
NER + EntrezGene summaries	49.0	66.7	56.5
NER + diseases	22.7	43.9	29.9
NER + functions	50.8	72.5	59.7
NER + keywords	53.0	53.6	53.3
NER + locations	74.2	14.8	24.7
NER + tissues	39.4	29.1	33.4
NER + immediate context filter (heuristics)	23.5	89.8	37.2
NER + immediate context filter (HMM)	52.9	80.8	63.4
NER + PMIDs	96.2	50.8	66.4

Starting from a baseline configuration (pure recognition of named entities; see text), each context type was evaluated separately. In addition, we present the impact of filtering by the immediate context: excluding genes from wrong species, abbreviations, and similar heuristics, and using an hidden Markov model (HMM) learned from the training data. Using PubMed IDs (PMIDs) curated for each gene (for instance, via GeneRIFs, Gene Ontology [GO] annotation, and UniProt) would be the best way to ensure high precision and F measure, although these data were not used for the BioCreative II evaluation. NER, named entity recognition.