Table 1.

Algorithms for Word Sense Disambiguation.

	publ.	Data	Background knowledge	Approach	Experiment	Accuracy
Established Knowledge	[12]	gene definition & abstract vector	5 human gen. dbs & MeSH	cosine similarity	52,529 Medline abstracts, 690 human gene symbols	92.7%
	[13]	free text	UMLS, Journal Descriptors	Journal Descriptor Indexing (JDI)	45 ambiguous UMLS terms (NLM WSD Collection)	78.7%
	[14]	Medline abstracts	BioCreative-2 GN lexicon & text, EntrezGene, UniProt, GOA	motifs from multiple sequence alignments	BioCreative-2 GN challenge	81%
	[15]	Medline abstracts	list of gene senses, EntrezGene	inverse co-author graph	BioCreative GN challenge	97%P
Supervised	[8]	XML tagged abstracts, positional info, PoS	-	naive Bayes, decision trees, inductive rule training	protein/gene/mRNA assignment: 9 million words (mol. biol. journals)	85%
	[49]	text	-	word count, word cooc	-	86.5%
	[9,50]	Medline abstracts	UMLS terms	UMLS term cooc	35 biomedical abbreviations	93%P
	[10]	abbreviations in Medline abstracts	-	SVM	build dictionary, use for abbreviations occurring with their long forms	98.5%
	[11]	gene symbol context (n words +/-)	-	SVM	-	85%
Unsupervised	[19,20]	document	-	LSA/LSI, 2nd order cooc	170,000 documents, 1013 terms (TREC-1) (Wall Street Journal)	↑ 7–14%
	[51]	word cooc, PoS tags	WordNet	average link clustering	13 words, ACL/DCI	73.4%
	[21]				Wall Street Journal Corpus
	[22]	-	-	1st, 2nd order context vectors (coocs within 5 positions)	24 Senseval-2 words, Line, Hard, Serve corpora	44%
	[23]	text	few tagged data, WordNet	co-training, collocations	12 common Engl. words × 4000 instances	96.5%
	[25]	-	-	co-training & majority voting	Senseval-2 generic English	↑ 9.8%
	[24]	-	WordNet	noun coocs, Markov clustering	-	-