Fast max-margin clustering for unsupervised word sense disambiguation in biomedical texts

BMC Bioinformatics

Table 1 Comparison with a supervised WSD system.

Keyword	All-in-1	L & R	SENSATIONAL
adjustment	0.62	0.57	0.56
blood pressure	0.54	0.46	0.49
degree	0.63	0.68	0.72
evaluation	0.50	0.57	0.57
growth	0.63	0.62	0.71
immunosuppression	0.59	0.63	0.59
man	0.58	0.80	0.51
mosaic	0.52	0.66	0.71
nutrition	0.45	0.48	0.42
radiation	0.61	0.72	0.65
repair	0.52	0.81	0.80
scale	0.65	0.84	0.68
sensitivity	0.49	0.70	0.74
weight	0.47	0.68	0.53
white	0.49	0.62	0.52
average	0.55	0.66	0.61

SENSATIONAL outperforms the baseline All-in-1 system by 6% on average, and comes within 5% of a supervised system for word sense disambiguation by Leroy & Rindflesch (L&R) on a standard NLM dataset. Average performance across all words is statistically significant using the Chi-square test with Yates' correction (two-tailed test with 1 degree of freedom; between All-in-1 and SENSATIONAL, Chi-square = 10.839, p = 0.0010; between SENSATIONAL and L&R, Chi-square = 7.875, p = 0.0050).

ISSN: 1471-2105