tESA: a distributional measure for calculating semantic relatedness

Table 3 Overview of the results for different experimental settings - corpus and benchmark pairs; ESA and tESA runs with M=10000 and DS (the method described in [28]) runs with M=200 and cutoff at 0,02 (robust parameters, that can be expected to provide decent results in different experimental settings)

Corpus	Method	umnsrsRelate	umnsrsSim	mayo101	mayo29ph	mayo29c
	ESA	0.608	0.621	0.546	0.835	0.734
Medline	tESA	0.649	0.639	0.549	0.783	0.687
	DS	0.46	0.438	0.511	0.483	0.493
	ESA	0.588	0.597	0.543	0.855	0.75
PMC	tESA	0.595	0.607	0.484	0.796	0.7
	DS	0.574	0.626	0.504	0.738	0.673
	ESA	0.501	0.5	0.548	0.822	0.722
Wiki	tESA	0.484	0.484	0.502	0.801	0.755
	DS	0.444	0.463	0.413	0.627	0.597
Best reported (citation)		0.54 [28]	0.58 [28]	0.6 [28]	0.84 [16]	0.9 [34]

The table row for best reference results has been compiled with results reported in the domain literature for the respective datasets, regardless of the type of method used to achieve those results. Best reported results for umnsrsRelate, umnsrsSim and mayo101 were attained with specific parameter combinations in our experiments (presented in [28]), whereas for the two smaller datasets the best results were previously obtained with knowledge-rich methods (distributional and IC-based respectively for mayo29ph and mayo 29c). Updated best results are highlighted with bold font

ISSN: 2041-1480