Skip to Main Content
Table 3: 
Hypernymy evaluation results for baselines and LexSub. LexSub considerably outperforms all the other methods and the Vanilla on nearly all hypernymy tasks. We attribute this performance to our novel loss function formulation for asymmetric relations and the separation of concerns imposed by the LexSub.
ModelsSimilarity (ρ)Directionality (Acc)Classification (Acc)
Hyperlexwblessbiblessblessledsevalweeds
Vanilla 0.1352 0.5101 0.4894 0.1115 0.7164 0.2404 0.5335 
 
Retrofitting 0.1055 0.5145 0.4909 0.1232 0.7279 0.2639 0.5547 
Counterfitting 0.1128 0.5279 0.4934 0.1372 0.7246 0.2900 0.5734 
LEAR 0.1384 0.5362 0.5024 0.1453 0.7399 0.2852 0.5872 
 
LexSub 0.2615 0.6040 0.4952 0.2072 0.8525 0.3946 0.7012 
ModelsSimilarity (ρ)Directionality (Acc)Classification (Acc)
Hyperlexwblessbiblessblessledsevalweeds
Vanilla 0.1352 0.5101 0.4894 0.1115 0.7164 0.2404 0.5335 
 
Retrofitting 0.1055 0.5145 0.4909 0.1232 0.7279 0.2639 0.5547 
Counterfitting 0.1128 0.5279 0.4934 0.1372 0.7246 0.2900 0.5734 
LEAR 0.1384 0.5362 0.5024 0.1453 0.7399 0.2852 0.5872 
 
LexSub 0.2615 0.6040 0.4952 0.2072 0.8525 0.3946 0.7012 
Close Modal

or Create an Account

Close Modal
Close Modal