Skip to Main Content
Table 14: 

Model size and time required for training and inference. All models use the RoBERTa encoder, whose 124.6 million parameters are not included here. Training times are approximate and include development set evaluation at every epoch.

ModelParamsTrain timeInfer speed
millionshourssents/s
Nonconstructive Classification 
 MLP_10 2.0 191 
 MLP_1 2.4 11 195 
 
Constructive: Sequential 
 K+19 11.8 120 0.3 
 RNN 4.8 68 135 
 
Constructive: Tree-structured 
 TreeRNN 8.3 10 125 
 AddrMLP 1.3 10 126 
ModelParamsTrain timeInfer speed
millionshourssents/s
Nonconstructive Classification 
 MLP_10 2.0 191 
 MLP_1 2.4 11 195 
 
Constructive: Sequential 
 K+19 11.8 120 0.3 
 RNN 4.8 68 135 
 
Constructive: Tree-structured 
 TreeRNN 8.3 10 125 
 AddrMLP 1.3 10 126 
Close Modal

or Create an Account

Close Modal
Close Modal