Hinglish

Tools and Data

Model Performance

Base LM	Dataset	Accuracy	Precision	Recall	F1	LM Perplexity
bert-base-multilingual-cased	Test	0.688	0.698	0.686	0.687	8.2
bert-base-multilingual-cased	Valid	0.62	0.592	0.605	0.55	8.2
distilbert-base-uncased	Test	0.693	0.694	0.703	0.698	6.51
distilbert-base-uncased	Valid	0.607	0.614	0.600	0.592	6.51
distilbert-base-multilingual-cased	Test	0.612	0.615	0.616	0.616	8.1
distilbert-base-multilingual-cased	Valid	0.55	0.531	0.537	0.495	8.1
roberta-base	Test	0.630	0.629	0.644	0.635	7.54
roberta-base	Valid	0.60	0.617	0.607	0.595	7.54
Ensemble	Test	0.714	0.718	0.718	0.718

Ensemble Performace

Model	Accuracy	Precision	Recall	F1	Config	Link to Model and output files
BERT	0.68866	0.69821	0.68608	0.6875	Batch Size - 16 Attention Dropout - 0.4 Learning Rate - 5e-07 Adam epsilon - 1e-08 Hidden Dropout Probability - 0.3 Epochs - 3	BERT
DistilBert	0.69333	0.69496	0.70379	0.6982	Batch Size - 16 Attention Dropout - 0.6 Learning Rate - 3e-05 Adam epsilon - 1e-08 Hidden Dropout Probability - 0.6 Epochs - 3	DistilBert
EnsembleBert1	0.69233	0.70236	0.69064	0.68952	Batch Size - 4 Attention Dropout - 0.7 Learning Rate - 5.01e-05 Adam epsilon - 4.79e-05 Hidden Dropout Probability - 0.1 Epochs - 3	EnsembleBert1
EnsembleBert2	0.691	0.7009	0.6889	0.68872	Batch Size - 4 Attention Dropout - 0.6 Learning Rate - 5.13e-05 Adam epsilon - 9.72e-05 Hidden Dropout Probability - 0.2 Epochs - 3	EnsembleBert2
EnsembleDistilBert1	0.70166	0.70377	0.70976	0.7061	Batch Size - 16 Attention Dropout - 0.8 Learning Rate - 3.02e-05 Adam epsilon - 9.35e-05 Hidden Dropout Probability - 0.4 Epochs - 3	EnsembleDistilBert1
EnsembleDistilBert2	0.689	0.691	0.69666	0.69335	Batch Size - 4 Attention Dropout - 0.6 Learning Rate - 5.13e-05 Adam epsilon - 9.72e-05 Hidden Dropout Probability - 0.2 Epochs - 3	EnsembleDistilBert2
EnsembleDistilBert3	0.69366	0.69538	0.70557	0.69905	Batch Size - 16 Attention Dropout - 0.4 Learning Rate - 4.74e-05 Adam epsilon - 4.09e-05 Hidden Dropout Probability - 0.6 Epochs - 3	EnsembleDistilBert3
Ensemble	0.71466	0.71867	0.71853	0.7182	NA	Ensemble