EURASIP Journal on Wireless Communications and Networking

Table 11 SED results on CoNLL-14

From: A multi-task learning framework for efficient grammatical error correction of textual messages in mobile communications

Method	Thres*	CoNLL-14 A1			CoNLL-14 A2			Merge
Method	Thres*	P	R	Acc	P	R	Acc	Acc
BERT	0.79	80.2	94.9	79.4	93.3	92.6	87.9	94.13
SpanBERT	0.89	79.2	96.3	79.0	92.3	94.1	88.1	94.05
BERT	0.89	76.7	96.6	76.4	90.5	95.6	87.6	92.0
RoBERTa	0.89	79.4	95.7	79.0	92.6	93.5	87.9	92.9
SpanBERT	0.90	75.6	98.1	75.8	89.9	97.8	88.6	92.5
LSTM	0.73	72.7	99.0	72.4	86.6	98.9	85.9	89.3
BERT	0.50	86.0	79.7	76.0	96.7	75.1	76.4	77.3
LSTM	0.50	76.2	85.0	70.0	89.4	83.6	77.4	78.2

P and R represent the precision and recall of incorrect sentences
* Threshold is determined by grid search in development set, except the last two rows

Back to article page