MIT Press

Skip to Main Content

Table 6:

MCC results for specific phenomena. Emb. is model embedding style; Enc. is model encoder training, Class. is model classifier training. R/F is real/fake, ELMo-St. is ELMo-style, and CoLA-Th. is threshold tuning on CoLA. LSTM LM WLPM is the LM with Lau et al. metrics Word LP Min-1.

Model	Emb.	Enc.	Class.	SVO	Wh	Causative	SV Agr.	Reflexive
LSTM LM WLPM	BNC	–	CoLA Th.	0.801	0.601	0.270	0.599	0.152

Pooling	ELMo-St.	CoLA	CoLA	0.637	0.102	0.633	0.128	0.075
Pooling	BNC	R/F	CoLA	0.381	0.184	0.463	0.098	0.043
Pooling	GloVe	R/F	CoLA	0.988	0.059	0.614	0.277	0.150
Pooling	ELMo-St.	R/F	CoLA	0.650	0.000	0.449	0.302	-0.020

Model	Emb.	Enc.	Class.	SVO	Wh	Causative	SV Agr.	Reflexive
LSTM LM WLPM	BNC	–	CoLA Th.	0.801	0.601	0.270	0.599	0.152

Pooling	ELMo-St.	CoLA	CoLA	0.637	0.102	0.633	0.128	0.075
Pooling	BNC	R/F	CoLA	0.381	0.184	0.463	0.098	0.043
Pooling	GloVe	R/F	CoLA	0.988	0.059	0.614	0.277	0.150
Pooling	ELMo-St.	R/F	CoLA	0.650	0.000	0.449	0.302	-0.020