MIT Press

Table 2:

Accuracy, RMSE, and parameter values of competing models for all data sets. An asterisk (*) indicates customization methods first introduced in this paper. A dash (-) indicates the model is too big to be trained in an NVIDIA 1080 Ti GPU. Boldface indicates that the performance of basis-customization is significantly better (p < 0.05) than that of a simple customization. Values colored

are performance weaker than that of the BiLSTM model, thus customization hurts the performance in those cases.

Models		Yelp 2013			AAPR		PolMed
Models		Accuracy	RMSE	Param	Accuracy	Param	Accuracy	Param
Base: BiLSTM		63.7	0.687	442k	61.70	188k	40.30	86k
bias vector (concat)	cust	66.3	0.661	1.3m	65.30	6.3m	40.57	121k
	basis-cust	66.9	0.654	653k	64.80	1.7m	40.92	95k
linear trasformation*	cust			4.6m	63.55	6.3m		379k
	basis-cust	67.1	0.662	655k	65.75	1.7m	41.89	96k
attention pooling	cust	65.4	0.674	1.3m	62.80	6.3m	40.93	119k
	basis-cust	66.0	0.671	652k	65.85	1.7m	41.73	95k
encoder weights*	cust	-	-	-	-	-		43.5m
	basis-cust	66.1	0.665	1.5m	66.15	2.1m	41.42	179k
word embedding*	cust			294m	-	-	40.84	46.0m
	basis-cust	66.1	0.666	1.0m	65.80	2.0m	41.58	455k

Models		Yelp 2013			AAPR		PolMed
Models		Accuracy	RMSE	Param	Accuracy	Param	Accuracy	Param
Base: BiLSTM		63.7	0.687	442k	61.70	188k	40.30	86k
bias vector (concat)	cust	66.3	0.661	1.3m	65.30	6.3m	40.57	121k
	basis-cust	66.9	0.654	653k	64.80	1.7m	40.92	95k
linear trasformation*	cust			4.6m	63.55	6.3m		379k
	basis-cust	67.1	0.662	655k	65.75	1.7m	41.89	96k
attention pooling	cust	65.4	0.674	1.3m	62.80	6.3m	40.93	119k
	basis-cust	66.0	0.671	652k	65.85	1.7m	41.73	95k
encoder weights*	cust	-	-	-	-	-		43.5m
	basis-cust	66.1	0.665	1.5m	66.15	2.1m	41.42	179k
word embedding*	cust			294m	-	-	40.84	46.0m
	basis-cust	66.1	0.666	1.0m	65.80	2.0m	41.58	455k

View Large

This Feature Is Available To Subscribers Only