MIT Press

Table 3:

Comparison of Accuracy and Macro-F₁ among different models. Results marked with ♯ are adopted from Chen et al. (2017), and those with * are adopted from the original papers. Performance improvements of the proposed TG-SAN model over the state-of-the-art, TNet (Li et al., 2018), are statistically significant at p ¡ 0.01.

Models		Tweet		Laptop		Restaurant
		Accuracy	Macro-F₁	Accuracy		Accuracy	Macro-F₁
Baselines	SVM (2014)	0.6340^♯	0.6330^♯	0.7049^*	−	0.8016^*	−
	AdaRNN (2014)	0.6630^*	0.6590^*	−	−	−	−
	TD-LSTM (2016a)	0.6662^♯	0.6401^♯	0.7183^♯	0.6843^♯	0.7800^♯	0.6673^♯

Attention-based	ATAE-LSTM (2016)	−	−	0.6870^*	−	0.7720^*	−
	IAN (2017)	−	−	0.7210^*	−	0.7860^*	−
	MemNet (2016b)	0.6850^♯	0.6691^♯	0.7033^♯	0.6409^♯	0.7816^♯	0.6583^♯
	RAM (2017)	0.6936^*	0.6730^*	0.7449^*	0.7135^*	0.8023^*	0.7080^*

State-of-the-art	TNet (2018)	0.7327	0.7132	0.7465	0.6985	0.8005	0.6901

Proposed Model	TG-SAN	0.7471	0.7365	0.7527	0.7118	0.8166	0.7259

Ablations	w/o CFU	0.7312	0.7141	0.7465	0.7042	0.8095	0.7189
	w/o SCU & CFU	0.7153	0.6975	0.7058	0.6559	0.8023	0.6960
	w/o TG	0.7269	0.7093	0.7324	0.6923	0.8131	0.6986

Models		Tweet		Laptop		Restaurant
		Accuracy	Macro-F₁	Accuracy		Accuracy	Macro-F₁
Baselines	SVM (2014)	0.6340^♯	0.6330^♯	0.7049^*	−	0.8016^*	−
	AdaRNN (2014)	0.6630^*	0.6590^*	−	−	−	−
	TD-LSTM (2016a)	0.6662^♯	0.6401^♯	0.7183^♯	0.6843^♯	0.7800^♯	0.6673^♯

Attention-based	ATAE-LSTM (2016)	−	−	0.6870^*	−	0.7720^*	−
	IAN (2017)	−	−	0.7210^*	−	0.7860^*	−
	MemNet (2016b)	0.6850^♯	0.6691^♯	0.7033^♯	0.6409^♯	0.7816^♯	0.6583^♯
	RAM (2017)	0.6936^*	0.6730^*	0.7449^*	0.7135^*	0.8023^*	0.7080^*

State-of-the-art	TNet (2018)	0.7327	0.7132	0.7465	0.6985	0.8005	0.6901

Proposed Model	TG-SAN	0.7471	0.7365	0.7527	0.7118	0.8166	0.7259

Ablations	w/o CFU	0.7312	0.7141	0.7465	0.7042	0.8095	0.7189
	w/o SCU & CFU	0.7153	0.6975	0.7058	0.6559	0.8023	0.6960
	w/o TG	0.7269	0.7093	0.7324	0.6923	0.8131	0.6986

View Large

This Feature Is Available To Subscribers Only