Enhancing Text Generation via Parse Tree Embedding

Table 1

Performance of the Tree-VAE and the four baselines on the Yelp review generation, e-mail generation, and COCO image captions generation tasks. Higher is better. “Tree-VAE (WA)” and “Tree-VAE (WOA)” represent Tree-VAE with attention mechanism and Tree-VAE without attention mechanism, respectively. BLEU-2, BLEU-3, BLEU-4, and BLEU-5 are the BLEU scores.

Yelp review	BLEU-2	BLEU-3	BLEU-4	BLEU-5
MLE	0.609	0.482	0.477	0.409
SVAE	0.663	0.654	0.587	0.571
SeqGAN	0.768	0.598	0.564	0.507
RankGAN	0.795	0.683	0.610	0.554
Tree-VAE (WA)	0.817	0.654	0.637	0.624
Tree-VAE (WOA)	0.806	0.697	0.664	0.631
E-mail	BLEU-2	BLEU-3	BLEU-4	BLEU-5
MLE	0.473	0.462	0.317	0.306
SVAE	0.548	0.517	0.441	0.353
SeqGAN	0.684	0.596	0.578	0.511
RankGAN	0.631	0.560	0.493	0.445
Tree-VAE (WA)	0.703	0.667	0.593	0.512
Tree-VAE (WOA)	0.711	0.670	0.600	0.527
COCO image captions	BLEU-2	BLEU-3	BLEU-4	BLEU-5
MLE	0.739	0.642	0.531	0.499
SVAE	0.801	0.784	0.668	0.633
SeqGAN	0.834	0.805	0.708	0.693
RankGAN	0.817	0.826	0.765	0.614
Tree-VAE (WA)	0.858	0.823	0.773	0.707
Tree-VAE (WOA)	0.853	0.819	0.794	0.730