Building Non-autoregressive (NAR) models in image captioning can fundamentally tackle the high inference latency of autoregressive models. However, existing NAR image captioning models are trained on maximum likelihood estimation, and suffer from their inherent multi-modality problem. Although constructing NAR models based on GANs can theoretically tackle this problem, existing GAN-based NAR models obtain poor performance when transferred to image captioning due to their incapacity of modeling complicated relations between images and text. To tackle this problem, we propose an Adversarial Non-autoregressive Transformer for Image Captioning (CaptionANT) by improving performance from two aspects: 1) modifying the model structure so as to be compatible with contrastive learning to effectively make use of unpaired samples; 2) integrating a reconstruction process to better utilize paired samples. By further combining with other effective techniques and our proposed lightweight structure, CaptionANT can better align input images and output text, and thus achieves new state-of-the-art performance for fully NAR models on the challenging MSCOCO dataset. More importantly, CaptionANT achieves a 26.72 times speedup compared to the autoregressive baseline with only 36.3% the number of parameters of the existing best fully NAR model for image captioning.

Neural response generative models have achieved remarkable progress in recent years but tend to yield irrelevant and uninformative responses. One of the reasons is that encoder-decoder based models always use a single decoder to generate a complete response at a stroke. This tends to generate high-frequency function words with less semantic information rather than low-frequency content words with more semantic information. To address this issue, we propose a content-aware model with two-stage decoding process named Two-stage Dialogue Generation (TSDG). We separate the decoding process of content words and function words so that content words can be generated independently without the interference of function words. Experimental results on two datasets indicate that our model significantly outperforms several competitive generative models in terms of automatic and human evaluation.