"Factual" or "Emotional": Stylized Image Captioning with Adaptive Learning and Attention
This addresses the problem of creating engaging, style-specific captions for images, which is incremental as it builds on existing stylized captioning methods.
The paper tackles generating stylized captions for images by proposing a model that uses a style-factual LSTM and adaptive learning to balance factual accuracy and style, achieving state-of-the-art performance on humorous/romantic and positive/negative caption datasets without extra supervision.
Generating stylized captions for an image is an emerging topic in image captioning. Given an image as input, it requires the system to generate a caption that has a specific style (e.g., humorous, romantic, positive, and negative) while describing the image content semantically accurately. In this paper, we propose a novel stylized image captioning model that effectively takes both requirements into consideration. To this end, we first devise a new variant of LSTM, named style-factual LSTM, as the building block of our model. It uses two groups of matrices to capture the factual and stylized knowledge, respectively, and automatically learns the word-level weights of the two groups based on previous context. In addition, when we train the model to capture stylized elements, we propose an adaptive learning approach based on a reference factual model, it provides factual knowledge to the model as the model learns from stylized caption labels, and can adaptively compute how much information to supply at each time step. We evaluate our model on two stylized image captioning datasets, which contain humorous/romantic captions and positive/negative captions, respectively. Experiments shows that our proposed model outperforms the state-of-the-art approaches, without using extra ground truth supervision.