Automatic Image Captioning

In this work, we present the design and implementation of a solution to the problem of modeling annotated data. We specifically target data with multiple types where an instance of one type of data serves as a description of another. We describe a hierarchical probabilistic mixture model correspondence latent Dirichlet allocation - that allows for variable representations to be associated with topics. We have used Gibbs sampling technique to perform posterior inference on the model. We then conducted experiments on 3 different datasets, assessing the models’ performance in terms of caption perplexity. Each dataset is made of pairs of data, one datatype being the images in the form of their features other being their respective captions.