Caption Generation from Keywords using Probabilistic Dual-LSTM Networks and Dynamic Programming

Abstract

Generation of natural languages has always been a keystone in AI research. In many modern challenges virtual assistants or caption generation it is often required to generate natural language texts from a set of annotations in the absence of a proper schema or language model. In this work we have used the principles of dynamic programming on probabilistic dual-LSTM Networks to generate sentences from a set of keywords. Validation against human judges showed that the system is able to provide quite satisfactory description of images from a set of keywords.