A light method for data generation: a combination of Markov Chains and Word Embeddings.

Abstract

Most of the current state-of-the-art Natural Language Processing (NLP) techniques are highly data-dependent. A significant amount of data is required for their training, and in some scenarios data is scarce. We present a hybrid method to generate new sentences for augmenting the training data. Our approach takes advantage of the combination of Markov Chains and word embeddings to produce high-quality data similar to an initial dataset. In contrast to other neural-based generative methods, it does not need a high amount of training data. Results show how our approach can generate useful data for NLP tools. In particular, we validate our approach by building Transformer-based Language Models using data from three different domains in the context of enriching general purpose chatbots.