Attribute-Driven Generation of Drug Reviews using Deep Learning

Abstract: In the last years, the demands on different models using deep learning to generate textual data conditionally have increased, where one would like to control what textual data to generate from a deep learning model. For this purpose, a couple of models have been developed and achieved state-of-art performance in the field of generating textual data conditionally. Therefore, the purpose of this study was to develop a new model that could outperform the relevant baseline models with respect to the BLEU metric. The alternative model combined some of the properties from the state-of-art models and was given the name the Variational Attribute-to-Sequence decoder model (shortened to the V-Att2Seq model) that paraphrases the name of one of the state-of-art models and "variational" refers to its application of variational recurrent autoencoders (VRAE). The data set used in this study contained drug reviews that were written by patients to express their opinion about the drug that they have used to treat a certain condition. The drug review texts were accompanied by the following attributes: the (name of the) drug, the condition, and the rating that the patient has given to the drug. The results in this study show that the V-Att2Seq model did not outperform all the baseline models, which concluded that the V-Att2Seq model did not satisfy the requirements imposed on the model itself. However, there are some future work that is suggested by this study to hopefully improve the performance of the V-Att2Seq model in the future such as including other mechanisms that are present in the state-of-art models, testing withe.g. other sizes and settings of the V-Att2Seq model, and testing different strategies forgenerating sequences since there is still potential that has been observed in the model that should be further investigated to improve its performance.