Complementing human effort in online reviews: A deep learning approach to automatic content generation and review synthesis

Abstract

Online product reviews are ubiquitous and helpful sources of information available to consumers for making purchase decisions. Consumers rely on both the quantitative aspects of reviews such as valence and volume as well as textual descriptions to learn about product quality and fit. In this paper we show how new achievements in natural language processing can provide an important assist for different kinds of review-related writing tasks. Working in the interesting context of wine reviews, we demonstrate that machines are capable of performing the critical marketing task of writing expert reviews directly from a fairly small amount of product attribute data (metadata). We conduct a kind of “Turing Test” to evaluate human response to our machine-written reviews and show strong support for the assertion that machines can write reviews that are indistinguishable from those written by experts. Rather than replacing the human review writer, we envision a workflow wherein machines take the metadata as inputs and generate a human readable review as a first draft of the review and thereby assist an expert reviewer in writing their review. We next modify and apply our machine-writing technology to show how machines can be used to write a synthesis of a set of product reviews. For this last application we work in the context of beer reviews (for which there is a large set of available reviews for each of a large number of products) and produce machine-written review syntheses that do a good job – measured again through human evaluation – of capturing the ideas expressed in the reviews of any given beer. For each of these applications, we adapt the Transformer neural net architecture. The work herein is broadly applicable in marketing, particularly in the context of online reviews. We close with suggestions for additional applications of our model and approach as well as other directions for future research.