Other people’s opinions are an important information source for both making informed decisions and acquiring knowledge. In offline sales negotiations, both the salesperson and the customer exchange product-related information about its features and their (personal) opinions on them. In online scenarios, customers commonly refer to product reviews to compare a limited number of options and make an informed decision. Similar to the offline sales channels, hearing about the subjective experience of other users is often more valuable than a purely fact-based question answering. A conversational agent tasked with online sales will have to enact the role of a salesperson and be capable of reacting to both product feature-related questions as well as opinions expressed by a (human) customer.
To date, researchers have mostly focused on summarizing the pros and cons of products from the respective reviews. In order for the conversational agent to hold a longer sales conversation, it is necessary to construct a large source corpus of product-related opinionated dialogues. To this end, we tap product reviews as a rich source of a mixture of feature-related statements that are both factual and opinionated. By mining such statements, this enables the template-based simulation of sales conversations with various properties.
In this work, we develop a framework, which encompasses the mining and generation pipeline that resulted in a novel conversational dataset of opinionated sales conversations. We further demonstrate the potential usefulness of synthetically generated sales conversations of this kind by augmenting the Multi-Domain Wizard-of-Oz dataset (MultiWOZ), and thus, the training data of state-of-the-art models for the dialog state tracking task, improving their overall joint accuracy on the original test data.