Subjectivity Aware Conversational Search Services

Online users are becoming increasingly dependent on Web services in choosing among products and services. This recent trend is motivated by the integration of conversational agents which took the human-machine interaction to unprecedented levels of ease, using natural language as a communication medium. Given the success of these systems, users are constantly switching to experiential search, providing utterances that are intrinsically subjective such as looking for a restaurant with a romantic ambiance, creative cooking or nice staff. Current Web services are unfortunately unable to decipher the subjective signals present in user utterances and only support objective attributes that are listed in service descriptions (e.g., restaurant address, cuisine, price range). To make the most of dialog systems, they must be able to detect subjective attributes in user utterances and filter responses according to user subjective preferences. This paper presents a framework and techniques that augment conversational search services with capabilities to understand and reason about subjective user utterances. We propose novel subjective tag-based indexing of information services. We discuss automatic subjective tag extraction from both user utterances and online reviews using state of the art machine learning techniques such as BERT, adversarial training and data programming. Experiments show that the proposed techniques outperform existing information retrieval systems and the search mechanisms provided by wellknown web search services such as Yelp.