Towards Mixed-Initiative Conversational Search
Published:
Title: Towards Mixed-Initiative Conversational Search
Date: 2020-10-02
Host: Language Technologies Institute, Carnegie Mellon University
Speaker: Hamed Zamani
Summary and thoughts:
Dr. Zamani presented at this week’s LTI colloquium on the topic “conversational search”, which has grown as an interest area due to the advances in automatic speech recognition, conversational agents such as Alexa, and the indispensable value of ‘search’ as a service in our daily lives. Historically, information retrieval systems and question answering agents were based on basic criterion required to get the desired search state. Conversational search, however, goes beyond the usual search paradigm where discourse and clarification become important needs of the system in supporting interactions between humans and the system. The speaker went through a thorough history of search paradigms leading upto the current work in clarifications in search queries, which his lab is heavily involved in.
Conversation and interaction in search queries depends on the end goal and information need of the user. The system must understand and adapt search results based on the initial query (in a single search query paradigm) and further clarifying inputs by the user. Clarifications from the user can take different forms such as disambiguation, preferences (personal, spatial, temporal information), topic and comparison. Clarification models can be created to generate clarifying questions using several components such as query aspect generation, question generation, and candidate answers generation, where each component is driven by the clarification utility. Dr. Zamani then presented the details of each component of the clarification model in detail. Overall, conversational search is already a major part of our life. Automating the process would be useful not just for general users, but may also lead to exploration of other topics in the area.