Designing the Context-based Topic Recommendation Strategy for the Q&A Agent

Abstract

Including conversational agents, the expectations of the intelligent agents are encreasing. However, the errors occurred by technical limitations and the agent's response for single turn-taking which only gives a proper answer to the single question cannot achieve the advanced goals like gaining advanced information. Regard this, this study's goal is to compare the process of the knowledge gaining of the users by the Q&A agent's context-based topic recommendation strategies. Two preliminary factors are required to design the Q&A agent. It's the categorization of databases based on the context and the design of the context-based recommendation strategy. With the categorization of the database, all the information of the Q&A agents represented in coordinates which has two axes of theme and level of the information. Which corresponded with the change of the subject and progress to the deeper knowledge. By representing the information on the coordinates, it became possible to the agent can figure out where the information it conveys to the user is located and recommend the topic in nearby subjects or levels. In this study, the topic recommendation strategy of the agent is referred to as how to constitute the question selection which suggested by each turn in using the Q&A agent with Menu/Button-based style. The strategy divided into a context-based and Random-shuffling model, and the context-based model divided into the Depth-oriented model and the Subject-expanding model. To evaluate the difference in the usage and information search behavior of the users by strategies, this study conducted a model comparison user test. The results of the test show the Depth-oriented strategy shows significant features from well-referred question suggestions to makes users gain the information easily. From the Subject-expanding model, it shows the intent of the users to explore the Q&A contents more from the number of turn-taking and increased number of interested-in subjects. Random-shuffling model, however, reported that it gives fatigues in usage while it changes the subjects and levels consistently even it exposure the information from various subjects and levels.