Rapid Lexical Alignment to a Conversational Agent

Abstract

Conversational partners modify their language to be more similar to each other during interactions. This phenomenon, known as alignment, has been shown in human-human interactions, but there is little work on lexical alignment in human-computer interactions. We investigate whether people lexically align to a conversational agent, and whether the degree of alignment depends on feedback from the agent. This study compared three feedback conditions for how the agent responded to users' word choice: (1) the agent only understood the specific words that it produced itself; (2) the agent understood the words that it produced as well as more appropriate synonyms; (3) the agent's understanding of words that it did not produce was random. Participants significantly aligned to the agent in all conditions, and aligned more when they learned that the agent's comprehension was contingent on their alignment. Thus, inducing lexical alignment may be an effective way to increase dialogue success.