Concept-level stance vectors improve viewpoint diversity in dialog systems

A new selection policy uses Perspectivized Stance Vectors to replace scalar scores, allowing dialog systems to select more diverse viewpoints in user engagement. While concept-level encoding produces more varied exposure sequences, user studies show it does not increase the depth of user justification.