"Love ya, jerkface": Using Sparse Log-Linear Models to Build Positive (and Impolite) Relationships with Teens
One challenge of implementing spoken dialogue systems for long-term interaction is how to adapt the dialogue as user and system become more familiar. We believe this challenge includes evoking and signaling aspects of long-term relationships such as rapport. For tutoring systems, this may additionally require knowing how relationships are signaled among non-adult users. We therefore investigate conversational strategies used by teenagers in peer tutoring dialogues, and how these strategies function differently among friends or strangers. In particular, we use annotated and automatically extracted linguistic devices to predict impoliteness and positivity in the next turn. To take into account the sparse nature of these features in real data we use models including Lasso, ridge estimator, and elastic net. We evaluate the predictive power of our models under various settings, and compare our sparse models with standard non-sparse solutions. Our experiments demonstrate that our models are more accurate than non-sparse models quantitatively, and that teens use unexpected kinds of language to do relationship work such as signaling rapport, but friends and strangers, tutors and tutees, carry out this work in quite different ways from one another.