UC San Diego

This thesis investigates behavioral intent signals in recommendation and natural language generation. Such signals can be obtained by including additional data correlated to user intent or by adding a preliminary model to infer intent. Leveraging intent signals not only improves a model’s predictive performance but can also add flexibility or enable it to to solve new, related problems. Explicitly modeling high-level intent results in safer models that are less likely to behave in unintended or harmful ways. We demonstrate the benefits of intent signals by describing five models with applications in task-oriented, weather-aware, B2B, and conversational recommendation, as well as in collaborative editing.