Learning the User’s Language: A Step Towards Automated Creation of User Models

Abstract

No matter how comprehensive the coverage of a natural language interlace, users will invariably provide unparsable input: idiosyncratic phrases, linguistically deviant utterances, or sentences simply beyond the linguistic sophistication of the interlace. Although individual users differ significantly from each other in their preferred linguistic expression, they show surprising consistency in multiple interactions over time. Addressing both observations, a general framework is developed for automated adaptation to the preferred and idiosyncratic linguistic behavior of individual users. In essence, a method for flexible recovery and completion of partial parses is developed, with subsequent individualized extensions to a base grammar. Thus, the system adapts to the user by growing its grammar dynamically to acquire her preferred forms of expression and recognize them directly in future interactions.