Routinization: What are the cognitive building blocks of syntactic structure?

Armin Buch, Christian Pietsch, Gerhard Jäger The psycholinguistic study of alignment phenomena in dialog on the one hand, and the application of machine learning techniques to natural language processing on the other hand, have revealed that the cognitive units of syntactic structure are larger than single lexical items or syntactic rules. The long term goal of the project is to contribute to a cognitively realistic architecture of dialog systems that is able to cope with the emergence and development of routines in human-machine interaction. The leading hypothesis of the project is that the non-local chunks of syntactic structure that have proven informative for machine learning purposes (for instance in data-oriented parsing) are exactly those items that are cognitively represented as a whole and play a role in routinization in dialogue. The project pursues three goals:

• We assume that cognitive units of syntactic structure are connected sub-graphs of a graph-theoretic representation of syntactic structure. Hence the predictive power of such an approach heavily depends on the chosen representation format. It is standard to use conventional phrase-structure trees for this purpose. We hypothesize that some variant of the proof nets of type-logical grammar are better suited for this task because grammatically interdependent items have a short geodesic distance in this format regardless of their linear distance in the text. This hypothesis will be tested by means of corpus studies.
• In a second step, we will study how routines emerge dynamically in a dialogue, using the representation format that was developed in the first phase.
• As long-term goal, the insights about the emergence of routines from the second phase will be employed to model syntactic alignment in human-machine interaction. As routinization plays an essential part in dialogs between humans, a dialog system that is capable to devolop or to pick up syntactic routines will display a more natural dialog behavior than extant systems.

[view:groupmembers==197]