The 1995 AAAI Spring Symposia Reports

The symposium entitled Empirical Methods in Discourse Interpretation and Generation had two goals: (1) to investigate the empirical methods that can be used in the development and evaluation of computational theories of discourse and (2) to develop a set of shared resources for the computational discourse community. The workshop was successful at achieving the first goal because of the large number of papers submitted (48) surveying a range of different empirical methods. Methods discussed included Wizard of Oz studies, machine learning on corpora tagged for discourse features, human subjective evaluation of the system output in a natural language–generation system using a grading scheme, simulation test beds, the running of the implemented systems with different discourse modules on subjects and the evaluation of the quality of the interaction by various methods, and the coding of independent and dependent variables and the statistical calculation of their relationships. The discussions were lively and addressed issues such as determining what methods to use for various problems, suiting statistical tests to different types of tagging scheme, and determining the generalizability of different methods. The discussions also focused on the second goal, but progress on this goal consisted of airing the issues and forming a committee to get funding for a follow-on workshop to focus on developing shared resources.