Following success in using machine learning (ML) techniques in areas such as speech recognition, part-of-speech tagging, word sense disambiguation, and parsing, there has been an increasing interest in applying ML to discourse processing. To date, there has been work in using machine learning techniques such as inductive learning methods (decision trees), statistical learning methods (HMMs), neural networks, and genetic algorithms to a number of discourse problems, e.g., dialogue act prediction, cue word usage, anaphora resolution, initiative tracking, and discourse segmentation.

In this symposium, we would like to bring together researchers with an interest in exploring the potential contribution of ML to problems in discourse interpretation and generation. Our goal is provide an opportunity for discussions among researchers in natural language discourse and in machine learning to facilitate collaboration between the two groups. We are interested in addressing the following issues:

• From the discourse processing point of view
• What tasks in discourse understanding/generation are most suitable for processing using ML-acquired models?
• What are the features of these tasks that make them particularly suitable for processing using ML-acquired models?
• Which ML approaches successfully adopted by other areas of natural language processing seem promising for use in discourse processing? And why?
• Is it possible to base the entire discourse processing component of a natural language system purely on ML-acquired models? If not, when should models acquired by traditional approaches come into play? And how should the two approaches be integrated?
• How can learning be performed during the discourse comprehension or generation process?
• How can knowledge acquired for discourse interpretation or generation be reused for the other?
• What types of pragmatic knowledge (e.g., discourse recipes, cue phrase classification) can be acquired by ML?
• What kinds of categories and features can be tagged automatically and/or reliably? How can useful features be identified?
• From the machine learning point of view
• What are the different ML techniques that may be suitable for acquiring knowledge for discourse processing?
• What are the features of these ML techniques that make them particularly suitable for application in discourse processing?
• How does the performance (e.g., accuracy, processing speed) of models for discourse processing based on ML techniques compare to those based on traditional methods?
• How do different ML techniques compare with one another in terms of accuracy, efficiency, amount of data needed for training, etc, for various problems in discourse processing?
• What discourse corpora are currently available for ML? What other corpora are needed for ML research?
• What characteristics of discourse processing cause problems for existing ML techniques?