SALS-SIG Research Seminar

Giving Voice to Plans

Mark Steedman & Bonnie Lynn Webber
Dept of Computer & Information Science, University of Pennsylvania

When: Thursday, 18th August 1994

Time: 4:00pm

Where: Microsoft Institute, Conference Room 1 (next to Bistro)

Abstract:

Planning systems can be used to many different ends: directly driving the action of an ``execution engine'', or assisting users in developing plans, or training them to do the right thing, or monitoring their requests for action, in order to detect deviations from a ``gold standard'' the system embodies in its own plans.

Designers of user-oriented planning systems face the problem that their users' activities may not be around a computer screen. In such cases, communication between system and user cannot rely on visual display: speech may be a more appropriate channel.

Natural speech employs both words and tunes (intonation). In the first part of this talk, we address two problems that arise in generating natural intonation for spoken language interfaces: (1) the fact that the phrases delimited by spoken intonation don't always match up with traditional syntactic structure, and (2) the lack of a formal semantic theory for intonation. Our solution relies on Combinatory Categorial Grammar (CCG), a novel grammar formalism that allows intonation structure (in the sense of Selkirk and others) to be entirely subsumed under a generalized notion of surface structure. Our system, called IBIS (for Information-based Intonation Synthesis) greatly simplifies the mapping between phonological-level phenomena and discourse semantic categories such as ``topic'', ``comment'' and ``focus'', allowing fully formal specifications of these notions. IBIS was developed with Scott Prevost at Penn.

The second part of this talk considers an application of IBIS in health care delivery. TraumAID is a decision-support system for the initial definitive management of multiple trauma. Initial definitive management scopes a complex range of diagnostic and therapeutic procedures. TraumAID's management support is based on its ability to map from the current set of patient findings and test results to a clinically-appropriate management plan. To do this, TraumAID employs (1) a rule-based reasoner able to draw diagnostic conclusions and identify the goals each of them implies; and (2) a planner able to identify a partially-ordered sequence of actions that satisfies those goals, while adhering to standard principles and practices of trauma management.

We describe both TraumAID and TraumaTIQ (TraumAID's information delivery module), noting how the latter decides what, if anything, to say to a physician during patient care and how it uses IBIS to say it. TraumAID and TraumaTIQ are joint work with John Clarke, M.D., Ron Rymon, Ph.D., Abigail Gertner, and Jonathan Kaye.

SPEAKERS:

Mark Steedman is Professor of Computer and Information Science at the University of Pennsylvania. He received his PhD from the University of Edinburgh in 1973. He joined the University of Pennsylvania in 1988, after teaching at the Universities of Warwich and Edinburgh. He was a Sloan Fellow and Visiting Professor at the University of Texas at Austin in 1981, and a Visiting Professor at Penn in 1986. He is a Fellow of AAAI.

His research interests cover a wide range of issues in the areas of computational linguistics, artificial intelligence, computer science and cognitive science, including syntax and semantics of natural languages and programming languages, parsing and comprehension of natural language discourse by humans and by machine, natural language generation, and intonation in spoken discourse.

Much of this work is based upon a novel theory of natural language syntax and semantics called Combinatory Categorial Grammar, and it is described in mumerous articles in learned journals and conference proceedings. Much of his current NLP research is addressed to issues in spoken discourse and dialogue, especially the meaning of intonation and prosody. He is currently working with colleagues in computer animation using these theories to guide the graphical animation of speaking virtual or simulated autonomous human agents.

Some of his research concerns the analysis of music by humans and machines. This work was recently the subject of a broadcast on the BBC World Service.
********************

Bonnie Lynn Webber is Professor of Computer and Information Science at the University of Pennsylvania. She received her PhD from Harvard University in 1978 and can type with all ten fingers.

Webber has research interests in both Natural Language processing (NLP) and medical applications of Artificial Intelligence. Her current NLP research is at the interface between language and behavior, working with colleagues in computer animation on the use of instructions to guide the behavior of animated human models. This work is described in ``Simulating Humans: Computer Graphics, Animation and Control'' by Norman Badler, Cary Phillips and Bonnie Webber, published recently by Oxford University Press. (For people with shorter attention spans, it is also described in the June 1994 issue of ``Discover'' magazine.) Webber has also co-edited several books, including ``Readings in Artificial Intelligence'' and ``Readings in Natural Language Processing''.

In the area of Medical Informatics, Webber is co-PI of the TraumAID project with Dr. John R. Clarke, M.D., F.A.C.S. The overall goal of this work is to improve the delivery of quality trauma care during the initial definitive phase of patient management. TraumAID exploits a range of AI techniques to do this, including rule-based reasoning, planning, and plan inference. TraumAID has been evaluated retrospectively and will soon be tested clinically in the Emergency Center at the Medical College of Pennsylvania.

Enquiries: Mark Dras