ICICLE (Interactive Computer Identification and Correction of Language Errors) is a tutoring system under development that instructs deaf users of American Sign Language on written English skills (see (McCoy & Masterman (Michaud), 1997 for a discussion of overall system architecture). The text generation module it employs produces original text to instruct the user on errors found in his or her writing, tailored to the user's understanding and learning style. The model I propose for planning this text composes it according to a four-tier response anatomy. It combines bottom-up and top-down planning approaches by working bottom-up from the error types themselves to group and order the errors and then top-down from the goal of tutoring on these errors, building and fleshing out a text plan. The planning process takes into account a detailed representation of user language proficiency and a history of interaction with a user in order to create text that is maximally understandable and informational to the individual.

The center goal of my work is refining the planning phase concerning the response anatomy, which is comprised of content, method, form, and manner. Content refers to the error(s) to be discussed; the method is the choice between possible pedagogical approaches to discussing the error; the form is the determination of how to structure the approach defined by the method; and the manner refers to the annotations to the text plan determined by the other parts, resulting in a cohesive, contextually-aware explanation. The top-down planning begins with the posting of the goal of instructing the user on an error and the selection of plan operators that represent approaches toward realizing this goal. My plan operator design is based largely on the work in (Moore & Paris, 1992); operators are selected according to the effect they have on the user's knowledge and a list of constraints which state when an operator is applicable, referencing multiple sources of knowledge including the user model, the history module, recent dialogue, and the domain knowledge base. When a selection is made from the first tier of operators representing method choices, subgoals are posted that help drive the further selection of form operators which contain schemata for structuring the text. A manner phase then applies operators to generate comparisons to established user knowledge and previous explanations. The contextually-aware final output is a theoretic text specification which can then be sent to a generator to be realized as English.

A longer discussion of my proposed planner can be found in (Michaud & McCoy, 1998). Future tasks include refining the concept of the user model proposed in (McCoy et al., 1996), developing a domain knowledge base about errors recognized by the system, and further specifying our planning operators. This project has been supported by NSF Grant # IRI-9416916, NSF Research Traineeship Grant #GER-9354869, and a Rehabilitation Engineering Research Center Grant from the National Institute on Disability and Rehabilitation Research of the U.S. Dept. of Education (#H133E30010).

References:

Michaud, Lisa N., and Kathleen F. McCoy. 1998. Planning Text in a System for Teaching English as a Second Language to Deaf Learners. In Proceedings of Integrating Artificial Intelligence and Assistive Technology, an AAAI '98 Workshop, Madison, Wisconsin, July 26.

McCoy, Kathleen F., and Lisa N. Masterman. 1997. A Tutor for Teaching English as a Second Language for Deaf Users of American Sign Language. In Proceedings of Natural Language Processing for Communication Aids, an ACL/EACL'97 Workshop, pages 160-164, Madrid, Spain, July.

McCoy, Kathleen F., Christopher A. Pennington, and Linda Z. Suri. 1996. English error correction: A syntactic user model based on principled mal-rule scoring. In Proceedings of User Modeling `96, Kailua-Kona, HI.

Moore, Johanna D. and Cecile L. Paris. 1992. Planning Text for Advisory Dialogues: Capturing Intentional and Rhetorical Information. In Computational Linguistics 19(4), 651-695.