Adaptive Parsing

As the computer gradually automates human-oriented tasks in multiple environ ments, the interface between computers and the ever-wider population of human users assumes progressively increasing importance. In the office environment, for instance, clerical tasks such as document filing and retrieval, and higher-level tasks such as scheduling meetings, planning trip itineraries, and producing documents for publication, are being partially or totally automated. The range of users for office oriented software includes clerks, secretaries, and businesspersons, none of whom are predominantly computer literate. The same phenomenon is echoed in the factory production line, in the securities trading floor, in government agencies, in educa tional institutions, and even in the home. The arcane command languages of yes teryear have proven too high a barrier for smooth acceptance of computerized func tions into the workplace, no matter how useful these functions may be. Computer naive users simply do not take the time to learn intimidating and complex computer interfaces. In order to place the functionality of modem computers at the disposition of diverse user populations, a number of different approaches have been tried, many meeting with a significant measure of success, to wit: special courses to train users in the simpler command languages (such as MS-DOS), designing point-and-click menu/graphics interfaces that require much less user familiarization (illustrated most clearly in the Apple Macintosh), and interacting with the user in his or her language of choice.

Inhalt

1. Introduction.- 1.1. Natural language interfaces.- 1.2. The dilemma.- 1.3. Towards a solution.- 1.4. The frequent user.- 1.5. Conclusions.- 1.6. Reader's guide.- 2. System Behavior in an Adaptive Environment.- 2.1. Foundations: Least-deviant-first parsing and MULTIPAR.- 2.2. The model.- 2.3. Assumptions and hypotheses.- 2.4. Adaptation vs. customizable and instructable interfaces.- 2.5. Prior research in adaptation.- 3. User Behavior in an Adaptive Environment.- 3.1. The behavioral hypotheses.- 3.2. The experimental condition.- 3.3. Control conditions.- 3.4. User profiles.- 3.5. Results and discussion.- 4. System Architecture and Knowledge Representation.- 4.1. Representing knowledge in CHAMP.- 4.2. Learning-related knowledge.- 4.3. Application-related knowledge.- 5. Understanding Non-Deviant Utterances.- 5.1. Segmentation.- 5.2. The Coalesce/Expand Cycle.- 5.3. A detailed example: the Coalesce/Expand Cycle for Can- 89 cel the 3 p.m. speech research meeting on June..- 6. Detecting and Recovering from Deviation.- 6.1. Parsing as least-deviant-first search.- 6.2. The cache.- 6.3. Error detection.- 6.4. A detailed example: Error detection during the parse of Schedule a meeting at 3 pm June 7..- 6.5. Error recovery.- 7. Resolution: Choosing Among Explanations.- 7.1. Converting meaning to action.- 7.2. Using default knowledge and inference.- 7.3. Interacting with the databases.- 7.4. Confirmation by effect.- 8. Adaptation and Generalization.- 8.1. Adapting to substitution deviations.- 8.2. Adapting to insertion deviations.- 8.3. Adapting to deletion deviations.- 8.4. Adapting to transposition deviations.- 8.5. A summary of adaptation in CHAMP.- 8.6. Controlling growth in the grammar through competition.- 8.7. The effects of adaptation on performance.- 9.Evaluating the Interface.- 9.1. CHAMP's performance on the hidden-operator data.- 9.2. CHAMP's performance in real user interactions.- 9.3. Summary of results.- 10. Critical Issues.- 10.1. The effect of kernel design on learning.- 10.2. The effect of kernel design on search.- 10.3. The effect of search constraints on system predictability.- 10.4. The lexical extension problem.- 10.5. Summary.- 11. Conclusions and Future Directions.- 11.1. Main results.- 11.2. Future directions.- References.