Page:Advanced Automation for Space Missions.djvu/24



2.2.4 Natural Language Interface

The primary reason for a natural language interface is to allow the largest possible number of novice users to access IESIS directly without need for "interpreters." The principal advantage for more knowledgeable customers is convenience, though many customers will likely require a more formal and precise interface language in addition to the natural language capability. There are major problems with current natural language interfaces that require careful consideration. The two primary difficulties are:

(1) The machine has only a very rudimentary notion of the subject of conversation during sentence interpretation even if it is quite competent to answer questions posed in a formal query language.

(2) The machine is not a social being and thus is deaf to most of the subtle information content in sentences generated by people.

Such flaws exist in all current natural language systems. For the anticipated "naive" user these obstacles must be minimized or IESIS will prove extremely uncomfortable and time-consuming. These flaws could possibly lead to system avoidance by less-sophisticated users, thus defeating one of the major mission goals.

Perhaps the simplest way to overcome such problems in natural language systems is to restrict the domain of discourse to a small set of possible concepts keyed to known individual human differences. Ultimately, the following may be the best approach for the "naive" user: Each would have a personal identification code known to IESIS, permitting the system to adjust its language to a compatible dialect. Knowledge of customer category could enable IESIS to employ reasonable default assumptions in restricting the domain of discourse (and thus the vocabulary) to a manageable subset of the overall system domain.

It is important that IESIS be able to communicate at the appropriate level of complexity and brevity. To accomplish this requires a system capability of modeling individual users. Some initial work in this area has been done (Rich, 1979), but no known current technique offers the level of performance necessary for IESIS. Natural language interfacing is one area that requires considerable advancement before it can hope to meet the IESIS system requirements: domain model, user model, dialogue model, reasonable default assumptions and common world knowledge, and explanatory capabilities.

2.2.5 Artificial Intelligence Problem-Solving

Clearly IESIS presents the usual difficulties in problem- solving typically involved in AI question-answering systems. But there is another new and important dimension added - effective combination of a world model database, world model theory, and a potentially resource-limited observational capability. The power of the problem-solvers and planners, and their capacity to execute plans in a dynamic and only partially known environment, will be instrumental in achieving a high-quality information delivery service at minimum cost.

Two specific areas where the quality of problem solving affects overall system efficiency and cost were considered for illustrative purposes. The first is communication link capacity. Given the goal of answering a large number of information requests, an intelligent planner able to isolate a parsimonious set of observations can considerably reduce ground link and intrasatellite link volumes. This set of observations is determined by consideration of individual requests, the ensemble of all current requests, and a prediction of expected requests.

A second area of concern is the number of satellites. If the system can employ a very sophisticated theory of observation, then it may be possible to shift most data-taking tasks to lower resolution. This system would allow data- taking by orbiters at higher altitudes having greater fields of view; thus a smaller total number of satellites would achieve the same frequency of coverage.

A major IESIS goal is to perform appropriate automatic data interpretation. System success in this activity demands a high-level capability to understand relationships between sensor readings and the actual state of the world as defined by human-oriented descriptors. This is precisely the problem in visual perception, an active current area in the field of artificial intelligence. Section 2.2.6 further discusses several aspects of the perception problem for an Earth- sensing system, and section 2.2.7 describes the need for flexibility and adaptability in IESIS. In both areas - perception, and system flexibility and adaptability - there is a tremendous need for development of fully autonomous techniques far more powerful than those presently available.

2.2.6 Theory of Observation

While the number of distinguishable states of the world of human interest (at a particular level of resolution and description) is extremely large, this figure is still dwarfed by the vast number of distinguishable ways the world may appear to rudimentary sensors. Machine sensors and human eyes see entirely different things when minor changes in the world state occur. For instance, in hilly terrain at low sun angle, satellites sensor readings vary rapidly as the shadows progress, but most of what is of interest to human beings is invariant.

To extract interesting information from sensors it is necessary first to understand the nature of the sensor as a transducer so that a mathematical inversion process can be performed on the readings. This involves computation of the electromagnetic reality at the image sensor location (i.e.,