Real-Time Natural Language Generation in NL-SOAR

NL-Soar is a computer system that performs language comprehension and generation within the framework of the Soar architecture [New90]. NL-Soar provides language capabilities for systems working in a real-time environment. Responding in real time to changing situations requires a flexible way to shift control between language and task operations. To provide this flexibility, NL-Soar organizes generation as a sequence of incremental steps that can be interleaved with task actions as the situation requires. This capability has been demonstrated via the integration of NL-Soar with two different independentlydeveloped Soar-based systems. 1 Real-time generation and NL-Soar NL-Soar is a language comprehension and generation facility designed to provide integrated real-time 1 language capabilities for other systems built within the Soar architecture [New90]. In particular, this requires integrating NL-Soar's generation subsystem with other task(s) that Soar is performing. 2 One possible way to achieve this integration would be to use generation as a kind of "back end" which other task(s) can call as a subroutine whenever they need to say something. This approach is widely used in applications such as database query systems or expert systems, where the main system invokes the generator to express the answer to a query or to explain some aspect of its reasoning or conclusions. In applications that need to provide real-time behavior, though, this "subroutine" approach is problematic. There is no way for the task to interrupt generation in order to handle some other (perhaps urgent) work. In addition, if generation is an unbounded process, it may proceed to complete an utterance that may have become unnecessary or even harmful because of changes in the situation; the task has no way to modify what it wants to say once generation has been invoked. While the task could of course simply stop NL-Soar in either of these cases, I NL-Soar is being used in applications that perform in both simulated and actual real-time environments. 2Similar issues arise in NL-Soar's language comprehension subsystem, which is not described here; see [LLN91, Lew93] for a discussion of this part of NL-Soar. there is no way to guarantee that generation will be interrupted in a state from which it can recover if reinvoked later. Furthermore, the problem isn't simply one of the speed of generation; using faster processors to run Soar won't eliminate the difficulties. It might seem that we could simply assume NL-Soar can run fast enough to finish constructing an utterance before the task has time to do anything else, this is not the case. First, generation is potentially unbounded; no matter how fast a computer is used, there will still be occasions when NL-Soar takes longer than the task can afford to wait. More significantly, this assumes that NL-Soar can absorb all the speedup; this is not reasonable. I f we have faster processors, we want all the tasks to share the speedup equally; a faster NL-Soar will be invoked by a task that can respond more quickly as well, and will thus want to interrupt NL-Soar more quickly. The underlying difficulty with the subroutine approach is that generation can take an unbounded amount of time; in a real-time situation, we need to guarantee that generation can't prevent the task from responding promptly to changes in the situation. Generation must be incremental and interruptible. NL-Soar accomplishes this by dividing generation into small steps that Can be interleaved with task steps. In cases where the small steps can' t be directly carried out and require more complex computation, the sub-steps are designed so that interruptions leave the system in a clean state (although some work may be lost and need to be redone). This allows NL-Soar to operate without limiting the task's ability to respond to things that happen during generation, and vice versa. 2 A Brief Introduction to Soar 3 Soar is an architecture for building cognitive models and AI systems that has been applied to a wide range of problems [RLN93]. Soar carries out a task by applying a sequence of operators to the state of a problem-space until it reaches a state that solves the goal Soar is working on. When Soar is unable to directly carry out some step (e.g. selecting or applying an operator or selecting a problem space), it creates a subgoal to resolve the impasse that is preventing it from proceeding. In response to this subgoal, Soar selects an appropriate problem 3For a more detailed description of Soar than is possible here, see [LNR87] or [New90"l.