Natural-language processing for computer-supported instruction

Introduction Many computer applications are concerned with interpreting or producing instructions and fostering education, for example, 1. Animated agents that execute instructions [Webber 1995], 2. Systems that automatically produce instructional text (see Excerpt A in Figure 1) [Paris 1995], 3. Intelligent tutoring systems (ITS) that help a student master a certain subject [Anderson 1995, Schulze 2000], and 4. Systems that facilitate student collaboration [Soller 2001]. Of these four examples, only example two is a true natural-language (NL) system: The animated agents in example one may be instructed via a menu; an ITS may provide feedback to the student via graphics. However, all the systems in examples one, three, and four potentially benefit from a natural-language interface. For instance, consider the learning gain, that is, how much a student learns in a certain setting (S). Generally, the learning gain is the difference between the student's score on the same test, before and after (S). It has been shown that the learning gain of students interacting with an ITS is halfway between the learning gain of students that were exposed to the material in the usual classroom setting (lowest) and students that interact with a human tutor (highest) [Anderson 1995]. The difference in learning gain between students interacting with an ITS and those interacting with a human tutor is attributed conversation between tutor and student Figure 1 (see page 24) presents two samples of NL instructions: The first is taken from an online help file, the second from a tutoring dialogue. These two examples illustrate some of the problems faced by systems that must support the interpretation and generation of instructions. Instructions in a technical manual, online help, or home repair manual such as those under (A) in teach how to perform a task mainly by describing the steps to be performed. They often include descriptions of what will happen as the result of a certain step (e.g., the window expands to show the Alarm options in step two) as a way to inform the user whether he or she is on the right track. The structure of the text closely reflects the structure of the task, as has long been known regarding task-oriented discourse [Grosz and Sidner 1986]. Tutorial dialogues such as (B) in Figure 1 present a completely different approach to instruction. (Note: This excerpt is taken from a dialogue in the BEESIM project corpus [Rosé, Di Eugenio, and Moore …