<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-0211">
  <Title>Discourse Processing for Explanatory Essays in Tutorial Applications</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Whereas most explanations are produced and adapted to benefit or inform a hearer, a self-explanation is produced for the benefit of the speaker. If there is a hearer he often already knows all about the topic as in a tutoring context. Self-explanation is a cognitively valuable pedagogical activity because it leads students to construct knowledge (Chi et al., 1994), and it can expose deep misconceptions (Slotta et al., 1995). But it is difficult to encourage self-explanation without giving the student substantive feedback on what they generate (Aleven and Koedinger, 2000; Chi et al., 2001). To give substantive feedback the system has to be able to understand student explanations to some degree.</Paragraph>
    <Paragraph position="1"> The Why-Atlas system presents students with qualitative physics problems and encourage them to write their answers along with detailed explanations for their answers. While physics misconceptions have proven to be particularly resistant to repair, practice with qualitative physics questions helps in overcoming some of these misconceptions (Hake, 1998).</Paragraph>
    <Paragraph position="2"> The student explanation shown in (1), which is from our corpus of human-human computer-mediated tutoring sessions, illustrates how challenging these explanations are for a system to understand. The problems we have examined require a short essay with an average of 6.9 sentences to fully explain to the satisfaction of experienced physics instructors. null  (1) Question: Suppose you are running in a straight line at constant speed. You throw a pumpkin straight up.</Paragraph>
    <Paragraph position="3">  Where will it land? Explain.</Paragraph>
    <Paragraph position="4"> Explanation: Once the pumpkin leaves my hand, the horizontal force that I am exerting on it no longer exists, only a vertical force (caused by my throwing it). As it reaches it's maximum height, gravity (exerted vertically downward) will cause the pumpkin to fall. Since no horizontal force acted on the pumpkin from the time it left my hand, it will fall at the same place where it left my hands.</Paragraph>
    <Paragraph position="5"> Statistical text classification approaches, such as latent semantic analysis (Landauer et al., 1998), have shown promise for classifying a student explanation into medium-grained good and bad categories (Graesser et al., 2000). For instance, a medium-Philadelphia, July 2002, pp. 74-83. Association for Computational Linguistics. Proceedings of the Third SIGdial Workshop on Discourse and Dialogue, grained category that should match (1) is the oftenobserved impetus misconception: If there is no force on a moving object, it slows down.</Paragraph>
    <Paragraph position="6"> Such medium-grained categories typically have multiple propositions and contain multiple content words. While successful with medium-grained classes, statistical approaches are not yet able to distinguish subtle but important differences between good and bad explanations. Statistical classification is insensitive to negations1, anaphoric references2, and argument ordering variations3 and its inferencing is weak4. To capture these subtle differences and to allow us to respond more directly to what the student actually said5, we need the precision possible so far only with symbolic approaches. So Why-Atlas parses each sentence into a propositional representation. null The PACT Geometry Tutor is an operational prototype that does a finer-grained symbolic classification (Aleven et al., 2001). PACT also parses a student explanation into a propositional representation but then uses LOOM to classify these into fine-grained categories that typically express one proposition. This approach looks promising (Aleven et al., 2001), but the system's goal is to elicit a justification for a step in a geometry proof and generally these can be expressed with a single sentence that succinctly translates into a small number of propositions. It isn't clear that this approach will work well for the longer, more complex explanations that the Why-Atlas system elicits.</Paragraph>
    <Paragraph position="7"> Instead of classifying propositions, the Why-Atlas system constructs abductive proofs of them.</Paragraph>
    <Paragraph position="8">  knowledge the correct statement and elicit more precision rather than continuing as if it were wrong. For example, if a student makes a correct statement about the velocity of an object but did not report it in terms of the horizontal and vertical components of the velocity, the tutor should ask which was intended. A proof-based approach gives more insight into the line of reasoning the student may be following across multiple sentences because proofs of the propositions share subproofs. Indeed, one proposition's entire proof may be a subproof of the next proposition. Moreover, subtle misconceptions such as impetus are revealed when they must be used to prove a proposition.</Paragraph>
    <Paragraph position="9"> Abductive inference has a long history in plan recognition, text understanding and discourse processing (Appelt and Pollack, 1992; Charniak, 1986; Hobbs et al., 1993; McRoy and Hirst, 1995; Lascarides and Asher, 1991; Rayner and Alshawi, 1992). We are using an extended version of SRI's Tacitus-lite weighted abductive inference engine (Hobbs et al., 1993) as our main tool for building abductive proofs. We had to extend it in order to use it for domain as well as language reasoning. As advised in (Appelt and Pollack, 1992), abductive inference requires some application specific engineering to become a practical technique.</Paragraph>
    <Paragraph position="10"> In this paper we describe how the system creates and utilizes a proof-based representation of student essays. We describe how it creates the proof given the output of sentence-level understanding, how it uses the proofs to give students feedback, some preliminary run-time measures, and the work we are currently doing to derive additional benefits from a proof-based approach for tutoring applications.</Paragraph>
    <Paragraph position="11"> First we give an overview of the Why-Atlas tutoring system architecture. Next we give some background on weighted abduction and Tacitus-lite+ and describe how it builds an abductive proof. Next we describe how the system uses the proofs to give students feedback on their essays. Finally, we discuss efficiency issues and our future evaluation plans.</Paragraph>
  </Section>
class="xml-element"></Paper>