<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0205">
  <Title>A Comparison of Tutor and Student Behavior in Speech Versus Text Based Tutoring</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Why2-Atlas and ITSPOKE Dialogue
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Systems
</SectionTitle>
      <Paragraph position="0"> Why2-Atlas is a text based intelligent tutoring dialogue system (Ros*e et al., 2002a; VanLehn et al., 2002). The goal of Why2-Atlas is to provide a platform for testing whether deep approaches to natural language processing elicit more learning than shallower approaches, for the task domain of qualitative physics explanation generation. Using Why2-Atlas, the activity in which students engage is answering deep reasoning questions involving topics in conceptual physics. One such question that we used is, A lightweight car and a massive truck have a head-on collision. On which vehicle is the impact force greater? Which vehicle undergoes the greater change in its motion? Explain. This is an appropriate task domain for pursuing questions about the bene ts of tutorial dialogue for learning because questions like this one are known to elicit robust, persistent misconceptions from students, such as heavier objects exert more force.</Paragraph>
      <Paragraph position="1"> (Hake, 1998; Halloun and Hestenes, 1985). We designed a set of 10 essay questions to use as training problems.</Paragraph>
      <Paragraph position="2"> Two physics professors and a computer science professor worked together to select a set of expectations (i.e., correct propositions that the tutors expected students to include in their essays) and potential misconceptions associated with each question. Additionally, they agreed on an ideal essay answer for each problem. In Why2-Atlas, a student rst types an essay answering a qualitative physics problem. A computer tutor then engages the student in a natural language dialogue to provide feedback, correct misconceptions, and to elicit more complete explanations. The rst version of Why2-Atlas was deployed and evaluated with undergraduate students in the spring of 2002; the system is continuing to be actively developed (Graesser et al., 2002).</Paragraph>
      <Paragraph position="3"> We are currently developing a speech-enabled version of Why2-ATLAS, called ITSPOKE (Intelligent Tutoring SPOKEn dialogue system), that uses the Why2-Atlas system as its back-end . To date we have interfaced the Sphinx2 speech recognizer (Huang et al., 1993) with stochastic language models trained from example user utterances, and the Festival speech synthesizer (Black and Taylor, 1997) for text-to-speech, to the Why2-Atlas backend. The rest of the needed natural language processing components, e.g. the sentence-level syntactic and semantic analysis modules (Ros*e, 2000), discourse and domain level processors (Makatchev et al., 2002), and a nite-state dialogue manager (Ros*e et al., 2001), are provided by a toolkit that is part of the Why2-Atlas backend. The student speech is digitized from microphone input, while the tutor's synthesized speech is played to the student using a speaker and/or headphone. We are now in the process of adapting the knowledge sources needed by the spoken language components to our application domain. For example, we have developed a set of dialogue dependent language models using the experimental human-computer typed corpus (4551 student utterances) obtained during the Why2-Atlas 2002 evaluation. Our language models will soon be enhanced using student utterances from our parallel human-human spoken language corpus.</Paragraph>
      <Paragraph position="4"> One goal of the ITSPOKE system is simply replacing text based dialogue interaction with spoken dialogue interaction and leaving the rest of the Why2-Atlas back-end unchanged, in order to test the hypothesis that student self-explanation (which leads to greater learning (Hausmann and Chi, 2002)) might be easier to achieve in spoken dialogues. This hypothesis is discussed further in Section 5. Although not the focus of this paper, another goal of the ITSPOKE system is to take full advantage of the speech modality. For example, speech contains rich acoustic and prosodic information about the speaker's current emotional state that isn't present in typed dialogue. Connections between learning and emotion have been well documented (Coles, 1999), so it seems likely that the success of computer-based tutoring systems could be greatly increased if they were capable of predicting and adapting to student emotional states, e.g. reinforcing positive states, while rectifying negative states (Evens, 2002). Preliminary machine learning experiments involving emotion annotation and automatic feature extraction from our corpus suggest that ITSPOKE can indeed be enhanced to automatically predict and adapt to student emotional states (Litman et al., 2003).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>