<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-4001">
  <Title>InfoMagnets: Making Sense of Corpus Data</Title>
  <Section position="4" start_page="254" end_page="255" type="metho">
    <SectionTitle>
3 Implementation
</SectionTitle>
    <Paragraph position="0"> As mentioned previously, InfoMagnets uses Latent Semantic Analysis (LSA) to relate documents to InfoMagnets. LSA is a dimensionality reduction technique that can be used to compute the semantic similarity between text spans of arbitrary size. For a more technical overview of LSA, we direct the reader to (Landauer et al., 1998).</Paragraph>
    <Paragraph position="1"> The LSA space is constructed using the corpus that the user desires to organize, possibly augmented with some general purpose text (such as newsgroup data) to introduce more domain-general term associations. The parameters used in building the space are set by the user during pre-processing, so that the space is consistent with the semantic granularity the user is interested in capturing.</Paragraph>
    <Paragraph position="2"> Because documents (or topic-segments) tend to cover more than one relevant topic, our clustering approach is based on what are determined heuristically to be the most important terms in the corpus, and not on whole documents. This higher granularity allows us to more precisely capture the topics discussed in the corpus by not imposing the assumption that documents are about a single topic.</Paragraph>
    <Paragraph position="3"> First, all terms that occur less than n times and in less than m documents are removed from consideration null  . Then, the remaining terms are clustered via average-link clustering, using their LSA-based vector representations and using cosine-correlation as a vector similarity measure. Our clustering algorithm combines top-down clustering (Bisecting K-Means) and bottom-up clustering (Agglomerative Clustering) (Steinbach et al., 2000). This hybrid  n and m are parameters set by the user.</Paragraph>
    <Paragraph position="4">  clustering approach leverages the speed of bisecting K-means and the greedy search of agglomerative clustering, thus achieving a nice effectiveness versus efficiency balance.</Paragraph>
    <Paragraph position="5"> Cluster centroids (InfoMagnets) and documents (or topic segments) are all treated as bag-of-words. Their vector-space representation is the sum of the LSA vectors of their constituent terms. When the user changes the topic-representation by removing or adding a term to an InfoMagnet, a new LSA vector is obtained by projecting the new bag-of-words onto the LSA space and re-computing the cosine correlation between all documents and the new topic.</Paragraph>
  </Section>
  <Section position="5" start_page="255" end_page="255" type="metho">
    <SectionTitle>
4 An Example of Use
</SectionTitle>
    <Paragraph position="0"> InfoMagnets was designed for easy usability by both computational linguistics and non-technical users. It has been successfully used by social psychologists working on on-line communities research as well as learning science researchers studying tutorial dialogue interactions (which we discuss in some detail here).</Paragraph>
    <Paragraph position="1"> Using InfoMagnets, a thermodynamics domain expert constructed a topic analysis of a corpus of human tutoring dialogues collected during classroom study focusing on thermodynamics instruction (Rose et al., 2005). Altogether each student's protocol was divided into between 10 and 25 segments such that the entire corpus was divided into approximately 379 topic segments altogether. Using InfoMagnets, the domain expert identified 15 distinct topics such that each student covered between 4 and 11 of these topics either once or multiple times throughout their interaction.</Paragraph>
    <Paragraph position="2"> The topic analysis of the corpus gives us a way of quickly getting a sense of how tutors divided their instructional time between different topics of conversation. Based on this topic analysis of the human-tutoring corpus, the domain expert designed 12 dialogues, which were then implemented using a dialogue authoring environment called TuTalk (Gweon et al., 2005). In a recent very successful classroom evaluation, we observed the instructional effectiveness of these implemented tutorial dialogue agents, as measured by pre and post tests.</Paragraph>
  </Section>
class="xml-element"></Paper>