<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-2009">
  <Title>Spkr ID Words Discourse Chunk</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Discourse chunking
</SectionTitle>
    <Paragraph position="0"> In order to accomplish a mutual goal (for example, two people trying to find a suitable appointment time), dialogue participants engage in predictable kinds of activity, structuring the conversation in a coherent way in order to accomplish their goals.</Paragraph>
    <Paragraph position="1"> Alexandersson et al. (1997) have noted that these conversations tend to follow certain patterns, particularly with regard to the way that topics get raised and dealt with: Hello The dialogue participants greet each other. They introduce themselves, unveil their affiliation, or the institution or location they are from.</Paragraph>
    <Paragraph position="2"> Opening The topic to be negotiated is introduced.</Paragraph>
    <Paragraph position="3"> Negotiation The actual negotiation, between opening and closing.</Paragraph>
    <Paragraph position="4"> Closing The negotiation is finished (all participants have agreed), and the agreed-upon topic is (sometimes) recapitulated.</Paragraph>
    <Paragraph position="5"> Good Bye The dialogue participants say good bye to each other.</Paragraph>
    <Paragraph position="6"> Within a conversation, the opening-negotiationclosing steps are often repeated in a cyclical pattern. null This work on discourse chunking combines the opening, negotiation, and closing sections into a single chunk. One reason for this is that these parts of the conversation tend to act as a single chunk; when they appear, they regularly appear together and in the same order. Also, some of these parts may be missing; a topic of negotiation is frequently brought up and resolved without an explicit opening or closing. Very often, the act of beginning a topic of negotiation defines the opening by itself, and the act of beginning a new negotiation entails the closing of the previous one.</Paragraph>
    <Paragraph position="7"> A slightly simplified model of conversation, then, appears in Figure 2.1.</Paragraph>
    <Paragraph position="8"> In this model, participants greet each other, engage in a series of negotiations, and finish the conversation when the goals of the dialogue are satisfied.</Paragraph>
    <Paragraph position="9"> These three parts of the conversation are dialogue chunks. These chunks are relevant from a DA tagging perspective. For example, the DA tags used in one of these chunks are often not used in other chunks. For an obvious example, it would be almost unheard of for the GREET tag to appear in the Good Bye chunk. Other DAs (such as FEEDBACK_POSITIVE) can occur in any of the three chunks. Knowing which chunk we are in, and where we are within a chunk, can facilitate the tagging task.</Paragraph>
    <Paragraph position="10"> Within chunks, some patterns emerge. Note that in the example from the Verbmobil-2 corpus (shown in Table 2.1), a negotiation topic is raised, and dealt with (by an ACCEPT speech act). Then there follows a sequence of FEEDBACK_POSITIVEs as the negotiation topic winds down. This winding down activity is common at the end of a negotiation chunk. Then a new topic is raised, and the process continues.</Paragraph>
    <Paragraph position="11"> One-word utterances such as okay or yeah are particularly problematic in this kind of task because they have rather general semantic content and they are commonly used in a wide range of contexts. The word yeah on its own, for example, can indicate acceptance of a proposition, mere Speaker ID Words DA Tag KNT some other time oh actually I see that I have got some free time in like the fifth sixth and seventh of  acknowledgement of a proposition, feedback, deliberation, or a few of these at once (Core &amp; Allen 1997). In Verbmobil-2, these utterances can be labeled either ACCEPT, FEEDBACK_POSITIVE, BACK-CHANNEL, or REQUEST_COMMENT. Without knowing where the utterance appears within the structure of the dialogue, these utterances are very difficult to classify.</Paragraph>
    <Paragraph position="12"> Some previous work has used prosody to solve this kind of problem (as with Stolcke 2000). I propose discourse chunks as an alternative method. It can pull information from the text alone, without the computational overhead that prosody can entail.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Chunk segmentation
</SectionTitle>
    <Paragraph position="0"> Just where do the discourse chunk boundaries lie? For this exercise, I have constructed a very simple set of rules to determine chunk boundaries. These rules come from my observations; future work will involve automatic chunk segmentation. However, these rules do arise from a principled assumption: the raising of a new topic shows the beginning of a discourse chunk. Therefore, a speech act that (according to the definitions in Alexandersson 1997) contains a topic or proposition represents the beginning of a discourse chunk.</Paragraph>
    <Paragraph position="1"> By definition, only four DAs contain or may contain a topic or proposition. These are INIT, EXCLUDE, REQUEST_SUGGEST, and SUGGEST.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Chunking rules
</SectionTitle>
      <Paragraph position="0"> The chunking rules are as follows:  1. The first utterance in a dialogue is always the start of chunk 1 (hello).</Paragraph>
      <Paragraph position="1"> 2. The first INIT or SUGGEST or REQUEST_SUGGEST or EXCLUDE in a dialogue is the start of chunk 2 (negotiation). 3. INIT, SUGGEST, REQUEST_SUGGEST, or EXCLUDE marks the start of a subchunk within chunk 2.</Paragraph>
      <Paragraph position="2"> 4. If the previous utterance is also the start of a  chunk, and if it is spoken by the same person, then this utterance is considered to be a continuation of the chunk, and is not marked.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5. The first BYE is the start of chunk 3 (good
</SectionTitle>
    <Paragraph position="0"> bye).</Paragraph>
    <Paragraph position="1"> Items within a chunk are numbered evenly from 1 (the first utterance in a chunk) to 100 (the last), as shown in Table 3.1. This normalizes the chunk distances to facilitate comparison between utterances. null</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 The case-based reasoning (CBR) tagger
</SectionTitle>
    <Paragraph position="0"> A thorough discussion of this CBR tagger goes beyond the scope of this paper, but a few comments are in order.</Paragraph>
    <Paragraph position="1"> Case-based reasoning (Kolodner 1993) is a form of machine learning that uses examples. In general, classification using a case-based reasoner involves comparing new instances (in this case, utterances) against a database of correctly-tagged instances. Each new instance is marked with the same tag of its nearest neighbour (that is, the closest match) from the database. A k-nearest neighbour approach selects the closest k matches from the database to be committee members, and the committee members vote on the correct classification. In this implementation, each committee member gets a vote equal to its similarity to the test utterance. Different values of k performed better in different aspects of the test, but this work uses k = 7 to facilitate comparison of results.</Paragraph>
    <Paragraph position="2">  with discourse chunks.</Paragraph>
    <Paragraph position="3"> The choice of features largely follows those of Samuel 2000, and are as follows:  analysis of previous DA tags. Both previous DA tag and 2-previous DA tag features use the best guess for previous utterances rather than the right answer, so this run allows us to test performance even with incomplete information.</Paragraph>
    <Paragraph position="4"> Discourse chunk tag Distances for this tag were computed by dividing the larger discourse chunk number from the smaller. Comparing two chunk starter utterances would give the highest similarity of 1, and comparing a chunk starter (1) to a chunk-ender (100) would give a lower similarity (.01).</Paragraph>
    <Paragraph position="5"> Not all features are equally important, and so an Evolutionary Programming algorithm (adapted from Fogel 1994) was used to weight the features. Weightings were initially chosen randomly for each member of a population of 100, and the 10 best performers were allowed to survive and mutate their weightings by a Gaussian random number. This was repeated for 10 generations, and the weightings from the highest performer were used for the CBR tagging runs.</Paragraph>
    <Paragraph position="6"> A total of ten stopwords were used (the, of, and, a, an, in, to, it, is, was), the ten most common words from the BNC (Leech, Rayson, &amp; Wilson 2001). These stopwords were removed when considering word similarity, but not n-gram similarity, since these low-content words are useful for distinguishing sequences of words that would otherwise be very similar.</Paragraph>
    <Paragraph position="7"> The database consisted of 59 hand-tagged dialogues (8398 utterances) from the Verbmobil-2 corpus. This database was also automatically tagged with discourse chunks according to the rules above. The test corpus consisted of 20 dialogues (2604 utterances) from Verbmobil-2. This corpus was tagged with correct information on discourse chunks; however, no information was given on the DA tags themselves.</Paragraph>
  </Section>
class="xml-element"></Paper>