<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1022">
  <Title>Addressee Identification in Face-to-Face Meetings</Title>
  <Section position="7" start_page="174" end_page="175" type="concl">
    <SectionTitle>
6 Conclusion and Future work
</SectionTitle>
    <Paragraph position="0"> We presented results on addressee classification in four-participants face-to-face meetings using Bayesian Network and Naive Bayes classifiers.</Paragraph>
    <Paragraph position="1"> The experiments presented should be seen as preliminary explorations of appropriate features and models for addressee identification in meetings.</Paragraph>
    <Paragraph position="2"> We investigated how well the addressee of a dialogue act can be predicted (1) using utterance, gaze and conversational context features alone as well as (2) using various combinations of these features. Regarding gaze features, classifiers' performances are measured using gaze directional cues of the speaker only as well as of all meeting participants. We found that contextual information aids classifiers' performances over gaze information as well as over utterance information.</Paragraph>
    <Paragraph position="3"> Furthermore, the results indicate that selected utterance features are the most unreliable cues for addressee prediction. The listeners' gaze direction provides useful information only in the situation where gaze features are used alone. Combinations of features from various resources increases classifiers' performances in comparison to performances obtained from each resource separately.</Paragraph>
    <Paragraph position="4"> However, the highest accuracies for both classifiers are reached by combining contextual and utterance features with speaker's gaze (BN:82.59%, NB:78.49%). We have also explored the effect of meeting context on the classification task.</Paragraph>
    <Paragraph position="5"> Surprisingly, addressee classifiers showed little gain from the information about meeting actions (BN:83.74%, NB:79.90%). For all feature sets, the Bayesian Network classifier outperforms the Naive Bayes classifier.</Paragraph>
    <Paragraph position="6"> In contrast to Vertegaal (1998) and Otsuka et al. (2005) findings, where it is shown that gaze can be a good predictor for addressee in four-participants face-to-face conversations, our results  show that in four-participants face-to-face meetings, gaze is less effective as an addressee indicator. This can be due to several reasons. First, they used different seating arrangements which is implicated in the organization of gaze. Second, our meeting environment contains attentional 'distracters' such as whiteboard, projector screen and notes. Finally, during a meeting, in contrast to an ordinary conversation, participants perform various meeting actions which may influence gaze as an aspect of addressing behavior.</Paragraph>
    <Paragraph position="7"> We will continue our work on addressee identification on the large AMI data collection that is currently in production. The AMI corpus contains more natural, scenario-based, meetings that involve groups focused on the design of a TV remote control. Some initial experiments on the AMI pilot data show that additional challenges for addressee identification on the AMI data are: roles that participants play in the meetings (e.g. project manager or marketing expert) and additional attentional 'distracters' present in the meeting room such as, the task object at first place and laptops.</Paragraph>
    <Paragraph position="8"> This means that a richer feature set should be explored to improve classifiers' performances on the AMI data including, for example, the background knowledge about participants' roles. We will also focus on the development of new models that better handle conditional and contextual dependencies among different types of features.</Paragraph>
  </Section>
class="xml-element"></Paper>