Menu
Home
Log in / Register

 
Home arrow Computer Science arrow Social Computing, Behavioral-Cultural Modeling and Prediction
< Prev   CONTENTS   Next >

2 Related Work

There is a vast literature on the detection of leaders from text-based communications, both manually and automatically. For manual analysis, a leading approach [4], based solely on written text communications, is to predict the leader based upon activity level and written linguistic quality as a surrogate for communication/expression. Other indicators that can be manually detected are: lengthier messages, more complex and rich language, more confidence, and more emotional affect [5].

For automated analysis, a number of leadership indicators have been employed, the most common of which is dominance (the most utterances or longest turns) [5], [6]. Including variation in the tone of voice and energy improved accuracy in detecting emergent leaders over solely using dominance measures [7]. Indicators of initiative are also common, including being the first to speak [8], introducing new topics [9], and making activity-related utterances [6]. Finally, evidence of behaviors that are used by those higher in a power hierarchy when interacting with subordinates, such as giving praise, indicates leadership status (see [10], [11]).

Dialogue act analysis is a method for automatically capturing the domainindependent processes of interactions, regardless of the domain-related content. Taxonomies of dialogue acts (DAs) range from 10-20 [12] up to hundreds [13]. DAs, in their various conceptual frameworks, have been related to a range of group performance metrics (see [14]–[16]). Since DAs are useful for assessing performance, a number of groups have developed methods for automatically classifying them. These systems generally take one of two forms. The featureor rule-based approach uses lists of text features (cue phrases) as indicators of a certain category [17]. Machine learning techniques predict the DA of each utterance typically using a wider range of features, but these technique require large amounts of labeled training data.

The rule-based system we use was first developed for extracting dialogue acts from chat data in an Air Force Dynamic Targeting Cell [2] in which the development of hand-tagged corpora was prohibitive in terms of access, time and cost, so a featurebased system was created. In military domains in which there is difficulty obtaining data (especially annotated data), this approach using specific cue phrases to classify dialogue acts has yielded state-of-the-art results on domain-specific corpora [18].

 
Found a mistake? Please highlight the word and press Shift + Enter  
< Prev   CONTENTS   Next >
 
Subjects
Accounting
Business & Finance
Communication
Computer Science
Economics
Education
Engineering
Environment
Geography
Health
History
Language & Literature
Law
Management
Marketing
Philosophy
Political science
Psychology
Religion
Sociology
Travel