Contrasting the Automatic Identification of Two Discourse Markers in Multiparty Dialogues
Sandrine Zufferey, Andréi Popescu-Belis · BORIS (University Library Bern) · 2007
The identification of occurrences of like and well that serve as discourse markers (DMs) is a classification problem which is studied here on a corpus of dialogue transcripts with more than 4,000 occurrences of each item. Decision trees using item-specific lexical, prosodic, positional and sociolinguistic features are trained using the C4.5 method. The results demonstrate improvement over past experiments, reaching the same range as inter-annotator agreement scores. DM identification appears to benefit from itemspecific classifiers, which perform better than general purpose ones, thanks to the differentiated use of lexical features. 1