Automatic Identification of Discourse Markers in Multiparty Dialogues

The lexical items that can serve as discourse markers (DMs) are often multi-functional. Like and well, in particular, play numerous other roles apart from DMs: for instance, the first one can also be a verb and the second one an adverb. The goal of the present study is the identification, on transcripts of multi-party dialogues, of the occurrences of like and well that play a discourse or pragmatic role. DM identification is a binary classification task over the set of all occurrences of tokens like and well. The importance of DMs to computational linguistics is first discussed, along with previous experiments in DM identification. Then, the data is briefly described, emphasizing the DM annotation procedure and an inter-annotator agreement study. The proposed method uses lexical, prosodic/positional and sociolinguistic features, together with machine learning algorithms, among which decision trees are preferred. The results obtained using a ten-fold cross-validation procedure are analysed at length, focussing first on overall performance, and then on the relevance of each type of features. Feature analysis using a range of techniques shows that lexical indicators are the most reliable features for DM identification, followed by prosodic/positional features. Sociolinguistic features are slightly correlated with the use of like as DM, while the dialogue act of the utterance containing a DM candidate does not seem relevant to DM identification. A differentiated treatment for each token appears to improve performance in almost all experiments. The methods and features used here improve performance over the past experiments, and suggest that DM identification is a tractable problem provided enough training data is available for each DM type, and that lexical features are used for each type.