Advanced search


Fabrizio Ghigi, Carlos D. Martínez-Hinarejos, José-Miguel Benedí. Active Learning to speed-up the process for Dialogue Act Labelling. Proceedings of 5th Language and Technology Conference, 2011, 2011. Zygmunt Vetulani (Editors). pp. 314-318. Adam Mickiewicz University.

The dialogue act labelling task is the process of splitting and annotating a dialogue into dialogue meaningful units; the labelling task can be performed semi-automatically by using statistical models trained from previously annotated dialogues. The appropriate selection of training dialogues can make the process faster, and Active Learning is one suitable strategy for this selection. In this work, Active Learning based on two different criteria (Weighted Number of Hypothesis and Entropy) has been tested for the task of dialogue act labelling by using the N-gram Transducers models. The framework was tested against two heterogeneous corpora, DIHANA and SwitchBoard. The results confirm the goodness of this kind of selection strategy.