Dialogue Act Recognition in Estonian Dialogues Using Artificial Neural Networks

This paper describes two experiments of applying dialogue act recognition to estonian dialogues. Two class systems were used in both of them—a general one (with 19 classes) and a detailed one (with 107 classes). In the first experiment the task was performed using learning vector quantization (LVQ). The preprocessing was done in WEBSOM (Self-Organizing Maps for Internet Exploration) style; the used method was originally designed for processing documents with self-organizing maps, with which LVQ shares its principle. The weighing method was tf×idf, which is the most popular method for term weight assignment. The first experiment wasn’t a success. The proposed explanation is that due to the difference between utterances and text documents the combination of LVQ and the used preprocessing method were too straightforward for this task, and a more sofisticated classifier was to be tested. In the second experiment multilayer perceptron (MLP) techniques were applied to data, preprocessed in the same way as in the first experiment. The tested networks had one and two hidden layers. The experiment ended successfully; networks with two hidden layers had shown the highest accuracies with both class systems.