Reconhecimento de actos de diálogo hierárquicos e multi-etiqueta em dados em Espanhol

Abstract

Dialog acts reveal the intention behind the utte-red words. Thus, their automatic recognition is im-portant for a dialog system trying to understand itsconversational partner. The study presented in thisarticle approaches that task on the DIHANA corpus,whose three-level dialog act annotation scheme posesproblems which have not been explored in recent stu-dies. In addition to the hierarchical problem, the twolower levels pose multi-label classification problems.Furthermore, each level in the hierarchy refers to adifferent aspect concerning the intention of the spea-ker both in terms of the structure of the dialog and thetask. Also, since its dialogs are in Spanish, it allows usto assess whether the state-of-the-art approaches onEnglish data generalize to a different language. Morespecifically, we compare the performance of differentsegment representation approaches focusing on bothsequences and patterns of words and assess the impor-tance of the dialog history and the relations betweenthe multiple levels of the hierarchy. Concerning thesingle-label classification problem posed by the top le-vel, we show that the conclusions drawn on Englishdata also hold on Spanish data. Furthermore, weshow that the approaches can be adapted to multi-label scenarios. Finally, by hierarchically combiningthe best classifiers for each level, we achieve the bestresults reported for this corpus.