Improving Unsegmented Dialogue Turns Annotation with N-gram Transducers

Improving Unsegmented Dialogue Turns Annotation with N-gram Transducers⋆ Carlos-D. Mart´ınez-Hinarejos, Vicent Tamarit, and Jos´e-Miguel Bened´ı Instituto Tecnol´ogico de Inform´atica, Universidad Polit´ecnica de Valencia, Camino de Vera, s/n, 46022, Valencia, Spain {cmartine,vtamarit,jbenedi}@dsic.upv.es Abstract. The statistical models used for dialogue systems need annotated data (dialogues) to infer their statistical parameters. Dialogues are usually annotated in terms of Dialogue Acts (DA). The annotation problem can be attacked with statistical models, that avoid annotating the dialogues from scratch. Most previous works on automatic statistical annotation assume that the dialogue turns are segmented into the corresponding meaningful units. However, this segmentation is not usually available. Most recent works tried the annotation with unseg- mented turns using an extension of the models used in the segmented case, but they showed a dramatical decrease in their performance. In this work we propose an enhanced annota- tion technique based on N-gram transducers that outperforms the accuracy of the classical HMM-based model for annotation and segmentation of unsegmented turns. Keywords: Dialogue annotation, statistical models 1 Introduction A dialogue system is usually defined as a computer system that interacts with a human user to fulfil a task (Dybkjær and Minker, 2008). These systems are of particular interest in many applications, like information systems that are accessed by telephone (Seneff and Polifroni, 2000; Aust et al., 1995) or assistant systems for people with special necessities (Wilks, 2006). All the systems define the way they react to user inputs with the so-called dialogue strategy. This strategy can be rule-based (Gorin et al., 1997) or data-based (Young, 2000). In any case, the strategies are based on the interpretation of the user input in terms of dialogue semantic units. These semantic units are usually coded in terms of Dialogue Acts (DA) (Bunt, 1994), which model the intention of the current user interaction along with its associated infor- mation. This concept can be extended to system responses. In an interaction, several dialogue meaningful units can be distinguished. These units are called segments (or utterances according to authors such as (Stolcke et al., 2000)), and each segment has associated only one DA label. The annotation of a dialogue corpus in terms of DA is an interesting problem for both the devel- opment of data-based dialogue systems and the study of discourse and dialogue structure. In the first case, the statistical models that implement the dialogue manager (Williams and Young, 2007; Meng et al., 2003; Stolcke et al., 2000) rely on annotated dialogues to estimate their parameters. This annotation process is developed by human experts and it is a hard and time-consuming task. The use of probabilistic models can provide a draft annotation of the corpus (Stolcke et al., 2000) that can make the manual annotation process faster. Most of the previous works on the use of probabilistic models for DA annotation use segmented dialogue turns (Stolcke et al., 2000; Webb and Wilks, 2005; Rangarajan et al., 2007). However, this segmentation is not usually available in the initial transcription of a dialogue corpus. Other works propose a …

Content truncated to 500 words