Improving Unsegmented Statistical Dialogue Act Labelling

International Conference RANLP 2009 - Borovets, Bulgaria, pages 434–440 Improving Unsegmented Statistical Dialogue Act Labelling ∗ Vicent Tamarit, Carlos-D. Mart´ınez-Hinarejos and Jos´e-M. Bened´ı Ru´ız Instituto Tecnol´ogico de Inform´atica Universidad Polit´ecnica de Valencia Camino de Vera s/n, 46022, Valencia, Spain {vtamarit,cmartine,jbenedi}@iti.upv.es Abstract An important part of a dialogue system is the correct labelling of turns with dialogue-related meaning. This meaning is usually represented by dialogue acts, which give the system seman- tic information about user intentions. Each di- alogue act gives the semantic of a segment of a turn, which can be formed by several segments. Probabilistic models that perform dialogue act labelling can be used on segmented or unseg- mented turns. The last option is the more realis- tic one, but provides poorer results. An hypoth- esis on the number of segments can be provided in this case to improve the results. We propose some methods to estimate the probability of the number of segments based on the transcription of the turn. The new labelling model includes the estimation of the probability of the number of segments in the turn. The results show that this inclusion significantly improves the labelling accuracy. Keywords dialogue systems, dialogue act, statistical labelling 1 Introduction A dialogue system is usually defined as a computer system that interacts with a human user to achieve a task using dialogue [6]. The computer system must interpret the user input, in order to obtain the mean- ing and the intention of the user turn. This is needed to give the appropriate answer to the user. The se- lection of this answer, along with other decisions that the system can take, is guided by the so-called dialogue strategy. This dialogue strategy can be rule-based [8] or data-based [17]. In the rule-based alternative, the dialogue manager selects the set of actions based on a set of production rules, usually implemented by an expert. In the data-based alternative, there are some ways to build the dialogue system. One option is us- ing a dialogue manager whose parameters have been estimated from annotated data using supervised ma- chine learning techniques, but this approach only take into account the strategies seen in the training data. For this reason simulated users [13] and reinforcement ∗Work supported by the EC (FEDER) and the Spanish MEC under grant TIN2006-15694-CO2-01 and by the Span- ish research programme Consolider Ingenio 2010: MIPRCV (CSD2007-00018). learning [15] are also used to obtain a more robust estimation of the dialogue manager parameters. In either case, the dialogue strategy needs the inter- pretation of user turns to achieve the aim of the user. This interpretation must only take into account the essential information for the dialogue process, which is usually represented by special labels called Dia- logue Acts (DA) [4]. With this approximation, each user turn can be assigned a sequence of DAs, where each DA is associated with non-overlapped sequences of words in the turn. These sequences of words are usually called segments (some authors refers to these sequences as ”utterances” [14]). …

Content truncated to 500 words