Exploiting multiple hypotheses for Multilingual Spoken Language Understanding

Proceedings of the Seventeenth Conference on Computational Natural Language Learning, pages 193–201, Sofia, Bulgaria, August 8-9 2013. c⃝2013 Association for Computational Linguistics Exploiting multiple hypotheses for Multilingual Spoken Language Understanding Marcos Calvo, Fernando Garc´ıa, Llu´ıs-F. Hurtado, Santiago Jim´enez, Emilio Sanchis Departament de Sistemes Inform`atics i Computaci´o Universitat Polit`ecnica de Val`encia, Val`encia, Spain {mcalvo,fgarcia,lhurtado,sjimenez,esanchis}@dsic.upv.es Abstract In this work, we present an approach for multilingual portability of Spoken Lan- guage Understanding systems. The goal of this approach is to avoid the effort of ac- quiring and labeling new corpora to learn models when changing the language. The work presented in this paper is focused on the learning of a specific translator for the task and the mechanism of transmitting the information among the modules by means of graphs. These graphs represent a set of hypotheses (a language) that is the input to the statistical semantic decoder that pro- vides the meaning of the sentence. Some experiments in a Spanish task evaluated with input French utterances and text are presented. They show the good behavior of the system, mainly when speech input is considered. 1 Introduction Spoken Language Understanding (SLU) is one of the key modules in many voice-driven human- computer interaction systems. Many successful SLU systems that have been developed in the last few years are based on statistical models automat- ically learned from semantically labeled corpora (Maynard and Lef`evre, 2001; Segarra et al., 2002; He and Young, 2006; Lef`evre, 2007; De Mori et al., 2008). One of the advantages of statistical models is the capability of representing the vari- ability of lexical realizations of concepts (mean- ings). On the other hand, they are usually plain models, that is, they can not represent a hierarchi- cal semantic dependency, although there are some works in this area (He and Young, 2003). How- ever, this is not a problem in most Spoken Dialog Systems since the semantic information to be ex- tracted is not very hierarchically structured. Another important aspect of these models is that they can be learned from corpora. The corpora used for training must be large enough to allow an accurate estimation of the probabilities, and it must represent the lexical and syntactic variabil- ity that is used in the language to express the se- mantics as much as possible. Although there are some approaches based on semi-supervised or un- supervised learning (T¨ur et al., 2005; Riccardi and Hakkani-T¨ur, 2005; Ortega et al., 2010), the most common approaches need to have a segmented and labeled training corpus. This is the case of discriminative models (like Conditional Random Fields (Hahn et al., 2010)), and generative models (such as Hidden Markov Models and Stochastic Finite State Automata (Segarra et al., 2002; Hahn et al., 2010)). In the case of supervised learn- ing, it is necessary to define a set of concepts that represent the semantic domain of the task and to associate these concepts to the corresponding se- quences of words in the sentences. This is the case of the French MEDIA corpus …

Content truncated to 500 words