R. Pieraccini, E. Tzoukermann, Z. Gorelov, J. Gauvainand E. Levin, C. Lee, and J. Wilpon. A speech understanding system based on statistical representation of semantics. In Proceedings of the International Conference on Acoustics Speech Signal Processing, San Francisco, USA, 1992.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....as a topic identification or document classification problem, and for every input query the system outputs a single identified topic using a vector based information retrieval technique. Other previous approaches include: grammar based parsing [4] 5] stochastic concept decoding with HMMs [6]; and probabilisitic recursive transition networks [7] We have also applied a Nave Bayesian approach previously [8] 3. TASK DOMAIN Our experiments are based on the ATIS (Air Travel Information Systems) corpus [9] We use the Class A sentences of ATIS 3, which have disjoint training and test ....

Pieraccini, R., E. Tzoukermann, Z. Gorelov, J. Gauvain, E. Levin, C. Lee and J. Wilpon, "A Speech Understanding System Based on Statistical Representation of Semantics," Proceedings of ICASSP, 1992, pp. I-193 to I-196.

....or implicitly. The explicit formulation usually needs to represent M by a finite number of topic classes. Approaches that change language models according to the estimated topic class M (e.g. 9] correspond to this formulation. Approaches using probabilistic state transition networks [10] or HMM [11] for forming semantic language models are also classified into this category. A cache model [12] is one of the approaches in which M is implicitly represented. In this paper, we consider that M is represented by a co occurrence of words based on the distributional hypothesis by Harris [13] ....

R. Pieraccini, et al., "A Speech Understanding System based on Statistical Representation of Semantics, " Proc. ICASSP'92, pp. I-193-196, 1992.

....by the semantic structure S. Given the observation sequence O, the most likely S has to be found. To pursue this goal, the a posterioriprobability has to be maximized. It can be transformed using the Bayes formula: 1) We choose to calculate the probabilities and by using stochastic methods [2] [5] [6] Due to the high variety of and the conditional probability can not be estimated directly from a set of training data. Therefore, additional representation levels are necessary. Clearly defined is the word level W, which can be used to calculate as follows: 2) is irrelevant to the ....

R. Pieraccini et al.: A Speech Understanding System Based on Statistical Representation of Semantics, Proc. ICASSP 1992 (San Francisco, USA), pp. I 193 - I 196

....units (simply called semuns ) s n : 1) Each semun corresponds to exactly one significant word and not more than one insignificant word out of W. It expresses a small semantic partition of the utterance (i.e. the semantic contribution of the significant word) similar to conceptual labels [7] [8] Each semun with is an (X 2) tupel of a type , a value and X particular successor semuns 1 : 2) 1) Currently, we use semuns with 1 X 5 successors. S s 1 s 2 . s n . s N , s n S 1 n N t s n [ v s n [ q 1 s n [ q X s n [ s 2 ....

R. Pieraccini et al.: A Speech Understanding System Based on Statistical Representation of Semantics, Proc. ICASSP 1992 (San Francisco, USA), pp. I.193-I.196

....parsing has focused on the problem of syntactic analysis rather than semantic interpretation. However, a number of groups participating in the ARPA sponsored ATIS benchmark for speech understanding have used learned rules to perform some semantic interpretation. The Chronus system from AT T (Pieraccini et al. 1992) used an approach based on stochastic grammars. Another approach employing statistical techniques is the Hidden Understanding Models of Miller, et al. 1994) Kuhn and De Mori (1995) have investigated an approach utilizing semantic classification trees, a variation on decision trees familiar in ....

Pieraccini, R.; Tzoukermann, E.; Z. Gorelov, J. L. G.; Levin, E.; Lee, C. H.; and Wilpon, J. 1992. A speech understanding system based on statistical representation of semantics. In Proceedings ICASSP 92. I--193--I--196.

....language application task. The task is to map a sequence of words from a retranscription of spoken airline reservations to a sequence of meanings of the words in terms of task slots corresponding to the task parameters (destination, airline, departure time, etc. We make the same hypothesis as [2] which is that the meaning of a sentence can be expressed by a sequence of basic meaning units, and that there is a sequential correspondence between each of these units and a sub sequence of the words pronounced. In this particular case, task constraints allow us to use task parameters as meaning ....

....the potential of this approach with a real life application. This paper suggests an original network architecture that bases it processing not only on the past, but on its predictions of the future. We thus obtain a non stationary system that differs from HMM models like the one used in [2] in that the states are continuous and not discrete, and they depend on past and future states. It obtains good performance on a conceptual segmentation task performed on real life data but could be applied to other sequence processing tasks as well. Acknowledgments Ad elaide St evenin Barbier ....

R. Pierracini, E. Tzoukermann, Z. Gorelov, J.L. Gauvain, E. Levin, C.H. Lee, J.G. Wilpon (1992) A Speech Understanding System Based on Statistical Representation of Semantics, International Conference on Acoustics, Speech and Signal Processing, ICASSP 92.

....The conceptual approach presents the following advantages : many ambiguities can be avoided since some interpretations do not make sense for the task (each parameter can only be filled in once) syntactic issues which do not affect the task can be ignored. We make the same hypothesis as [3] : The meaning of a sentence can be expressed by a sequence of basic meaning units and there is a sequential correspondence between each of these units and a sub sequence of the words pronounced . We have conceived our system both for the association of concepts to words and for prediction of ....

....Concept prediction is much easier than word prediction because in a limited domain task, the number of concepts is often finite even though the number of instances may be very large. 2.1. Current approaches We quote below two approaches representative of important research directions in this area. [3] proposes a speech understanding system based on statistical representation of task specific semantic knowledge which uses Hidden Markov Models. This system presents many similarities with ours. Both rely on statistical feature extraction and allow an easy integration of successive processing ....

[Article contains additional citation context not shown here]

Pieraccini R., Tzoukermann E., Gorelov Z., Gauvain J.L, Levin E., Lee C.H, Wilpon J.G(92)- "A Speech Understanding System Based on Statistical Representation of Semantics", ICASSP

No context found.

R. Pieraccini, E. Tzoukermann, Z. Gorelov, J. Gauvainand E. Levin, C. Lee, and J. Wilpon. A speech understanding system based on statistical representation of semantics. In Proceedings of the International Conference on Acoustics Speech Signal Processing, San Francisco, USA, 1992.

No context found.

Pieraccini R., Tzoukermann E., GorelovZ., Gauvain J.L, Levin E., Lee C.H, Wilpon J.G(92)-"A Speech Understanding System Based on Statistical Representation of Semantics", ICASSP 92, San Francisco CA.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC