Gunther Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on architectures for integrated speech/language systems in verbmobil. In The 16th International Conference on Computational Linguistics, volume 1, pages 484--489, Copenhagen, Danmark, August5--9 1996.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Many Organization Project funded by the German Federal Ministery for Research and Technology (BMFT 01 IV 101 H 9) involving many partners (22 universities, 7 companies) Status Experimental, working version available. Phase I completed) Home page http: www.dfki.de verbmobil References [GKSW96] TD15) Purpose Verbmobil: Translation of spontaneous speech in face to face situations. TP15 subproject: parallel and distributed parsing for spontaneous speech (amongst others) Domain Coverage The probabilistic unification grammar that is used is comprised of 700 lexical entries and 60 ....

Gunther Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on architectures for integrated speech/language systems in verbmobil. In The 16th International Conference on Computational Linguistics, volume 1, pages 484--489, Copenhagen, Danmark, August5--9 1996.

....possible, and . the total communication between processors should be as small as possible. IMPACT NLI 1998 6 8 2 Basics IMPACT The requirements on communication are based on observations showing that communication tends to be expensive, and can slow down parsing considerably [Tho91] Tho94] [GKSW96], dV93] In addition to the temporal aspects of the problem, we also pose the requirement that the total storage used by all processors be of the same order as the storage used by sequential implementations. Most practical natural language parsers require a vast amount of memory. It would simply ....

....routing policies IMPACT NLI 1998 6 IMPACT 3.2 Deriving Justification Graphs 17 into the simulation is whether the communication bottlenecks that are often encountered in practice can indeed be explained at this higher level of abstraction. We believe this is the case. All parsers reported in [GKSW96, dV93] displayed a high degree of imbalance between the processors. These bottlenecks would very likely also show up when only simulating at the level of distribution schemes. Also the large amount of communication reported in [Tho91] would show up in such simulations, because these aspects are largely ....

Gunther Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on architectures for integrated speech/language systems in Verbmobil. In The 16th International Conference on Computational Linguistics, volume 1, pages 484--489, Copenhagen, Danmark, August5--9 1996. 8, 17

....the potential overhead of the system. In addition, this fine granularity makes it hard to implement on distributed systems. Nevertheless, despite all arguments in favor of the first approach, most attempts to exploit parallelism at the contextfree level were not very successful (e.g. 5, 6] [7], 8] Experiments in [9] suggest that many of the failed attempts to come to an efficient parallel parser can be attributed to the nature of the parallelism inherent in unification based parsers. In this paper, we will present some of these experimental results. After pinpointing the problems, ....

....using a distributed sharedmemory model. In addition to being very specific to parsing, this solution exploits coarse grained parallelism of the kind that proved unusable for the Deltra system. We suspect that many other NLP systems cannot benefit from this approach either (see, e.g. remarks in [7]) Our approach provides a solution in case fine grained parallelism is required. 3 Analysis of Parallelism The Deltra system was used as the basis of our experiments. It is based on a DCG like formalism, as found in Prolog. The grammar used in the experiments covers a wide range of linguistic ....

Gunther Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on architectures for integrated speech/language systems in Verbmobil. In The 16th International Conference on Computational Linguistics, volume 1, pages 484--489, Copenhagen, Danmark, August5--9 1996.

....to concentrate on the more coarse grained solution of parallelizing the control structure, taking unification as an atomic operation. Often, however, the size of unifications within a parsing operation vary wildly. This thwarts an even distribution of unification operations amongst processors. [GKSW96] mentions that in some cases it might even be possible that one unification takes up half of the total parsing time. Therefore, it might still be profitable to investigate the possibility of parallel unification. As we mentioned before, it is not easy to find an efficient parallel algorithm for ....

Gunther Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on architectures for integrated speech/language systems in verbmobil. In The 16th International Conference on Computational Linguistics, volume 1, pages 484--489, Copenhagen, Danmark, August5--9 1996.

....Environment (Amtrup, 1995) is an environment for the development of distributed AI systems . As part of the Verbmobil real time speech to speech translation project ICE has addressed two key problems for this type of system, viz. distributed processing and incremental interpretation (Gorz et al. 1996): distribution to contribute to processing speed in what is a very compute intensive application area; incremental interpretation both for speed reasons and to facilitate feedback of results from downstream modules to upstream ones (e.g. to inform the selection of word interpretations from phone ....

Gorz, G., M. Kessler, J. Spilker, and H. Weber. 1996. Research on Architectures for Integrated Speech/Language Systems in Verbmobil. In Proceedings of COLING-96, Copenhagen.

....knowledge by allowing any kind of feature structure description while simultaneously retaining the common ground of results, namely time. Two successful existing incremental systems are TDMT [22] which takes an example based approach to translation on large scale parallel machines and INTARC [16], which at least partly uses a chart based method for the analysis phase. The transfer and generation modules, however, deviate from this schema by being oriented at dialogue act transfer and schema based generation [19] One recent attempt to design an architectural framework for NLP systems, ....

Gunther Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on Architectures for Integrated Speech/Language Systems in Verbmobil. In Proc. of the 16 th COLING, pages 484--489, Copenhagen, Denmark, August 1996.

....Environment (Amtrup, 1995) is an environment for the development of distributed AI systems . As part of the Verbmobil real time speech to speech translation project ICE has addressed two key problems for this type of system, viz. distributed processing and incremental interpretation (Gorz et al. 1996): distribution to contribute to processing speed in what is a very compute intensive application area; incremental interpretation both for speed reasons and to facilitate feedback of results from downstream modules to upstream ones (e.g. to inform the selection of word interpretations from phone ....

Gorz, G., M. Kessler, J. Spilker, and H. Weber. 1996. Research on Architectures for Integrated Speech/Language Systems in Verbmobil. In Proceedings of COLING-96, Copenhagen.

....mass for them. Procedures like the one above have efficient formulations for context free grammars, which are polynomial ( Bak79] The procedure has been successfully used within the Verbmobil project to supply a given linguistic grammar with extra rules for spontaneous speech phenomena ( GKJW96] 5.4 Learning Grammars from Plain Text Learning grammars from plain text is a harder variant of adapting grammars to new data as described above. The difference is that the initial grammar G is empty. It is possible to construct a grammar without using linguistic constraints based on ....

G. Gorz, M. Kesseler, Spilker J., and H. Weber. Research on Architectures for Integrated Speech/Language Systems in Verbmobil. In Proceedings of COLING, 1996.

....depending on the dialogue. After optimization of the algorithm even higher rates are expected. 4. SYNTAX PARSER One of the main benefits of prosody in the INTARC system is the use of prosodic phrase boundaries inside the word lattice search. The incremental probabilistic search engine based on [6] receives word hypotheses and phrase boundary hypotheses as an input. The input is represented as a chart (i.e. a well formed substring table) where frames correspond to chart vertices and word hypotheses are edges which map to pairs of vertices. Word boundary hypotheses (WBHs) are mapped to ....

G. Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on architectures for integrated speech/language systems in verbmobil. In Proc. Int. Conf. on Computational Linguistics, Kopenhagen, 1996.

....of this study lies with the authors. 2 Tomita s GLRP is exponential in some rare cases. Schmid 1994] or [Hauenstein and Weber 1994] The latter two approaches use unification grammars in conjunction with N Gram models to search and parse the lattice simultaneously. In [Weber 1994b] and [Gorz et al. 1996] a stochastic unification grammar is used for the same task. The lattice is being parsed to prduce deep linguistic descriptions, using a model of the unification grammar s rule and type operations together with a word based N Gram model, while pruning is performed by a beam search strategy. The ....

....on an agenda in every chart vertex, which results in a beam search. Agendas in each vertex keep pairs of chart edges to be processed next. Those pairs are being scored using a linear combination of the acoustic, N Gram, and grammar model scores. It is described in more detail in [Weber 1994a] and [Gorz et al. 1996]. The grammar trees which are found in that parsing step are passed to a second processing step using the original grammar G. In that step the tree found is re parsed deterministically using unification. If the approximation of G by the conjunction of G 0 and the grammar model would be very ....

Gorz, G., M. Kesseler, J. Spilker, and H. Weber. 1996. Research on Architectures for Integrated Speech/Language Systems in VERBMOBIL. In Proceedings of COLING-96, Kopenhagen, August.

....plays a decisive role. Integration of prosody into a speech to speech translator as an additional speech language interface is a current topic of research. Within the Verbmobil project, the experimental system INTARC was designed which performs simultaneous speech to speech translation (cf. [6, 4]) In INTARC, particular emphasis has been put on the issues of incrementality and (top down) component interaction in order to take into account expectations and predictions from higher level linguistic components for lower level components. For this purpose time synchronous versions of ....

Gorz, G., Kesseler, M., Spilker, J. and Weber, H.: Research on Architectures for Integrated Speech/Language Systems in Verbmobil. Proc. of COLING-96, Kopenhagen, August 1996

No context found.

Gunther Gorz, Marcus Kesseler, Jorg Spilker, and Hans Weber. Research on architectures for integrated speech/language systems in Verbmobil. In The 16th International Conf. on Computational Linguistics, volume 1, pp. 484--489, Copenhagen, DK, 1996.

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