... connectionist1
Sometimes connectionist networks are also called artificial neural networks. From now on we will use only the term ``connectionist networks'', and the term ``hybrid connectionist architecture'' to refer to an architecture which emphasizes the use of connectionist networks but does not rule out the use of symbolic representations on higher levels where they might be needed.
... SCREEN2
Symbolic Connectionist Robust EnterprisE for Natural language
... word graph3
The speech input in the form of test word graphs was taken from the so-called Blaubeuren Meeting Corpus. The particular word graphs we used here were provided by project partners for general test purposes in the Verbmobil project. They were particularly generated for testing parsing strategies. Therefore the speech recognizer was fine-tuned to produce relatively small word graphs with a relatively high word accuracy of 93%. The vocabulary size for the HMM recognizer is 628. The average number of hypotheses per word was 6.3 over 10 dialogs.
... eliminate4
This means that repaired utterance parts are actually only marked as deleted.
... certain signal5
The HMM-speech recognizer used for generating word hypotheses in our domain has a word accuracy of about 93% for the best match between the word graph and the desired transcript utterance. This recognizer was particularly optimized for this task and domain in order to be able to examine the robustness at the language level. An unoptimized version for this task and domain currently has 72% word accuracy.
... weighted equally6
This integration of speech, syntax, and semantics confidence values provided better results than just using one or two of these three knowledge sources.
... phrase7
In Figure 7 we show the influence of the phrase start delimiter on the abstract syntactic and semantic categorization with dotted lines.
... input stream8
Pauses and interjections can sometimes provide clues for repairs (Nakatani & Hirschberg, 1993) although currently we do not use these clues for repair detection. Compared to the lexical, syntactic, and semantic equality of constituents, interjections and pauses provide relatively weak indicators for repairs since they also occur relatively often at other places in a sentence. However, since we just mark interjections and pauses as deleted we could make use of this knowledge in the future if necessary.
... (upper right square)9
The dialog acts we use are: accept (ACC), query (QUERY), reject (REJ), request-suggest (RE-S), request-state (RE-S), state (STATE), suggest (SUG), and miscellaneous (MISC). Since this paper focuses on the syntactic and semantic aspects of SCREEN we do not further elaborate on the implemented dialog part here. Further details on dialog act processing have been described previously (Wermter & Löchel, 1996).
... request10
In the snapshots in Figure 12 the abstract syntactic and semantic categories have not yet been computed and therefore are represented as NIL. In the next processing step this computation will be performed which can be seen in next Figure 13.
... kept11
In our experiments low values (n=10) provided the best overall performance.
... (BAS-SYN-DIS and BAS-SEM-DIS)12
This was explained in more detail in Section 4.3.2

SCREEN (screen@nats5.informatik.uni-hamburg.de)
Mon Dec 16 15:33:13 MET 1996