Interactive Spoken-Language Processing in a Hybrid Connectionist System SCREEN

There has been a renewed interest in the interactive processing of spoken natural language, which has already proven useful for improving interactions between humans and computers and for translating dia1ogues.l To process spoken language from imeractions between humans or between humans and computers, we must take into account the “interactive noise” involved in spontaneous spoken language. Interactive noise refers to the interjections, pauses, repetitions, corrections of mistakes, false starts, uncommon syntactic or semantic constructions, and so on that occur in spontaneous spoken language but not in written language. Speech recognizers, which we use to identify spoken words for spokenlanguage processing, cannot perform optimally and often create an additional type of noise. They thus may make incorrect word hypotheses and produce ungrammatical sentences. To continue processing spoken language despite these problems, spoken-language analysis must be fault-tolerant and robust. Traditional and newer symbolic methods have been used for the analysis ofwritten language. However, more robust methods must be used for spoken-language analysis. Previous approaches have combined statistical and symbolic method^.^^^ We examined methods that are hybrids of connectionist and symbolic methods because connectionist networks are robust for the type of unpredictable input found in spoken language. -Earlier work* demonstrated how language processing can be achieved in connectionist networks. However, these early models could be used only for relatively small tasks, such as prepo,sitional phrase attachment, and only in restricted domains, sometimes only with artificially generated sentences. Our SCREEN (Symbolic Connectionist Robust Enterprise for Natural Language) system analyzes real-world utterances and can be used for relatively large and complex tasks. SCREEN learns a robust flat syntax, semantics, and pragmatics representation. (Pragmatics relates to an utterance’s intention.) The system also deals with uncommon syntactic and semantic language irregularities. The system is able to produce many utterance hypotheses based on spoken input and can determine which hypotheses are most likely. SCREEN’S ability to analyze spoken language, despite encountering mistakes and uncertainties, demonstrates the system’s robustness and potential. Because SCREEN is a German spokerclanguage system, the example of spoken language that we will analyze will be in German. We will provide a literal and, where appropriate, a more easily understood English translation of any German words and sentences that we use.

@Article{WW96, 
 	 author =  {Wermter, Stefan and Weber, Volker},  
 	 title = {Interactive Spoken-Language Processing in a Hybrid Connectionist System SCREEN}, 
 	 journal = {Computer},
 	 number = {7},
 	 volume = {29},
 	 pages = {65--74},
 	 year = {1996},
 	 month = {Jul},
 	 publisher = {IEEE},
 	 doi = {}, 
 }