LANGUAGE PROCESSING FOR SPOKEN DIALOGUE SYSTEMS: IS SHALLOW PARSING ENOUGH? Ian Lewin Ralph Becket Johan Boye David Carter Manny Rayner and Mats Wir´ en SRI International, 23 Millers Yard, Cambridge CB2 1RQ United Kingdom Telia Research AB, Vitsandsgatan 9, S-12386 FARSTA, Sweden ABSTRACT With maturing speech technology, spoken dialogue systems are increasingly moving from research proto- types to fielded systems. The fielded systems how- ever generally employ much simpler linguistic and di- alogue processing strategies than the research proto- types. We describe an implemented spoken-language dialogue system for a travel planning domain which supports a mixed initiative dialogue strategy. The sys- tem accesses a commercially available travel informa- tion web-server. The system architecture combines both shallow and deep linguistic processors, partly so that a robust if shallow analysis is always available to the dialogue manager, and partly so that we can be- gin to examine where significant gains can be made by employing more advanced linguistic processing. We present the results of a preliminary investigation using data from a Wizard of Oz experiment. The results lend limited support to our original hypothesis that deep linguistic processing will prove useful at points where the user takes the initiative in driving the dialogue for- ward. 1. INTRODUCTION With maturing speech technology, spoken dialogue sys- tems are increasingly moving from research prototypes to fielded systems. The fielded systems however gen- erally employ much simpler linguistic and dialogue processing strategies than the research prototypes (for a range of example systems, see, amongst others, [2], [1], [10], [11] and [3]). For example, in the fielded systems, domain-specific keyword/phrase spotting and slot-filling techniques are preferred for utterance inter- pretation. At the dialogue level, these systems tend to keep the dialogue initiative to themselves by treating the user simply as an answer-supplier. Particular sys- tems may also implement particular instances of more sophisticated processing. However, the simple meth- ods do dovetail simply because the more expectations that a system can impose on a dialogue, then the more those expectations can be used to aid interpretation of user utterances. Currently, there is little work which attempts to examine at what points deep linguistic pro- cessing might prove significantly useful in the sorts of spoken language dialogue system that are currently be- ing fielded. SRI International and Telia Research AB are de- veloping a Swedish language spoken dialogue system for accessing a web-based travel database. The sys- tem is being built by adaptation of existing general- purpose speech recognition and language understand- ing components including the Nuance toolkit ([13]) and the Core Language Engine (CLE) with a domain independent Swedish grammar ([4]). The Swedish ver- sion of the CLE was originally built with a machine translation application in mind ([14]). The system also includes a dialogue manager whose role is to progress the dialogue as a whole, deciding on the best interpre- tation of user utterances and deciding what it should do and say next. We have also added a parallel, faster but very simple linguistic processing path. This ensures the existence of a fallback “robust” analysis. It also provides our dialogue manager with interesting strate- gic choices concerning the two input paths. Finally, it enables us to begin evaluating wherein lie the advan- tages in deep linguistic processing and when shallow analysis may be reliably used. We do not here con- sider system development issues. For example, one of the objectives in our current project was to exam- ine how easily our Swedish grammar, designed to be domain independent, could be adapted to the new ap- plication. We present the results of some preliminary investigations into the relative contributions of shallow and deep analysis in our travel scenario. The results lend limited support to our original hypothesis that