HIGGINS Error handling strategies in a spoken dialogue system

1. HIGGINS Error handling strategies in a spoken dialogue system Rolf Carlson, Jens Edlund and Gabriel Skantze • Error handling research issues • The long term research goals of the project include these items, which permeate shorter term research issues: • How do different levels of feedback and detail in system responses affect robustness and efficiency of the dialogue? • Which factors should govern the dialogue strategy regarding feedback and information detail? • How can the different system modules calculate confidence measures? • How can we categorise errors regarding the types of error handling they demand? • How should the system act when it doesn't understand, or only partially understands, what the user says? • How does the user's beliefs about the system’s understanding correlate to its actual understanding, depending on error handling strategy? • The HIGGINS domain • The primary domain of HIGGINS is city navigation for pedestrians. • Secondarily, HIGGINS is intended to provide simple information about the immediate surroundings. • Methods • We endeavour to follow these design principles: • Versatility • The dialogue system is distributed and module based, which makes switches between different system configurations easier. • The modules are built to be versatile and it is simple to test different module internal configurations. • Empirical iterations • Modules are based on experiences drawn from empirical data, when possible. • Modules are built to be testable against empirical data. Scenario User: I want to go to the closest subway station. System: Ok, to the closest subway station. Can you describe where you are now? User: I have an ATM to my left and a pedestrian crossing in front of me. System: Can you see some trees to your right and a white building in front of you? User: Yes System: Ok, take left after the large building which you have on your left and follow the street until you reach a crossing. User: Ok, there is a bus station here. System: That’s right. Take left again after the bus station. • Current research issues • A sample of the issues we are currently working on: • Which robust parsing techniques are suitable for more complex utterances? • The HIGGINS parser presently achieves robust parsing by using grammars allowing insertions, fragments and non-congruence. • How does a dialogue manager that is able to vary its feedback level make its choices? • The HIGGINS parser produces confidence measures which may be combined with ASR confidence to provide a basis for selecting feedback level. • How can incremental speech recognition be utilised to improve dialogue? • The HIGGINS parser supports incremental parsing that facilitates fast feedback, which could help the user detect and correct errors. • What demands on interpretation and error handling are specific for the chosen domain? • Pedestrian navigation dialogues and human error handling strategies has been collected in a modified Wizard of Oz setting. • Spatial descriptions in a simulated 3D environment are currently being collected. • Domain challenges • The system should be able to: • Represent complex properties of objects as well as relations between objects. • Make morphological distinctions between singular/plural, definite/indefinite. • Engage in a dialogue to find out the user’s destination and then iteratively update a hypothesis of the user’s position based on the user’s descriptions of the surroundings. • Compute the user’s position using spatial and temporal reasoning. • Engage in an extended dialogue in case the user’s descriptions are insufficient for determining the position. • Generate route directions in appropriately sized chunks, providing the optimal path and using grounded concepts in the directions. http://www.speech.kth.se/higgins/