Abstract Egges
CarSim is a prototype system to visualize and animate a 3D scene from a written description. The text corpus consists of 87 car accident reports written in French and provided by the MAIF insurance company. The CarSim's objective is to replay these accidents symbolically in a virtual world. The CarSim architecture is divided into two parts that communicate using a formal representation of the accident.The representation structure is a template similar to that of information extraction systems (MUC-3 1991). It has been designed so that it contains the information necessary to reproduce and animate the accident entities. CarSim's first part is a linguistic module that extracts information from the report and fills the template slots. It uses a partial parser, a co-reference resolver, a small function detector, and semantic information. The second part is a visualizer that takes the filled template as an input, creates the visual entities, and animates them.
A first baseline version of CarSim (Dupuy et al., 2001) could convert successfully 17% of the texts into a plausible symbolic animation. We modified the prototype to improve the detections of the actors and the determination of the initial directions of the cars. Using these new algorithms, CarSim could synthesize a credible scene and animation for 34% of the texts.
References:
Dupuy S, A. Egges, V. Legendre, P. Nugues, Generating a 3D Simulation of
a Car Accident From a Written Description
in Natural Language: The CarSim System, Proc. of The Workshop on
Temporal and Spatial Information Processing, ACL
2001 Conference, Toulouse, July 2001.
MUC-3, Proceeding of the Third Message Understanding Conference, Morgan Kaufmann, 1991.
Last modified $Date: 2001/10/04 13:39:44 $ by Parlevink Webmaster