Section: New Results

GIVE

For the Generation of Instruction in Virtual Environment challenge edition 2.5 (GIVE), we developed two systems. The first system is the successor of the system that we presented to the GIVE 2 challenge (2010). We solved two known problems of this system, namely the indefinite presupposition problem and the ambiguity arising from underspecified referring expressions [26] . The GIVE 2.5 challenge proved that these improvements were efficient, and showed an increase of 21% in terms of task success (47% in GIVE 2, 68% in GIVE 2.5). The second system, developed in collaboration with the University of Cordoba is the first to our knowledge that uses a human-human corpus to provide whole utterances thanks to plan matching techniques [22] , [21] . The system ranked fifth in terms of task success (58%), but second in terms of referring targets identification. The naturalness of instructions and the simplified system development makes it an interesting research track to follow.