Because they represent schematic situations, semantic frames (Fillmore, 82), as intantiated into FrameNet (Baker, Fillmore and Petruck, 83) are an appealing level of generalization over the eventualities described in texts. Although advances in machine learning sometimes seem to diminish the need to use as input sophisticated structured representations of sentences, the enthusiasm for interpreting trained neural networks somewhat seems to reaffirm that need.
In this talk, I will present some feedback from the development of a French FrameNet, including analysis of the main difficulties we faced during annotation. I will describe how linking generalizations can be extracted from the frame-annotated data, using deep syntactic annotations. I will then investigate what kind of input is most effective for FrameNet parsing, from no syntax at all to deep syntactic representations.
The work I’ll present is a joint work with Marianne Djemaa, Philippe Muller, Laure Vieu, G. de Chalendar, B. Sagot, P. Amsili (for the French FrameNet), C. Ribeyre, D. Seddah, G. Perrier, B. Guillaume (deep syntax), and Olivier Michalon, Alexis Nasr (semantic parsing).