Invariants and variability of synonymy networks: Self mediated agreement by confluence

Résumé En

Edges of graphs that model real data can beseen as judgements whether pairs of objectsare in relation with each other or not. So,one can evaluate the similarity of two graphswith a measure of agreement between judgesclassifying pairs of vertices into two cate-gories (connected or not connected). Whenapplied to synonymy networks, such measuresdemonstrate a surprisingly low agreement be-tween various resources of the same language.This seems to suggest that the judgementson synonymy of lexemes of the same lexi-con radically differ from one dictionary ed-itor to another. In fact, even a strong dis-agreement between edges does not necessarilymean that graphs model a completely differ-ent reality: although their edges seem to dis-agree, synonymy resources may, at a coarsergrain level, outline similar semantics. To in-vestigate this hypothesis, we relied on sharedcommon properties of real world data net-works to look at the graphs at a more globallevel by using random walks. They enabledus to reveal a much better agreement betweendense zones than between edges of synonymygraphs. These results suggest that althoughsynonymy resources may disagree at the levelof judgements on single pairs of words, theymay nevertheless convey an essentially simi-lar semantic information.