Invariants and Variability of Synonymy Networks: Self Mediated Agreement by Confluence

Edges of graphs that model real data can be seen as judgements whether pairs of objects are in relation with each other or not. So, one can evaluate the similarity of two graphs with a measure of agreement between judges classifying pairs of vertices into two categories (connected or not connected). When applied to synonymy networks, such measures demonstrate a surprisingly low agreement between various resources of the same language. This seems to suggest that the judgements on synonymy of lexemes of the same lexicon radically differ from one dictionary editor to another. In fact, even a strong disagreement between edges does not necessarily mean that graphs model a completely different reality: although their edges seem to disagree, synonymy resources may, at a coarser grain level, outline similar semantics. To investigate this hypothesis, we relied on shared common properties of real world data networks to look at the graphs at a more global level by using random walks. They enabled us to reveal a much better agreement between dense zones than between edges of synonymy graphs. These results suggest that although synonymy resources may disagree at the level of judgements on single pairs of words, they may nevertheless convey an essentially similar semantic information.