Separating Disambiguation from Composition in Distributional Semantics
Dimitri Kartsaklis, Mehrnoosh Sadrzadeh, Stephen Pulman · 2013
Most compositional-distributional models of meaning are based on ambiguous vector representations, where all the senses of a word are fused into the same vector. This paper provides evidence that the addition of a vector disambiguation step prior to the actual composition would be beneficial to the whole process, producing better composite representations. Furthermore, we relate this issue with the current evaluation practice, showing that disambiguation-based tasks cannot reliably assess the quality of composition. Using a word sense disambiguation scheme based on the generic procedure of Schütze (1998), we first provide a proof of concept for the necessity of separating disambiguation from composition. Then we demonstrate the benefits of an “unambiguous” system on a composition-only task. 1