Verb sense induction with tensor decomposition
SZTAKI, Lágymányosi u. 11.
I submitted a paper to ACL main conf with the following abstract, and plan to submit it to workshops as well. The main question of the work in progress talk is now how to evaluate the decomposition quantitatively.
We present verb representations based on dependency relations and tensor-decomposition applying the rather recent Orth-ALS decomposition algorithm to a subject × preverb × verb × object tensor in Hungarian, where preverbs (verb-modifying particles) interfere with argument structure and polisemy. The work-in-progress paper evaluates the obtained representations qualitatively, by clustering the verb embedding vectors, and showing some which intuitively fit to some semantic verb class.