https://www.youtube.com/watch?v=7v1gfOuYbpc

We generalize normalized PMI(Bouman 2009) for the three-variable case (van de Cruys 2011) and populate a subject-verb-object tensor with this association measure. The concatenation of subject, verb, and object embedding vectors was tested on SVO triple similarity data (Kartsaklis and Sadrzadeh, 2014). The results are better than with other association measures. For an earlier version of the work, see [my poster at this year's MSZNY (publ / makrai20-verbtensor-mszny). That paper provides only motivation and literature, see my own calculations in the last section of the poster, which is unfortunately in Hungarian.

Fun fact: the original motivation was that the cohesion of my dissertation, this project should have linked my two previous topics: verbs and vectors. If you like, this has come true, but in the meantime I have realized that there is another way to ideologize what my dissertation is about: Lexical relations.

github.com/makrai/verb-tensor