17 Feb 2017 - Mehrnoosh Sadrzadeh: Seminar

Distributional models of meaning represent words by vectors, obtained from the frequency of their occurrences  in corpora of text. Similarity is the main relation applied to word vectors, but one can also do entailment using inclusion relations over  features of these vectors.  This has indeed been studied before, e.g. see Kotlerman et al, Herbelot and Ganesalingam, and Rimell.  In recent joint work with D. Kartsaklis, presented at COLING and LACL 2016, we showed how inclusion relations can be lifted compositionally from word features to phrase and sentence features. We considered and compared  performances of different compositional models: from simple addition and multiplication, to min and max, to matrix multiplication and cube contraction. In an earlier version (joint with E. Balkir) we also worked with density matrices. We experimented on datasets containing verb  and  adjective noun phrases, and SVO sentences. I will present the setting and our results. I will then discuss how one can develop  logic for distributional semantics based on this and a few other possible entailment relations. The logic part is on going work with J. Hedges, M. Dostal and R. Muskens.