Syntax and word meaning contribute tothe compositional construction of the meaning of sentences. In studyingcompositional distributional semantics, we discovered that generally modelsbased on vectors tend to forget the syntactic structure of the sentences. Then,we started to investigate how we can partially preserve syntactic structures invectors. We came across an interesting idea. We discovered a way to representtree structured data in small vectors.

In this talk, I report on this idea thatwe called distributed tree kernels (DTK). DTKs are a novel method to reducetime and space complexity of tree kernels. Using a linear complexity algorithmto compute vectors for trees, we embed feature spaces of tree fragments inlow-dimensional spaces where the kernel computation is directly done with dotproduct. We show that DTKs are faster, correlate with tree kernels, and obtaina statistically similar performance in two natural language processing tasks. Ifinally speculate on the relation between DTKs, the models on the compositionaldistributional semantics, and the use of distributional semantics in treekernels.

Fabio Massimo Zanzotto is an AssociateProfessor at Department of Enterprise Engineering of the University of Rome “TorVergata”. Since 1998, he has interests in the research endeavour of ArtificialIntelligence. He is the author of more than 100 publications in the area ofMachine Learning for Natural Language Processing, mainly focusing on:recognizing textual entailment, syntactic parsing for Italian, and, recently,distributed/distributional models for NLP where he investigates the relationbetween classical machine learning techniques and advanced graph representationof meaning.