Sonderforschungsbereich 732:

Project D10-N (2014-2018)

Incrementality in Compositional Distributional Semantics

Principal Investigator: Sebastian Padó
Researchers: Jason Utt, Abhijeet Gupta, Diego Frassinelli

The principle of compositionality is fundamental to model-theoretic semantics in the tradition of (Montague, 1973). Distributional models of lexical semantics are starting to account for compositionality as well: Novel syntax-semantics interfaces map syntactic categories onto algebraic types -- e.g. nouns to vectors, adjectives to matrices, etc. -- and define general operations to combine objects of these types (Baroni & Zamparelli, 2010; Grefenstette et al., 2013). This approach offers an attractive framework for distributional semantics construction, but research has so far been limited.

The goal of Project D10 is to contribute to research on tensor-based compositional distributional semantic models by developing a syntax-semantics interface with three properties:

(a) it will be dependency-based rather than based on constituents;

(b) it will be incremental, that is, construct semantics in a left-to-right manner;

(c) it will incorporate a notion of plausibility for (partial) analyses based on expectations at the level of individual composition operations.

The first property is important to develop syntax-semantics interfaces for languages with a more free word order. The second and third are well-known properties of human sentence processing