Probabilistic dynamic semantics

Develop probabilistic semantic grammar fragments and convert them to Stan code.

Overview

Semantic research in the Montagovian tradition has taken an experimental turn in the last decade, as the empirical goals of semanticists have gotten more ambitious. Large-scale inference datasets—and the statistical models they require—have become a more standard part of the semanticist's toolkit. pds is an approach to developing models of inference datasets that involves seamlessly deriving them from semantic grammar fragments. As a library, it encodes the ideas described in Probabilistic Dynamic Semantics by Julian Grove and Aaron Steven White. Applications of the framework can also be found in Grove and White (2025a) and Grove and White (2025b).

Components of the library

• Framework:
  • An encoding of Combinatory Categorial Grammar (CCG)-based grammar fragments, as well as a CKY-style parser with string-based (instead of span-based) memoization.
  • A framework for typed λ-calculus based on δ-rules, which transform complex semantic representations into simpler (but equivalent) typed probabilistic models.
  • Code for transforming typed λ-terms into models in the Stan programming language.
• Theory:
  • An encoding of theories of formal pragmatics, including common grounds and Questions Under Discussion (QUDs).
• Analysis:
  • Analyses of particular natural language phenomena, including factivity and vagueness.

Documentation

See both the text Probabilistic Dynamic Semantics and the ESSLLI 2025 Summer School Course Notes.

Installation

You can install the library and open a repl by cloning this repository and doing:

cd pds
cabal v2-repl --repl-options="-ghci-script .ghci"

A Nix-based installation is also possible by doing:

cd pds
nix-shell --run "cabal v2-repl --repl-options=\"-ghci-script .ghci\""

Quick start

There is a test-suite under active development here. To load the examples which are currently implemented in Analysis, install the library and load the test-suite:

ghci> :l test/Spec.hs

There are three models in the test-suite which can be printed out. The model on factivity can be gotten by doing:

stanOutput factivityExample

It yields the following result:

parameters {
  real<lower=0.0, upper=1.0> v;
  real<lower=0.0, upper=1.0> w;
  real<lower=0.0, upper=1.0> u;
}

model {
  v ~ logit_normal(0.0, 1.0);
  w ~ logit_normal(0.0, 1.0);

  target += log_mix(v, truncated_normal_lpdf(y | 1.0, u, 0.0, 1.0), truncated_normal_lpdf(y | w, u, 0.0, 1.0));
}

The other two models are scaleNormingExample and likelihoodExample.