On the Learnability of Programming Language Semantics

TL;DR

This paper explores the learnability of programming language semantics using game semantics and neural networks, demonstrating how LSTM models can approximate complex semantic interactions in Idealised Algol.

Contribution

It introduces a novel approach of applying LSTM neural networks to learn and analyze game-semantic models of programming languages, bridging formal semantics and machine learning.

Findings

LSTM models can effectively learn game-semantic representations.

Model accuracy depends on term complexity and number of free variables.

Learned models enable semantic analysis between different language variants.

Abstract

Game semantics is a powerful method of semantic analysis for programming languages. It gives mathematically accurate models ("fully abstract") for a wide variety of programming languages. Game semantic models are combinatorial characterisations of all possible interactions between a term and its syntactic context. Because such interactions can be concretely represented as sets of sequences, it is possible to ask whether they can be learned from examples. Concretely, we are using long short-term memory neural nets (LSTM), a technique which proved effective in learning natural languages for automatic translation and text synthesis, to learn game-semantic models of sequential and concurrent versions of Idealised Algol (IA), which are algorithmically complex yet can be concisely described. We will measure how accurate the learned models are as a function of the degree of the term and the number of free variables involved. Finally, we will show how to use the learned model to perform latent semantic analysis between concurrent and sequential Idealised Algol.