"Natural language as an Action Language"

Samuel Corey, Yuliya Lierler

Keywords: Natural language understanding, Action languages, ALM, Answer set programming, Neuro-symbolic reasoning, Knowledge representation, Dynamic domains, Executable models, Large language models

TL;DR: We transform natural language narratives into executable ALM transition systems using LLM-based canonicalization, enabling verifiable symbolic reasoning over dynamic domains.

Abstract:

Large Language Models (LLMs) can interpret procedural text and generate plausible reasoning chains, but they lack explicit representations of state transitions and causal structure (Valmeekam et al., 2022). This limits their ability to provide verifiable, model-based reasoning in dynamic domains (Gelfond & Kahl, 2014, Ch. 8). We investigate a neuro-symbolic architecture that treats natural language narratives as executable specifications of dynamic domains.

Our approach builds on TEXT2ALM (Olson & Lierler, 2019), a system that translates natural language action narratives into ALM (Action Language Modular) system descriptions (Inclezan & Gelfond, 2016), enabling formal reasoning via the CALM compiler (Wertz, Chandrasekan, & Zhang, 2018). However, TEXT2ALM requires highly constrained input and is brittle when applied directly to unrestricted text. Inspired by recent work coupling large language models with answer set programming for robust reasoning from text (Yang, Ishay, & Lee, 2023), we introduce an LLM-based canonicalization layer that preprocesses natural language into a schema-aligned intermediate representation compatible with TEXT2ALM.

The LLM decomposes complex sentences, normalizes sentence structure, resolves implicit state changes, and translates verbs in the text into verbs belonging to action classes accepted by TEXT2ALM. The resulting pipeline (Natural Language → LLM Canonicalization → ALM → Executable Transition System) is designed to compile narratives and procedural descriptions into formally grounded dynamic domain models. This architecture supports executable reasoning over state transitions extracted from natural language.

We argue that this framework provides a pathway for integrating statistical language understanding with verifiable symbolic reasoning. In the context of agentic AI, our approach explores how natural language can be transformed into executable models of action and state change.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 37