LLM Agents Implement an NLG System from Scratch: Building Interpretable Rule-Based RDF-to-Text Generators

2025-12-20

AI Summary

To address the reliance on supervised data, severe hallucination, and inefficient inference in RDF-to-text generation, this paper proposes a neuro-symbolic approach based on collaborative multi-role LLM agents performing automated programming. Without human annotation or model fine-tuning, LLM agents autonomously parse RDF semantics, induce logical rules, and generate an interpretable Python-based rule engine—enabling zero-shot, purely rule-driven text generation. Our key contribution is replacing backpropagation with LLM-agent collaboration that “simulates training” through iterative reasoning and code synthesis, yielding the first fully interpretable, rule-based generator with millisecond-scale CPU-only inference (100× speedup) and near-zero resource overhead. Evaluated on WebNLG and OpenDialKG, our method significantly reduces hallucination; while fluency slightly lags behind fine-tuned models, it exhibits superior generalization across unseen RDF structures and domains.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Neuro-Symbolic Learning
• Knowledge Representation and Reasoning: Automated Reasoning and Theorem Proving

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

We present a novel neurosymbolic framework for RDF-to-text generation, in which the model is "trained" through collaborative interactions among multiple LLM agents rather than traditional backpropagation. The LLM agents produce rule-based Python code for a generator for the given domain, based on RDF triples only, with no in-domain human reference texts. The resulting system is fully interpretable, requires no supervised training data, and generates text nearly instantaneously using only a single CPU. Our experiments on the WebNLG and OpenDialKG data show that outputs produced by our approach reduce hallucination, with only slight fluency penalties compared to finetuned or prompted language models

Problem

• Develops a neurosymbolic framework for RDF-to-text generation
• Uses LLM agents to create rule-based generators without human texts
• Produces interpretable, fast systems that reduce hallucination in outputs

Innovation

• LLM agents collaboratively generate rule-based Python code
• No supervised training data required for domain adaptation
• Neurosymbolic framework reduces hallucination in text generation