[Paper Review] Planning with Logical Graph-based Language Model for Instruction Generation

Fan Zhang, Kebing Jin+1 more|arXiv (Cornell University)|Aug 26, 2023

Topic ModelingComputer Science3 citations

TL;DR

This paper proposes Logical-GLM, a graph-based language model that integrates AI planning with large language models to generate more logically consistent and executable instructional texts. By constructing logical Bayesian graphs from PDDL-converted instructions and alternately optimizing the graph and language model via an EM-style framework, the method achieves higher reasoning accuracy and interpretability with fewer parameters and less training data compared to standard language models.

ABSTRACT

Despite the superior performance of large language models to generate natural language texts, it is hard to generate texts with correct logic according to a given task, due to the difficulties for neural models to capture implied rules from free-form texts. In this paper, we propose a novel graph-based language model, Logical-GLM, to infuse logic into language models for more valid text generation and interpretability. Specifically, we first capture information from natural language instructions and construct logical bayes graphs that generally describe domains. Next, we generate logical skeletons to guide language model training, infusing domain knowledge into language models. Finally, we alternately optimize the searching policy of graphs and language models until convergence. The experimental results show that Logical-GLM is both effective and efficient compared with traditional language models, despite using smaller-scale training data and fewer parameters. Our approach can generate instructional texts with more correct logic owing to the internalized domain knowledge. Moreover, the usage of logical graphs reflects the inner mechanism of the language models, which improves the interpretability of black-box models.

Motivation & Objective

• To address the lack of logical consistency in instructional text generation by large language models.
• To improve reasoning and executability of generated instructions in complex, knowledge-intensive tasks.
• To enhance model interpretability by internalizing domain logic through structured representations.
• To reduce reliance on large-scale data and high-parameter models while maintaining performance.
• To enable mutual guidance between language models and logical graphs for better policy learning.

Proposed method

• Convert natural language instructions into PDDL-formatted action sequences to extract structured, rule-based domain knowledge.
• Construct logical Bayesian graphs from filtered PDDL sequences to represent domain logic and constraints.
• Train a language model on the cleaned, structured instructional data to improve factual and logical consistency.
• Implement an EM-style iterative optimization framework where the language model and logical graph guide each other.
• Use heuristic search over the logical graph to generate logical skeletons that inform and refine the language model’s output.
• Alternate between refining the graph structure and fine-tuning the language model until convergence.

Experimental results

Research questions

• RQ1 Can a graph-based language model improve the logical consistency and executability of generated instructional texts?
• RQ2 How does integrating AI planning with language models affect performance with reduced training data and model size?
• RQ3 To what extent does the mutual optimization of language models and logical graphs enhance interpretability of model decisions?
• RQ4 How robust is the model to variations in expected program length or task complexity?
• RQ5 Can the model generate high-quality, task-specific instructions without relying on large-scale pretraining?

Key findings

• Logical-GLM achieves higher executability and correctness in generated instructions compared to standard language models, even with smaller model size and less training data.
• The model demonstrates improved consistency between sentences and reduced generation of irrelevant or incoherent actions, as shown in the case study on 'Watch movie'.
• Executability increases significantly with longer expected program lengths, indicating robustness to task complexity.
• The model outperforms both vanilla and fine-tuned GPT-2-xl in generating coherent, task-relevant instruction sequences.
• The use of logical Bayesian graphs enhances interpretability by exposing the internal reasoning mechanism of the language model.
• The EM-style iterative optimization framework successfully aligns language generation with domain logic, leading to more reliable and structured outputs.