[Paper Review] Causal Graph Discovery with Retrieval-Augmented Generation based Large Language Models

Yuzhe Zhang, Yipeng Zhang+3 more|arXiv (Cornell University)|Feb 23, 2024

Data Quality and Management4 citations

TL;DR

This paper proposes LACR, a retrieval-augmented generation (RAG)-based method that leverages scientific literature to discover causal graphs without relying on statistical data. By retrieving relevant research snippets and using large language models to extract and aggregate associational relationships, LACR constructs high-quality causal graphs that outperform existing statistical and LLM-based methods on benchmark datasets like SACHS and BIOLOGIST.

ABSTRACT

Causal graph recovery is traditionally done using statistical estimation-based methods or based on individual's knowledge about variables of interests. They often suffer from data collection biases and limitations of individuals' knowledge. The advance of large language models (LLMs) provides opportunities to address these problems. We propose a novel method that leverages LLMs to deduce causal relationships in general causal graph recovery tasks. This method leverages knowledge compressed in LLMs and knowledge LLMs extracted from scientific publication database as well as experiment data about factors of interest to achieve this goal. Our method gives a prompting strategy to extract associational relationships among those factors and a mechanism to perform causality verification for these associations. Comparing to other LLM-based methods that directly instruct LLMs to do the highly complex causal reasoning, our method shows clear advantage on causal graph quality on benchmark datasets. More importantly, as causality among some factors may change as new research results emerge, our method show sensitivity to new evidence in the literature and can provide useful information for updating causal graphs accordingly.

Motivation & Objective

• To address limitations in traditional causal discovery methods that depend on biased or low-quality observational data.
• To overcome the reliance of existing LLM-based causal inference on the model's inherent reasoning ability by grounding inference in external literature.
• To develop a generic, data-driven method for causal graph recovery that does not require task-specific prompts or knowledge bases.
• To improve the reliability and accuracy of causal graph construction by systematically aggregating evidence from a large corpus of scientific literature.
• To demonstrate that LLMs enhanced with retrieval from scientific literature can achieve expert-level causal reasoning in complex biological systems.

Proposed method

• Retrieves relevant text chunks from a large, curated corpus of scientific literature using dense vector retrieval.
• Employs a large language model (LLM) to identify and label associational relationships between factors in the retrieved text.
• Aggregates extracted associational relationships across multiple documents to infer potential causal connections.
• Constructs a causal graph where nodes represent factors and directed edges represent inferred causal relationships.
• Uses a RAG framework to enhance LLM reasoning by grounding it in factual scientific literature, reducing hallucination and bias.
• Applies a multi-step prompting strategy to guide the LLM in extracting and validating relationships without requiring fine-tuning.

Experimental results

Research questions

• RQ1 Can LLMs effectively extract causal relationships from scientific literature without relying on statistical data or domain-specific prompts?
• RQ2 How does the performance of LACR compare to statistical estimation-based causal discovery methods on real-world biological datasets?
• RQ3 Does using a broader literature pool improve or degrade the quality of causal graph recovery, and why?
• RQ4 To what extent does the LLM’s reasoning capability limit the accuracy of causal orientation compared to association detection?
• RQ5 Can retrieval-augmented generation enhance LLMs’ causal reasoning to match expert-level performance in causal graph construction?

Key findings

• LACR achieved an F1 score of 75.00% on the SACHS causal graph, outperforming statistical baseline methods such as FASK (F1: 51.61%) and Sachs’s Model (F1: 75.00%).
• On the BIOLOGIST consensus graph, LACR achieved an F1 score of 70.97% when using the P+S+F pool, significantly outperforming FASK (F1: 64.86%) and other baselines.
• The method demonstrated higher precision and recall in detecting associational relationships (e.g., 50.00% precision, 60.00% recall for P+S+F(SQ)) compared to causal orientation tasks.
• Performance declined when using larger literature pools, likely due to distributional shift and lack of model exposure to newer or less common literature sources.
• Causation existence verification (identifying associations) outperformed causation orientation (inferring directionality), indicating that LLMs still struggle with complex causal reasoning.
• LACR’s performance was robust across different literature pools, with the best results achieved using a balanced combination of PubMed, SACHS, and BIOLOGIST literature.