Source-linked AI summary
Can large language models build causal graphs?
Stephanie Long, Tibor Schuster, Alexandre Piché
TL;DR
Building medical causal diagrams can require substantial literature review and expert input, and experts may miss confounding pathways. This study tests GPT-3’s edge scoring across prompt and wording choices; it exceeds 50% accuracy on each tested DAG in at least one setting, but the authors recommend expert verification.
Problem
Building comprehensive medical DAGs takes time and resources, and expert diagrams may omit important confounding pathways.
Source: §1 · p. 1Method
The study tests GPT-3 by scoring statements for edge presence or absence across ordered variable pairs and varying prompt language and linking verbs.
Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 5Results
GPT-3 reached much higher than 50% accuracy on all tested DAGs in at least one setting, while accuracy varied with prompts and linking verbs.
Source: §4 · p. 6Takeaways & Limitations
The authors propose using LLMs to extract common knowledge from medical text alongside expert knowledge, with expert verification of DAG construction.
Source: §5 · p. 6The edge-probing approach assumes causal connections are well established in the text corpus, and lagging updates may limit usefulness for novel diseases.
Source: §4 · p. 6
Abstract
from arXiv · showhide
Building causal graphs can be a laborious process. To ensure all relevant causal pathways have been captured, researchers often have to discuss with clinicians and experts while also reviewing extensive relevant medical literature. By encoding common and medical knowledge, large language models (LLMs) represent an opportunity to ease this process by automatically scoring edges (i.e., connections between two variables) in potential graphs. LLMs however have been shown to be brittle to the choice of probing words, context, and prompts that the user employs. In this work, we evaluate if LLMs can be a useful tool in complementing causal graph development.
1 Introduction
Causal diagrams are needed to address causal questions, but building comprehensive medical DAGs can be laborious and costly. This paper evaluates whether GPT-3 can help identify edges and whether prompt wording affects its accuracy.
1 Introduction
- Causal questions about treatments and outcomes require causal diagrams, and observed data alone cannot answer them.The paper notes that external knowledge is needed to understand data-generating mechanisms.Source: §1 · p. 1
- Medical DAG construction can be constrained by the volume of literature and experts’ time, while experts may miss confounding pathways.The paper frames LLMs as a possible way to leverage existing text data that expresses causal sentiments.Source: §1 · p. 1
- The study asks whether LLMs can help researchers build medical causal diagrams from existing text data and under what conditions GPT-3 answers accurately.It considers prompt engineering, alternative language, and limitations.Source: §1 · pp. 1–2
- The contributions include testing whether GPT-3 can identify edge presence or absence and whether prompt language or linking verbs improve classification accuracy.The paper also explores GPT-3’s limitations in understanding medical causal relationships.Source: §1 · p. 2
2 Background
The background describes LLMs as repositories of learned relationships whose performance may depend on prompts and training text. It also introduces DAGs as tools for representing causal assumptions and relationships.
2.1 Large language models
- LLMs can encode learned relationships, but their performance is sensitive to prompts, and training-text style may affect how causal language is used.Prompt-based learning can use in-domain examples or task descriptions; the internet training corpus spans casual and formal writing.Source: §2.1 · p. 2
2.2 Causal diagram overview
- DAGs illustrate qualitative assumptions, represent causal relationships, identify potential confounders to control when estimating causal effects, and classify relationship types that may give rise to selection bias.Source: §2.2 · p. 3
- A DAG contains measured and unmeasured variables connected by directed edges; an arrow points from a parent to a child, while no arrow indicates no direct relationship.DAGs are directed and acyclic, with no loops.Source: §2.2 · p. 3
- A causal DAG represents direct effects and includes all common causes, while its arrows exclude mediators relative to the chosen level of abstraction.Its directed paths represent causal pathways from ancestors to descendants, without feedback loops.Source: §2.2 · p. 3
3 Experiments
The experiments tested GPT-3’s edge classification on four medical DAGs, comparing prompts, linking verbs, and variable specificity. Accuracy varied with prompt authority and wording, while added specificity often worsened it.
3.1 Experimental details
- The evaluation used four medical DAGs varying in complexity and context, scoring edge-present and edge-absent statements for each ordered variable pair; accuracy required the correct statement to score higher.Source: §3.1 · p. 4
3.2 Results
- Medical-doctor prompts significantly improved GPT-3 accuracy in two DAGs, but baseline prompts outperformed alternatives in Diabetes and Obesity, while Big Pharma lowered accuracy versus baseline in three of four.Source: Table 1 · p. 4, §3.2 · p. 4
- “Increases risk” had the highest accuracy for three of four DAGs but not Alcohol; “cause” yielded decent results across all DAGs, with no verb consistently improving accuracy.Source: Table 2 · p. 5, §3.2 · p. 5
- More specific statements did not improve accuracy overall and often worsened it across linking verbs, despite improved accuracy for excessive-alcohol wording on the Alcohol DAG.Source: Table 3 · p. 5, §3.2 · p. 5
4 Discussion
GPT-3 scored above chance on every tested DAG in at least one setting, though accuracy varied with prompts and linking verbs. The discussion frames medical causal-graph development as an emerging LLM application with important scope limits.
4 Discussion
- Above 50% accuracy (random guessing) was reached on every tested DAG in at least one prompt or linking-verb setting, and accuracy varied with wording.The study describes its tested DAGs as relatively simple and well studied.Source: §4 · pp. 5–6
- To the authors’ knowledge, this was the first study to examine LLMs for causal-diagram development in the medical context.Source: §4 · p. 6
- Earlier medical NLP studies reported promising LLM performance on medical exam questions, while context-specific models such as BioBert outperformed GPT-3 on medical NLP tasks.Source: §4 · p. 6
- The discussion identifies training-data lag for novel diseases, casual internet language, and the assumption that probed causal links are established in the corpus as limitations.Source: §4 · p. 6
- Future work proposes testing medical-context models with PubMed or medical terminology and controlling for acyclicity, which preliminary evaluations did not examine.Source: §4 · p. 6
5 Conclusion
GPT-3’s edge accuracy depends on how variable relationships are worded, and the authors recommend expert verification when using LLMs to build DAGs. They see a possible role for LLMs in combining common knowledge from medical text with expert knowledge, while calling for further research on limitations.
5 Conclusion
- GPT-3’s accuracy in confirming an edge between two DAG variables depends on the language used to describe their relationship.Source: §5 · p. 6
- Because expert opinion can also produce imperfect DAGs by omitting confounders, the authors recommend expert verification when LLMs are used to build DAGs.Source: §5 · p. 6
- LLMs may extract common knowledge from medical text that, paired with expert knowledge, could make comprehensive DAG creation more efficient; the authors call for further research on reported limitations.Source: §5 · p. 6