[Paper Review] Uncertainty Weighted Causal Graphs

Eduardo C. Garrido‐Merchán, Cristina Puente+2 more|arXiv (Cornell University)|Feb 2, 2020

Misinformation and Its Impacts16 references4 citations

TL;DR

This paper proposes a method to generate causal graphs from text that represent uncertainty in causal relationships using Bayesian inference with probability distributions over linguistic hedges (e.g., 'often', 'rarely'). By modeling the strength of causal links as posterior distributions derived from adverb frequency and prior beliefs, the approach produces weighted causal graphs that capture imprecision and confidence levels, validated through synthetic and real-world lung cancer text analysis with coherent, interpretable uncertainty distributions.

ABSTRACT

Causality has traditionally been a scientific way to generate knowledge by relating causes to effects. From an imaginery point of view, causal graphs are a helpful tool for representing and infering new causal information. In previous works, we have generated automatically causal graphs associated to a given concept by analyzing sets of documents and extracting and representing the found causal information in that visual way. The retrieved information shows that causality is frequently imperfect rather than exact, feature gathered by the graph. In this work we will attempt to go a step further modelling the uncertainty in the graph through probabilistic improving the management of the imprecision in the quoted graph.

Motivation & Objective

• To address the limitation of point estimates in causal graph edge weights, which fail to represent uncertainty in real-world causal knowledge.
• To model causal relationships as probability distributions rather than crisp values, reflecting degrees of belief and imprecision in natural language.
• To develop a system that generates interpretable, uncertainty-aware causal graphs from domain-specific text, such as medical literature.
• To validate the approach using both synthetic data and real-world medical texts, demonstrating coherent uncertainty quantification.
• To lay the foundation for future integration into knowledge systems, question-answering, and machine consciousness architectures.

Proposed method

• Extract causal relations from text using linguistic parsing, identifying causes, effects, and qualifying adverbs (e.g., 'frequently', 'rarely') as uncertainty indicators.
• Assign prior probability distributions to adverbs based on their semantic meaning and linguistic context, modeling initial beliefs about causal strength.
• Apply Bayesian inference to update priors using observed causal relation frequencies, computing posterior distributions over causal link strength.
• Use KL divergence to select the most representative adverb for each edge, minimizing discrepancy between inferred and actual posterior distributions.
• Approximate posterior distributions using a discrete grid over the probability simplex, with computational complexity O(NR), where N is the number of causal relations and R is grid size.
• Visualize the resulting causal graph with edge weights as probability distributions, where edge size reflects relation frequency and label shows the most likely adverb.

Experimental results

Research questions

• RQ1 How can causal relationships in text be modeled with uncertainty rather than point estimates?
• RQ2 To what extent can linguistic hedges (e.g., 'often', 'rarely') be used to quantify the strength and reliability of causal links?
• RQ3 Can Bayesian inference effectively learn posterior distributions over causal link strength from sparse or imprecise textual data?
• RQ4 How do different prior distributions and grid resolutions affect the accuracy and interpretability of the learned uncertainty?
• RQ5 Can the system produce coherent, domain-valid uncertainty representations in real-world medical text, such as lung cancer causality?

Key findings

• The system successfully models causal relationships as probability distributions, capturing uncertainty features such as skewness and kurtosis, which point estimates cannot.
• In synthetic experiments, the method learned spiky posterior distributions when adverbs were consistent, indicating high confidence, and broad distributions when data was sparse, reflecting high uncertainty.
• In the real lung cancer domain experiment, the system correctly inferred that heavy smoking has a high probability of causing lung cancer (near 1), while workplace exposure showed high uncertainty, aligning with medical intuition.
• The posterior distribution for 'inhaling radon gas' as a cause of lung cancer was concentrated near high probability, consistent with known medical evidence.
• The system demonstrated that edge weights reflect both the frequency of causal relations and the semantic strength of adverbs, with edge size proportional to relation count.
• The computational cost scales quadratically with the number of causal relations and grid size, making it feasible for moderate-scale text corpora.