[Paper Review] Learning Causal Bayesian Networks from Text

Farhad Moghimifar, Afshin Rahimi+2 more|arXiv (Cornell University)|Nov 26, 2020

Topic Modeling18 references4 citations

TL;DR

This paper proposes a novel method to extract causal relationships between high-level concepts from text, using linguistic variables and formal concept analysis to build a Causal Bayesian Network (CBN), achieving 69.4% accuracy—13-20% higher than state-of-the-art methods—on a psychological text dataset, with capability to detect non-adjacent and bidirectional causalities.

ABSTRACT

Causal relationships form the basis for reasoning and decision-making in Artificial Intelligence systems. To exploit the large volume of textual data available today, the automatic discovery of causal relationships from text has emerged as a significant challenge in recent years. Existing approaches in this realm are limited to the extraction of low-level relations among individual events. To overcome the limitations of the existing approaches, in this paper, we propose a method for automatic inference of causal relationships from human written language at conceptual level. To this end, we leverage the characteristics of hierarchy of concepts and linguistic variables created from text, and represent the extracted causal relationships in the form of a Causal Bayesian Network. Our experiments demonstrate superiority of our approach over the existing approaches in inferring complex causal reasoning from the text.

Motivation & Objective

• To address the limitation of existing NLP methods that only extract causal relations between low-level, adjacent events.
• To enable causal reasoning at the conceptual level, independent of sentence proximity in text.
• To represent extracted causal knowledge in a structured, probabilistic form (Causal Bayesian Network) for downstream AI applications.
• To develop a hierarchical linguistic variable framework that supports inference of causal relations between concepts not directly co-occurring in text.
• To release the PsyCaus dataset for evaluating causal relation extraction in psychology, supporting reproducible research.

Proposed method

• Extract linguistic variables and values from text using IsA (hypernym-hyponym) relations via a probabilistic method (Wu et al., 2012).
• Construct a formal concept lattice using Formal Concept Analysis (Ganter and Wille, 2012) to model hierarchical relationships between linguistic variables and their values.
• Apply a novel causality detection function (Equation 1) that computes asymmetric similarity scores between variable pairs, with a threshold μ to determine causal direction.
• Use the hierarchical structure to infer causal links between concepts even when they are not directly mentioned together in the text.
• Represent the final causal knowledge as a Causal Bayesian Network (CBN), enabling probabilistic inference and counterfactual reasoning.
• Leverage the CBN to model bidirectional causality by detecting symmetric absolute values in the causality score function, though the directed nature of CBNs requires post-processing (e.g., decision trees) for such relations.

Experimental results

Research questions

• RQ1 Can causal relationships between high-level concepts be reliably extracted from text, even when the cause and effect are not in adjacent sentences?
• RQ2 Can a hierarchical linguistic variable framework improve the detection of non-adjacent causal relations in text?
• RQ3 How effective is the proposed method in identifying bidirectional causal relationships between concepts (e.g., emotion and eating disorder)?
• RQ4 Can a Causal Bayesian Network be effectively constructed from textual data to support complex causal and counterfactual reasoning?
• RQ5 To what extent does the proposed method outperform existing feature-based, distribution-based, and heuristic baselines in causal relation extraction?

Key findings

• The proposed method achieved a test accuracy of 0.694 on causal relation classification, significantly outperforming state-of-the-art feature-based methods (0.56) and distribution-based methods (0.50).
• The method improved by 13% over the best feature-based baseline and 20% over the majority-class baseline, demonstrating robustness and generalization.
• The model achieved a macro-averaged F1 score of 0.66 at threshold μ = 0.05, indicating optimal sensitivity to the causality detection threshold.
• The model successfully identified bidirectional causal relations (e.g., between emotion and eating disorder), even when no explicit causal connectives were present.
• The hierarchical structure enabled detection of causal links between concepts not co-occurring in the same sentence, such as delusional disorder and displeasure, which were correctly identified as causally related.
• The model’s performance was sensitive to the threshold μ, with F1 peaking at μ = 0.05, confirming the importance of hyperparameter tuning for optimal causality detection.