On Building a Semi-Automated Framework for Generating Causal Bayesian Networks from Raw Text

Solat J. SheikhIJCAI 2023 · DOI

TL;DR

This paper introduces SCANER, a semi-automated framework to convert unstructured text into Causal Bayesian Networks (CBNs). This approach aims to overcome the limitations of purely manual or fully automated methods, enabling more robust decision-support tools across diverse domains.

What Problem Does This Paper Address?

In today's data-rich world, vast amounts of unstructured text from sources like social media, news, and medical records hold valuable information for decision-making. However, extracting actionable insights from this raw text is challenging due to its inherent complexity and ambiguity. While manual construction of causal models (like Causal Bayesian Networks) is accurate, it's labor-intensive. Existing automated methods often struggle with domain generalization and can produce less interpretable or less comprehensive models. This research tackles the need for a generalized, semi-automated framework that can efficiently and accurately build causal models from raw text, making them accessible for decision support in critical domains.

Methods

The proposed framework, SCANER (Semi-automated Causal Network Extraction from Raw text), combines manual and automated techniques. The workflow involves:

  • Manual Simplification of Text: Initial human review to prepare the text.
  • Preprocessing: Standard NLP techniques like lemmatization.
  • Triples Extraction: Identifying subject-verb-object (SVO) triples using libraries and dependency parsing.
  • Acquiring Causal Trigger Words: Utilizing a dictionary of words that indicate causality.
  • Filtering Causal Triples: Selecting triples that represent causal relationships.
  • Causal Network Generation: Transforming the filtered causal triples into a Directed Acyclic Graph (DAG).
  • Causal Bayesian Network Generation: Populating the DAG with conditional probability tables (CPTs) to create a full CBN.
  • Utilization as Decision Support Tool: Applying the generated CBNs for decision-making.

The novelty lies in the semi-automated nature, aiming for generalization across domains and a balance between human oversight and automated efficiency.

Datasets

The paper mentions testing the framework on text from three domains: political narratives, food insecurity, and medical science. No specific datasets are detailed, and the authors note the unavailability of datasets with ground truth for validation.

Key Insights and Ideas

  • Bridging the Gap: The research highlights the critical need for a middle ground between purely manual and fully automated causal model generation, addressing the trade-offs in accuracy, labor, and generalizability.
  • Domain Generalization is Key: Existing automated methods are often domain-specific. SCANER aims to be a generalized framework, applicable across various text types.
  • Semi-Automation for Trust: By incorporating manual simplification and potentially human annotation for validation, the framework aims to build trust in the generated models, especially for critical decision-making.
  • From Text to Actionable Models: The core idea is to transform unstructured text into structured, probabilistic models (CBNs) that can directly inform decisions.

Key Results

The paper reports that the component for extracting causal networks (without CPTs) has been developed and a prototype is available on a cloud platform for visualization. Preliminary experiments on three domains show that the proposed approach generates "dense and accurate causal networks" when compared to the popular causality extraction system Eidos. These findings have been submitted for publication in an impact factor journal.

Limitations and Future Work

The authors acknowledge several limitations and outline future directions:

  • Validation Data: The lack of ground truth datasets for validation is a significant challenge. The current approach relies on feedback from human annotators.
  • CPT Generation: A key focus for future work is populating the causal networks with conditional probability tables (CPTs) to create full Causal Bayesian Networks.
  • Model Modification: The authors plan to explore modifying generated models, particularly for political narratives, to reduce aggressiveness or extremity in the text without causing conflict.
  • Domain Application: While tested on three domains, the framework's broader applicability and performance across a wider range of domains will be a subject of future investigation.

Summary updated

BibTeX
@inproceedings{sheikh2023building,
  author = {Solat J. Sheikh},
  title = {{On Building a Semi-Automated Framework for Generating Causal Bayesian Networks from Raw Text}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/822},
}

Loading related content...

Figure