Causal Semantics (S_C)

A framework for extracting and analyzing causal attributions

Overview

Causal Semantics is a computational framework for extracting and analyzing causal relations from natural language text. It bridges three research traditions:

  • Logical theories differentiating between monocausal and polycausal structures
  • Linguistic analyses identifying semantic dimensions like promoting vs. inhibiting influences
  • Computational methods that scale, but tend to reduce relations to binary cause-effect pairs

In short, S_C aims to extract and represent causal attributions:

Climate change causes species extinction.

and represent them as graphs:

\text{Climate change} \xrightarrow{+1} \text{Species extinction}

Core Innovation

The framework models causal relations as (C, E, I) tuples, where:

  • C (Cause): The causing entity
  • E (Effect): The affected entity
  • I (Influence): A signed scalar \in [-1, +1] encoding:
    • Polarity (sign): Promoting (+) vs. inhibiting (−) influence
    • Salience (magnitude): Monocausal (|I|=1.0) vs. polycausal (|I|<1.0) attribution

This representation enables:

  1. Semantic precision: Capturing the direction and strength of a causal attribution
  2. Quantitative aggregation: Accumulating attributions into weighted causal networks
  3. Graph-based analysis: Visualizing discourse dynamics as Attributional Causal Graphs (ACGs)

Example

Consider these sentences from environmental discourse:

Climate change causes species extinction.
(C_\text{climate change}, E_\text{species extinction}, I_{+1.0})

Conservation measures reduce forest dieback.
(C_\text{conservation measures}, E_\text{forest dieback}, I_{-0.5})

The first example expresses a promoting, monocausal relation (I=+1.0), while the second expresses an inhibiting, contributory relation (I=-0.5).

Architecture

The framework consists of three main modules:

graph LR
    A[Text Input] --> B1[Annotation]
    A[Text Input] --> B2[C-BERT]

    B1 --> C[Tuple-Construction]
    B2 --> C

    C --> D[Aggregation]
    D --> E[ACG]
    
    E --> D2[Visualization]
    E --> D3[Analysis]

  1. Extraction: Identifying causal relations in text through indicators, annotation schemes, and the C-BERT transformer
  2. Processing: Converting annotations into formal (C,E,I) tuples and aggregating them

Applications

This framework has been applied to a German Environmental corpus (1990-2022) to:

  • analyze responsibility attributions in biodiversity debates [1]
  • disambiguate references to forest diebacks [2]

Citation

If you use this framework in your research, please cite:

@phdthesis{johnson2026causalsemantics,
  title={Kausalsemantik. Eine Operationalisierung der -sterben Komposita im Umweltdiskurs},
  author={Patrick Johnson},
  school={Technical University of Darmstadt},
  year={forthcoming}
}
@misc{cbert,
  title={C-BERT: Factorized Causal Relation Extraction},
  author={Patrick Johnson},
  doi={10.26083/tuda-7797},
  year={2026}
}

References

[1]
Johnson P. Causal semantics. An operationalization using the example of -sterben compounds in german environmental discourse. PhD thesis. Technical University of Darmstadt, forthcoming.
[2]
Johnson P. Causal attribution and semantic change over time in waldsterben (“forest dieback”) within german public discourse 2026 (in pr.).