Research map / Probabilistic and scientific modeling / Causal discovery
To Know the Causes of Things: Text Mining for Causal Relations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A doctoral-consortium overview of the author's research on end-to-end causal text mining: causal corpus creation (CNC, UniCausal), BERT/GNN-based causal extraction, counterfactual data augmentation, and causal knowledge graph construction from news.
Problem
Extracting, representing, and applying causal relations from unstructured text, given inconsistent causal datasets and models that fail on negated or strengthened causal sentences.
Approach
Treats causal text mining end-to-end: builds unified causal corpora, trains transformer/GNN extractors, augments training with counterfactual causal sentences as contrast sets, and converts extracted relations into clustered causal knowledge graphs.
Key results
- Model in official FinCausal 2021 run reached 95.56% precision, 95.56% recall and 95.57% F1, placing 1st.
- UniCausal consolidates six causal corpora into 58,720, 12,144 and 69,165 examples for its three tasks.
- Causal KG construction reduced 15,686 disconnected subgraphs to one connected graph via argument clustering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Include a mixture of counterfactual edits in training | Causal sentence classification models · Within- and out-of-corpus domains | performance | Helped |
| Cluster arguments when constructing the causal knowledge graph | News causal knowledge graph · News | connected components: 15,686 -> 1 | Helped |
Surprising findings
- Models misclassify augmented causal sentences that are negated or strengthened in causal meaning, despite minor linguistic differences.
Limitations
- Current work focuses on extracting causal relations from reliable news sources, not fake or subjective content.
- Causal strength, direction and context-dependence (non-transitivity) are not yet modelled.
Experimental setup
Datasets: Causal News Corpus UniCausal FinCausal 2021
Benchmarks: Causal News Corpus UniCausal FinCausal 2021
Metrics: precision recall F1
Builds on: BERT UniCausal Causal News Corpus
Closest papers by meaning
- A Joint Framework with Heterogeneous-Relation-Aware Graph and Multi-Channel Label Enhancing Strategy for Event Causality Extraction (AAAI 2024 · similarity 0.84)
- Revealing Multimodal Causality with Large Language Models (NeurIPS 2025 · similarity 0.84)
- Integrating Background Knowledge for Scalable Causal Discovery (NeurIPS 2026 · similarity 0.83)
- CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.83)
- CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.83)
- Causality can systematically address the monsters under the bench(marks) (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.