| | --- |
| | license: gpl-3.0 |
| | language: |
| | - en |
| | annotations_creators: |
| | - expert-annotated |
| | task_categories: |
| | - text-classification |
| | - token-classification |
| | size_categories: |
| | - 1k<n<10k |
| | pretty_name: Causal Sentence and Cause–Effect Span Annotation |
| | --- |
| | |
| | # Causal Sentence and Cause–Effect Span Annotation |
| |
|
| | ## Dataset Summary |
| | This dataset contains 3,014 annotated sentences for causal language analysis. Half of the sentences are labeled as causal, and half as non-causal. For causal sentences, annotations include cause and effect spans, along with explicit cause–effect links. |
| |
|
| | The annotations were created using the Doccano platform in a three-step process: |
| | 1. **Causality classification**: Label each sentence as causal or non-causal. |
| | 2. **Span extraction**: Mark the spans corresponding to cause(s) and effect(s). |
| | 3. **Linking**: Connect each cause span to its corresponding effect span. |
| |
|
| | ## Annotation Process |
| | A pilot study was conducted to assess inter-annotator reliability. Three authors independently annotated 60 sentences (20 per source). Agreement was measured using **Krippendorff’s alpha (α)** for each task. |
| | - For span extraction, agreement was computed with **partial matching** (e.g., “poor sleeping during the night” vs. “poor sleeping” were considered a match). |
| | - Average Krippendorff’s α across all tasks was **0.80**, exceeding the standard reliability threshold (Krippendorff, 2004). |
| |
|
| | After validation, the lead author annotated the full dataset following the same guidelines. |
| |
|
| | ## Dataset Composition |
| | - **Total sentences**: 3,014 |
| | - **Non-causal sentences**: ~50% |
| | - **Causal sentences**: ~50% (with span and link annotations) |
| |
|
| | ## Data Splits |
| | The dataset is divided as follows: |
| | - **Training**: 70% |
| | - **Validation**: 15% |
| | - **Test**: 15% |
| |
|
