README.md

---
datasets:
- CausalNewsCorpus
language: en
library_name: transformers
license: mit
metrics:
- f1
- precision
- recall
tags:
- token-classification
- roberta
- causal-narrative
- cause-effect-extraction
- span-extraction
- ner
---

# RoBERTa Causal Span Extractor

This model is a fine-tuned version of `roberta-base` for **causal span extraction**
(token classification). It identifies **cause** and **effect** text spans in sentences.

## Model Description

- **Base Model**: roberta-base
- **Task**: Token classification (BIO tagging)
- **Labels**: O, B-CAUSE, I-CAUSE, B-EFFECT, I-EFFECT
- **Training Data**: CausalNewsCorpus V2 (sentences with exactly 1 causal pair)
- **Training Samples**: 1105
- **Dev Samples**: 133

## Training Results

See the training notebook for detailed metrics.

## Usage

```python
from transformers import RobertaTokenizerFast, RobertaForTokenClassification
import torch

model_name = "causal-narrative/roberta-causal-span-extractor"
tokenizer = RobertaTokenizerFast.from_pretrained(model_name, add_prefix_space=True)
model = RobertaForTokenClassification.from_pretrained(model_name)

text = "The heavy rain caused flooding in the city."
words = text.split()
inputs = tokenizer(words, is_split_into_words=True, return_tensors="pt",
                   truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    preds = torch.argmax(outputs.logits, dim=2)[0]

id2label = model.config.id2label
word_ids = tokenizer(words, is_split_into_words=True).word_ids()
prev = None
for wid in word_ids:
    if wid is not None and wid != prev:
        print(f"{words[wid]:20s} {id2label[preds[word_ids.index(wid)].item()]}")
        prev = wid
```

## Labels

| Label | Description |
|-------|-------------|
| O | Non-causal token |
| B-CAUSE | Beginning of cause span |
| I-CAUSE | Inside cause span |
| B-EFFECT | Beginning of effect span |
| I-EFFECT | Inside effect span |