---
---
license: mit
language:
- multilingual
tags:
- zero-shot-classification
- text-classification
- pytorch
metrics:
- accuracy
- f1-score
---
# xlm-roberta-large-legislative-cap-v3
## Model description
An `xlm-roberta-large` model finetuned on multilingual training data containing texts of the `legislative` domain labelled with [major topic codes](https://www.comparativeagendas.net/pages/master-codebook) from the [Comparative Agendas Project](https://www.comparativeagendas.net/).

## How to use the model
#### Loading and tokenizing input data
```python
import pandas as pd
import numpy as np
from datasets import Dataset
from transformers import (AutoModelForSequenceClassification, AutoTokenizer, 
                          Trainer, TrainingArguments)

CAP_NUM_DICT = {0: '1', 1: '2', 2: '3', 3: '4', 4: '5', 5: '6', 
6: '7', 7: '8', 8: '9', 9: '10', 10: '12', 11: '13', 12: '14', 
13: '15', 14: '16', 15: '17', 16: '18', 17: '19', 18: '20', 19: 
'21', 20: '23', 21: '999'}

tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-large')
num_labels = len(CAP_NUM_DICT)

def tokenize_dataset(data : pd.DataFrame):
    tokenized = tokenizer(data["text"],
                          max_length=MAXLEN,
                          truncation=True,
                          padding="max_length")
    return tokenized

hg_data = Dataset.from_pandas(data)
dataset = hg_data.map(tokenize_dataset, batched=True, remove_columns=hg_data.column_names)
```

#### Inference using the Trainer class
```python
model = AutoModelForSequenceClassification.from_pretrained('poltextlab/xlm-roberta-large-legislative-cap-v3',
                                                           num_labels=22,
                                                           problem_type="multi_label_classification",
                                                           ignore_mismatched_sizes=True
                                                           )

training_args = TrainingArguments(
    output_dir='.',
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8
)

trainer = Trainer(
    model=model,
    args=training_args
)

probs = trainer.predict(test_dataset=dataset).predictions
predicted = pd.DataFrame(np.argmax(probs, axis=1)).replace({0: CAP_NUM_DICT}).rename(
    columns={0: 'predicted'}).reset_index(drop=True)

```

### Fine-tuning procedure
`xlm-roberta-large-legislative-cap-v3` was fine-tuned using the Hugging Face Trainer class with the following hyperparameters:
```python
training_args = TrainingArguments(
    output_dir=f"../model/{model_dir}/tmp/",
    logging_dir=f"../logs/{model_dir}/",
    logging_strategy='epoch',
    num_train_epochs=10,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    learning_rate=5e-06,
    seed=42,
    save_strategy='epoch',
    evaluation_strategy='epoch',
    save_total_limit=1,
    load_best_model_at_end=True
)
```
We also incorporated an EarlyStoppingCallback in the process with a patience of 2 epochs.

## Model performance
The model was evaluated on a test set of 163947 examples (10% of the available data).<br>
Model accuracy is **0.89**.
| label        |   precision |   recall |   f1-score |   support |
|:-------------|------------:|---------:|-----------:|----------:|
| 0            |        0.83 |     0.82 |       0.83 |      7095 |
| 1            |        0.82 |     0.8  |       0.81 |      3326 |
| 2            |        0.9  |     0.93 |       0.92 |     10975 |
| 3            |        0.9  |     0.91 |       0.91 |      5783 |
| 4            |        0.85 |     0.84 |       0.85 |      5451 |
| 5            |        0.9  |     0.94 |       0.92 |      6858 |
| 6            |        0.87 |     0.88 |       0.88 |      5495 |
| 7            |        0.88 |     0.92 |       0.9  |      4597 |
| 8            |        0.87 |     0.89 |       0.88 |      1737 |
| 9            |        0.89 |     0.91 |       0.9  |      8849 |
| 10           |        0.89 |     0.9  |       0.9  |     11791 |
| 11           |        0.86 |     0.87 |       0.87 |      6234 |
| 12           |        0.88 |     0.82 |       0.85 |      3593 |
| 13           |        0.88 |     0.85 |       0.87 |     10178 |
| 14           |        0.88 |     0.88 |       0.88 |     10644 |
| 15           |        0.92 |     0.91 |       0.91 |      4381 |
| 16           |        0.91 |     0.92 |       0.91 |      5067 |
| 17           |        0.84 |     0.86 |       0.85 |      4896 |
| 18           |        0.9  |     0.87 |       0.88 |     19029 |
| 19           |        0.88 |     0.92 |       0.9  |     10958 |
| 20           |        0.74 |     0.72 |       0.73 |       225 |
| 21           |        0.99 |     0.96 |       0.97 |     16785 |
| macro avg    |        0.88 |     0.88 |       0.88 |    163947 |
| weighted avg |        0.89 |     0.89 |       0.89 |    163947 |

## Inference platform
This model is used by the [CAP Babel Machine](https://babel.poltextlab.com), an open-source and free natural language processing tool, designed to simplify and speed up projects for comparative research.  

## Cooperation
Model performance can be significantly improved by extending our training sets. We appreciate every submission of CAP-coded corpora (of any domain and language) at poltextlab{at}poltextlab{dot}com or by using the [CAP Babel Machine](https://babel.poltextlab.com).

## Debugging and issues
This architecture uses the `sentencepiece` tokenizer. In order to run the model before `transformers==4.27` you need to install it manually.

If you encounter a `RuntimeError` when loading the model using the `from_pretrained()` method, adding `ignore_mismatched_sizes=True` should solve the issue.