metadata

license: mit
language:
  - multilingual
tags:
  - zero-shot-classification
  - text-classification
  - pytorch
metrics:
  - accuracy
  - f1-score
extra_gated_prompt: >-
  Please provide some basic information that confirms the scholarly use of our
  models.

  If you use our models for your work or research, please cite this paper:
  Sebők, M., Máté, Á., Ring, O., Kovács, V., & Lehoczki, R. (2024). Leveraging
  Open Large Language Models for Multilingual Policy Topic Classification: The
  Babel Machine Approach. Social Science Computer Review, 0(0).
  https://doi.org/10.1177/08944393241259434
extra_gated_fields:
  Name: text
  Country: country
  Institution: text
  E-mail: text

xlm-roberta-large-judiciary-cap-v3

Model description

An xlm-roberta-large model finetuned on multilingual training data containing texts of the judiciary domain labelled with major topic codes from the Comparative Agendas Project.

How to use the model

Loading and tokenizing input data

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

model = AutoModelForSequenceClassification.from_pretrained('poltextlab/xlm-roberta-large-judiciary-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-judiciary-cap-v3 was fine-tuned using the Hugging Face Trainer class with the following hyperparameters:

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 1833 examples (10% of the available data).
Model accuracy is 0.77.

label	precision	recall	f1-score	support
0	0.39	0.26	0.32	34
1	0.82	0.8	0.81	296
2	0.64	0.62	0.63	34
3	0.7	0.78	0.74	9
4	0.87	0.77	0.82	171
5	0.48	0.41	0.44	29
6	0.68	0.65	0.67	20
7	0.79	0.88	0.83	56
8	0.65	0.79	0.71	33
9	0.68	0.8	0.74	81
10	0.88	0.83	0.85	489
11	0.71	0.79	0.75	28
12	0.67	0.89	0.76	9
13	0.77	0.82	0.79	251
14	0.59	0.86	0.7	37
15	0.65	0.62	0.64	24
16	0.64	0.43	0.51	21
17	1	0.14	0.25	7
18	0.59	0.67	0.63	139
19	0.77	0.86	0.81	63
20	0	0	0	2
21	0	0	0	0
macro avg	0.64	0.62	0.61	1833
weighted avg	0.78	0.77	0.77	1833

Inference platform

This model is used by the CAP Babel Machine, 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.

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.