xlm-roberta-large-hungarian-cap-v3

Model description

An xlm-roberta-large model finetuned on hungarian training data 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-hungarian-cap-v3',
                                                           num_labels=22,
                                                           problem_type="multi_label_classification",
                                                           ignore_mismatched_sizes=True
                                                           )

training_args = TrainingArguments(
    output_dir='.',
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16
)

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-hungarian-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=16,
    per_device_eval_batch_size=16,
    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 67749 examples (10% of the available data).
Model accuracy is 0.83.

label	precision	recall	f1-score	support
0	0.76	0.77	0.76	5815
1	0.64	0.6	0.62	1534
2	0.85	0.82	0.84	2217
3	0.82	0.81	0.81	1789
4	0.67	0.71	0.69	1635
5	0.91	0.88	0.9	2812
6	0.75	0.68	0.71	847
7	0.76	0.71	0.73	821
8	0.71	0.66	0.68	351
9	0.85	0.83	0.84	1489
10	0.74	0.77	0.76	2991
11	0.78	0.7	0.73	1476
12	0.72	0.67	0.7	1120
13	0.74	0.71	0.72	2129
14	0.82	0.76	0.79	1227
15	0.87	0.81	0.84	1104
16	0.66	0.55	0.6	456
17	0.64	0.7	0.67	3163
18	0.72	0.68	0.7	6056
19	0.76	0.8	0.78	1418
20	0.71	0.76	0.74	616
21	0.94	0.96	0.95	26683
macro avg	0.76	0.74	0.75	67749
weighted avg	0.83	0.83	0.83	67749

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.