Iker/NoticIA

"A Clickbait Article Summarization Dataset in Spanish."

We present NoticIA, a dataset consisting of 850 Spanish news articles featuring prominent clickbait headlines, each paired with high-quality, single-sentence generative summarizations written by humans.

• 📖 Paper: NoticIA: A Clickbait Article Summarization Dataset in Spanish
• 💻 Baseline Code: https://github.com/ikergarcia1996/NoticIA
• 🤖 Pre Trained Models https://huggingface.co/collections/Iker/noticia-and-clickbaitfighter-65fdb2f80c34d7c063d3e48e
• 🔌 Online Demo: https://iker-clickbaitfighter.hf.space/

For example, given the following headline and web text:

# ¿Qué pasará el 15 de enero de 2024?
Al parecer, no todo es dulzura en las vacaciones de fin de años, como lo demuestra la nueva intrig....

The summary is:

Que los estudiantes vuelven a clase.

Data explanation

• web_url (int): The URL of the news article
• web_headline (str): The headline of the article, which is a Clickbait.
• web_text (int): The body of the article.
• summary (str): The summary written by humans that answers the clickbait headline.

Dataset Usage

1. We are working on implementing NoticIA on the Language Model Evaluation Harness library: https://github.com/EleutherAI/lm-evaluation-harness

2. If you want to train an LLM or reproduce the results in our paper, you can use our code. See the repository for more info: https://github.com/ikergarcia1996/NoticIA

3. If you want to manually load the dataset and run inference with an LLM: You can load the dataset with the following command:

from datasets import load_dataset
dataset = load_dataset("Iker/NoticIA")

In order to perform inference with LLMs, you need to build a prompt. The one we use in our paper is:

def clickbait_prompt(
    headline: str,
    body: str,
) -> str:
    """
    Generate the prompt for the model.
    Args:
        headline (`str`):
            The headline of the article.
        body (`str`):
            The body of the article.
    Returns:
        `str`: The formatted prompt.
    """
    return (
        f"Ahora eres una Inteligencia Artificial experta en desmontar titulares sensacionalistas o clickbait. "
        f"Tu tarea consiste en analizar noticias con titulares sensacionalistas y "
        f"generar un resumen de una sola frase que revele la verdad detrás del titular.\n"
        f"Este es el titular de la noticia: {headline}\n"
        f"El titular plantea una pregunta o proporciona información incompleta. "
        f"Debes buscar en el cuerpo de la noticia una frase que responda lo que se sugiere en el título. "
        f"Responde siempre que puedas parafraseando el texto original. "
        f"Usa siempre las mínimas palabras posibles. "
        f"Recuerda responder siempre en Español.\n"
        f"Este es el cuerpo de la noticia:\n"
        f"{body}\n"
    )

Here is a practical end-to-end example using the text generation pipeline.

from transformers import pipeline
from datasets import load_dataset

generator = pipeline(model="google/gemma-2b-it",device_map="auto")
dataset = load_dataset("Iker/NoticIA")

example = dataset["test"][0]
prompt = clickbait_prompt(headline=example["web_headline"],body=example["web_text"])
outputs = generator(prompt, return_full_text=False,max_length=4096)
print(outputs)

# [{'generated_text': 'La tuitera ha recibido un número considerable de comentarios y mensajes de apoyo.'}]

Here is a practical end-to-end example using the generate function

from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset

tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b-it")
model = AutoModelForCausalLM.from_pretrained("google/gemma-2b-it",device_map="auto",quantization_config={"load_in_4bit": True})
dataset = load_dataset("Iker/NoticIA")

example = dataset["test"][0]
prompt = clickbait_prompt(headline=example["web_headline"],body=example["web_text"])
prompt = tokenizer.apply_chat_template(
        [{"role": "user", "content": prompt}],
        tokenize=False,
        add_generation_prompt=True,
    )
model_inputs = tokenizer(
        text=prompt,
        max_length=3096,
        truncation=True,
        padding=False,
        return_tensors="pt",
        add_special_tokens=False,
    )

outputs = model.generate(**model_inputs,max_length=4096)
output_text = tokenizer.batch_decode(outputs)

print(output_text[0])

# La usuaria ha comprado un abrigo para su abuela de 97 años, pero la "yaya" no está de acuerdo.

Uses

This dataset is intended to build models tailored for academic research that can extract information from large texts. The objective is to research whether current LLMs, given a question formulated as a Clickbait headline, can locate the answer within the article body and summarize the information in a few words. The dataset also aims to serve as a task to evaluate the performance of current LLMs in Spanish.

Out-of-Scope Use

You cannot use this dataset to develop systems that directly harm the newspapers included in the dataset. This includes using the dataset to train profit-oriented LLMs capable of generating articles from a short text or headline, as well as developing profit-oriented bots that automatically summarize articles without the permission of the article's owner. Additionally, you are not permitted to train a system with this dataset that generates clickbait headlines.

This dataset contains text and headlines from newspapers; therefore, you cannot use it for commercial purposes unless you have the license for the data.

Dataset Creation

The dataset has been meticulously created by hand. We utilize two sources to compile Clickbait articles:

• The Twitter user @ahorrandoclick1, who reposts Clickbait articles along with a hand-crafted summary. Although we use their summaries as a reference, most of them have been rewritten (750 examples from this source).
• The web demo ⚔️ClickbaitFighter⚔️, which operates a pre-trained model using an early iteration of our dataset. We collect all the model inputs/outputs and manually correct them (100 examples from this source).

Who are the annotators?

The dataset was annotated by Iker García-Ferrero and validated by Begoña Altuna. The annotation took ~40 hours.

Citation

@misc{noticia2024,
      title={NoticIA: A Clickbait Article Summarization Dataset in Spanish}, 
      author={Iker García-Ferrero and Begoña Altuna},
      year={2024},
      eprint={2404.07611},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}