metadata

dataset_info:
  - config_name: Behaviour
    features:
      - name: text
        dtype: string
      - name: label
        dtype: int64
    splits:
      - name: train
        num_bytes: 608966
        num_examples: 5000
      - name: test
        num_bytes: 128067
        num_examples: 1000
    download_size: 455378
    dataset_size: 737033
  - config_name: Synth
    features:
      - name: text
        dtype: string
      - name: label
        dtype: int64
    splits:
      - name: train
        num_bytes: 681703
        num_examples: 7014
      - name: test
        num_bytes: 199047
        num_examples: 2908
    download_size: 498443
    dataset_size: 880750
configs:
  - config_name: Behaviour
    data_files:
      - split: train
        path: Behaviour/train-*
      - split: test
        path: Behaviour/test-*
  - config_name: Synth
    data_files:
      - split: train
        path: Synth/train-*
      - split: test
        path: Synth/test-*

Automatic Misogyny Identification (AMI)

Original Paper: https://amievalita2020.github.io

Task presented at EVALITA-2020

This task consists of tweet classification, specifically, categorization of the level of misogyny in a given text.

We taken both subtasks, raw_dataset uploaded as Behaviour (3 class classification) and synthetic uploaded as Synth (2 class classification).

Example

Here you can see the structure of the single sample in the present dataset.

Behaviour

{
  "text": string, # text of the tweet
  "label": int, # 0: Non Misogino, 1: Misogino, 2: Misogino Aggressivo
}

Synth

{
  "text": string, # text of the tweet
  "label": int, # 0: Non Misogino, 1: Misogino
}

Statitics

AMI Behaviour	Non Misogino	Misogino	Misogino Aggressivo
Training	2663	554	1783
Test	500	324	176

AMI Synth	Non Misogino	Misogino
Training	3670	3344
Test	1454	1454

Proposed Prompts

Here we will describe the prompt given to the model over which we will compute the perplexity score, as model's answer we will chose the prompt with lower perplexity. Moreover, for each subtask, we define a description that is prepended to the prompts, needed by the model to understand the task.

Behaviour

Description of the task: "Indica il livello di misoginia presente nei seguenti tweet.\n\n"

Cloze Style:

Label (Non Misogino): "Tweet: '{{text}}'.\nIl tweet non presenta caratteristiche misogine."

Label (Misogino): "Tweet: '{{text}}'.\nIl tweet presenta caratteristiche misogine."

Label (Misogino Aggressivo): "Tweet: '{{text}}'.\nIl tweet presenta caratteristiche misogine aggressive."

MCQA Style:

Tweet: '{{text}}'.\nDomanda: che livello di misoginia è presente nel tweet?\nA. Nessuno\nB. Misogino\nC. Misogino Aggressivo\nRisposta:

Synth

Description of the task: "Indica se i seguenti tweet presentano caratteristiche misogine.\n\n"

Cloze Style:

Label (Non Misogino): "Tweet: '{{text}}'.\nIl tweet non presenta caratteristiche misogine."

Label (Misogino): "Tweet: '{{text}}'.\nIl tweet presenta caratteristiche misogine."

MCQA Style:

Tweet: '{{text}}'.\nDomanda: Il tweet contiene elementi misogini? Rispondi sì o no:

Results

The following results are given by the Cloze-style prompting over some english and italian-adapted LLMs.

AMI Synth	ACCURACY (5-shots)
Gemma-2B	53.78
QWEN2-1.5B	60.72
Mistral-7B	71.59
ZEFIRO	74.69
Llama-3-8B	74.55
Llama-3-8B-IT	78.47
ANITA	82.66

Acknowledge

We would like to thank the authors of this resource for publicly releasing such an intriguing benchmark.

Additionally, we extend our gratitude to the students of the MNLP-2024 course, whose first homework explored various interesting prompting strategies.

The original dataset is freely available for download link.

License

The data come under license Creative Commons Attribution Non Commercial Share Alike 4.0 International.