Responsibility Perspective Transfer for Italian Femicide News (ACL 2023 Findings)

Abstract: Different ways of linguistically expressing the same real-world event can lead to different perceptions of what happened. Previous work has shown that different descriptions of gender-based violence (GBV) influence the reader's perception of who is to blame for the violence, possibly reinforcing stereotypes which see the victim as partly responsible, too. As a contribution to raise awareness on perspective-based writing, and to facilitate access to alternative perspectives, we introduce the novel task of automatically rewriting GBV descriptions as a means to alter the perceived level of responsibility on the perpetrator. We present a quasi-parallel dataset of sentences with low and high perceived responsibility levels for the perpetrator, and experiment with unsupervised (mBART-based), zero-shot and few-shot (GPT3-based) methods for rewriting sentences. We evaluate our models using a questionnaire study and a suite of automatic metrics.

Repo overview

This repository contains all of the data and code necessary for reproducing our experiments.

Key files

• input data: our dataset is based on the ones used/created in the SocioFillmore project, see https://aclanthology.org/2022.acl-demo.24/ (base dataset) & https://aclanthology.org/2022.aacl-main.79/ (perspective prediction dataset) and the repositories linked to it. The file data/questionnaire-combined.jsonl contains the averaged gold perception scores from the perspective prediction project.
• aligned data: the aligned dataset (sentences referring to the same femicide event and the same facts inside the event) can be found in data/alignment-annotations/final_aligned_version.xslx. output/align contains different ways of rearranging this data, see the script make_evaluation_pairs.py
• gpt3 predictions: see output/gpt3_predictions for the outputs used to do the analysis for the paper. See output/extra_runs for additional runs that did not make it into the paper.
• evaluation: * automatic evaluation: see the script autoeval_perspective.py and the files in output/auto-eval. * human evaluation: see the notebooks notebooks/human_eval.ipynb for the analysis of the main evaluation survey and notebooks/human_eval.fewshot_prompt.ipynb for the analysis of the survey for evaluating few-shot examples. See the questionnaire/ folder for files used to prepare Qualtrics surveys, and output/questionnaire/responses.processed.*.csv for anonymized + aggregated ratings. For privacy reasons, we cannot release the raw filled questionnaires output.

mBART-based model

How to use

from transformers import MBartForConditionalGeneration, MBart50TokenizerFast

model = MBartForConditionalGeneration.from_pretrained("laihuiyuan/RPT")
tokenizer = MBart50TokenizerFast.from_pretrained("laihuiyuan/RPT", src_lang="it_IT")

source = "Provaglio d'Iseo , donna trovata morta in casa : si sospetta il compagno"
meta_info = "Simona Simonini, Elio Cadei, partner, percosse, Provaglio d'Iseo, c." #<victim name, perpetrator name, relationship, weapon, municipality, place>

inputs = meta_info + ' ' +source
inputs = tokenizer(inputs, return_tensors="pt")
decode_start_id =tokenizer.lang_code_to_id['it_IT']
output = model.generate(input_ids=inputs['input_ids'], num_beams=5, max_length=80, forced_bos_token_id=decode_start_id)
transferred_text = tokenizer.decode(output[0].tolist(), skip_special_tokens=True, clean_up_tokenization_spaces=False)

Citation

@inproceedings{minnemaa-etal-2023-responsibility,
    title = "Responsibility Perspective Transfer for Italian Femicide News",
    author = "Minnemaa, Gosse and Lai, Huiyuan and Muscato, Benedetta and Nissim, Malvina",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2023",
    month = July,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
}