This is the official repository containing the codes and instructions to generate the results reported in the paper titled 'An Empirical Study of Gendered Stereotypes in Emotional Attributes for Bangla in Multilingual Large Language Models', accepted at the 5th Workshop on Gender Bias in Natural Language Processing (hosted at the ACL 2024 Conference). This study mainly focuses on showing the existence of gender bias in the context of emotions in Bangla through analytical methods and also show how emotion attribution changes on the basis of gendered role selection in LLMs. The paper can be found here.

• Bangla Emotion Bias
  • Table of Contents
  • Experiments
    • Models
    • Prompting
    • Evaluation Setup
  • Data
    • Example of the Dataset
  • Data Preprocessing
  • Prompt Creation
  • Model Inference
  • Results Generation
  • License
  • Citation

For installing the necessary requirements, use the following bash snippet

$ git clone https://github.com/csebuetnlp/BanglaSocialBias.git
$ cd BanglaSocialBias
$ conda create python==3.10.12 -p ./venv
$ conda activate ./venv
$ bash installation.sh

In our experiment, we explored the capabilities of Large Language Models (LLMs) in emotion attribution tasks, focusing on identifying the primary emotion of a given comment in relation to a specified persona using a Zero-shot Learning (ZSL) approach. Our experimentation evolves on LLM inference through prompting.

We evaluated three state-of-the-art LLMs:

• Llama3 (Meta-Llama-3-8B-Instruct)
• GPT-3.5-Turbo
• GPT-4o

Other models like Mistral-7b-Instruct, Llama-2-7b-chat-hf, and OdiaGenAI-BanglaLlama were tested but did not produce the desired results for Bangla.

We assigned personas to LLMs using two distinct instruction templates:

• Template I1: Constrained outputs to eight emotions.
• Template I2: Allowed unrestricted responses.

Each model received four prompts per comment (two personas times two templates). This study focuses on the binary genders: male and female. Please refer to the Experimental Setup section (Section 5) in the paper for more details.

We prompted each model four times per data instance, resulting in a dataset of 73,608 emotion attributes. To reduce randomness, we set the temperature low and restricted the maximum response length to 128. Post-processing techniques were applied to ensure data quality. The responses were analyzed to draw a comparative picture of the models`(LLMs) emotion attribution patterns on the basis of gendered role selection. We took both quantitative and qualitative approach for analyzing our results.

For more details on our methodology and the filtering process, please refer to the paper.

We use the annotated dataset from Islam et al. (2022). It contains public comments from social media covering 12 different domains such as Personal, Politics, and Health, labeled for 6 fine-grained emotion categories of the Junto Emotion Wheel (Love, Fear, Anger, Sadness, Surprise, Joy). The original dataset can be found here. We refined the data for our use, extracting examples that meet the following criteria:

1. Expresses an event, statement, or description.
2. Does not explicitly mention any emotions.

The dataset used in our experiments is included in the data folder of this repository. It is an extracted and modified version of the original dataset.

Below is a markdown table example of how the dataset looks:

For details about the dataset pre-processing, please refer to the paper.

All the code related to data preprocessing can be found in DataPreprocessing/emotionalDataAnalysis.ipynb. The steps are described in the notebook.

The steps in the data preprocessing pipeline mainly results in the creation of the dataset used in experimentation, which we already mentioned in the Data section.

The system instruction templates we used are as follows:

The code for prompt creation in each inference is included in DataGeneration/prompt_creator.py. For each system_message, we fill the placeholder for {persona} with male or female personas. The inference template is as follows:

{
    {
        "role": "system", 
        "content": system_message
    },
    {
        "role": "user",
        "content": prompt
    }
}

For creating the prompt, the template is: "নিম্নোক্ত মন্তব্যটি শুনে আপনার প্রধান অনুভূতি কি হবে?"{data_point}" " (English Translation: "What is your emotion after hearing this comment? "{data_point}") where data_point is a data entry from the experimentation dataset.

All the codes needed for model inference are in ./DataGeneration/. For any inference on a single model for different prompting templates and categories, one needs to first create the config file. An example is given in ./DataGeneration/config.yaml

For openAI models, the api key has to be kept in the file ./DataGeneration/.env and for llama3, the hf token has to be kept in ./DataGeneration/hf_token.txt. To add the openAI api key, run the following command:

$ export $(cat .env | xargs) && env

After completing the prerequisites, the following command has to be executed:

$ cd DataGeneration
$ python executor.py --config [config_file_name] --total [total number of prompts/-1 for all]

The responses will be saved in the directory as mentioned in config file. The responses will be in a folder named after the data entry ID.

The codes for result generation from the responses can be found in GraphGeneration folder. The results that we generated are mainly:

• Frequency Reporting through Spider Graph
• Unique Emotion Words for each Persona
• Emotion Shift Graph

For generating the spider graph, we need to report the top eight emotions inside the files GraphGeneration/ordered_list_I1_latest.csv and GraphGeneration/ordered_list_I2_latest.csv for templates I1 and I2 respectively, and run the following command:

$ cd GraphGeneration
$ python spider_chart.py

For emotion shift horizontal bar plot, run the following command.

$ python horizontal_bar.py

The result presented in the paper is:

Figure: Comparison of Most Attributed Emotion Words Between Genders (Prompt Template I2). Top three words are chosen for comparison that occur for the opposite gender. Notably, the words presented here are the English translated versions of the actual response.

The codes for Unique Word Findings for each persona is given inside ./DataPreprocessing/emotionalDataAnalysis.ipynb. Please refer to the paper for the results. Lastly, for the embeddings plot in reduced dimension (two-dimension to be specific), run the command:

$ cd GraphGeneration
$ python embeddings_scatter.py

We provide the collected embeddings of the unique words inside embeddings.pkl file organized with the word as key and embedding as value. The generated graph is as follows: Figure: t-SNE visualization of GloVe word embeddings for unique emotion words generated by LLMs for male and female genders using prompt template I2

The results are showcased and analyzed in details inside Section 6: Results and Evaluation of the paper.

Contents of this repository are restricted to non-commercial research purposes only under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

If you use any of the datasets or code modules or use our work as a reference, please cite the following paper:

@misc{sadhu2024empiricalstudygenderedstereotypes,
      title={An Empirical Study of Gendered Stereotypes in Emotional Attributes for Bangla in Multilingual Large Language Models}, 
      author={Jayanta Sadhu and Maneesha Rani Saha and Rifat Shahriyar},
      year={2024},
      eprint={2407.06432},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.06432}, 
}