Linguistic Ethical Interventions (LEI) 🌺

Is it possible to use natural language to intervene in a model’s behavior and alter its prediction in a desired way? Linguistic Ethical Interventions (or, LEI 🌺) is a dataset designed for studying this question. In this task, the goal is to amend a question answering (QA) model’s unethical behavior by communicating context-specific principles of ethics and equity to it.

This repository contains the data and the baselines used in our experiments. See the details below.

Getting the data

In the paper, we propose the LEI challenge based on UnQover by adding interventions to the QA example to verify if existing models can understand and follow the instructions we provided. Here we briefly talk about the training and test dataset used in our paper:

• Easiest way: You can download the data we used for train/test in the paper here. It contains the data for both the specified and underspecified cases.
• From Scratch
  • We add interventions to a subset of UnQover (covers religion, ethnicity and gender). Please using the activities and nouns provided under word_lists folder.
  • The interventions are under interventions folder.
  • For underspecified case:
    • generate the questions:
    1. follow the instructions in UnQover.

    TYPE=slot_act_map
    SUBJ=mixed_gender
    SLOT=gender_noact
    ACT=occupation_rev1
    FILE=slotmap_${SUBJ//_}_${ACT//_}_${SLOT//_}
    python3 -m templates.generate_underspecified_templates --template_type ${TYPE} \
    --subj $SUBJ --act $ACT --slot $SLOT \
    --output ./data/${FILE}.source.json
    

    2. Do the predictions. The interventions will be added to the QA context.

    intv_type=ethical
    python3 -u -m qa_hf.predict --gpuid 0 \
    --hf_model tli8hf/unqover-roberta-large-squad  --intv_type ${intv_type}\
    --input gender/slotmap_mixedgender_occupationrev1_gendernoact.source.json --output ./data/gender/robertalarge_gender-${intv_type}.output.json --attribute gender --converted 0
    

  • For specified case (only for gender):
    1. Replace the /templates folder in UnQover with the templates folder provided.
    2. Repeat the above procedure

Reproducing the Models and Evaluation

Our model is based on RoBERTa-SQuAD model, which can be obtained from UnQover repo (See Section 2).

remember to install the required packages as in requirement.txt

We fine tune the aforementioned model on religion dataset and verify its generalization to gender/ethnicity. To do that, simply run ./run.sh 1e-5 data/underspecified outputs 64 4 4 1 "--doadversarial --doirrelevant --squad" U_ US theta4 1 0.5 1. Note that this also includes the model selection part, and feel free to comment that out.

---

Todo:

• spotted typos in interventions
• run the cmd to verify the code

Relevant Paper

Ethical-Advice Taker:Do Language Models Understand Natural Language Interventions? (ACL 2021 Findings) Jieyu Zhao, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Kai-Wei Chang