Are-Large-Language-Models-Attentive-Readers

This project is a framework to create adversarial paragraphs for the HotpotQA dataset, by treating each question in HotpotQA as a two-hop question . It extracts relevant information from each of these hops using Stanza, and uses several techniques to create fake sub-questions which are then fed into GPT-4 to create the fake paragraphs.

For citing the paper and the repo:

@inproceedings{bhuiya2024seemingly,
  title={Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?},
  author={Bhuiya, Neeladri and Schlegel, Viktor and Winkler, Stefan},
  booktitle={Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing},
  pages={2514--2528},
  year={2024}
}

Prerequisites

• Python 3.10.12

Install the required packages from the requirements.txt file:

pip install -r requirements.txt

Put your OpenAI API key in the .env file.

How to Use

Run pipeline.sh to execute all the preprocessing steps.

Note

For cases where OpenAI sends no responses, it is stored as an empty array in the file. These should be handled by rerunning the requests either using fill_in_failed_req.py or your own script.

Once pipeline.sh is run successfully, use create_prompts_with_fake_para.py to create the adversarial datasets.

python create_prompts_with_fake_para.py [-h] [-t TYPE] [--related | --no-related] [-c COUNT] [-m MODEL] [--same_type | --no-same_type]

The parameters correspond to the section 5.3 in the paper. The MODEL parameter is for defining the prompt format. If no model is specified, it creates the dataset in the HotpotQA format in $FINAL_PROMPTS/hotpotqa_format, where FINAL_PROMPTS is defined in the .env file.

The Pipeline

Extracting the Modifiable Parts

getting_modifiable_parts.py parses the sub-questions to find the main object and modifiable portions (converted to "other" and "named-entities"). To pass in your own sub-questions, you need to modify this file.

Creating the Fake Sub-questions

create_prompts_for_fake_named.py creates OpenAI prompts to generate fake named entities to replace the named-entities extracted from the previous step. create_fake_entities.py uses the prompts and sends requests to OpenAI, and extract_fake_named_entities.py extracts the fake named entities from the OpenAI responses.

create_real_tuples is a preprocessing step where tuples are created from the sub-questions. create_fake_questions.py uses the modifiable portions from the first step along with the fake named-entities extracted to create fake questions, which will later be used to create the adversarial paragraphs.

Creating the Adversarial Paragraphs

create_fake_question_tuples.py combines different fake sub-questions with the same base question to generate multiple fake sub-question pairs. This script also creates OpenAI prompts to create adversarial paragraphs based on these.

create_fake_para.py sends the requests, and extract_fake_paragraphs.py extracts the fake/adversarial paragraphs.

Dataset

The folder adversarial_datasets contains the the final datasets in the distractor setting with the plausible paragraphs generated using the above method, The format of the files is exactly the same as that of HotpotQA.