Towards Consistent Natural Language Explanations via Explanation-Consistency Finetuning

This is the implementation of the paper Towards Consistent Natural Language Explanations via Explanation-Consistency Finetuning.

Table of Contents

• Overview
• Requirements
• Code Structure
• Demo
• How to Cite

Overview

Large language models (LLMs) often generate convincing, fluent explanations. However, their explanations are often inconsistent on different inputs. For example, an LLM may generate the explanation "all birds can fly" when answering the question "Can sparrows fly?", but answer "no" to the question "Can penguins fly?". Explanations should be consistent on different inputs so that they should allow a human to simulate the LLM's decision process on multiple inputs based on the LLM's explanation for a single input.

We propose Explanation-consistency Finetuning (EC Finetuning), a method that adapts LLMs so that they generate more consistent natural-language explanations. EC Finetuning involves finetuning an LLM on synthetic data that is carefully constructed to contain consistent explanations. Across a variety of question-answering datasets, EC Finetuning improves the consistency of natural-language explanations, and generalizes to datasets which were not seen during finetuning.

You could find more details of this work in our paper.

Requirements

To run our code, please install all the dependency packages by using the following command:

pip install -r requirements.txt

Code Structure

Code is within the src/ directory:

• cf_gen.py generates counterfactuals relevant to an explanation.

• cf_ans.py generates answers to counterfactuals consistent with the model's explanations on other inputs.

• consistency_data_augmentation.py generates EC training data.

• expl_generation.py finetunes CoT QA systems on (input, explanation, output) tuples.

• score_consistency.py scores the consistency of the model's explanations.

We include the demonstration examples we use to generate EC training data in ec_dems/.

Demo

As an example, we provide a code file src/main.py that runs the entire pipeline (generate EC training data, finetune CoT models on EC data, score the consistency of explanations).

This file can be run simply with cd src && python main.py

Questions?

If you have any questions related to the code or the paper, feel free to reach out to us at yanda.chen@cs.columbia.edu.

Citation

@misc{chen2024consistent,
      title={Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning}, 
      author={Yanda Chen and Chandan Singh and Xiaodong Liu and Simiao Zuo and Bin Yu and He He and Jianfeng Gao},
      year={2024},
      eprint={2401.13986},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}