alsace08 / sumcot Goto Github PK

Let's elicit LLMs summarize step by step following the professional communication theory!

In this work, you can use directly or get inspired by:

• A fine-grained generic summary data annotation protocol (combining micro and macro demands)

• An expert-aligned generic summary test set (rewrite CNN/DailyMail and BBC XSum)

• An expandable CoT-based open-end generation path (not only SumCoT)

Our annotation protocol is mainly based on Lasswell Communication Model --- a famous communication theory proposed by Lasswell(1948). Additionally, we removed as much noise as possible from the original data set and performed data analysis (See paper for more details).

Case comparisons for our Element-aware summary and original dataset-specific summary:

• See our constructed datasets in /data. The two json files hold the data of CNN/DailyMail and BBC/XSum, respectively.

• Each json file contains 200 samples, each containing the source document, original summary, element-aware summary, GPT-3 summary (standard), and GPT-3 summary (under SumCoT). You can use these data for more fine-grained evaluation.

Dataset format:

{
  "dataset_name": [
        {
        "id": 0,
        "src": "xxx",
        "original_summary": "xxx",
        "element-aware_summary": "xxx",
        "gpt3_summary": "xxx",
        "gpt3_cot_summary": "xxx"
        },
        {
        "id": 1,
        "src": "xxx",
        "original_summary": "xxx",
        "element-aware_summary": "xxx",
        "gpt3_summary": "xxx",
        "gpt3_cot_summary": "xxx"
        },
        ...
        {
        "id": 199,
        "src": "xxx",
        "original_summary": "xxx",
        "element-aware_summary": "xxx",
        "gpt3_summary": "xxx",
        "gpt3_cot_summary": "xxx"
        }
    ]
}

If you want to generate summaries (w/ or w/o SumCoT) by GPT-3, run generation.py:

python generation.py \
--dataset cnndm
--start_id 0
--end_id 199

where cot_true indicates if you want to use the SumCoT technique for generation, start_id and end_id determine the range of source documents in the dataset that you want to request. If you only want to generate the summary of the $a$-th sample, set start_id=end_id=a.

To facilitate the subsequent evaluation, the summaries just generated will be stored into $dataset_name$_output.json.

Note: Don't forget to modify your openai_key in the Decoder().

We also provide code in /evaluation/eva.py to reproduce the results:

python eva.py \
--dataset cnndm \
--start_id 0 \
--end_id 199 \
--bs_true false

where bs_true indicates if you want to use the bert score, start_id and end_id are the same as above.

You can also build your own dataset json file, adjust start_id and end_id to perform your own evaluation.

Experimental results in this paper are obtained during October-December 2022 via the OpenAI API interface (basically before the ChatGPT release). If your reproduction results are slightly different from the results of the paper, you can report any of them (just indicate).

@inproceedings{wang-etal-2023-element,
    title = "Element-aware Summarization with Large Language Models: Expert-aligned Evaluation and Chain-of-Thought Method",
    author = "Wang, Yiming  and
      Zhang, Zhuosheng  and
      Wang, Rui",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.482",
    doi = "10.18653/v1/2023.acl-long.482",
    pages = "8640--8665",
}