AutoCAP: Towards Automatic Cross-lingual Alignment Planning for Zero-shot Chain-of-Thought

📷 This is the code repository for the paper: AutoCAP: Towards Automatic Cross-lingual Alignment Planning for Zero-shot Chain-of-Thought. ACL 2024 Findings.[PDF]

The overall workflow of AutoCAP, which consist of Automatic Language Selection Prompting and Automatic Weight Allocation Prompting.

Preparation steps: environment installation

Environment installation command:

pip install -r requirements.txt

💻 Stage 1: Automatic Language Selection Prompting

First modify the output path /temperature in manage_res_request.py:

INPUT_DIR = "mgsm/input"
OUTPUT_DIR = "auto-clsp-exp/mgsm/l6-01-tp02-tp02"
TEMPERATURE_1 = 0.2
TEMPERATURE_2 = 0.2

---

【OPENAI KEY modification request.py】

--request_url [end url]
--api_key [api key]

---

Run the following command to request Automatic Language Selection Prompting and Automatic Weight Allocation Prompting.

python manage_res_request.py

💻 Stage 2: Request missing CLP/evaluation metrics

Modify SELECTION_PATH to f'{OUTPUT_DIR}/output-2' from the previous stage:

---

【OPENAI KEY modification tool.py】

openai.api_key [api key]
openai.api_base [end url]

---

Run the following command to request Automatic Language Selection Prompting and Automatic Weight Allocation Prompting.

python metric4auto.py

💯 Model Performance

💬 Contact

Please create Github issues here or email Yongheng Zhang or Qiguang Chen or Libo Qin if you have any questions or suggestions.

📲 Reference

@inproceedings{zhang-etal-2024-autocap,
    title = "{A}uto{CAP}: Towards Automatic Cross-lingual Alignment Planning for Zero-shot Chain-of-Thought",
    author = "Zhang, Yongheng  and
      Chen, Qiguang  and
      Li, Min  and
      Che, Wanxiang  and
      Qin, Libo",
    booktitle = "Findings of the Association for Computational Linguistics ACL 2024",
    month = aug,
    year = "2024",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-acl.546",
    doi = "10.18653/v1/2024.findings-acl.546",
    pages = "9191--9200",
}