LANKA
This is the source code for paper: Knowledgeable or Educated Guess? Revisiting Language Models as Knowledge Bases (ACL 2021, long paper)
Reference
If this repository helps you, please kindly cite the following bibtext:
@inproceedings{cao-etal-2021-knowledgeable,
title = "Knowledgeable or Educated Guess? Revisiting Language Models as Knowledge Bases",
author = "Cao, Boxi and
Lin, Hongyu and
Han, Xianpei and
Sun, Le and
Yan, Lingyong and
Liao, Meng and
Xue, Tong and
Xu, Jin",
booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
month = aug,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.acl-long.146",
pages = "1860--1874",
Usage
To reproduce our results:
1. Create conda environment and install requirements
git clone https://github.com/c-box/LANKA.git
cd LANKA
conda create --name lanka python=3.7
conda activate lanka
pip install -r requirements.txt
2. Download the data
-
Download the data using terminal
pip install gdown gdown https://drive.google.com/uc?id=1oQ7TXrZ7aQXpZnENu2Sytc8A0D3yvqkP unzip data.zip rm data.zip
-
Or you can acquire the data using the following Google Drive link.
https://drive.google.com/file/d/1oQ7TXrZ7aQXpZnENu2Sytc8A0D3yvqkP/view?usp=sharing
3. Run the experiments
If your GPU is smaller than 24G, please adjust batch size using "--batch-size" parameter.
3.1 Prompt-based Retrieval
-
Evaluate the precision on LAMA and WIKI-UNI using different prompts:
-
Manually prompts created by Petroni et al. (2019)
python -m scripts.run_prompt_based --relation-type lama_original --model-name bert-large-cased --method evaluation --cuda-device [device] --batch-size [batch_size]
-
Mining-based prompts by Jiang et al. (2020b)
python -m scripts.run_prompt_based --relation-type lama_mine --model-name bert-large-cased --method evaluation --cuda-device [device]
-
Automatically searched prompts from Shin et al. (2020)
python -m scripts.run_prompt_based --relation-type lama_auto --model-name bert-large-cased --method evaluation --cuda-device [device]
-
-
Store various distributions needed for subsequent experiments:
python -m scripts.run_prompt_based --model-name bert-large-cased --method store_all_distribution --cuda-device [device]
-
Calculate the average percentage of instances being covered by top-k answers or predictions (Table 1):
python -m scripts.run_prompt_based --model-name bert-large-cased --method topk_cover --cuda-device [device]
-
Calculate the Pearson correlations of the prediction distributions on LAMA and WIKI-UNI (Figure 3, the figures will be stored in the 'pics' folder):
python -m scripts.run_prompt_based --model-name bert-large-cased --method prediction_corr --cuda-device [device]
-
Calculate the Pearson correlations between the prompt-only distribution and prediction distribution on WIKI-UNI (Figure 4):
python -m scripts.run_prompt_based --model-name bert-large-cased --method prompt_only_corr --cuda-device [device]
-
Calculate the KL divergence between the prompt-only distribution and golden answer distribution of LAMA (Table 2):
python -m scripts.run_prompt_based --relation-type [relation_type] --model-name bert-large-cased --method cal_prompt_only_div --cuda-device [device]
3.2 Case-based Analogy
-
Evaluate case-based paradigm:
python -m scripts.run_case_based --model-name bert-large-cased --task evaluate_analogy_reasoning --cuda-device [device]
-
Detailed comparison for prompt-based and case-based paradigms (precision, type precision, type change, etc.) (Table 4):
python -m scripts.run_case_based --model-name bert-large-cased --task type_precision --cuda-device [device]
-
Calculate the in-type rank change (Figure 6):
python -m scripts.run_case_based --model-name bert-large-cased --task type_rank_change --cuda-device [device]
3.3 Context-based Inference
-
For explicit answer leakage (Table 5 and 6):
python -m scripts.run_context_based --model-name bert-large-cased --method explicit_leak --cuda-device [device]
-
For implicit answer leakage (Table 7):
python -m scripts.run_context_based --model-name bert-large-cased --method implicit_leak --cuda-device [device]