BERT is to NLP what AlexNet is to CV
This is the official implementation of BERT is to NLP what AlexNet is to CV: Can Pre-Trained Language Models Identify Analogies? (the camera-ready version of the paper is here) which has been accepted by the ACL 2021 main conference. We evaluate pretrained language models (LM) on five analogy tests that follow SAT-style format as below.
QUERY word:language
OPTION
(1) paint:portrait
(2) poetry:rhythm
(3) note:music <-- the answer!
(4) tale:story
(5) week:year
We devise a new class of scoring functions, referred to as analogical proportion (AP) scores, to solve word analogies in an unsurpervised fashion and investigate the relational knowledge that LM learnt through pretraining.
Please see our paper for more information and discussion.
Get started
git clone https://github.com/asahi417/analogy-language-model
cd analogy-language-model
pip install -e .
Run Experiments
The following scripts reproduce our results in the paper.
# get result for our main AP score
python experiments/experiment_ppl_variants.py
# get result for word embedding baseline
python experiments/experiment_word_embedding.py
# get result for other scoring function such as vector difference, etc
python experiments/experiment_scoring_comparison.py
Here's the result summary that can be attained by running those scripts.
Dataset
The datasets used in our experiments can be downloaded from the following link:
Please see the Analogy Tool for more information about the dataset and baselines.
Citation
Please cite our reference paper if you use our data or code:
@inproceedings{ushio-etal-2021-bert,
title = "{BERT} is to {NLP} what {A}lex{N}et is to {CV}: Can Pre-Trained Language Models Identify Analogies?",
author = "Ushio, Asahi and
Espinosa Anke, Luis and
Schockaert, Steven and
Camacho-Collados, Jose",
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.280",
doi = "10.18653/v1/2021.acl-long.280",
pages = "3609--3624",
abstract = "Analogies play a central role in human commonsense reasoning. The ability to recognize analogies such as {``}eye is to seeing what ear is to hearing{''}, sometimes referred to as analogical proportions, shape how we structure knowledge and understand language. Surprisingly, however, the task of identifying such analogies has not yet received much attention in the language model era. In this paper, we analyze the capabilities of transformer-based language models on this unsupervised task, using benchmarks obtained from educational settings, as well as more commonly used datasets. We find that off-the-shelf language models can identify analogies to a certain extent, but struggle with abstract and complex relations, and results are highly sensitive to model architecture and hyperparameters. Overall the best results were obtained with GPT-2 and RoBERTa, while configurations using BERT were not able to outperform word embedding models. Our results raise important questions for future work about how, and to what extent, pre-trained language models capture knowledge about abstract semantic relations.",
}
Please also cite the relevant reference papers if using any of the analogy datasets.
138 Dec 30, 2022
203 Dec 20, 2022
27 Dec 12, 2022
197 Dec 25, 2022
4 Jan 22, 2022
3 Jan 12, 2022
82 Dec 19, 2022
63 Dec 25, 2022
3.1k Jan 8, 2023
18 Dec 20, 2022
281 Dec 15, 2022
2 Jan 06, 2022
1.6k Dec 29, 2022
2 Dec 20, 2021
475 Jan 04, 2023
27 Jul 09, 2022
64 May 08, 2022
0 Nov 29, 2022
7 Sep 20, 2022
4 Nov 11, 2021
121 Dec 27, 2022
8.8k Jan 01, 2023
8 Nov 26, 2022
60 Dec 31, 2022
138 Dec 27, 2022
1 May 11, 2022
23 Sep 05, 2022
23 Dec 25, 2022
2k Jan 04, 2023