LAMA: LAnguage Model Analysis

LAMA ia a probe for analyzing the factual and commonsense knowledge contained in pretrained language models.

The dataset for the LAMA probe is available at https://dl.fbaipublicfiles.com/LAMA/data.zip

LAMA contains a set of connectors to pretrained language models.
LAMA exposes a transparent and unique interface to use:

• Transformer-XL (Dai et al., 2019)
• BERT (Devlin et al., 2018)
• ELMo (Peters et al., 2018)
• GPT (Radford et al., 2018)
• RoBERTa (Liu et al., 2019)

Actually, LAMA is also a beautiful animal.

Reference:

The LAMA probe is described in the following paper:

@inproceedings{petroni2019language,
  title={Language Models as Knowledge Bases?},
  author={F. Petroni, T. Rockt{\"{a}}schel, A. H. Miller, P. Lewis, A. Bakhtin, Y. Wu and S. Riedel},
  booktitle={In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019},
  year={2019}
}

The LAMA probe

To reproduce our results:

1. Create conda environment and install requirements

(optional) It might be a good idea to use a separate conda environment. It can be created by running:

conda create -n lama37 -y python=3.7 && conda activate lama37
pip install -r requirements.txt

2. Download the data

wget https://dl.fbaipublicfiles.com/LAMA/data.zip
unzip data.zip
rm data.zip

3. Download the models

DISCLAIMER: ~55 GB on disk

Install spacy model

python3 -m spacy download en

Download the models

chmod +x download_models.sh
./download_models.sh

The script will create and populate a pre-trained_language_models folder. If you are interested in a particular model please edit the script.

4. Run the experiemnts

python scripts/run_experiments.py

results will be logged in output/ and last_results.csv.

LAMA-UHN

This repository also provides a script (scripts/create_lama_uhn.py) to create the data used in (Poerner et al., 2019).

Negated-LAMA

This repository also gives the option to evalute how pretrained language models handle negated probes (Kassner et al., 2019), set the flag use_negated_probes in scripts/run_experiments.py.

What else can you do with LAMA?

1. Encode a list of sentences

and use the vectors in your downstream task!

pip install -e git+https://github.com/facebookresearch/LAMA#egg=LAMA

import argparse
from lama.build_encoded_dataset import encode, load_encoded_dataset

PARAMETERS= {
        "lm": "bert",
        "bert_model_name": "bert-large-cased",
        "bert_model_dir":
        "pre-trained_language_models/bert/cased_L-24_H-1024_A-16",
        "bert_vocab_name": "vocab.txt",
        "batch_size": 32
        }

args = argparse.Namespace(**PARAMETERS)

sentences = [
        ["The cat is on the table ."],  # single-sentence instance
        ["The dog is sleeping on the sofa .", "He makes happy noises ."],  # two-sentence
        ]

encoded_dataset = encode(args, sentences)
print("Embedding shape: %s" % str(encoded_dataset[0].embedding.shape))
print("Tokens: %r" % encoded_dataset[0].tokens)

# save on disk the encoded dataset
encoded_dataset.save("test.pkl")

# load from disk the encoded dataset
new_encoded_dataset = load_encoded_dataset("test.pkl")
print("Embedding shape: %s" % str(new_encoded_dataset[0].embedding.shape))
print("Tokens: %r" % new_encoded_dataset[0].tokens)

2. Fill a sentence with a gap.

You should use the symbol [MASK] to specify the gap. Only single-token gap supported - i.e., a single [MASK].

python lama/eval_generation.py  \
--lm "bert"  \
--t "The cat is on the [MASK]."

_{^{source: https://commons.wikimedia.org/wiki/File:Bluebell_on_the_phone.jpg}}

Note that you could use this functionality to answer cloze-style questions, such as:

python lama/eval_generation.py  \
--lm "bert"  \
--t "The theory of relativity was developed by [MASK] ."

Install LAMA with pip

Clone the repo

git clone [email protected]:facebookresearch/LAMA.git && cd LAMA

Install as an editable package:

pip install --editable .

If you get an error in mac os x, please try running this instead

CFLAGS="-Wno-deprecated-declarations -std=c++11 -stdlib=libc++" pip install --editable .

Language Model(s) options

Option to indicate which language model(s) to use:

• --language-models/--lm : comma separated list of language models (REQUIRED)

BERT

BERT pretrained models can be loaded both: (i) passing the name of the model and using huggingface cached versions or (ii) passing the folder containing the vocabulary and the PyTorch pretrained model (look at convert_tf_checkpoint_to_pytorch in here to convert the TensorFlow model to PyTorch).

• --bert-model-dir/--bmd : directory that contains the BERT pre-trained model and the vocabulary
• --bert-model-name/--bmn : name of the huggingface cached versions of the BERT pre-trained model (default = 'bert-base-cased')
• --bert-vocab-name/--bvn : name of vocabulary used to pre-train the BERT model (default = 'vocab.txt')

RoBERTa

• --roberta-model-dir/--rmd : directory that contains the RoBERTa pre-trained model and the vocabulary (REQUIRED)
• --roberta-model-name/--rmn : name of the RoBERTa pre-trained model (default = 'model.pt')
• --roberta-vocab-name/--rvn : name of vocabulary used to pre-train the RoBERTa model (default = 'dict.txt')

ELMo

• --elmo-model-dir/--emd : directory that contains the ELMo pre-trained model and the vocabulary (REQUIRED)
• --elmo-model-name/--emn : name of the ELMo pre-trained model (default = 'elmo_2x4096_512_2048cnn_2xhighway')
• --elmo-vocab-name/--evn : name of vocabulary used to pre-train the ELMo model (default = 'vocab-2016-09-10.txt')

Transformer-XL

• --transformerxl-model-dir/--tmd : directory that contains the pre-trained model and the vocabulary (REQUIRED)
• --transformerxl-model-name/--tmn : name of the pre-trained model (default = 'transfo-xl-wt103')

GPT

• --gpt-model-dir/--gmd : directory that contains the gpt pre-trained model and the vocabulary (REQUIRED)
• --gpt-model-name/--gmn : name of the gpt pre-trained model (default = 'openai-gpt')

Evaluate Language Model(s) Generation

options:

• --text/--t : text to compute the generation for
• --i : interactive mode
  one of the two is required

example considering both BERT and ELMo:

python lama/eval_generation.py \
--lm "bert,elmo" \
--bmd "pre-trained_language_models/bert/cased_L-24_H-1024_A-16/" \
--emd "pre-trained_language_models/elmo/original/" \
--t "The cat is on the [MASK]."

example considering only BERT with the default pre-trained model, in an interactive fashion:

python lamas/eval_generation.py  \
--lm "bert"  \
--i

Get Contextual Embeddings

python lama/get_contextual_embeddings.py \
--lm "bert,elmo" \
--bmn bert-base-cased \
--emd "pre-trained_language_models/elmo/original/"

Troubleshooting

If the module cannot be found, preface the python command with PYTHONPATH=.

If the experiments fail on GPU memory allocation, try reducing batch size.

Acknowledgements

• https://github.com/huggingface/pytorch-pretrained-BERT
• https://github.com/allenai/allennlp
• https://github.com/pytorch/fairseq

Other References

• (Poerner et al., 2019) Nina Poerner, Ulli Waltinger, and Hinrich Schütze. BERT is Not a Knowledge Base (Yet): Factual Knowledge vs. Name-Based Reasoning in Unsupervised QA. arXiv preprint arXiv:1911.03681, 2019.

• (Dai et al., 2019) Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G. Carbonell, Quoc V. Le, and Ruslan Salakhutdi. Transformer-xl: Attentive language models beyond a fixed-length context. CoRR, abs/1901.02860.

• (Peters et al., 2018) Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. NAACL-HLT 2018

• (Devlin et al., 2018) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805.

• (Radford et al., 2018) Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training.

• (Liu et al., 2019) Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv:1907.11692.

Licence

LAMA is licensed under the CC-BY-NC 4.0 license. The text of the license can be found here.