lm-scorer 0.6.0

Language Model based sentences scoring library

lm-scorer

📃 Language Model based sentences scoring library

Synopsis

This package provides a simple programming interface to score sentences using different ML language models.

A simple CLI is also available for quick prototyping.
You can run it locally or on directly on Colab using this notebook.

Do you believe that this is useful? Has it saved you time? Or maybe you simply like it?
If so, support this work with a Star ⭐️.

Install

pip install lm-scorer

Usage

import torch
from lm_scorer.models.auto import AutoLMScorer as LMScorer

# Available models
list(LMScorer.supported_model_names())
# => ["gpt2", "gpt2-medium", "gpt2-large", "gpt2-xl", distilgpt2"]

# Load model to cpu or cuda
device = "cuda:0" if torch.cuda.is_available() else "cpu"
batch_size = 1
scorer = LMScorer.from_pretrained("gpt2", device=device, batch_size=batch_size)

# By default the bos and eos (<|endoftext|>) tokens are added when scoring, so the
# score includes the probability that the sentence ends where it does. Pass
# eos=False (and/or bos=False) to exclude them:
# scorer = LMScorer.from_pretrained("gpt2", device=device, eos=False)

# Return token probabilities (provide log=True to return log probabilities)
scorer.tokens_score("I like this package.")
# => (scores, ids, tokens)
# scores = [0.018321, 0.0066431, 0.080633, 0.00060745, 0.27772, 0.0036381]
# ids    = [40,       588,       428,      5301,       13,      50256]
# tokens = ["I",      "Ġlike",   "Ġthis",  "Ġpackage", ".",     "<|endoftext|>"]

# Compute sentence score as the product of tokens' probabilities
scorer.sentence_score("I like this package.", reduce="prod")
# => 6.0231e-12

# Compute sentence score as the mean of tokens' probabilities
scorer.sentence_score("I like this package.", reduce="mean")
# => 0.064593

# Compute sentence score as the geometric mean of tokens' probabilities
scorer.sentence_score("I like this package.", reduce="gmean")
# => 0.013489

# Compute sentence score as the harmonic mean of tokens' probabilities
scorer.sentence_score("I like this package.", reduce="hmean")
# => 0.0028008

# Get the log of the sentence score.
scorer.sentence_score("I like this package.", log=True)
# => -25.835

# Score multiple sentences.
scorer.sentence_score(["Sentence 1", "Sentence 2"])
# => [1.1508e-11, 5.6645e-12]

# NB: Computations are done in log space so they should be numerically stable.

CLI

The pip package includes a CLI that you can use to score sentences.

usage: lm-scorer [-h] [--model-name MODEL_NAME] [--tokens] [--log-prob]
                 [--reduce REDUCE] [--batch-size BATCH_SIZE]
                 [--significant-figures SIGNIFICANT_FIGURES] [--cuda CUDA]
                 [--debug]
                 sentences-file-path

Get sentences probability using a language model.

positional arguments:
  sentences-file-path   A file containing sentences to score, one per line. If
                        - is given as filename it reads from stdin instead.

optional arguments:
  -h, --help            show this help message and exit
  --model-name MODEL_NAME, -m MODEL_NAME
                        The pretrained language model to use. Can be one of:
                        gpt2, gpt2-medium, gpt2-large, gpt2-xl, distilgpt2.
  --tokens, -t          If provided it provides the probability of each token
                        of each sentence.
  --log-prob, -lp       If provided log probabilities are returned instead.
  --reduce REDUCE, -r REDUCE
                        Reduce strategy applied on token probabilities to get
                        the sentence score. Available strategies are: prod,
                        mean, gmean, hmean.
  --batch-size BATCH_SIZE, -b BATCH_SIZE
                        Number of sentences to process in parallel.
  --significant-figures SIGNIFICANT_FIGURES, -sf SIGNIFICANT_FIGURES
                        Number of significant figures to use when printing
                        numbers.
  --cuda CUDA           If provided it runs the model on the given cuda
                        device.
  --debug               If provided it provides additional logging in case of
                        errors.

Development

You can install this library locally for development using the commands below. If you don't have it already, you need to install poetry first.

# Clone the repo
git clone https://github.com/simonepri/lm-scorer
# CD into the created folder
cd lm-scorer
# Create a virtualenv and install the required dependencies using poetry
poetry install

You can then run commands inside the virtualenv by using poetry run COMMAND.
Alternatively, you can open a shell inside the virtualenv using poetry shell.

If you wish to contribute to this project, run the following commands locally before opening a PR and check that no error is reported (warnings are fine).

# Run the code formatter
poetry run task format
# Run the linter
poetry run task lint
# Run the static type checker
poetry run task types
# Run the tests
poetry run task test

Citation

If you use lm-scorer in your research, please cite it using the metadata in CITATION.cff, or the following BibTeX entry:

@software{primarosa_lm_scorer,
  author    = {Primarosa, Simone},
  title     = {{lm-scorer}: Language Model based sentences scoring library},
  year      = {2020},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20584992},
  url       = {https://github.com/simonepri/lm-scorer}
}

The concept DOI 10.5281/zenodo.20584992 always resolves to the latest release; each release also has its own version DOI on Zenodo.

Used in research

lm-scorer has been used in the following peer-reviewed publications:

• Schuster, S., & Linzen, T. (2022). When a sentence does not introduce a discourse entity, Transformer-based models still sometimes refer to it. NAACL 2022. https://doi.org/10.18653/v1/2022.naacl-main.71
• Gupta, P., Tsvetkov, Y., & Bigham, J. P. (2021). Synthesizing Adversarial Negative Responses for Robust Response Ranking and Evaluation. Findings of the ACL: ACL-IJCNLP 2021. https://doi.org/10.18653/v1/2021.findings-acl.338
• Kasner, Z., & Dušek, O. (2020). Data-to-Text Generation with Iterative Text Editing. INLG 2020. https://aclanthology.org/2020.inlg-1.9
• Don-Yehiya, S., Choshen, L., & Abend, O. (2022). PreQuEL: Quality Estimation of Machine Translation Outputs in Advance. EMNLP 2022. https://doi.org/10.18653/v1/2022.emnlp-main.767
• Alqahtani, A., Sarioglu Kayi, E., Hamidian, S., Compton, M., & Diab, M. (2022). A Quantitative and Qualitative Analysis of Schizophrenia Language. CLPsych @ NAACL 2022. https://arxiv.org/abs/2201.10430
• Zhang, Z., Mita, M., & Komachi, M. (2023). ClozEx: A Task toward Generation of English Cloze Explanation. Findings of EMNLP 2023. https://aclanthology.org/2023.findings-emnlp.347
• Krause, L., Sommerauer, P., & Vossen, P. (2022). Towards More Informative List Verbalisations. KGSum @ ISWC 2022. https://ceur-ws.org/Vol-3257/paper14.pdf
• Zhang, J., Mishra, A., Avinesh, P. V. S., Patwardhan, S., & Agarwal, S. (2022). Can Open Domain Question Answering Systems Answer Visual Knowledge Questions? arXiv:2202.04306. https://arxiv.org/abs/2202.04306
• Harel, R., Elboher, Y., & Pinter, Y. (2024). Protecting Privacy in Classifiers by Token Manipulation. PrivateNLP @ ACL 2024. https://arxiv.org/abs/2407.01334
• Rodrigues, R. C., Inuzuka, M. A., Gomes, J. R. S. A., et al. (2021). Zero-shot Hashtag Segmentation for Multilingual Sentiment Analysis. arXiv:2112.03213. https://arxiv.org/abs/2112.03213
• Li, J., Ren, M., Gao, Y., & Yang, Y. (2023). Ask to Understand: Question Generation for Multi-hop Question Answering. CCL 2023. https://arxiv.org/abs/2203.09073

Authors

• Simone Primarosa - simonepri

See also the list of contributors who participated in this project.

License

This project is licensed under the MIT License - see the license file for details.