bpe_knockout 2026.3.24

Implementation of BPE-knockout, a morphologically informed post-processing step for BPE tokenisers.

Repo hosting the implementation of BPE-knockout and ReBPE tokenisation, as well as the intrinsic evaluations for the NAACL 2024 paper "BPE-knockout: Pruning Pre-existing BPE Tokenisers with Backwards-compatible Morphological Semi-supervision".

(Extrinsic evaluations for that paper were done with RobBERT's framework, and the pre-trained model checkpoints are available on the HuggingFace Hub.)

Table of contents:

• Installation
• Usage:
  • HuggingFace compatibility
  • Saving and loading
• Paper results:
  • Running experiments from the paper
  • Using your own data
• Data licenses
• BibTeX citation

HuggingFace compatibility

Are you used to working with the HuggingFace suite for language modelling and tokenisation? No problem! You can incorporate BPE-knockout anywhere you're already using a BPE tokeniser loaded from HuggingFace, with only 2 extra imports and 2 more lines of code. For example, if you're using roberta-base's English tokeniser, you would run:

# Load HuggingFace object
from transformers import AutoTokenizer
hf_bpe_tokeniser = AutoTokenizer.from_pretrained("roberta-base")

# Construct TkTkT object  FIXME: No longer works because string-based loading is banned in TkTkT.
from tktkt.models.bpe.knockout import BPEKnockout
tktkt_bpek_tokeniser = BPEKnockout.fromHuggingFace(hf_bpe_tokeniser, "English")

# Convert back to HuggingFace
from tktkt.interfaces.huggingface import TktktToHuggingFace
hf_bpek_tokeniser = TktktToHuggingFace(tktkt_bpek_tokeniser, specials_from=hf_bpe_tokeniser)

The resulting object is indeed a HuggingFace tokeniser, but internally it works using BPE-knockout.

Installing

Minimal package

If you are only interested in using the BPE-knockout and ReBPE tokenisers, I recommend you to just install the TkTkT package. It has this package as a dependency.

Warning: If you do decide to manually install this package, don't forget to add the [all] suffix (pip install bpe_knockout[all]) in case you don't have an installation for any of my other packages.

Reproducing the paper

If you want to run experiments from the paper (and/or have access to the word count files), this means you want to download everything in this repository, and tell Python to use the folder into which you cloned for the package code, rather than copying the code to your global or virtual site-packages directory.

The once-working experiments run for the camera-ready version of the paper can be recovered from commit c08a7. For this purpose, run:

git clone https://github.com/bauwenst/BPE-knockout
cd BPE-knockout
git checkout c09a7
pip install -e .[all]

Warning:

• If you're using conda or venv, don't forget to activate your environment before running any calls to pip install.
• If you have an editable installation of my other packages TkTkT and/or Fiject and would like to keep it, do not include the [all] suffix.

More usage examples

Saving and loading tokeniser after knockout

Although knockout takes under 3 minutes to complete and lasts as long as the tokeniser's lifetime, you may still want to cache the resulting BPE-knockout tokeniser for using it again without re-doing knockout. Starting from a HuggingFace tokeniser, that would look like this:

from pathlib import Path
from transformers import AutoTokenizer
from tktkt.models.bpe.knockout import BPEKnockout
from tktkt.models.huggingface.wrapper import HuggingFacePreprocessor

# Load old tokeniser and apply knockout. FIXME: No longer works.
hf_base = AutoTokenizer.from_pretrained("roberta-base")
tktkt_knockout = BPEKnockout.fromHuggingFace(hf_base, language="English")

# Save new tokeniser and reload it.
save_path = tktkt_knockout.save(folder=Path(".") / "roberta-knockout")
tktkt_knockout_loaded = BPEKnockout.load(save_path, preprocessor=HuggingFacePreprocessor(hf_base))

The objects tktkt_knockout and tktkt_knockout_loaded have the exact same internals. (If you want to give them the HuggingFace interface, wrap them with a call to TktktToHuggingFace.)

Notice how .load() requires a preprocessor. That's because .save() does not store the preprocessor with the tokeniser, as it assumes that you already have a code snippet that loads the preprocessor. In this case, because the original tokeniser was a HuggingFace tokeniser, its preprocessor was a HuggingFace preprocessor as well.

Loading from the HuggingFace hub

The BPE-knockout tokeniser used for the (continued) pre-training of language models in the paper were saved as above and uploaded to the HuggingFace hub. Rather than downloading them manually, you can load them with a from_pretrained call. For convenience, there is one that has the TkTkT interface and one that converts it to the HuggingFace interface:

from tktkt.models.bpe.knockout import BPEKnockout
# FIXME: No longer works.
dutch_bpe_knockout_hf    = BPEKnockout.from_pretrained("Bauwens/RoBERTa-nl_BPE_30k_BPE-knockout_9k")
dutch_bpe_knockout_tktkt = BPEKnockout.from_pretrained_tktkt("Bauwens/RoBERTa-nl_BPE_30k_BPE-knockout_9k")

Running experiments from the paper

Given that you have an editable install, follow these steps to reproduce the paper results:

1. Unzip the .rar file under data/compressed/.
2. Run py tst/main.py or python tst/main.py in a terminal.

Using your own data

It is possible to use other datasets (even other languages) than the ones used for the paper. All files read by the package are declared in a globally accessible ProjectConfig object. We ship three default ProjectConfigs with the package for English, German and Dutch (see the bpe_knockout.project.config file).

To add a config, you'll need the following files:

1. A dataset of morphological decompositions. BPE-knockout is built on top of MoDeST for supplying morphological data, so your data format must be compatible with those supported by MoDeST. (Otherwise, you can always write your own class.)
2. Optional: if you want to run all experiments, you also need a tab-separated file of words and their frequencies from a sufficiently large corpus (morphologies without such a frequency will get frequency 1, and words that don't have a morphology will be ignored);
3. Optional: if you don't want to generate a new BPE tokeniser from your word counts, the file(s) that specify your existing BPE tokeniser.

Now write a function akin to setupEnglish() to return your new config. If you want to run the experiments on these new data, import it in tst/main.py. If you want to just use these new data for knockout, import your function in tktkt.models.bpe.knockout and add it to the CONFIGS dictionary.

Data licenses

All data is included in the repo, because it is obtainable for free elsewhere and free of license too.

• Morphological decompositions were derived from WebCelex at the Max Plank Institute.
• Language modelling data is derived from OSCAR on HuggingFace.

Citation

If you use BPE-knockout in your own work, cite the paper using e.g.:

@inproceedings{bauwens-delobelle-2024-bpe,
    title = "{BPE}-knockout: Pruning Pre-existing {BPE} Tokenisers with Backwards-compatible Morphological Semi-supervision",
    author = "Bauwens, Thomas  and  Delobelle, Pieter",
    booktitle = "Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = jun,
    year = "2024",
    address = "Mexico City, Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.naacl-long.324",
    pages = "5810--5832"
}