Lately, propelled by the phenomenal advances around the transformer architecture, the legal NLP field has enjoyed spectacular growth. To measure progress, well curated and challenging benchmarks are crucial. However, most benchmarks are English only and in legal NLP specifically there is no multilingual benchmark available yet. Additionally, many benchmarks are saturated, with the best models clearly outperforming the best humans and achieving near perfect scores. We survey the legal NLP literature and select 11 datasets covering 24 languages, creating LEXTREME. To provide a fair comparison, we propose two aggregate scores, one based on the datasets and one on the languages. The best baseline ( XLM-R large) achieves both a dataset aggregate score a language aggregate score of 61.3. This indicates that LEXTREME is still very challenging and leaves ample room for improvement. To make it easy for researchers and practitioners to use, we release LEXTREME on huggingface together with all the code required to evaluate models and a public Weights and Biases project with all the runs.
The final LEXTREME score is computed using the harmonic mean of the dataset and the language aggregate score, thus weighing datasets and languages equally, promoting model fairness and robustness following Shavrina and Malykh (2021) and Chalkidis et al,.
We evaluated multilingual models as well as monolingual models. The multilingual models are the following:
| Model | Source | Parameters | Vocabulary Size | Pretraining Specs | Pretraining Corpora | Pretraining Languages |
|---|---|---|---|---|---|---|
| MiniLM | Wang et al. (2020) | 118M | 250K | 1M steps / BS 256 | 2.5TB CC100 | 100 |
| DistilBert | Sanh (2019) | 135M | 120K | BS up to 4000 | Wikipedia | 104 |
| mDeberta-v3 | He et al. (2020, 2021) | 278M | 128K | 500K steps / BS 8192 | 2.5TB CC100 | 100 |
| XLM-R base | Conneau et al. (2020) | 278M | 250K | 1.5M steps / BS 8192 | 2.5TB CC100 | 100 |
| XLM-R large | Conneau et al. (2020) | 560M | 250K | 1.5M steps / BS 8192 | 2.5TB CC100 | 100 |
| Legal-XLM-R-base | Niklaus et al. 2023 | 184M | 128K | 1M steps / BS 512 | 689GB MLP | 24 |
| Legal-XLM-R-large | Niklaus et al. 2023 | 435M | 128K | 500K steps / BS 512 | 689GB MLP | 24 |
| Legal-XLM-LF-base | Niklaus et al. 2023 | 208M | 128K | 50K steps / BS 512 | 689GB MLP | 24 |
| Legal-mono-R-base | Niklaus et al. 2023 | 111M | 32K | 200K steps / BS 512 | 689GB MLP | 1 |
| Legal-mono-R-large | Niklaus et al. 2023 | 337M | 32K | 500K steps / BS 512 | 689GB MLP | 1 |
In the following, we will provide the results on the basis of the multilingual models.
We compute the dataset aggregate score by taking the successive harmonic mean of (1.) the languages inside the configurations (e.g., de,fr,it within SJP), (2.) the configurations inside the datasets (e.g., OTS-UL, OTS-CT within OTS), and (3.) the datasets inside LEXTREME (BCD, GAM, etc.).
| Model | BCD | GAM | GLC | SJP | OTS | C19 | MEU | GLN | LNR | LNB | MAP | Agg. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MiniLM | 53.0 | 73.3 | 42.1 | 67.7 | 44.1 | 5.0 | 29.7 | 74.0 | 84.5 | 93.6 | 57.8 | 56.8 |
| DistilBERT | 54.5 | 69.5 | 62.8 | 66.8 | 56.1 | 25.9 | 36.4 | 71.0 | 85.3 | 89.6 | 60.8 | 61.7 |
| mDeBERTa-v3 | 60.2 | 71.3 | 52.2 | 69.1 | 66.5 | 29.7 | 37.4 | 73.3 | 85.1 | 94.8 | 67.2 | 64.3 |
| XLM-R-base | 63.5 | 72.0 | 57.4 | 69.3 | 67.8 | 26.4 | 33.3 | 74.6 | 85.8 | 94.1 | 62.0 | 64.2 |
| XLM-R-large | 58.7 | 73.1 | 57.4 | 69.0 | 75.0 | 29.0 | 42.2 | 74.1 | 85.0 | 95.3 | 68.0 | 66.1 |
| Legal-XLM-R-base | 62.5 | 72.4 | 68.9 | 70.2 | 70.8 | 30.7 | 38.6 | 73.6 | 84.1 | 94.1 | 69.2 | 66.8 |
| Legal-XLM-R-large | 63.3 | 73.9 | 59.3 | 70.1 | 74.9 | 34.6 | 39.7 | 73.1 | 83.9 | 94.6 | 67.3 | 66.8 |
| Legal-XLM-LF-base | 72.4 | 74.6 | 70.2 | 72.9 | 69.8 | 26.3 | 33.1 | 72.1 | 84.7 | 93.3 | 66.2 | 66.9 |
We compute the language aggregate score by taking the successive harmonic mean of (1.) the configurations inside the datasets, (2.) the datasets for the given language (e.g., MAP and MEU for lv), and (3.) the languages inside LEXTREME ( bg,cs, etc.).
| Model | bg | cs | da | de | el | en | es | et | fi | fr | ga | hr | hu | it | lt | lv | mt | nl | pl | pt | ro | sk | sl | sv | Agg. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MiniLM | 52.7 | 48.6 | 42.8 | 54.6 | 50.3 | 34.3 | 40.1 | 46.3 | 42.2 | 39.0 | 42.8 | 29.7 | 29.6 | 40.5 | 44.2 | 40.8 | 40.8 | 29.5 | 22.7 | 61.6 | 59.6 | 44.3 | 30.0 | 43.4 | 40.5 |
| DistilBERT | 54.2 | 48.6 | 46.0 | 60.1 | 58.8 | 48.0 | 50.0 | 48.8 | 49.6 | 47.9 | 51.4 | 35.9 | 31.2 | 50.1 | 51.9 | 41.5 | 44.4 | 34.6 | 34.5 | 63.2 | 63.8 | 51.3 | 36.2 | 50.1 | 46.7 |
| mDeBERTa-v3 | 54.1 | 51.3 | 51.7 | 63.6 | 57.7 | 50.7 | 53.3 | 50.8 | 54.6 | 49.2 | 54.9 | 37.4 | 37.5 | 55.1 | 53.9 | 47.0 | 52.5 | 42.1 | 41.0 | 65.7 | 65.3 | 55.4 | 37.5 | 56.1 | 50.5 |
| XLM-R-base | 56.4 | 48.3 | 48.3 | 60.6 | 57.6 | 50.1 | 47.2 | 46.7 | 48.6 | 49.4 | 50.1 | 33.6 | 32.8 | 53.4 | 50.0 | 44.1 | 43.8 | 35.2 | 41.3 | 66.1 | 63.7 | 45.3 | 33.7 | 50.0 | 47.1 |
| XLM-R-large | 59.9 | 56.0 | 56.3 | 65.4 | 60.8 | 56.2 | 56.6 | 56.5 | 56.9 | 51.4 | 55.4 | 42.5 | 38.1 | 58.5 | 58.1 | 49.9 | 53.9 | 39.5 | 46.4 | 68.6 | 66.8 | 57.9 | 42.4 | 59.1 | 53.7 |
| Legal-XLM-R-base | 55.6 | 58.8 | 50.4 | 63.6 | 63.7 | 66.8 | 56.3 | 57.0 | 52.6 | 50.1 | 56.6 | 38.7 | 56.5 | 56.1 | 57.2 | 49.1 | 56.0 | 41.6 | 43.9 | 68.2 | 66.1 | 55.6 | 38.6 | 54.9 | 53.5 |
| Legal-XLM-R-large | 57.8 | 55.6 | 50.4 | 65.7 | 60.7 | 69.3 | 55.7 | 54.5 | 56.6 | 53.3 | 57.2 | 39.7 | 39.1 | 58.1 | 60.6 | 48.4 | 57.2 | 39.4 | 45.5 | 67.3 | 65.5 | 49.3 | 39.7 | 56.4 | 53.6 |
| Legal-XLM-LF-base | 54.4 | 49.3 | 48.1 | 64.0 | 60.5 | 52.8 | 49.2 | 52.2 | 48.2 | 48.5 | 55.4 | 33.0 | 34.7 | 54.6 | 54.8 | 45.2 | 52.5 | 40.1 | 40.6 | 68.3 | 64.1 | 48.4 | 33.0 | 51.3 | 48.9 |
| NativeLegalBERT | - | - | - | - | - | 53.1 | 46.9 | - | - | - | - | - | - | 45.3 | - | - | - | - | - | - | 59.0 | - | - | - | 51.1 |
| NativeBERT | 54.8 | 57.3 | 51.2 | 63.0 | 62.3 | 52.0 | 42.6 | 47.2 | 52.4 | 49.4 | 50.1 | - | 37.4 | 47.1 | - | - | - | 37.0 | 40.5 | 66.5 | 63.1 | 44.8 | - | 55.1 | 50.2 |
| Legal-mono-R-base | 55.9 | 49.5 | 51.5 | 61.3 | 61.3 | 50.5 | 52.1 | 53.5 | 53.6 | 51.1 | 52.2 | 44.1 | 54.1 | 51.8 | 55.5 | 50.0 | 59.1 | 54.3 | 34.4 | 67.1 | 61.5 | 48.8 | 53.4 | 58 | 53.5 |
LEXTREME consist of three classification task types:
- Single Label Text Classification (SLTC)
- Multi Label Text Classification (MLTC)
- Named Entity Recognition (NER)
The dataset consists of 11 diverse multilingual legal NLU (natural language understanding) datasets. Six datasets have one single configuration and five datasets have two or three configurations. This leads to a total of 18 tasks (8 SLTC, 5 MLTC and 5 NER).
We use the existing train, validation, and test splits if present. In the other cases we split the data ourselves (80% train, 10% validation and test each).
For a detailed description of each task and dataset, see Niklaus et al. (2023). Datasets are abbreviated by three capital letters. Configurations of datasets, in case they exist, are indicated by an additional letter separated by a hyphen.
| Task | Type | Train / Dev / Test Examples | Train / Dev / Test Labels |
|---|---|---|---|
| BCD-J (brazilian_court_decisions_judgment) | SLTC (Judgment Prediction) | 3234 / 404 / 405 | 3 / 3 / 3 |
| BCD-U (brazilian_court_decisions_unan) | SLTC (Judgment Prediction) | 1715 / 211 / 204 | 2 / 2 / 2 |
| GAM (german_argument_mining) | SLTC (Argument Mining) | 19271 / 2726 / 3078 | 4 / 4 / 4 |
| GLC-V (greek_legal_code_volume) | SLTC (Topic Classification) | 28536 / 9511 / 9516 | 47 / 47 / 47 |
| GLC-C (greek_legal_code_chapter) | SLTC (Topic Classification) | 28536 / 9511 / 9516 | 386 / 377 / 374 |
| GLC-S (greek_legal_code_subject) | SLTC (Topic Classification) | 28536 / 9511 / 9516 | 2143 / 1679 / 1685 |
| SJP (swiss_judgment_prediction) | SLTC (Judgment Prediction) | 59709 / 8208 / 17357 | 2 / 2 / 2 |
| OTS-UL (online_terms_of_service_unfairness_levels) | SLTC (Unfairness Classification) | 2074 / 191 / 417 | 3 / 3 / 3 |
| OTS-CT (online_terms_of_service_clause_topics) | MLTC (Unfairness Classification) | 19942 / 1690 / 4297 | 9 / 8 / 9 |
| C19 (covid19_emergency_event) | MLTC (Event Classification) | 3312 / 418 / 418 | 8 / 8 / 8 |
| MEU-1 (multi_eurlex_level_1) | MLTC (Topic Classification) | 817239 / 112500 / 115000 | 21 / 21 / 21 |
| MEU-2 (multi_eurlex_level_2) | MLTC (Topic Classification) | 817239 / 112500 / 115000 | 127 / 126 / 127 |
| MEU-3 (multi_eurlex_level_3) | MLTC (Topic Classification) | 817239 / 112500 / 115000 | 500 / 454 / 465 |
| GLN (greek_legal_ner) | NER | 17699 / 4909 / 4017 | 17 / 17 / 17 |
| LNR (legalnero) | NER | 7552 / 966 / 907 | 11 / 9 / 11 |
| LNB (lener_br) | NER | 7828 / 1177 / 1390 | 13 / 13 / 13 |
| MAP-C (mapa_coarse) | NER | 27823 / 3354 / 10590 | 13 / 11 / 11 |
| MAP-F (mapa_fine) | NER | 27823 / 3354 / 10590 | 44 / 26 / 34 |
It works best with python 3.9 and torch==1.10.0+cu113. Otherwise, we experienced problems with fp16 training and evaluation.
# install torch like this to avoid fp16 problems
pip install torch==1.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txtIn case you get the
error AttributeError: module 'distutils' has no attribute 'version' (pytorch/pytorch#69894)
Run
pip install setuptools==59.5.0In order to log the training results, we used Weights & Biases. When running the script below, you will be asked if you want to use Weights & Biases or not. In case you want to use Weights & Biases too, you should log in to your Weights & Biases account beforehand, by typing the following command in the terminal:
wandb login {WANDB_API_KEY}
You can find WANDB_API_KEY in your profile setting on Weights & Biases after signing up or login.
We provide access to LEXTREME at https://huggingface.co/datasets/joelito/lextreme.
For example, to load the swiss_judgment_prediction (Niklaus et al. 2021) dataset, you first simply install the datasets' python library and then make the following call:
from datasets import load_dataset
dataset = load_dataset("joelito/lextreme", "swiss_judgment_prediction")It is possible to reproduce the results of the paper by running the finetuning for each dataset separately. Alternatively, you can run main.py which, in a nutshell, will generate bash scripts for each dataset with the necessary hyperparameters and run them on every available GPU in your system (if available).
The following command will make sure that you run most experiments as described in the paper:
python main.py
It allows a certain degree of customizability by specifying the following arguments:
| short argument name | full argument name | description | default value |
|---|---|---|---|
| -as | --accumulation_steps | Define the number of accumulation_steps. | Generated automatically depending on the batch size and the size of the pretrained model |
| -bz | --batch_size | Define the batch size. | Generated automatically depending on the size of the pretrained model |
| -gn | --gpu_number | Define which GPU you would like to use. If you want to specify multiple GPUs, seperate the integers by a comma. | Available GPUs are detected automatically. If no GPU is available, the CPU is used instead. |
| -gm | --gpu_memory | Define how much memory your GPUs have. Depending on that the batch size will be calculated automatically. In case the batch size is too big, you can change it manually with -bz. |
11 |
| -hier | --hierarchical | Define whether you want to use a hierarchical model or not. Caution: this will not work for every task. | Defined automatically depending on the dataset |
| -lang | --language | Define if you want to filter the training dataset by language. | all_all_all; only important for multlingual datasets; per default the entire dataset is used |
| -lc | --lower_case | Define if lower case or not. | False |
| -lmt | 50 | Define which kind of language model you would like to use; you can choose between small,base and large language models or all of them. | all_all_all = all pretrained language models as decribed in the paper |
| -los | --list_of_seeds | Define the number of training epochs. | Three seeds (1,2,3) are used |
| -lr | --learning_rate | Define the learning rate. | 1e-05 |
| -nte | --num_train_epochs | Define the number of training epochs. | 50 |
| -rmo | --running_mode | Define whether you want to run the finetungin on all available training data or just a small portion for testing purposes. | default = the entire dataset will be used for finetuning. The other option is experimental which will only take a small fraction of the dataset for experimental purposes. |
| -t | --task | Choose a specific task or all of them. | all. The other options are: brazilian_court_decisions_judgment, brazilian_court_decisions_unanimity, covid19_emergency_event, german_argument_mining, greek_legal_code_chapter_level, greek_legal_code_subject_level, greek_legal_code_volume_level, greek_legal_ner, legalnero, lener_br, mapa_ner_coarse_grained, mapa_ner_fine_grained, multi_eurlex_level_1, multi_eurlex_level_2, multi_eurlex_level_3, online_terms_of_service_unfairness_category, online_terms_of_service_unfairness_level, swiss_judgment_prediction |
| -dmo | --download_mode | Choose if you want to redownload the datasets or use load them from cache. | force_redownload. The other options are reuse_dataset_if_exists, reuse_cache_if_exists. For more information, visit: https://huggingface.co/docs/datasets/v1.4.1/loading_datasets.html. |
| -od | --output_dir | Specify the output directory for the logs. | Generated automatically with a time stamp. |
| -rev | -revision | The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a git-based system for storing models and other artifacts on huggingface.co, so revision can be any identifier allowed by git. | main |
For example, if you want to finetune on swiss_judgment_prediction with the seeds [1,2,3], 10 epochs, and all
pretrained language models as described in the paper, you can type the following:
python main.py --task swiss_judgment_prediction --list_of_seeds 1,2,3 --num_train_epochs 10
Temporary bash files will be created and saved in the
folder temporary_scripts and they will be run
immediately. These bash files will be overwritten the next time you run main.py.
If you want to finetune only on, let's say, xlm-roberta-large, you can type the following command.
python main.py --task swiss_judgment_prediction --list_of_seeds 1,2,3 --num_train_epochs 10 --language_model_type xlm-roberta-large
If, additionally, you don't want to make use of a hierarchical model (swiss_judgment_prediction makes use of
hierarchical models due to the length of the input documents), you type the following.
python main.py --task swiss_judgment_prediction --list_of_seeds 1,2,3 --num_train_epochs 10 --language_model_type xlm-roberta-large --hierarchical False
Not all tasks support the use of hierarchical types. For example, the code for the named entity recognition tasks has
not been optimized to make use of both the non-hierarchical and the hierarchical variants. Therefore,
setting -hierarchical to True will cause an error.
In case you want to perform hyperparameter search this is possible via the argument do_hyperparameter_search. The
values for metric_for_best_model (and accordingly for greater_is_better,
see the huggingface documentation)
will stay the same, i.e., the hyperparameters will be searched by searching for the lowest evaluation loss. In the
following, we provide the command for hyperparameter search for the finetuning task german_argument_mining with the
model distilbert-base-multilingual-cased.
python main.py -gn 1 -gm 80 --task german_argument_mining --do_hyperparameter_search True -lmt distilbert-base-multilingual-cased -ld hyperparameter_tuning
Automatically, you will create a new project in wandb which is the same as the name of the logging directory, in the
case above hyperparameter_tuning. The runs in wandb will be named according to this pattern: finetuning task
__num_train_epochs_X__weight_decay_X__batch_size_X__seed_X__learning_rate_X__num_train_epochs_actually_trained_X.Xin this context means the actual values of hyperparameters;num_train_epochs_actually_traineddepicts the actual number of training epochs, because the training will stop earlier if no improvements are to be detected during training.
If you want to change the value for metric_for_best_model, add it to the bash command like this:
python main.py -gn 1 -gm 80 --task german_argument_mining --do_hyperparameter_search True -lmt distilbert-base-multilingual-cased -ld hyperparameter_tuning --metric_for_best_model macro-f1
Automatically, greater_is_better will change to true.
You can choose between three types of search methods, i.e., grid, random, bayes (see
the wandb documentation). Per default, you will
use grid. You can change this by passing search_type_method to the command, like this:
python main.py -gn 1 -gm 80 --task german_argument_mining --do_hyperparameter_search True -lmt distilbert-base-multilingual-cased -ld hyperparameter_tuning --metric_for_best_model macro-f1 --search_type_method bayes
The hyperparamters to search for will be loaded from the json
file hyperparameter_search_config.json
.
Note that when using the search type method grid you will not be able to use learning rate as an hyperparamter to
tune; wandb would through this error: "Invalid sweep config: Parameter learning_rate is a disallowed type with grid
search. Grid search requires all parameters to be categorical, constant, int_uniform, or q_uniform. Specification of
probabilities for categorical parameters is disallowed in grid search". Therefore, if grid is chosen, the values for
learning_rate
in hyperparameter_search_config.json
are ignored; instead the default learning rate (1e-5) or the one that you specify will be taken.
If you want to extend the benchmark with your own datasets, you can do so by following the following instructions. Once these steps are finished, make a merge request, and we merge the changes into the main branch.
- Make sure your dataset is available on the huggingface hub and has a train, validation and test split.
- Make sure that the structure of your dataset is in compliance with the other datasets of LEXTREME.
- Create a pull request to the lextreme repository by adding the following to the LEXTREME.py file:
- Create a dict _{YOUR_DATASET_NAME} (similar to _BRAZILIAN_COURT_DECISIONS_JUDGMENT) containing all the necessary information about your dataset (task_type, input_col, label_col, etc.)
- Add your dataset to the BUILDER*CONFIGS
list:
LextremeConfig(name="{your_dataset_name}", \*\**{YOUR_DATASET_NAME}) - Test that it works correctly by loading your subset with
load_dataset("lextreme", "{your_dataset_name}")and inspecting a few examples.
The following instructions will suffice only if
- your dataset is in compliance with the other datasets of LEXTREME and
- the tasks of your dataset belong to these classes:
token classification,single-label text classification,multi-label text classification.
- Navigate to the folder
utilsand open the filemeta_infos.json. - The file contains several fields with important information about each dataset and finetuning task. Some of this information is essential to run the code. The fields are:
dataset_jurisdiction: Not important for the code. Nevertheless, important to assess the jurisdictional coverage of LEXTREME.dataset_abbreviations: Not important for the code. Nevertheless, important to add the results of the finetuning to the existing tables.task_abbreviations: Not important for the code. Nevertheless, important to add the results of the finetuning to the existing tables.task_type_mapping: Important for the code. Specify to which type of task your dataset, e.g. the respective finetuning task, belongs to. Choose one of the following abbreviations:- NER (token classification/ named entity recognition)
- SLTC (single-label text classification)
- MLTC (multi-label text classification)
task_language_mapping: Important for the code. Provide a list of languages that your finetuning task covers. Use only two-letter lowercase abbreviation. You can find an overview here.config_to_dataset: Not important for the code. Nevertheless, this information is useful.dataset_to_config: Not important for the code. Nevertheless, this information is useful.task_default_arguments: Important for the code. Here, you can define the arguments that will be served for finetuning. Have a look at the existing examples. Essentially, what you need to provide ismax_seq_lengthandhierarchical.max_segmentsandmax_seg_lengthare only needed ifhierarchicalis set totrue.language_models: Important for the code. If your dataset covers a new language, you might want to add a new monolingual language model for that language. Provide the name as given on huggingface.
Navigate to the utils folder and open the file meta_infos.json. The key field 'language_models' contains all the language models available for use. If your language model is a general-purpose one, proceed to the key field 'general'. If it is a language model pre-trained on legal data, navigate to the key field 'legal'. Next, specify the language(s) on which your model has been trained, such as 'en', 'de', 'multilingual', etc. Finally, determine the category of your model, i.e., whether it is 'small', 'base', or 'large'. Insert the Hugging Face name of your model into the appropriate field. Remember to save the changes to the JSON file.
Please cite the following preprint:
@misc{niklaus2023lextreme,
title={LEXTREME: A Multi-Lingual and Multi-Task Benchmark for the Legal Domain},
author={Joel Niklaus and Veton Matoshi and Pooja Rani and Andrea Galassi and Matthias Stürmer and Ilias Chalkidis},
year={2023},
eprint={2301.13126},
archivePrefix={arXiv},
primaryClass={cs.CL}
}