BERT Got a Date: Introducing Transformers to Temporal Tagging
Satya Almasian*, Dennis Aumiller*, and Michael Gertz
Heidelberg University
Contact us via: <lastname>@informatik.uni-heidelberg.de
Code and data for the paper BERT Got a Date: Introducing Transformers to Temporal Tagging. Temporal tagging is the task of identification of temporal mentions in text; these expressions can be further divided into different type categories, which is what we refer to as expression (type) classification. This repository describes two different types of transformer-based temporal taggers, which are both additionally capable of expression classification. We follow the TIMEX3 schema definitions in their styling and expression classes (notably, the latter are one of TIME, DATE, SET, DURATION
). The available data sources for temporal tagging are in the TimeML format, which is essentially a form of XML with tags encapsulating temporal expressions.
An example can be seen below:
Due to lockdown restrictions, 2020 might go down as the worst economic year in over <TIMEX3 tid="t2" type="DURATION" value="P1DE">a decade</TIMEX3>.
For more data instances, look at the content of data.zip
. Refer to the README file in the respective unzipped folder for more information.
This repository contains code for data preparation and training of a seq2seq model (encoder-decoder architectured initialized from encoder-only architectures, specifically BERT or RoBERTa), as well as three token classification encoders (BERT-based).
The output of the models discussed in the paper is in the results
folder. Refer to the README file in the folder for more information.
Data Preparation
The scripts to generate training data is in the subfolder data_preparation. For more usage information, refer to the README file in the subfolder. The data used for training and evaluation are provided in zipped form in data.zip
.
Evaluation
For evaluation, we use a slightly modified version of the TempEval-3 evaluation toolkit (original source here). We refactored the code to be compatible with Python3, and incorporated additional evaluation metrics, such as a confusion matrix for type classification. We cross-referenced results to ensure full backward-compatibility and all runs result in the exact same results for both versions. Our adjusted code, as well as scripts to convert the output of transformer-based tagging models are in the evaluation subfolder. For more usage information, refer to the README file in the respective subfolder.
Temporal models
We train and evaluate two types of setups for joint temporal tagging and classification:
- Token Classification: We define three variants of simple token classifiers; all of them are based on Huggingface's BertForTokenClassification. We adapt their "token classification for named entity recognition script" to train these models. All the models are trained using
bert-base-uncased
as their pre-trained checkpoint. - Text-to-Text Generation (Seq2Seq): These models are encoder-decoder architectures using BERT or RoBERTa for initial weights. We use Huggingface's EncoderDecoder class for initialization of weights, starting from
bert-base-uncased
androberta-base
, respectively.
Seq2seq
To train the seq2seq models, use run_seq2seq_bert_roberta.py
. Example usage is as follows:
python3 run_seq2seq_bert_roberta.py --model_name roberta-base --pre_train True \
--model_dir ./test --train_data ./data/seq2seq/train/tempeval_train.json \
--eval_data ./data/seq2seq/test/tempeval_test.json --num_gpu 2 --num_train_epochs 1 \
warmup_steps 100 --seed 0 --eval_steps 200
Which trains a roberta2roberta model defined by model_name
for num_train_epochs
epochs on the gpu with ID num_gpu
. The random seed is set by seed
and the number of warmup steps by warmup_steps
. Train data should be specified in train_data
and model_dir
defines where the model is saved. set eval_data
if you want intermediate evaluation defined by eval_steps
. If the pre_train
flag is set to true it will load the checkpoints from the hugginface hub and fine-tune on the dataset given. If the pre_train
is false, we are in the fine-tuning mode and you can provide the path to the pre-trained model with pretrain_path
. We used the pre_train
mode to train on weakly labeled data provided by the rule-based system of HeidelTime and set the pre_train
to false for fine-tunning on the benchmark datasets. If you wish to simply fine-tune the benchmark datasets using the huggingface checkpoints you can set the pre_train
to ture, as displayed in the example above. For additional arguments such as length penalty, the number of beams, early stopping, and other model specifications, please refer to the script.
Token Classifiers
As mentioned above all token classifiers are trained using an adaptation of the NER script from hugginface. To train these models use
run_token_classifier.py
like the following example:
python3 run_token_classifier.py --data_dir /data/temporal/BIO/wikiwars \
--labels ./data/temporal/BIO/train_staging/labels.txt \
--model_name_or_path bert-base-uncased \
--output_dir ./fine_tune_wikiwars/bert_tagging_with_date_no_pretrain_8epochs/bert_tagging_with_date_layer_seed_19 --max_seq_length 512 \
--num_train_epochs 8 --per_device_train_batch_size 34 --save_steps 3000 --logging_steps 300 --eval_steps 3000 \
--do_train --do_eval --overwrite_output_dir --seed 19 --model_date_extra_layer
We used bert-base-uncased
as the base of all our token classification models for pre-training as defined by model_name_or_path
. For fine-tuning on the datasets model_name_or_path
should point to the path of the pre-trained model. labels
file is created during data preparation for more information refer to the subfolder. data_dir
points to a folder that contains train.txt
, test.txt
and dev.txt
and output_dir
points to the saving location. You can define the number of epochs by num_train_epochs
, set the seed with seed
and batch size on each GPU with per_device_train_batch_size
. For more information on the parameters refer to the Hugginface script. In our paper, we introduce 3 variants of token classification, which are defined by flags in the script. If no flag is set the model trains the vanilla BERT for token classification. The flag model_date_extra_layer
trains the model with an extra date layer and model_crf
adds the extra crf layer. To train the extra date embedding you need to download the vocabulary file and specify its path in date_vocab
argument. The description and model definition of the BERT variants are in folder temporal_models. Please refer to the README file for further information. For training different model types on the same data, make sure to remove the cached dataset, since the feature generation is different for each model type.
Load directly from the Huggingface Model Hub
We uploaded our best-performing version of each architecture to the Huggingface Model Hub. The weights for the other four seeding runs are available upon request. We upload the variants that were fine-tuned on the concatenation of all three evaluation sets for better generalization to various domains. Token classification models are variants without pre-training. Both seq2seq models are pretrained on the weakly labled corpus and fine-tuned on the mixed data.
Overall we upload the following five models. For other model configurations and checkpoints please get in contact with us:
- satyaalmasian/temporal_tagger_roberta2roberta: Our best perfoming model from the paper, an encoder-decoder architecture using RoBERTa. The model is pre-trained on weakly labeled news articles, tagged with HeidelTime, and fined-tuned on the train set of TempEval-3, Tweets, and Wikiwars.
- satyaalmasian/temporal_tagger_bert2bert: Our second seq2seq model , an encoder-decoder architecture using BERT. The model is pre-trained on weakly labeled news articles, tagged with HeidelTime, and fined-tuned on the train set of TempEval-3, Tweets, and Wikiwars.
- satyaalmasian/temporal_tagger_BERT_tokenclassifier: BERT for token classification model or vanilla BERT model from the paper. This model is only trained on the train set of TempEval-3, Tweets, and Wikiwars.
- satyaalmasian/temporal_tagger_DATEBERT_tokenclassifier: BERT for token classification with an extra date embedding, that encodes the reference date of the document. If the document does not have a reference date, it is best to avoid this model. Moreover, since the architecture is a modification of a default hugginface model, the usage is not as straightforward and requires the classes defined in the
temporal_model
module. This model is only trained on the train set of TempEval-3, Tweets, and Wikiwars. - satyaalmasian/temporal_tagger_BERTCRF_tokenclassifier :BERT for token classification with a CRF layer on the output. Moreover, since the architecture is a modification of a default huggingface model, the usage is not as straightforward and requires the classes defined in the
temporal_model
module. This model is only trained on the train set of TempEval-3, Tweets, and Wikiwars.
In the examples
module, you find two scripts model_hub_seq2seq_examples.py
and model_hub_tokenclassifiers_examples.py
for seq2seq and token classification examples using the hugginface model hub. The examples load the models and use them on example sentences for tagging. The seq2seq example uses the pre-defined post-processing from the tempeval evaluation and contains rules for the cases we came across in the benchmark dataset. If you plan to use these models on new data, it is best to observe the raw output of the first few samples to detect possible format problems that are easily fixable. Further fine-tuning of the models is also possible. For seq2seq models you can simply load the models with
tokenizer = AutoTokenizer.from_pretrained("satyaalmasian/temporal_tagger_roberta2roberta")
model = EncoderDecoderModel.from_pretrained("satyaalmasian/temporal_tagger_roberta2roberta")
and use the DataProcessor
from temporal_models.seq2seq_utils
to preprocess the json
dataset. The model can be fine-tuned using Seq2SeqTrainer
(same as in run_seq2seq_bert_roberta.py
). For token classifiers the model and the tokenizers are loaded as follows:
tokenizer = AutoTokenizer.from_pretrained("satyaalmasian/temporal_tagger_BERT_tokenclassifier", use_fast=False)
model = BertForTokenClassification.from_pretrained("satyaalmasian/temporal_tagger_BERT_tokenclassifier")
Classifiers need a BIO-tagged file that can be loaded using TokenClassificationDataset
and fine-tuned with the hugginface Trainer
. For more information on the usage of these models refer to their model hub page.
Citation
If you use our models in your work, we would appreciate attribution with the following citation:
@article{almasian2021bert,
title={{BERT got a Date: Introducing Transformers to Temporal Tagging}},
author={Almasian, Satya and Aumiller, Dennis and Gertz, Michael},
journal={arXiv},
year={2021}
}