Edit model card
YAML Metadata Warning: empty or missing yaml metadata in repo card (https://huggingface.co/docs/hub/model-cards#model-card-metadata)

RoBERTa2RoBERTa temporal tagger

Seq2seq model for temporal tagging of plain text using RoBERTa language model. The model is introduced in the paper BERT got a Date: Introducing Transformers to Temporal Tagging and release in this repository.

Model description

RoBERTa is a transformers model pretrained on a large corpus of English data in a self-supervised fashion. We use RoBERTa in an encoder-decoder architecture for text generation, where the input is raw text and the output is the temporally annotated text. The model is pre-trained on a weakly annotated dataset from a rule-based system (HeidelTime) and fine-tuned on the temporal benchmark datasets (Wikiwars, Tweets, Tempeval-3).

Intended uses & limitations

This model is best used accompanied with code from the repository. Especially for inference, the direct output might be noisy and hard to decipher, in the repository we provide cleaning functions for the output and insert the temporal tags from the generated text in the input text. If you have temporally annotated data you can fine-tune this model.

How to use

you can load the model as follows:

tokenizer = AutoTokenizer.from_pretrained("satyaalmasian/temporal_tagger_roberta2roberta")
model = EncoderDecoderModel.from_pretrained("satyaalmasian/temporal_tagger_roberta2roberta")

for inference use:

model_inputs = tokenizer(input_text, truncation=True, return_tensors="pt")
out = model.generate(**model_inputs)
decoded_preds = tokenizer.batch_decode(out, skip_special_tokens=True)

for an example with post-processing, refer to the repository. to further fine-tune, use the Seq2SeqTrainer from hugginface. An example of a similar fine-tuning can be found here.

trainer = Seq2SeqTrainer(


where the training_args is an instance of Seq2SeqTrainingArguments. #Training data We use four data sources: For Pretraining :1 million weakly annotated samples from heideltime. The samples are from news articles between the 1st January 2019 and the 30th July. Fine-tunning: Tempeval-3, Wikiwars, Tweets datasets. For the correct data versions please refer to our repository.

#Training procedure The model is pre-trained on the weakly labeled data for $3$ epochs on the train set, from publicly available checkpoints on huggingface (roberta-base), with a batch size of 12. We use a learning rate of 5e-05 with an Adam optimizer and linear weight decay. Additionally, we use 2000 warmup steps. We fine-tune the 3 benchmark data for 8 epochs with 5 different random seeds, this version of the model is the only seed=4. The batch size and the learning rate is the same as the pre-training setup, but the warm-up steps are reduced to 100. For training, we use 2 NVIDIA A100 GPUs with 40GB of memory. For inference in seq2seq models, we use Greedy decoding, since beam search had sub-optimal results.

Downloads last month
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.