DFKI-NLP / sherlock

State-of-the-art Information Extraction
3 stars 1 forks source link

Binary RC Konfigs #53

Open leonhardhennig opened 2 years ago

leonhardhennig commented 2 years ago

Vorlage wäre das Relex-Projekt (der Vorgänger von Sherlock): https://github.com/DFKI-NLP/RelEx/tree/master/configs/relation_classification/tacred . Im Prinzip müssten wir die Konfigs nur übernehmen und ggf. anpassen (falls sich bei AllenNLP von 0.9.0 auf 2.8.0 was an den Konfigs geändert hat). die "relex/baseline_self_attention_tacred_bert.jsonnet" entspricht der Konfig "transformer.jsonnet"

leonhardhennig commented 2 years ago

in branch "config" baseline_boe.jsonnet. Getestet auf tacrev, ok.

TODO: Erweitern auf mehr Modelle wenn sinnvoll

phucdev commented 2 years ago

We will stick to using https://github.com/DFKI-NLP/sherlock/blob/master/scripts/cluster/binary_relation_clf_en.sh and perform some hyperparameter tuning to get a good enough model for the prediction

phucdev commented 2 years ago

I kept the batch size fixed at 32 and trained the model using different learning rates (2e-5, 3e-5, 4-5, 5e-5) for 5 epochs.

The best bert-base-uncased model with a learning rate of 4e-5 achieved a F1 score of 0.883 on the test split of the unionized relation dataset. The best roberta-base model with a learning rate of 4e-5 achieved a F1 score of 0.892 on the test split of the unionized relation dataset.