ncbi / bert_gt Goto Github PK

We have our dataset in which we have annotated the entities, relations, and coreferences. So how do we train a custom dataset such that it can be implemented using the bert-gt model?

Hello, can you please provide the specific virtual environment configuration, I am not sure if it is a version issue that is causing me to get strange errors when running your code, I hope you can provide the details of the environment configuration.

Hi @potinglai, thank you for making these models available! I am interested in running the bert_gt model on new PubMed abstracts for a class assignment to show the latest methods in relation extraction. I ran the model on the BioRED data using the steps in the README file but am not sure how I can use the trained model for processing new abstracts and showing the outputs. I also tried downloading the PubMedBERT model from the BioRED repository (https://github.com/ncbi/BioRED) but was not able to load with tensorflow successfully as the documentation isn't complete. Any help would be appreciated. Thank you!

Hello, how long have you run ‘biored_exp’ on theNVIDIA Tesla V100 SXM2 server?

Hello @ptlai !

Thank you so much for the avaibility of the code and the data !

While I was writing my own code for a custom dataset, I was trying to understand if the good results obtained by the nary dataset were obtained due to their dependency graph (obtained with Stanford) compared to the others (obtained with spacy), as it is the biggest differences with the other datasets, excluding the fact that it is also balanced.

However, I discovered that the nary dataset after preprocessing that the tagging part, where you replace the entities by their tag, does not look right.

For example, for the sentence :

“As compared to the large number of secondary resistance mutations noted in acquired imatinib resistance in CML , in the case of EGFR-TK , there are currently only several documented resistance point mutations to gefitinib and erlotinib , including T790M [ , ] , L747S [ ] and D761Y [ ] .”

where gefitinib and T790M are the drug and variant of interest, it should be transformed to

“As compared to the large number of secondary resistance mutations noted in acquired imatinib resistance in CML , in the case of EGFR-TK , there are currently only several documented resistance point mutations to @DRUG$ and erlotinib , including @VARIANT$ [ , ] , L747S [ ] and D761Y [ ] .”

However, in this case, in the data I have

“As compared to the large number of secondary resistance mutations noted in acquired imatinib resistance in CML , in the case of EGFR-TK , there are currently only several documented resistance point mutations to @DRUG$ gefitinib and erlotinib , including @VARIANT$ T790M [ , ] , L747S [ ] and D761Y [ ] .”

Is it normal ? I am afraid that this could be a source of leaked information during the learning process and could result in overfitting.

Thank you for your attention !

Hey, Could you please tell how to use bert_gt for predicting new data without the need for training. If possible please provide the sample input, output files, all version requirements, and pretrained_model and other related codes to run. Thank you

What information is depicted by a column after processing the CDR dataset is not clear. kindly explain

I am running only the run_cdr_exp.sh script with do_predict=true and train=false and eval=false. However, I notice that the prediction probabilities differ for the same data in each iteration. Could you please help me understand the reason behind this issue and suggest a way to resolve it? I would appreciate any help you can provide.

Hi there, thank you very much for making the code and data available! I am interested in testing the model with my own dataset for Nary cross-sentence RE task. I am wondering what are the functions in convert_nary_2_bert_gt focus on converting the dependency (arcs/edges) info that original dataset have into the bert_gt neighbors column content. or if you could elaborate the concept how to do such conversion, that would be much appreciated. Thank you in advance!