How to get .source and .target file at comet_atomic2020_bart

[yongho94]

Hello sir.

I tried to run your codes that use BART model to generate knowledge triples.

In your codes, "models/comet_atomic2020_bart/finetune.py" requires "train.source" file and "train.target" file...

However, I couldn't figure out how to get these files.

How can I get these files?

Thanks.

[RubenBranco]

Hi @yongho94,

I'm not one of the authors but I might be able to help here. The code expects a .source .target format that used to be the standard format for huggingface libraries before datasets came about. Here's the example page: https://github.com/huggingface/transformers/tree/master/examples/legacy/seq2seq

To produce this for comet, you iterate over the csv file and for each row you concatenate the head with the relation "{head} {rel}" and write to a "train.source" file, and then the tail is written to a train.target file, such that each row in the files correspond to each other.

Might be wrong though.

[yongho94]

Hi @yongho94,

I'm not one of the authors but I might be able to help here. The code expects a .source .target format that used to be the standard format for huggingface libraries before datasets came about. Here's the example page: https://github.com/huggingface/transformers/tree/master/examples/legacy/seq2seq

To produce this for comet, you iterate over the csv file and for each row you concatenate the head with the relation "{head} {rel}" and write to a "train.source" file, and then the tail is written to a train.target file, such that each row in the files correspond to each other.

Might be wrong though.

Thanks @RubenBranco !!

It seems to be work.

I think i need to try it.

Thanks.

[keisks]

Hi @yongho94,

Thank you for your question. Regarding the data format for BART, @RubenBranco is correct.

The src and trg dataset (for BART) is available here. If you are also interested in the model we trained, you can get it from here

I hope this helps!

[Kelaxon]

@keisks, Sorry for re-opening this, and thanks for the fantastic work.

I have another question regarding the data format:

I saw there are some "none" targets in the training, validating, and testing set. Why do you introduce them? Are they used to prevent over-fitting? If so, how do you determine the ratio and sampling method?

Thanks!

[keisks]

The none targets mean that annotators answered there are no tails for given head and relation. In the dataset, we include all the annotations (=no sampling). As you can see, they are sometimes redundant because multiple annotators give same answers.

[Kelaxon]

Thanks for the explaination👍🏻