A question about adjusting reciprocal relations

[gxxxzhang]

Hello, may I ask why adjusting reciprocal relations in your work? And the function get_fixed_relation in you code.

[bigdante]

Hello, may I ask why adjusting reciprocal relations in your work? And the function get_fixed_relation in you code.

The reason is that the dataset is not fully annotated. For example, with the relation "has part": "part of", the original example might have only one annotated, missing the other, which can be supplemented. The function get_fixed is used to effectively correct errors in the model's output. For instance, for the label "has part", it might output "has parts". Although the likelihood is low, I do this to better match the labels.

Best~

[gxxxzhang]

Thank you for your answer. I would also like to ask whether training analysis_data can improve model performance and whether it is possible to provide an analysis_data version of the fine-tuning model.

[bigdante]

Thank you for your answer. I would also like to ask whether training analysis_data can improve model performance and whether it is possible to provide an analysis_data version of the fine-tuning model.

Unfortunately, I haven't achieved results that surpass the current approach. It could be due to some issues with my data or possibly not fine-tuning the parameters correctly. However, I find this method quite interesting. As for the checkpoint (ckpt), I am currently unable to provide it. Perhaps I will be able to provide it later.

Best~