Pre-processing and Training

[BettyHcZhang]

Thank your for sharing your code. I followed your steps to pre-processing and training, but I found Overall F1 score is 69, lower than paper report. So I think I must missed some important part in the experiment, could you give me some suggestions? Thank you again.

[taoshen58]

I've never encountered such a problem. I re-runed this code from scratch 2 months ago and got a similar result.

How did u search the full logical form? Due to there is a timeout parameter, too much burden on one CPU would decrease the successful searching rate and further affect the model training.

[BettyHcZhang]

I searched the logical form for training and dev, the success ratio is listed in the picture, maybe the figure here is already affected by the timeout parameter?

[taoshen58]

The rates are reasonable. It's weird to get only 69% overall score.

[BettyHcZhang]

I think your model architecture is reasonable and brilliant, and thank you very much for your quick reply.

[taoshen58]

Sorry for not helping. If you have any further questions, welcome asking anytime.

[sixlife]

The data link is invalid, can you help me send the data link? thank you