search

yangheng95 / PyABSA

Sentiment Analysis, Text Classification, Text Augmentation, Text Adversarial defense, etc.;
https://pyabsa.readthedocs.io
MIT License
923 stars 159 forks source link

Not Predicting Negative Label #139

Closed anilkumar9912 closed 2 years ago

anilkumar9912 commented 2 years ago

Hi @yangheng95 , I'm trying out the atepc on English checkpoint. pyabsa version: 1.8.30 Used checkpoint: fast_lcf_atepc_English_cdw_apcacc_80.16_apcf1_78.34_atef1_75.39.zip The model is predicting Positive and Neutral sentiments only even when there are negative aspects. I've performed inference on Tshirt and SemEval dataset but I did not find any negative polarity tag in any of the examples. Can you please explain how to get the three classes?(Positive,Negative and Neutral) Thank you.

yangheng95 commented 2 years ago

Thanks for your report, I confirm there is a problem. However, I am not available for fixing it right now, maybe later. And you can tell me if you find the reason for this error to fix it.

anilkumar9912 commented 2 years ago

My initial guess is to check the number of classes the checkpoint is trained on I checked the config of the trained checkpoint fast_lcf_atepc_English_cdw_apcacc_80.16_apcf1_78.34_atef1_75.39.zip for polarity classes: it's given 3. So it's a 3 class (0:Negative,1:Neutral,2:Positive) classifiation. While inferring, Class 0 is neglected. Maybe there's an issue while decoding softmax scores to class labels. Where's this decoding is happening? (mapping scores to class labels).

yangheng95 commented 2 years ago

Sorry for late reply, please try 1.8.32 which may fixes this problem

anilkumar9912 commented 2 years ago

Thanks @yangheng95 .It works with the new 1.8.32 version.