difference between implementation and LCF-paper

[yangheng95]

difference between implementation and LCF-paper

[fhamborg]

Not sure if relevant, but the implementation at https://github.com/songyouwei/ABSA-PyTorch/blob/master/models/lcf_bert.py is different to the implementation here and different to the description in the paper. It uses also only one self attention (compared to the three mentioned in the paper) but after applying linear_double (https://github.com/songyouwei/ABSA-PyTorch/blob/b925abcdf43c51e475a64ed04984d4911b88676d/models/lcf_bert.py#L116) whereas in this repository the self attention is applied before linear_double (https://github.com/yangheng95/LC-ABSA/blob/c945a94e0f86116c5578245aa9ad36c46c7b9c4a/models/lc_apc/lcf_bert.py#L56)

[yangheng95]

Thank you for your concern. I modified the model when I cleaned and refactored the code because I found that the model could achieve better results in some cases without adding the last MHSA, as well as reduce some computation. Since the paper has been published, we have merely access to revise some minor problems in the aforementioned paper.

[fhamborg]

Alright, thanks for letting me know. Since you mentioned the current implementation seems to be better than the one described in the paper, I'll also stick with the current implementation. FYI: Note the other issue opened in ABSA-PyTorch repository, where there is also a difference between their implementation and your implementation here.