songyouwei
different LCF-BERT implementations
The implementation of LCF-BERT in this repository is different to both the description of LCF-BERT in the paper (https://www.mdpi.com/2076-3417/9/16/3389/htm) as well as different to the original implementation of LCF-BERT by the paper's author (https://github.com/yangheng95/LC-ABSA/blob/master/models/lc_apc/lcf_bert.py).
Specifically, the paper mentions three self attentions (see paper Figure 3), whereas the implementation applies only one self attention (to the merged outputs, called "feature interactive learning layer" in the paper), but not to the global bert output and also not to the local bert output (see paper Figure 3). In comparison to the original LCF-BERT implementation, there is also a difference, i.e., there, the self attention (also only one instead of three somehow) is applied to the local bert output.
@yangheng95 fyi, this issue here mentions in addition to https://github.com/yangheng95/LC-ABSA/issues/10 that there is also a difference between LCF-BERT implementation in ABSA-PyTorch compared to LC-ABSA repositories (specifically, where the single self attention is placed is different when comparing the two repositories)
This issue mentions the same problem as #128. BTW, the latest code for LCF-BERT is tracked at LC-ABSA. Some of the structures changed in LC-ABSA so the latest codes are not mergeable.
