newly trained parse_xfm_bart models not generating first character

[bjascob]

For newly trained parse_xfm models using bart-X, when looking at the raw graph output the first character is missing. This does not impact the released models and I've only confirmed it on bart-base. It likely exists with bart-large but not with the t5 models.

The issue is due to the fact that the huggingface model config.json file has changed. There is now a line "forced_bos_token_id": 0 present. This changes the behavior. The change to the model config happened around the end of Feburary 2022. It looks like it is from this https://github.com/huggingface/transformers/issues/15559 issue.

Adding the line {..., "forced_bos_token_id" : null} to the config's model_args section appears to fix this. The fix needs to be tested with both bart models and t5-base needs to be verified to work correctly without any changes.

Note that it's not obvious that this is happening because the first character is always a "(" for all graphs and the deserializer can generally handle the missing start paren without issue. This means the bug may not introduce any visible changes, but could cause the smatch score to change a very small amount.

[bjascob]

Bart configs updated in commit https://github.com/bjascob/amrlib/commit/f25fe16998f1c14879e793a225c0174f4d7ac4f3. T5-base works with no changes. Bart-x and T5 re-trained models yield the following (new models were not uploaded)...

Training Results

All models trained with transformers 4.20.0 on 6/24/2022 Scores are without :wiki tags

• model_parse_xfm_t5_base: raw model output correctly begins with a paren (ie.. no missing first character) Epoch 16: (dev, beams=1) SMATCH -> P: 0.844, R: 0.818, F: 0.831 Epoch 16: (test, beams=4) SMATCH -> P: 0.835, R: 0.803, F: 0.819 parse_t5-v0.2.0 is scored at 81.9 (test, beams=4) but this is the older parse code/config params
• model_parse_xfm_bart_base: raw model output correctly begins with a paren (ie.. no missing first character) Epoch 16: (dev, beams=1) SMATCH -> P: 0.843, R: 0.818, F: 0.830 Epoch 16: (test, beams=4) SMATCH -> P: 0.833, R: 0.810, F: 0.821 parse_xfm_bart_base-v0.1.0 is scored at 82.3 (test, beams=4)
• model_parse_xfm_bart_large: raw model output correctly begins with a paren (ie.. no missing first character) Epoch 16: (dev, beams=1) SMATCH -> P: 0.856, R: 0.841, F: 0.848 Epoch 16: (test, beams=4) SMATCH -> P: 0.844, R: 0.832, F: 0.838 parse_xfm_bart_large-v0.1.0 is scored at 83.7 (test, beams=4)