Consider Adding a New Parse Model

[bjascob]

The amrlib/parse_T5 model scores 81 on AMR-3. There are two publicly available models that have slightly better performance..

• SapienzaNLP/SPRING
  • 83.0 on AMR-3 and 84.5 on AMR-2
  • Only works with transformers < 3 (transformers 4.12 is current release)
  • Has code and pretrained models models on GitHub
• IBM/Structrued-BART
  • 82.7 on AMR-3 and 84.7 on AMR-2
  • Doesn't use transformers (trains with FB fairseq lib)
  • Has code on GitHub but no pretrained model is publicly available
  • Currently code does not support stand-alone parsing.
  • Is a "transition parser". (Is this faster than the auto-regressive transformer style models?)
  • The paper Hierarchical Curriculum learning claims to be able to improve Bart (in SPRING model) by 1 point to 84.1 on AMR-3
  • The Instance Curricular portion of this has not proved to be useful in improving the T5 model's scores

Both of these models are based on BART-large which has roughly 2X the parameters of the T5-base model used in amrlib. This may cause issues training on older/smaller GPUs and could be slower for inference.

[bjascob]

Added model parse_spring based on the SapienzaNLP/SPRING code. This model achieves an 83.5 smatch, pre-wikification.

Revised the parse_t5 code to train and infer significantly faster and retrained the model to 16 epochs. This pushes up the smatch score to 81.9 pre-wikification. The new model is model_parse_t5-v0_2_0. The parse_t5 code is compatible with both the new and old models.