amr-reentrancies

Scripts for the analysis of reentrancies in AMR discussed in [1].

Causes of reentrancies

Given an AMR corpus (<amr_file>), follow the instructions to quantify the main causes of reentrancies in your data:

• Preprocess the corpus with the preprocessing scripts of AMREager (https://github.com/mdtux89/amr-eager-multilingual):

  ./preprocessing.sh -f <amr_file> 
  python preprocessing.py --amrs -f <amr_file>
  

• Compute the coreferences in the data with coref.py. It requires python3 and neuralcoref (https://github.com/huggingface/neuralcoref):

  python3 coref.py <amr_file>
  

• Run the reen_stats.py script:

  python reen_stats.py -prefix <amr_file>
  

• When using the option -verbose the script also outputs the examples where one of more reentrancies could not be classified. See [1] for details.

Oracle experiments

Given an AMR test set (<test_set>) and the output of a parser (<test_set_parsed>):

• To get the results for all actions in isolation:

  ./oracle_script.sh <test_set> <test_set_parsed> (--silent)
  

• To get the results for the combination of all actions:

  ./all_actions_byscore.sh <test_set> <test_set_parsed> (--silent)
  

[1] "The Role of Reentrancies in Abstract Meaning Representation Parsing", Ida Szubert, Marco Damonte, Shay B. Cohen and Mark Steedman. Findings of EMNLP (2020).