Light Coreference Resolution for Russian with Hierarchical Discourse Features

This is our solution for RuCoCo-23 shared task: Coreference Resolution in Russian (only single antecedent resolution).

1. Set up the syntax & NER parser

• (Option 1) With Docker

  • Run the container locally or remotely using the following command:

       docker run --rm -d -p 3334:3333 --name spacy_ru tchewik/isanlp_spacy:ru
    

  • Connect to it from Python:

    from isanlp.processor_remote import ProcessorRemote
     
    spacy_address = ['0.0.0.0', 3334]
    spacy_processor = (ProcessorRemote(spacy_address[0], spacy_address[1], '0'),
                       ['tokens', 'sentences'],
                       {'lemma': 'lemma',
                        'postag': 'postag',
                        'morph': 'morph',
                        'syntax_dep_tree': 'syntax_dep_tree',
                        'entities': 'entities'})

• (Option 2) Locally

  • Download the model

    python -m spacy download ru_core_news_lg
    

  • Initialize in Python using ProcessorSpaCy

    from isanlp.processor_spacy import ProcessorSpaCy
    
    spacy_processor = (ProcessorSpaCy(model_name='ru_core_news_lg'),
                      ['tokens', 'sentences'],
                      {'lemma': 'lemma',
                       'postag': 'postag',
                       'morph': 'morph',
                       'syntax_dep_tree': 'syntax_dep_tree',
                       'entities': 'entities'})

2. Set up the RST parser (only for model_rh)

• (Only option) With Docker

  • Run the container locally or remotely using the following command:

    docker run --rm -d -p 3335:3333 --name rst_ru tchewik/isanlp_rst:2.1-rstreebank
    

  • Connect to it from Python:

    from isanlp.processor_remote import ProcessorRemote
    
    rst_address = ['0.0.0.0', 3335]
    rst_processor = (ProcessorRemote(rst_address[0], rst_address[1], 'default'),
                     ['text', 'tokens', 'sentences', 'postag', 'morph', 'lemma', 'syntax_dep_tree'],
                     {'rst': 'rst'})

3. Set up the coreference resolver

There are two models from the test leaderboard of RuCoCo-23: base and Rh-enhanced. The latter requires RST parsing which makes it slow. There are also two options for running: with Docker or locally.

name	F1 (dev)	F1 (test)	time (example, CPU only)	for local run (place into models/)	docker image
base	74.3	72.8	~883 ms	model_base.tar.gz	tchewik/corefhd:base
base+rh	74.6	73.3	~19 s	model_rh.tar.gz	tchewik/corefhd:rh

• (Option 1) With Docker

  • Run the container locally or remotely using the following command using selected tag (base or rh):

       docker run --rm -d -p 3336:3333 --name corefhd tchewik/isanlp_corefhd:<tag>
    

  • Connect to it from Python:

    from isanlp.processor_remote import ProcessorRemote
    
    coref_address = ['0.0.0.0', 3336]
    
    # Base model
    corefhd = (ProcessorRemote(coref_address[0], coref_address[1], 'default'),
               ['text', 'tokens', 'sentences',
                'lemma', 'postag', 'syntax_dep_tree', 'entities'],
               {'entity_clusters': 'entity_clusters'})
    
    # Rh model
    corefhd = (ProcessorRemote(coref_address[0], coref_address[1], 'default'),
               ['text', 'tokens', 'sentences',
                'lemma', 'postag', 'syntax_dep_tree', 'entities', 'rst'],
               {'entity_clusters': 'entity_clusters'})

• (Option 2) Locally

  • Download the model as models/model_base.tar.gz or models/model_rh.tar.gz (link in the table).
  • Find the python path for allennlp and update for LUKE (see load_custom_allennlp_scripts.bash)
  • Initialize in Python using ProcessorCorefHD:

    from processor_corefhd import ProcessorCorefHD
    
    # Base model
    corefhd_processor = (ProcessorCorefHD(cuda_device=-1, use_discourse=False),
               ['text', 'tokens', 'sentences',
                'lemma', 'postag', 'syntax_dep_tree', 'entities'],
               {0: 'entity_clusters'})
    
    # Rh model
    corefhd_processor = (ProcessorCorefHD(cuda_device=-1, use_discourse=True),
               ['text', 'tokens', 'sentences',
                'lemma', 'postag', 'syntax_dep_tree', 'entities', 'rst'],
               {'entity_clusters': 'entity_clusters'})

4. Process the texts

• Construct the pipeline from initialized processors:

  • For base model

      from isanlp import PipelineCommon
      from isanlp.processor_razdel import ProcessorRazdel
    
      ppl = PipelineCommon([
         (ProcessorRazdel(), ['text'],
          {'tokens': 'tokens',
           'sentences': 'sentences'}),
         spacy_processor,
         corefhd_processor
      ])

  • For Rh model

      from isanlp import PipelineCommon
      from isanlp.processor_razdel import ProcessorRazdel
    
      ppl = PipelineCommon([
         (ProcessorRazdel(), ['text'],
          {'tokens': 'tokens',
           'sentences': 'sentences'}),
         spacy_processor,
         rst_processor,
         corefhd_processor
      ])

• Run the constructed pipeline:

  text = open('text_example.txt', 'r').read().strip()
  result = ppl(text)

  The result is given in token spans:

     >>> result['entity_clusters']
     [[[0, 1], [7, 7], [19, 19], [103, 104], [126, 126]],
      [[23, 27], [30, 30]],
      [[68, 69], [72, 72]],
      [[78, 83], [132, 132]],
      [[44, 53], [138, 138], [152, 152]],
      [[133, 134], [140, 140], [149, 149]],
      [[89, 90], [142, 142]]]

  Example finding the corresponding text spans:

  def print_coreference_clusters(text, tokens, entity_clusters):
     def mention_to_str(mention):
         return text[tokens[mention[0]].begin: tokens[mention[1]].end]
     for entity in entity_clusters:
         print(f'{mention_to_str(entity[0])} ::: {[mention_to_str(mention) for mention in entity[1:]]}')
     
  >>> print_coreference_clusters(result['text'], result['tokens'], result['entity_clusters'])
  Иоганн Шильтбергер ::: ['он', 'отрок', 'сам Иоганн', 'он']
  рыцаря по имени Леонгарт Рихартингер ::: ['его']
  венгерские крестоносцы ::: ['которым']
  24-летним сыном герцога Бургундии Жаном Бесстрашным ::: ['Жана']
  венгерский король и будущий император Священной Римской империи Сигизмунд I ::: ['Сигизмунда', 'Сигизмунд']
  бургундские рыцари ::: ['Они', 'им']
  турецкой армией ::: ['турок']

Cite

Further information and examples can be found in our paper:

@INPROCEEDINGS{chistova2023light,
      author = {Chistova, E. and Smirnov, I.},
      title = {Light Coreference Resolution for Russian with Hierarchical Discourse Features},
      booktitle = {Computational Linguistics and Intellectual Technologies. Papers from the Annual International Conference "Dialogue" (2023)},
      year = {2023},
      number = {22},
      pages = {34--41}
}