IsaNLP SRL FrameBank

This Python 3 library provides SRL parser for Russian based on neural network models trained on FrameBank corpus -- semantically annotated corpus of texts in Russian. The parser should be used in conjunction with IsaNLP library and can be considered its module.
Disclaimer: the code is in alpha stage.

Installation

Note: the library is considered for Python 3.6. If you do not have one, checkout pyenv or this docker container powered with the data science tools: jupyter, keras, sklearn, gensim, tensorflow, pytorch, etc.

Quick (few dependencies for usage with docker)

1. Install IsaNLP and its dependencies:

pip install grpcio
pip install git+https://github.com/IINemo/isanlp.git

2. Install IsaNLP SRL FrameBank library:

pip install git+https://github.com/IINemo/isanlp_srl_framebank.git

Full

1. Do the same steps for Quick installation.
2. Install gensim, tensorflow, numpy, sklearn:

pip install gensim==3.6.0 smart-open==1.7.0 tensorflow==1.12.0 numpy sklearn

3. Load the neural models from Git LFS

curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
sudo apt-get install git-lfs
git lfs install
git-lfs pull

Quick start with docker (< 10 mins)

1. Deploy docker containers for morphology, syntax, and SRL parsing:

docker run --rm -p 3333:3333 inemo/isanlp
docker run --rm --shm-size=1024m -ti -p 3334:9999 inemo/syntaxnet_rus server 0.0.0.0 9999
docker run --rm -p 3335:3333 inemo/isanlp_srl_framebank

2. Connect from python using PipelineDefault:

#
from isanlp_srl_framebank.pipeline_default import PipelineDefault  
#
ppl = PipelineDefault(address_morph=('localhost', 3333),
                      address_syntax=('localhost', 3334),
                      address_srl=('localhost', 3335))
res = ppl('Мы поехали на дачу.')
#

4. The res variable should contain all annotations including SRL annotations stored in res['srl']:

{'text': 'Мы поехали на дачу.',
 'tokens': [<isanlp.annotation.Token at 0x7f2290211240>,
  <isanlp.annotation.Token at 0x7f22902112e8>,
  <isanlp.annotation.Token at 0x7f2290211320>,
  <isanlp.annotation.Token at 0x7f22902113c8>,
  <isanlp.annotation.Token at 0x7f2290211438>],
 'sentences': [<isanlp.annotation.Sentence at 0x7f2290211470>],
 'mystem_postag': [['SPRO,мн,1-л=им',
   'V,сов,нп=прош,мн,изъяв',
   'PR=',
   'S,жен,неод=вин,ед',
   '']],
 'lemma': [['мы', 'поехать', 'на', 'дача', '.']],
 'morph': [[{'fPOS': 'PRON', 'Number': 'Plur', 'Person': '1', 'Case': 'Nom'},
   {'fPOS': 'VERB',
    'Aspect': 'Perf',
    'Valency': 'INTR',
    'Tense': 'Past',
    'Number': 'Plur',
    'VerbForm': 'Fin'},
   {'fPOS': 'ADP'},
   {'fPOS': 'NOUN',
    'Gender': 'Fem',
    'Animacy': 'Inan',
    'Case': 'Acc',
    'Number': 'Sing'},
   {}]],
 'postag': [['PRON', 'VERB', 'ADP', 'NOUN', '']],
 'syntax_dep_tree': [[<isanlp.annotation.WordSynt at 0x7f2290211630>,
   <isanlp.annotation.WordSynt at 0x7f2290211668>,
   <isanlp.annotation.WordSynt at 0x7f22902116d8>,
   <isanlp.annotation.WordSynt at 0x7f2290211710>,
   <isanlp.annotation.WordSynt at 0x7f2290211748>]],
 'srl': [[<isanlp.annotation.Event at 0x7f229021f278>]]}

5. The variable res['srl'] can be visualized as:
   

Package overview

1. The semantic role labeler. The SRL parser is implemented in ProcessorSrlFramebank class. Path: src/isanlp_srl_framebank/processor_srl_framebank.py.
2. Trained neural network models for SRL parser: model for "known" predicates and model for "unknown" predicates. Path: models.
3. Docker container inemo/isanlp_srl_framebank that contains preinstalled libraries and models. The container provides gRPC service for SRL parsing. Path: docker/parser. The container can be obtained with the command:
   docker run --rm -p 3335:3333 inemo/isanlp_srl_framebank
4. Dockerized web demo for visualization of SRL annotations via brat annotation tool: inemo/isanlp_srl_framebank_demo. Path: docker/demo. For example, the container can be obtained with the command (note, you need NLP services started on the corresponding ports, see Quick start section):
   docker run --rm -p 1111:80 -e IP_ADDRESS=10.0.0.9 -e MORPH_PORT=3333 -e SYNTAX_PORT=3334 -e SEM_PORT=3335 inemo/isanlp_srl_framebank_demo

Usage

Parsing

For version of parsing with less code (using docker) see Quick start section.
The processor for parsing is implemented in ProcessorSrlFramebank class.

from isanlp_srl_framebank.processor_srl_framebank import ProcessorSrlFramebank

The object should be initialized by a path to models and dictionaries of the parser. Example:

srl_proc = ProcessorSrlFramebank(model_dir_path = '~/srl_model_path')

The model_dir_path should contain:

• embeddings.vec -- a model of word embeddings that can be loaded with gensim.models.word2vec.KeyedVectors.load_word2vec_format function.
• known_preds.json -- the list of "known" predicates serialized into JSON. These predicates would be processed by the model for "known" predicates.
• known_predicates -- (mandatory) directory with model files for "known" predicates.
• unknown_predicates -- (optional) directory with model files for "unknown" predicates.
• The directory with model files must contain:
  • feature_encoder.pckl -- pickled sklearn.preprocessing.DictVectorizer object that was used for categorical feature conversion during training.
  • label_encoder.pckl -- pickled sklearn.preprocessing.LabelBinarizer object that was used to encode labels into vectors during training.
  • feature_model.pckl -- pickled object that was used for extraction of features from linguistic annotations during training. The object should have extract_features member function.
  • neural_model.h5 -- saved keras model (with structure and weights).

After initialization the object is ready for parsing using __call__ function:

  res = srl_proc(tokens, postag, morph, lemma, syntax_dep_tree)

The input arguments are:

• tokens -- a list of tokens.
• postag -- a list of lists of postags in Universal Dependencies format (the first list means sentences, the second list means words in the sentence).
• morph -- a list of lists of morphological features in Universal Dependencies format.
• lemma -- a list of lists of lemmas.
• syntax_dep_tree -- a list of syntax dependency trees of sentences in Universal Dependecies format.

The proper annotations can be obtained by other components of IsaNLP library. Example:

from isanlp.processor_remote import ProcessorRemote
from isanlp.ru.converter_mystem_to_ud import ConverterMystemToUd
from isanlp import PipelineCommon
# - To start morphology processor:
#$ docker run -d --rm -p 3333:3333 inemo/isanlp
# - To start ProcessorSyntaxNetRemote:
#$ docker run --shm-size=1024m -ti --rm -p 3334:9999 inemo/syntaxnet_rus server 0.0.0.0 9999
syntax_ppl = PipelineCommon([(ProcessorRemote('10.0.0.9', 3333, 'default'), 
                              ['text'], 
                              {'sentences' : 'sentences', 
                               'tokens' : 'tokens',
                               'postag' : 'mystem_postags',
                               'lemma' : 'lemma'}),
                              (ProcessorSyntaxNetRemote('10.0.0.9', 3334), 
                               ['tokens', 'sentences'], 
                               {'syntax_dep_tree' : 'syntax_dep_tree', 
                                'morph' : 'morph_syntaxnet'}),
                              (ConverterMystemToUd(), 
                               ['mystem_postags'], 
                               {'morph' : 'morph',
                                'postag' : 'postag'})])
#
annots = syntax_ppl('Мама мыла раму.')
annots = srl_proc(annots['postag'],   
                  annots['morph'], 
                  annots['lemma'], 
                  annots['syntax_dep_tree'])

Finally, you get the SRL annotations:

print(annots['srl'])

[[<isanlp.annotation.Event at 0x7f229021f978>]]

SRL data structures

The results of SRL parser are stored in a list of lists of isanlp.annotation.Event objects. The first list corresponds to sentences, the second list corresponds to annotations inside a sentence. Event objects have two members:

• pred(tuple): predicate address -- tuple of begin end positions in sentence.
• args(list): list of arguments -- tuples of isanlp.annotation.TaggedSpan objects. Each tagged span contains:
  • begin(int) -- starting position (in tokens in a sentence) of the argument in a sentence.
  • end(int) -- ending position of an argument in a sentence.
  • tag(string) -- semantic role.

With lemma annotations it is possible to print roles like this:

def print_roles(lemma, role_annot):
    for sent_num, ann_sent in enumerate(role_annot):
        for event in ann_sent:
            print('=====Pred: {}'.format(lemma[sent_num][event.pred[0]]))
            for arg in event.args:
                print('Arg({}): {}'.format(arg.tag, lemma[sent_num][arg.begin]))

Example:

print_roles(annots['lemma'], annots['srl'])

The result for 'Дети убежали из дома и заблудились в лесу.':

=====Pred: убегать
Arg(агенс): ребенок
Arg(начальная точка): дом
=====Pred: заблудиться
Arg(место): лес

Model training

1. Download and preprocess dataset. python ./run_download_preproc.py --workdir=<existing_workdir>

2. Do the linguistic preprocessing of the dataset (postagging, parsing, etc.)

• Start linguistic preprocessing services:

docker run --rm -d -p 3333:3333 inemo/isanlp
docker run --rm -d -p 3334:3333 inemo/isanlp_udpipe

• Run preprocessing (change 192.168.1.69 to address of the machine, on which you started the linguistic processing services). python ./run_ling_parse.py --workdir=existing_workdir --isanlp_proc=192.168.1.69:3333 --udpipe_proc=192.168.1.69:3334

3. Generate features, embed with ELMo, and train models: python ./run_training_pipeline.py --data_dir=existing_workdir --workdir=existing_workdir

Finally you can take trained models from directories 'known_preds' and 'unknown_preds' and place them in <project root>/models directory to create updated docker containers (or use them in the processor).

Starting web demo

1. To start web demo, you should start NLP services for morphology, syntax, and SRL parsing first:

docker run --rm -p 3333:3333 inemo/isanlp
docker run --rm --shm-size=1024m -ti -p 3334:9999 inemo/syntaxnet_rus server 0.0.0.0 9999
docker run --rm -p 3335:3333 inemo/isanlp_srl_framebank

2. Start demo web application by the following command:

docker run --rm -d -p 1111:80 -e IP_ADDRESS=10.0.0.9 -e MORPH_PORT=3333 -e SYNTAX_PORT=8111 -e SEM_PORT=3334  inemo/isanlp_srl_framebank_demo

Note: instead of 10.0.0.9 you should use your host ip address (but not localhost or 127.0.0.1).
3. After web server started you can acesses demo interface at http://10.0.0.9:1111/demo/wui.py

Compatability

The library should be compatible at least with Python 3.7.
Tested with gensim==3.6.0 tensorflow==1.15.0 .

Cite / Reference

1. Models were published in RANLP proceedings

• GOST: Larionov D., Shelmanov A., Chistova E., Smirnov I. Semantic role labeling with pretrained language models for known and unknown predicates // Proceedings of RANLP. — 2019. — P. 245–256.
• BibTex:

@INPROCEEDINGS{devshelm2019ranlp,
      author = {Larionov D., Shelmanov A., Chistova E., Smirnov I.},
      title = {Semantic role labeling with pretrained language models for known and unknown predicates},
      booktitle = {Proceedings of Recent Advances in Natural Language Processing},
      year = {2019},
      pages = {620--628}
}

2. Original models for SRL were published in Dialog proceedings.

• GOST:
  Shelmanov A., Devyatkin D. Semantic role labeling with neural networks for texts in Russian // Computational Linguistics and Intellectual Technologies. Papers from the Annual International Conference "Dialogue" (2017). — Vol. 1. — 2017. — P. 245–256.
• BibTex:

@INPROCEEDINGS{devshelm2017dialog,
      author = {Shelmanov, A.O. and Devyatkin, D.A.},
      title = {Semantic role labeling with neural networks for texts in {Russian}},
      booktitle = {Computational Linguistics and Intellectual Technologies. Papers from the Annual International Conference "Dialogue" (2017)},
      year = {2017},
      volume={1},
      number = {16},
      pages = {245--256}
}