DraymondHIT / syntactic-constituent-extraction

Extract (subject, verb, object) , (subject, mod) and (verb, mod) relations from a sentence.

syntactic-constituent-extraction

• Extract (subject, verb, object) , (subject, mod) and (verb, mod) relations from a sentence using Spacy.
• Substitute some pronouns by coreference resolution using longformer_coreference_ontonotes.
• Detect and extract (if exists) cause-and-effect relationship from a sentence using CiRA and Multilabel Model.

Installation

Uses Python 3.5+ and Spacy for its parser.

pip install -r requirements

# use spacy to download its small model
python -m spacy download en_core_web_sm

Parameters Download

For CiRA, the authors did not supply model parameters. Of course, you could train the model by yourself, but if you would like to run the following demo quickly, you can download parameters for CiRA by this link. After download parameters, you can put them under the ./models/ directory. For Multilabel Models, you can download parameters by run

gdown --id 1NizNkSdah8Jcs8wtDTD5dEnxUWrCP4Zi -O roberta_dropout_linear_layer_multilabel.ckpt

For longformer_coreference_ontonotes, you can download parameters by run

gdown --id 1CQxUq2zvCHc1mJUEZ_Zy6WSJQqFz76Pw

After that, you can also put the parameters under the same directory.

Run the Demo

python demo.py

Example

For the sentence below as an example.

If a user signs up, he will receive a confirmation email.

Outputs includes the original sentence, a list of SVO tuples, a list of SM tuples, a list of VM tuples and cause-and-effect relationship if it exists.

text: If a user signs up, he will receive a confirmation email.
SVO:  [('a user', 'will receive', 'a confirmation email')]
SM:   [('a user', '')]
VM:   [('will receive', ['If a user signs up'])]
CE:   {'cause': ['a user signs up'], 'effect': ['he will receive a confirmation email']}

Notes

• When there is no object, the program will return just SV parts.
• When there is no mod for subject or object, the program will return (subject, '') or (object, '').
• When there is no cause-and-effect relationship in the sentence, the program will return None. Otherwise, a dictionary containing keys named cause and effect will be returned.

TODO

• Realize batch processing

• Improve the accuracy of cause-and-effect relationship extraction