SPO Extractor using Stanza

A dependency parsing-based tool for extracting Subject-Predicate-Object (SPO) triples from biomedical text using Stanford CoreNLP and Stanza.

Table of Contents

• Overview
• Features
• Requirements
• Installation
• Usage
• Example Sentences and Universal Dependencies
• Algorithm
• Testing
• Future Enhancements

Overview

This tool extracts structured SPO (Subject-Predicate-Object) relationships from natural language sentences, particularly useful for biomedical text mining. It uses dependency parsing patterns and noun phrase chunking to identify semantic relationships between entities.

Features

• Trigger-based extraction with dependency parsing
• Neural pipeline for accurate dependency analysis
• Support for complex grammatical structures (passive voice, coordinating conjunctions, relative clauses)
• Integration with Stanford CoreNLP's Tregex for chunking
• Stanza's neural models for dependency parsing

Requirements

• Python 3.x
• Stanford CoreNLP 4.0.0 or later
• Stanza library with English models

Installation

1. Download Stanford CoreNLP

Download and extract the following:

• Stanford CoreNLP
• English models

Place the model JAR files in the CoreNLP distribution folder.

2. Set Environment Variables

export CORENLP_HOME=/path/to/stanford-corenlp-full-2020-04-20
export DATA_DIR=/path/to/data/

3. Install Python Dependencies

pip install -r requirements.txt

4. Download Stanza English Model

python -c 'import stanza; stanza.download("en")'

Usage

Extract SPO from Input Directory

PYTHONPATH=. python bin/run_spo.py -i input_directory -o output_file

Parameters:

• -i, --input: Directory containing input text files
• -o, --output: Output file for extracted SPO triples

Test with Example Sentences

PYTHONPATH=. python tests/test_SPOs.py

Example Sentences and Universal Dependencies

• The encapsulation of rifampicin leads to a reduction of the Mycobacterium smegmatis inside macrophages.
• nsubj (nominal subject) <= VERB => obl (oblique nominal)

• The Norwalk virus is the prototype virus that causes epidemic gastroenteritis infecting predominantly older children and adults.
• acl:relcl (adjectival clause) => VERB => obj (object)

• It is widely agreed that the exposure to ambient air pollution may cause serious respiratory illnesses and that weather conditions may also contribute to the seriousness.
• nsubj <= VERB => obj

• In this report, ribavirin was shown to inhibit SARS coronavirus replication in five different cell types of animal or human origin at therapeutically achievable concentrations.
• nsubj:pass <= xcomp => VERB => obj

• Chronic hepatitis virus infection is a major cause of chronic hepatitis, cirrhosis, and hepatocellular carcinoma worldwide.
• nsubj <= NOUN => nmod => conj
• coordinating conjunctions

Supported Dependency Patterns

The extractor handles various dependency patterns including:

• nsubj => VERB => obj: Basic subject-verb-object
• nsubj => VERB => obl: Verb with oblique nominal
• acl:relcl => VERB => obj: Relative clauses
• nsubj:pass => xcomp => VERB => obj: Passive voice constructions
• nsubj => NOUN => nmod => conj: Coordinating conjunctions

What is Stanza?

Stanza is a Python wrapper that combines Stanford CoreNLP and PyTorch-based NLP models:

• Tregex for noun phrase chunking
• Neural pipeline for dependency parsing

Algorithm

The extraction process follows these steps:

1. Input Processing: Accept a sentence and a list of trigger words
2. Trigger Detection: Check if any trigger word appears in the sentence
3. Parsing: If triggered, run dependency parser and chunker on the sentence
4. Head Word Identification: Use dependency relations from the trigger to identify syntactic head words
5. NP Extraction: Extract noun phrases by merging dependency relations and chunks based on head words

Testing

Run All Tests

cd /path/to/project-directory
pytest tests/test_SPOs.py

Run Data Reader Tests

export DATA_DIR="$(pwd)/data/tests"
pytest tests/test_data_reader.py

Test with Example Sentences

PYTHONPATH=. python tests/test_SPOs.py

Future Enhancements

• Biomedical NER Integration: Biomedical Named Entity Recognizers can improve NP chunking and identify semantic roles of noun phrases
• Entity Type Classification: Distinguish between different types of biomedical entities (diseases, drugs, proteins, etc.)
• Relation Classification: Categorize extracted relationships by type (causation, association, treatment, etc.)