medspacy / nlp_preprocessor

SpaCy component for modifying the string of a doc before tokenizing.

nlp_preprocessor

This package implements a SpaCy component for modifying the string of a doc before tokenizing.

Overview

This component is useful in clinical text processing for cleaning up text formatting, normalizing abbreviations or misspellings, and removing problematic semi-structured template texts.

The Preprocessor class is instantiated with the tokenizer of a spaCy Language model. To add the preprocessor to your pipeline, you must set it to be the tokenizer:

from nlp_preprocessor import Preprocessor

preprocessor = Preprocessor(nlp.tokenizer)
nlp.tokenizer = preprocessor

Processing steps are then added to the preprocessor by calling preprocessor.add(rules). The preprocessor then executes each of the processing steps and finally returns tokenizer(text) to return a spaCy Doc.

You can access the original tokenizer through preprocessor.tokenizer.

The PreprocessingRule class offers one way to define preprocessing steps. This class takes a compiled regular expression pattern as well as an optional replacement string (default is a blank string) and description. When a rule is executed on the text it will replace any occurrences of the pattern with the replacement string.

from nlp_preprocessor import PreprocessingRule
import re

rule = PreprocessingRule(re.compile("there"), repl="world!")
preprocessor.add([rule])
doc = preprocessor("Hello, there")
print(doc)
>>> Hello, world!

You can use any Python callable as a rules, such as lambda functions or other objects, as long as they take the text as input and return the text as output.

Usage

Installation

You can install nlp_preprocessor using pip:

pip install nlp_preprocessor

Or clone this repository install using the setup.py script:

$ python setup.py install

Once you've installed the package and spaCy, make sure you have a spaCy language model installed (see https://spacy.io/usage/models):

$ python -m spacy download en_core_web_sm

Example

In this short example below, we'll use the preprocessor to delete templated text, lower-case the document, and expand an abbreviation.

from nlp_preprocessor import Preprocessor, PreprocessingRule
import spacy
import re

text = "TOBACCO SCREENING: The patient does/does not smoke tobacco daily. Purpose of visit: R/O pneumonia."

nlp = spacy.blank("en")
preprocessor = Preprocessor(nlp.tokenizer)
nlp.tokenizer = preprocessor

rules = [
    PreprocessingRule(re.compile("TOBACCO SCREENING: The patient does/does not smoke tobacco daily."),
                                desc="Remove tobacco screening."),
    lambda x: x.lower(),
    PreprocessingRule(re.compile("r/o"), repl="rule out", desc="Normalize 'rule out' abbreviation."),
]
preprocessor.add(rules)

doc = nlp(text)
print(doc)
>>> purpose of visit: rule out pneumonia.