Terminology

EDS-NLP simplifies the terminology matching process by exposing a eds.terminology pipeline that can match on terms or regular expressions.

The terminology matcher is very similar to the generic matcher, although the use case differs slightly. The generic matcher is designed to extract any entity, while the terminology matcher is specifically tailored towards high volume terminologies.

There are some key differences:

1. It labels every matched entity to the same value, provided to the pipeline
2. The keys provided in the regex and terms dictionaries are used as the kb_id_ of the entity, which handles fine-grained labelling

For instance, a terminology matcher could detect every drug mention under the top-level label drug, and link each individual mention to a given drug through its kb_id_ attribute.

Usage

Let us redefine the pipeline :

import spacy

nlp = spacy.blank("fr")

terms = dict(
    covid=["coronavirus", "covid19"],  # 
    flu=["grippe saisonnière"],  # 
)

regex = dict(
    covid=r"coronavirus|covid[-\s]?19|sars[-\s]cov[-\s]2",  # 
)

nlp.add_pipe(
    "eds.terminology",
    config=dict(
        label="disease",
        terms=terms,
        regex=regex,
        attr="LOWER",
    ),
)

This snippet is complete, and should run as is.

Configuration

The pipeline can be configured using the following parameters :

Parameter	Explanation	Default
label	Top-level label.	Required
terms	Terms patterns. Expects a dictionary.	None (use regex only)
regex	RegExp patterns. Expects a dictionary.	None (use terms only)
attr	spaCy attribute to match on (eg NORM, LOWER)	"TEXT"
ignore_excluded	Whether to skip excluded tokens during matching	False

Patterns, be they terms or regex, are defined as dictionaries where keys become the kb_id_ of the extracted entities. Dictionary values are a either a single expression or a list of expressions that match the concept (see example).

Authors and citation

The eds.terminology pipeline was developed by AP-HP's Data Science team.