Tool to identify ableist language in job descriptions.
Developed jointly by the Department of Labor Office of Disability Employment Policy ("DOL ODEP"), xD | U.S. Census Bureau, and the Presidential Innovation Fellows.
What is ableist language?
Ableist language is language that is offensive to people with disability. It can also refer to language that is derogatory, abusive or negative about disability. Ableism is the systemic exclusion and oppression of people with disability, often expressed and reinforced through language. [source]
Why is this tool important?
Ableist language in job descriptions can cause people with disabilities to feel excluded from jobs that they are qualified for. This typically occurs when a description references abilities or enduring attributes of an individual that are unnecessary for the job or for which accommodations can be proactively offered instead of focusing on developed skills that can be acquired to succeed in the role. By identifying ableist language and suggesting alternatives, this tool will support more inclusive hiring practices.
This package requires Python >= 3.8.
Clone the repo and install the package (preferably in a virtual environment):
git clone git@github.com:USDepartmentofLabor/ableist-language-detector.git
python -m pip install ableist-language-detector/
Download spaCy dependencies.
python -m spacy download en_core_web_sm
Developer Installation
If you plan on contributing to the repo, complete these additional steps:
Install the dev requirements.
python -m pip install -r requirements_dev.txt
extract_onet_terms.py
: Extract representative terms for abilities and skills from O*Net data. Used as one of our sources for our ableist lexicon.detector.py
: Main module that identifies ableist language in a job description.model_api.py
: Creates custom mlflow model to serve detector for REST API access.
There are four ways to access the tool.
To run a local instance of the tool as a web application, see the instructions in this respository.
The command line tool allows you to check a single job description for ableist language by passing a .txt
file containing the job description text to the main detector.py
script.
The script will print out any ableist language that was detected in the job description, along with the location of the language (token index position in the text), the root form of the terms, suggested alternative verbs, and an example of how to use the alternative phrasing.
Usage: detector.py [OPTIONS]
Extract ableist terms from a job description.
Options:
-j, --job_description_file TEXT
Path to file containing the job description
text. [required]
--help Show this message and exit.
Example usage:
>>> python ableist_language_detector/detector.py -j sample_job_descriptions/short_job_description.txt
Found 4 instances of ableist language.
Match #1
PHRASE: lifting | LEMMA: lift | POSITION: 21:22 | ALTERNATIVES: ['move', 'install', 'operate', 'manage', 'put', 'place', 'transfer', 'transport'] | EXAMPLE: Transport boxes from shipping dock to truck
Match #2
PHRASE: bend | LEMMA: bend | POSITION: 37:38 | ALTERNATIVES: ['lower oneself', 'drop', 'move to', 'turn'] | EXAMPLE: Install new ethernet cables under floor rugs
Match #3
PHRASE: move your hands | LEMMA: move your hand | POSITION: 7:10 | ALTERNATIVES: ['observe', 'operate', 'transport', 'transfer', 'activate'] | EXAMPLE: Operates a machine using a lever
Match #4
PHRASE: move your wrists | LEMMA: move your wrist | POSITION: 31:34 | ALTERNATIVES: ['observe', 'operate', 'transport', 'transfer', 'activate'] | EXAMPLE: Operates a machine using a lever
The main functionality is also available directly via detector.find_ableist_language()
for those who would like a more flexible way to integrate the functionality into existing pipelines/applications. The detector.find_ableist_language()
function returns a collection of AbleistLanguageMatch
objects, which contain the same information listed above as attributes.
AbleistLanguageMatch Attribute |
Type | Description |
---|---|---|
text |
str |
Raw form of matched phrase |
lemma |
str |
Lemma (i.e. root form) of matched phrase |
start |
int |
The starting token index of the matched phrase within the document |
end |
int |
The ending token index (exclusive) of the matched phrase within the document |
data.verb |
str |
The lemma form of the matched verb from the ableist lexicon |
data.alternative_verbs |
List[str] |
The list of suggested alternative verbs from the ableist lexicon |
data.example |
str |
An example of an alternative verb used in a phrase/sentence from the ableist lexicon |
data.object_dependent |
bool |
True if the data.verb is considered ableist depending on the verb's object; else False |
data.objects |
List[str] |
If data.object_dependent == True , a list of objects that, combined with data.verb , is considered ableist language |
Example usage:
>>> from ableist_language_detector import detector
>>> sample_job_description = """
requirements
- must be able to move your hands repeatedly
- type on a computer
- comfortable with lifting heavy boxes
- excellent communication skills
- move your wrists in circles and bend your arms
"""
>>> ableist_language = detector.find_ableist_language(sample_job_description)
>>> print(ableist_language)
[lifting, bend, move your hands, move your wrists]
# Accessing attributes
def print_results(result):
"""Convenience function to print attributes."""
print(f"Found {len(result)} instances of ableist language.\n")
if len(result) > 0:
for i, ableist_term in enumerate(result):
print(
f"Match #{i+1}\n"
f"PHRASE: {ableist_term} | LEMMA: {ableist_term.lemma} | "
f"POSITION: {ableist_term.start}:{ableist_term.end} | "
f"ALTERNATIVES: {ableist_term.data.alternative_verbs} | "
f"EXAMPLE: {ableist_term.data.example}\n"
)
>>> print_results(ableist_language)
Found 4 instances of ableist language.
Match #1
PHRASE: lifting | LEMMA: lift | POSITION: 22:23 | ALTERNATIVES: ['move', 'install', 'operate', 'manage', 'put', 'place', 'transfer', 'transport'] | EXAMPLE: Transport boxes from shipping dock to truck
Match #2
PHRASE: bend | LEMMA: bend | POSITION: 38:39 | ALTERNATIVES: ['lower oneself', 'drop', 'move to', 'turn'] | EXAMPLE: Install new ethernet cables under floor rugs
Match #3
PHRASE: move your hands | LEMMA: move your hand | POSITION: 8:11 | ALTERNATIVES: ['observe', 'operate', 'transport', 'transfer', 'activate'] | EXAMPLE: Operates a machine using a lever
Match #4
PHRASE: move your wrists | LEMMA: move your wrist | POSITION: 32:35 | ALTERNATIVES: ['observe', 'operate', 'transport', 'transfer', 'activate'] | EXAMPLE: Operates a machine using a lever
A custom MLflow model accessible via an API can be created with the model_api.py
script.
To build the model, which will create a local directory called detector_model
, run:
python ableist_language_detector/model_api.py
You can optionally use the same script to run the MLflow detector model on an input job description file:
python ableist_language_detector/model_api.py \
--train_only False \
-j sample_job_descriptions/short_job_description.txt
Serve the model to port 1234 after making the following script executable:
./serveModel.sh detector_model
The model can now be used via a REST API and will return key value pairs
of ableist term location and properties.
The predictAPI.sh
script contains an example curl command for using the model API.
>>>cat sample_job_descriptions/short_job_description.txt | predictAPI.sh
{"21": {"lemma": "lift", "text": "lifting", "start": "21", "end": "22", "alternative_verbs": "['move', 'install', 'operate', 'manage', 'put', 'place', 'transfer', 'transport']", "example": "Transport boxes from shipping dock to truck"},
"37": {"lemma": "bend", "text": "bend", "start": "37", "end": "38", "alternative_verbs": "['lower oneself', 'drop', 'move to', 'turn']", "example": "Install new ethernet cables under floor rugs"},
"7": {"lemma": "move your hand", "text": "move your hands", "start": "7", "end": "10", "alternative_verbs": "['observe', 'operate', 'transport', 'transfer', 'activate']", "example": "Operates a machine using a lever"},
"31": {"lemma": "move your wrist", "text": "move your wrists", "start": "31", "end": "34", "alternative_verbs": "['observe', 'operate', 'transport', 'transfer', 'activate']", "example": "Operates a machine using a lever"}}
The tool checks for job descriptions against an ableist language lexicon. To view the language that's currently in our lexicon, see the ableist_language_detector/ableist_word_list.csv file. This lexicon is constantly evolving and we appreciate any feedback or requests for changes. To do so, please open an issue.
The lexicon was developed based on the following data sources in consultation with subject matter experts at DOL ODEP.