Table of Contents
This repository contains code and trained models corresponding to the paper "MoralStrength: Exploiting a Moral Lexicon and Embedding Similarity for Moral Foundations Prediction".
Run Predictor.ipynb to see a functioning version of the moral foundations predictor. Keep reading for some examples of use below.
On an new work, we have generated two new versions of the Liberty/oppression moral foundation lexicon: the LibertyMFD lexicon.
The lexicons are accessible in this repository, in the liberty/2nd_version folder (link here).
We expect to update this lexicon soon.
If you use this lexicon, please cite the following publication:
@inproceedings{10.1145/3524458.3547264,
author = {Araque, Oscar and Gatti, Lorenzo and Kalimeri, Kyriaki},
title = {LibertyMFD: A Lexicon to Assess the Moral Foundation of Liberty.},
year = {2022},
isbn = {9781450392846},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3524458.3547264},
doi = {10.1145/3524458.3547264},
booktitle = {Proceedings of the 2022 ACM Conference on Information Technology for Social Good},
pages = {154–160},
numpages = {7},
keywords = {lexicon, natural language processing, word embeddings, liberty, moral foundations theory, moral values},
location = {Limassol, Cyprus},
series = {GoodIT '22}
}
We have generated a new lexicon that contains the Liberty/oppression moral foundation.
To access the lexicon, see the liberty/1st_version folder (link here).
This lexicon will be updated regularly.
If you use the liberty lexicon, please cite the following paper:
DOI: https://doi.org/10.1145/3442442.3452351
The software is written in Python 3. For installing, please use pip:
pip install moralstrength
This repository is intended for users that are willing to use the software through Python. Alternatively, we have published a Graphical Interface that works on Linux, MacOS, and Windows. Please visit this repository.
This repository contains the MoralStrength lexicon, which enables researchers to extract the moral valence from a variety of lemmas.
It is available under the directory moralstrength_annotations.
An example of use of the lexicon with Python is:
>>> import moralstrength
>>> moralstrength.word_moral_annotations('care')
{'care': 8.8, 'fairness': -1, 'loyalty': -1, 'authority': -1, 'purity': -1}Also, this repository contains several already-trained models that predict the presence of a certain moral trait. That is, whether the analyzed text is relevant for a moral trait, or not. A minimal example of use:
import moralstregnth
text = "PLS help #HASHTAG's family. No one prepares for this. They are in need of any assistance you can offer"
moralstrength.string_moral_value(text, moral='care')You can check the available moral traits using the moralstrength.lexicon_morals method.
The complete list of methods that can be used is shown in the next section.
This package offers a function to perform unsupervised prediction over a list of texts, giving the prediction in a organized fashion. For example:
from moralstrength.moralstrength import estimate_morals
texts = '''My dog is very loyal to me.
My cat is not loyal, but understands my authority.
He did not want to break the router, he was fixing it.
It is not fair! She cheated on the exams.
Are you pure of heart? Because I am sure not.
Will you take care of me? I am sad.'''
texts = texts.split('\n')
result = estimate_morals(texts, process=True) # set to false if text is alredy pre-processed
print(result)The result of this short script would be as follows. The estimation is given in a pandas.DataFrame format.
care fairness loyalty authority purity
0 NaN NaN 8.875 5.1250 NaN
1 NaN NaN 8.875 6.9625 NaN
2 NaN NaN NaN NaN NaN
3 NaN 9.0 NaN NaN NaN
4 NaN NaN NaN NaN 9.0
5 8.8 NaN NaN NaN NaN
The original version of the MoralStrength lexicon is described here:
Oscar Araque, Lorenzo Gatti, Kyriaki Kalimeri,
MoralStrength: Exploiting a moral lexicon and embedding similarity for moral foundations prediction,
Knowledge-Based Systems,
Volume 191,
2020,
105184,
ISSN 0950-7051,
https://doi.org/10.1016/j.knosys.2019.105184.
(http://www.sciencedirect.com/science/article/pii/S095070511930526X)
which is also open in arXiv https://arxiv.org/abs/1904.08314.
A new improved version of the lexicon can be used to predict moral values. By default, the software uses the last version. to use the original version, you can do:
from moralstrength import lexicon_use
lexicon_use.select_version("original")
# predict here moral values using the original MoralStrengthIf at any moment you want to use the new version of the lexicon again, just do:
lexicon_use.select_version("latest")The methods that are under moralstrength.moralstrength are the following:
get_available_lexicon_traits()
Returns a list of traits that were annotated and can be queried
by word_moral_value().
care: Care/Harm
fairness: Fairness/Cheating
loyalty: Loyalty/Betrayal
authority: Authority/Subversion
purity: Purity/Degradation
get_available_models()
Returns a list of available models for predicting texts.
Short explanation of names:
unigram: simple unigram-based model
count: number of words that are rated as closer to moral extremes
freq: distribution of moral ratings across the text
simon: SIMilarity-based sentiment projectiON
or a combination of these.
For a comprehensive explanation of what each model does and how it performs on
different datasets, see https://arxiv.org/abs/1904.08314
(published at Knowledge-Based Systems).
get_available_prediction_traits()
Returns a list of traits that can be predicted by string_moral_value()
or file_moral_value().
care: Care/Harm
fairness: Fairness/Cheating
loyalty: Loyalty/Betrayal
authority: Authority/Subversion
purity: Purity/Degradation
non-moral: Tweet/text is non-moral
string_average_moral(text, moral)
Returns the average of the annotations for the words in the sentence (for one moral).
If no word is recognized/found in the lexicon, returns -1.
Words are lemmatized using spacy.
string_moral_value(text, moral, model='unigram+freq')
Returns the estimated probability that the text is relevant to either a vice or
virtue of the corresponding moral trait.
The default model is unigram+freq, the best performing (on average) across all
dataset, according to our work.
For a list of available models, see get_available_models().
For a list of traits, get_available_prediction_traits().
string_moral_values(text, model='unigram+freq')
Returns the estimated probability that the text is relevant to vices or virtues
of all moral traits, as a dict.
The default model is unigram+freq, the best performing (on average) across all
dataset, according to our work.
For a list of available models, see get_available_models().
For a list of traits, get_available_prediction_traits().
word_moral_value(word, moral)
Returns the association strength between word and moral trait,
as rated by annotators. Value ranges from 1 to 9.
1: words closely associated to harm, cheating, betrayal, subversion, degradation
9: words closely associated to care, fairness, loyalty, authority, sanctity
If the word is not in the lexicon of that moral trait, returns -1.
For a list of available traits, get_available_lexicon_traits()
word_moral_values(word)
Returns a dict that gives the association strength between word and every
moral trait, as rated by annotators. Value ranges from 1 to 9.
1: words closely associated to harm, cheating, betrayal, subversion, degradation
9: words closely associated to care, fairness, loyalty, authority, purity/sanctity
If the word is not in the lexicon of that moral trait, returns -1.
The moralstrength_raw folder contains the raw annotations collected from figure-eight.
The folder all_annotators_except_failed contains all the annotations collected, except for the annotators that failed the task (see the paper for details on the control questions, which were based on valence ratings from Warriner et al.).
The folder filtered_annotators contains the annotations after the annotators with low inter-annotator agreement were removed.
The filename is RAW_ANNOTATIONS_[MORAL], where MORAL is the moral trait considered and can either be AUTHORITY, CARE, FAIRNESS, LOYALTY or PURITY.
The fields in each file are:
- WORD the word to be annotated
- ANNOTATOR_ID the unique ID of each annotator
- VALENCE the valence rating of WORD, on a scale from 1 (low) to 9 (high)
- AROUSAL the arousal rating of WORD, on a scale from 1 (low) to 9 (high)
- RELEVANCE whether WORD is related to the MORAL
- EXPRESSED_MORAL the moral strength of WORD, i.e. whether it is closer to one or the other extremes pertaining the MORAL trait.
The numbers for EXPRESSED_MORAL range from 1 to 9, and the extremes of the scales are:
- 1=Subversion, 9=Authority for AUTHORITY
- 1=Harm, 9=Care for CARE
- 1=Proportionality, 9=Fairness for FAIRNESS
- 1=Disloyalty, 9=Loyalty for LOYALTY
- 1=Degradation, 9=Purity for PURITY
For privacy reason, the annotator ID has been salted and hashed, so that going back to the original annotator ID is not possible, but it is still possible to track each annotator's ratings across the different morals.
In the folder moralstrength/tasks we also include the original description of the annotation tasks for the crowdsourcing process.
The interested reader can consult the instructions given to the human annotators.