These AIs Can Predict Your Moral Principles

The death penalty, abortion, gun legislation: There's no shortage of controversial topics that are hotly debated today on social media. These topics are so important to us because they touch on an essential underlying force that makes us human, our morality.

Researchers in Brazil have developed and analyzed three models that can describe the morality of individuals based on the language they use. The results were published last month in IEEE Transactions on Affective Computing.

Ivandre Paraboni is an associate professor at the School of Arts, Sciences and Humanities at the University of Sao Paulo who led the study. His team choose to focus on a theory commonly used by social scientists called Moral foundations theory. It postulates several key categories of morality including care, fairness, loyalty, authority, and purity.

The aim of the new models, according to Paraboni, is to infer values of those five moral foundations just by looking at their writing, regardless of what they are talking about. They may be talking about their everyday life, or about whatever they talk about on social media," Paraboni says. And we may still find underlying patterns that are revealing of their five moral foundations."

To develop and validate the models, Paraboni's team provided more than 500 volunteers with questionnaires. Participants were asked to rate eight topics (e.g., same sex marriage, gun ownership, drug policy) with sentiment scores (from 0 = totally against' to 5 = totally in favor'). They were also asked to write out explanations of their ratings.

Human judges then gave their own rating to a subset of explanations from participants. The exercise determined how well humans could infer the intended opinions from the text. Knowing the complexity of the task from a human perspective in this way gave us a more realistic view of what the computational models can or cannot do with this particular dataset," says Paraboni.

Using the text opinions from the study participants, the research team created three machine learning algorithms that could assess the language used in each participant's statement. The models analyzed psycholinguistics (emotional context of words), words, and word sequences, respectively.

All three models were able to infer an individual's moral foundations from the text. The first two models, which focus on individual words used by the author, were more accurate than the deep learning approach that analyzes word sequences.

Paraboni adds, Word counts-such as how often an individual uses words like sin' or duty'-turned out to be highly revealing of their moral foundations, that is, predicting with higher accuracy their degrees of care, fairness, loyalty, authority, and purity."

He says his team plans to continue to incorporate other forms of linguistic analysis into their models. They are, he says, exploring other models that focus more on the text (independent of the author) as a way to analyze Twitter data.