Researchers use machine learning models to detect moods of White English speakers from Facebook posts
Researchers at the University of Virginia and NIH have created machine learning models from Facebook posts by White and Black English speakers that accurately detect the mood of White English speakers. Interestingly, the models do not work with Black English speakers even when the ML models are trained on only Black English speaker texts.
Machine learning is an AI method to create a computer model from large amounts of data that can be used for prediction. For this study, four models were trained to recognize mood based on text input. The researchers did this in order to determine whether the posts written by participating White and Black English speakers, specifically their use of first-person pronouns (I-usage) like “I” or “We”, could predict their low mood.
First, the researchers looked at whether I-usage increased with depression by race, and they found this to be true. However, I-usage varied far less among Black participants and was used more often. The researchers also tested the relationships between clusters of words from 5 topics related to negative emotions, such as worthlessness, and found that increasing topic word usage increased with dour mood among Whites, but not Blacks.
Next, the researchers examined the relationship between I-usage and mood by applying two machine learning methods on Black speaker texts, then White speaker texts to create four models.
The Black language and White language models were able to reliably detect low-mood in White, but not Black, participant Facebook posts. The researchers had a few ideas why this occured, such as the dual-identity of Blacks, but more research is needed.
This finding is an exciting step towards building humanoid robots that can perceive how we feel, then try to cheer us up. Before that day comes, know that daily exercise helps improve mood and a workout machine like this at Amazon can help.