Fig. 2. Complete or partial words on which RoBERTa models fine-tuned on research assistants relied most for generating personal quality scores.

Font size is proportional to word importance. Darker words are more common. Token “gru” is a fraction of the word “grueling,” and token “unte” is a fraction of the word “volunteer.” Words importance is not invariant across essays, it depends on word context. Word importance and frequency were largely independent (r = −0.03 and P < 0.001). For instance, for intrinsic motivation, the model relied more on the word “pleasure” then the word “fun,” but essays were more likely to contain the word “fun” then the word “pleasure.”