Neural Networks Recognize Emotions: From GoEmotions to Claude

⚡ The gist in 5 seconds - The point: neural networks talk about feelings quite well, but that's algorithmic work, not an expression of emotions; 17.2% of 500 respondents turned to ChatGPT for support. - Where it's available: models like Claude and ChatGPT are already used in conversations, for example in response to the phrase "I got laid off." - The limitation: the number next to the "anxiety" label describes the model's prediction, not a measurement of a person's actual anxiety level. ### 🔍 What was found Emotion recognition is based on labeled datasets. GoEmotions contains 58,000 English comments from Reddit, labeled across 27 emotions plus a neutral label. Each comment was evaluated by three annotators; in case of disagreement, two more were brought in, and for training, labels chosen by only one person were removed, leaving about 93% of the examples. In the StudEmo corpus, 5,182 reviews were rated by 25 people each, with every individual's answers preserved — this shows where a text allows different interpretations. A survey of 500 people showed that 17.2% had turned to ChatGPT for support with mental health issues. Technically, a BERT-based classifier splits a message into tokens, and self-attention layers relate tokens to one another: in the phrase "I was waiting for the talk, but it got postponed," the representation of "postponed" can take into account "waiting" and "talk." Generative models like Claude don't have a fixed list of emotions: when asked "what is he feeling?" they construct the answer token by token, taking the instruction and context into account. A study with 1,001 English-language descriptions of events showed: direct text classification yields a micro-F1 of 0.60; if intermediate features are set by people — 0.66; and if they are predicted by another model — it drops to 0.48. ### 💡 Why it matters AI empathy is computation, not feelings: models can attribute an emotion to the wrong participant, as in the example of the cat that fell asleep on the phone, where the classifier might decide that the contentment was felt by the cat, not by the girl waiting for a reply. Understanding how the labels are structured and where they come from helps developers build support systems properly: accounting for the experiencer, the cause, and the object of the emotion, rather than relying on names or individual words. Observer labels describe the observers' opinion, not the author's inner state — something worth remembering when using AI in sensitive scenarios. ### 🧩 Context The piece was published on the BotHub company blog on Habr and opens with the saddest short story ever: "I got laid off," a person writes to Claude, and it replies: "It sounds like you're worried about the future." The article examines research on emotion classification: from the GoEmotions corpus, where labels were assigned by readers, to experiments with author hashtags and the distribution of opinions in StudEmo. It turns out that the questions "what did the author report about themselves?" and "what did an outsider read into the post?" are only superficially similar, and a person's name is a poor substitute for a story about what happened to them.

🔗 Read on habr.com