Emotion recognition in chat is not yet understanding feelings

CORTEX combines two languages

The researchers supplemented an existing English-language, emotion-labeled corpus with neutral texts. They then used machine translation to create a Polish version. This resulted in a parallel bilingual dataset with three classes for overall mood and nine emotion classes.

The structure is useful for comparative research: the same content examples can be examined in both languages. At the same time, the Polish version carries characteristics of translation. It is not automatically the same as spontaneously written everyday Polish.

For a real chat, this difference is important. People write with typos, dialect, irony, abbreviations, and personal idioms. A cleanly labeled dataset always captures only a part of the story the user chooses to share.

BERT achieved the best scores

Naive Bayes, Support Vector Machines, fastText, and BERT were compared. The best BERT-based models achieved more than 90 percent accuracy in three-part sentiment classification and almost 80 percent for nine emotions.

The Polish results were slightly below the English ones in accuracy and F1 score. According to the study, the largest gap was seen with BERT. Language transfer is therefore not just a matter of identical categories but measurably affects classification.

Almost 80 percent is interesting for a research dataset. For a conversational decision, this value also means that a relevant proportion of examples is assessed differently than in the gold standard. The more a function relies on the label, the more important the consequences of such deviations become.

A label does not yet say what is needed

Two messages can both sound sad and yet require completely different responses. One person wants to tell a story, another wants to check information, and a third contradicts an earlier interpretation. The emotion class does not automatically describe the communicative intention.

The same wording can also mean something different depending on the course of the conversation. "It's okay" can express relief, withdrawal, irony, or an abrupt end to the conversation. Without prior turns and without a reaction to corrections, classification remains uncertain.

A supportive chat should therefore treat emotion recognition as a hint, not as a diagnosis or fixed truth. A cautious response such as "Sounds like this is on your mind" leaves room for correction. A statement like "You are sad" tends to close it.

Quality does not arise from the system naming as many feelings as possible. It shows in whether the system holds its assumption lightly and adjusts its response when the user disagrees.

Empathetic language is socially evaluated

In two experiments, Liu and Sundar examined how people perceive emotional expressions from a health chatbot. In one study, 158 participants read a dialogue, while in the other, 88 participants interacted with a real chatbot.

Sympathy, cognitive empathy, and affective empathy were preferred over purely emotionless information. This was especially pronounced among people who were initially skeptical about whether machines possess social abilities.

The result shows that tone is not merely decoration. People respond to linguistic signs of concern even though they know the counterpart is technically generated. However, it does not follow that the most intense empathy is always better.

Excessive certainty can seem artificial or overbearing. A system does not directly know the inner experience. A more credible response is one that is appropriate and limited, building on the visible content without claiming a closeness that the dialogue does not support.

More channels mean more signals and more sources of error

According to the project description, Empath.ai combines text, facial, and speech recognition to capture emotions contextually. The idea is understandable: intonation, facial expressions, and word choice together can yield a richer picture than text alone.

However, more data does not automatically resolve ambiguity. A tired face can indicate lack of sleep, concentration, illness, or simply an unfavorable camera moment. Voice and facial expressions also vary culturally and individually.

Multimodal capture also intensifies data protection concerns. Audio and facial data are more sensitive than a voluntarily entered sentence. A product must therefore justify what additional information is truly necessary and what benefit it provides to the user.

For many low-threshold conversations, text can be a strength precisely for this reason: the user decides what they express. The system works with fewer signals but also claims less insight.

Conversation quality requires different test tasks

A classification metric checks whether a model selects the same label as the dataset. It does not check whether a response in the course of the conversation listens, accepts a no, or truly lets go of a previously corrected assumption.

For conversational systems, additional test cases are therefore needed. A user may later relativize an emotional account. The system should not only acknowledge this correction in a friendly manner but also incorporate it into subsequent responses.

The chosen direction of the conversation is equally important. In storytelling mode, a precisely recognized distress can still lead to a poor response if a plan is immediately suggested. In next-steps mode, a cautious structure may be appropriate.

Emotion recognition and dialogue management should remain separately measurable. Otherwise, it is impossible to determine whether an error stems from incorrect interpretation or from an inappropriate response rule.

Language differences must be tested in real-world use

The fact that Polish performed somewhat worse in the experiment cautions against the assumption that a successful English system can be transferred without loss. German, in turn, brings its own colloquial language, sentence structures, and regional variants.

A translation of English examples can serve as a starting point. For robust quality, it additionally requires original German-language utterances from the intended usage situations and a comprehensible, ideally repeated, labeling process.

Disagreement among annotators is not merely a disruptive factor. It shows which utterances are ambiguous even for humans. In such cases, a system should not be forced to maximum certainty but should take uncertainty into account in its response format.

The sensible use remains supportive

CORTEX demonstrates that automatic mood and emotion recognition can achieve good results even across languages. The study provides a dataset and a technical comparison, not clinical evidence of efficacy for a complete chat offering.

In a responsible architecture, the classification can provide indications for tone, further tests, or downstream quality analysis. It should neither generate a diagnosis nor alone decide what help a person receives.

A conversation does not become good because the system finds the correct emotion label. It becomes better when it handles this assumption cautiously, reads along with the course of the conversation, and allows the user to correct the assessment at any time.

Emotion recognition is thus one tool among several. Only together with dialogue memory, clear role boundaries, data protection, and test-specific conversation logic can it become an offering that seems not only technically attentive but also humanly connectable.