Two related signals
Sentiment describes positive or negative language. Emotion categories offer more specific clues such as anger or confidence. The implementation combines Apple NaturalLanguage sentiment scoring with keyword-based emotion cues and can use semantic review for uncertain cases.
Search using the clue you have
If you remember that someone sounded frustrated, a mood-related search may help narrow your retained conversations. Add a topic or date to reduce ambiguity. Some potentially ambiguous terms require explicit emotion syntax so an ordinary topic search is not accidentally reinterpreted.
Context changes the meaning
“We discussed anger” is not the same as “I am angry.” Quoted speech, sarcasm, cultural differences, and transcription errors can confuse the signal. The implementation includes confidence and source information and attempts to distinguish quoted or discussed emotions from the speaker’s own expression.
Review the evidence
Treat a label as a retrieval aid. Read the actual words and, where retained and appropriate, listen to the original audio. Do not treat an emotion tag as a diagnosis, a measurement of someone’s internal state, or proof of intent.
Use tone to be more considerate
A useful outcome is returning to a difficult conversation with context and care. A poor outcome is declaring that a person definitely felt a particular way because an automated label said so.
Try it with a sample
- Try a topic-and-time search first.
- Add an emotion clue and compare the results.
- Check the wording around any matched segment.
About this explanation
Reviewed against Modules/Assistant/Shared/Services/SentimentAnalysisService.swift. Examples use fictional people and content. Availability depends on the build, edition, permissions, and selected services. Source changes trigger a review notice and removal from website AI retrieval until reviewed.