Baxnet Ideas · Technical explainer
A 👍 Is Not a Sentiment Label
TL;DR: Reactions are useful conversational evidence, but turning a thumbs-up directly into positive sentiment throws away the context that gave it meaning.
“The parcel is downstairs.” 👍
“Dad’s appointment moved to Thursday.” 👍
“Fine. Do whatever you want.” 👍
The symbol is identical. Its job in the conversation is not.
The first thumbs-up probably means got it. The second may be a quick acknowledgement from someone who does not have time to write. The third could mean agreement, sarcasm, resignation, or simply that the person wants the exchange to end.
That ambiguity matters when software turns reactions into data. It is tempting to translate 👍 into positive, count it, and move on. Doing so produces tidy numbers by throwing away the part that made the reaction interpretable.
Positive-looking reactions are often doing social work
A WASSA 2026 paper tested the assumption at scale. The researchers analysed 647,879 reacted-to messages from 993 public, English-language cryptocurrency channels on Telegram. They automatically classified the sentiment and emotion of each message, then compared those labels with the standard emoji reactions it received.
More than 84% of messages classified as negative or neutral received reactions with positive emoji sentiment. Fewer than 0.3% of negative messages received negative emoji feedback. Joy-related reactions appeared so broadly that removing them often improved the measured alignment between a message’s emotion and its reactions.
The researchers interpret many of these positive-looking reactions as social signals. They can express support for the person, agreement with the point, group loyalty, encouragement, gratitude, banter, or acknowledgement. The reaction is part of the interaction around the message, not necessarily a mirror of the emotion inside it.
There are important limits. These were public crypto channels, not private chats between friends or partners. The message and reaction labels came from automated classifiers that were not validated on this dataset. Channel administrators could also restrict which reactions were available, often leaving people with a choice set already tilted towards approval. That makes the paper a warning about simplistic labels, with no claim to decode reactions everywhere.
A separate ACL 2026 study reached the problem from another direction. It asked language models to infer indirect intent from short dialogues where the response consisted only of silence, a facial-expression emoji, or a described movement. Across its constructed examples, models averaged 51% accuracy on the facial-expression cases; people scored 89%. Some models lost as much as 60 percentage points when an indirect response became non-verbal rather than verbal.
That study was not about platform reactions, and its benchmark contained only 50 scenarios per response type. Still, it demonstrates the underlying difficulty: once words disappear, meaning has to be recovered from context, shared norms, and the relationship between the participants. A symbol alone does not carry all of that.
Record the tap before interpreting it
A careful personal intelligence system should first preserve the observable event: Sam used 👍 on this message at 14:03. The reaction happened; its sentiment still has to be inferred.
To make even a tentative interpretation, the system may need the target message, the turns around it, the available reaction menu, who reacted, when they reacted, and how the same people usually use that symbol. Even then, acknowledgement likely; sentiment unclear may be a better result than positive.
This matters because a small labelling shortcut can distort a larger account of a relationship. A summary might report that someone responds positively to difficult news, that a tense group is unusually supportive, or that a conversation ended warmly. If the metric was built by counting thumbs-ups, the report would look more certain than its underlying evidence.
For Mimoto, reactions should remain connected to their source and their conversational surroundings. They can still help show who participates, who acknowledges updates, and which habits belong to a group. They should not quietly become ground-truth emotion labels just because they are easy to extract.
A thumbs-up is easy to count. What it meant is a different question.
Further reading: Emoji Reactions on Telegram: Unreliable Indicators of Emotional Resonance and Unveiling the Limits of Large Language Models in Inferring Pragmatic Meaning from Non-Verbal Responses.