The ranking measures emotional dominance, not polarity
The paper's central object is an emoji–emotion ranking system. For each emoji, the authors aggregate the weighted emotion scores of the messages in which that emoji appears, producing an emoji–emotion association distribution. The ranking then orders emojis by relative emotional dominance — which of the five detected emotions (Happy, Angry, Sad, Fear, Surprise) dominates that emoji's observed contexts.
This is a different measurement from a sentiment lexicon score. A lexicon assigns an emoji a polarity value, such as positive or negative. The ranking here assigns a distribution over five emotions and reports which one leads. A reader comparing this work with polarity-based emoji lexicons should not treat the two outputs as interchangeable: one is a single valence label, the other is a multi-emotion profile with a dominant tendency.
A reply-only corpus of 100,000 emoji-bearing tweets
The dataset was assembled through an automated search-based pipeline. For each target emoji, a dedicated Twitter (X) search query was issued over the selected time interval, with filters restricting results to English-language replies and excluding posts containing links. The stated purpose of those filters was to focus the corpus on short, conversational, emoji-containing responses.
The resulting corpus contains 100,000 emoji-containing replies collected between January 1, 2020 and November 1, 2025, and comprises 1,393 distinct emojis. Two scope conditions follow directly from this design and should travel with any result: the data are English-language, and they are replies rather than standalone posts. The authors note in their limitations that the English-only restriction limits cross-cultural generalization.
Each message becomes five weighted emotion scores
After preprocessing and text cleaning, the authors applied text-to-emotion classification to detect five primary emotions for each message: Happy, Angry, Sad, Fear, and Surprise. The output per message is a set of weighted emotion scores rather than a single hard label.
Aggregation happens at the emoji level. Every context in which a given emoji appears contributes its message-level emotion scores, and those scores are pooled to estimate that emoji's emotion association distribution. This is why the paper describes the profiles as probabilistic: the unit of analysis is the distribution over contexts, not a hand-assigned meaning for the emoji itself.
A second view places emojis in valence–arousal space
Beyond the discrete ranking, the authors project emojis into the Russell valence–arousal space to enable continuous affective interpretation. This is a separate representation from the five-emotion ranking: the ranking reports relative dominance among Happy, Angry, Sad, Fear, and Surprise, while the Russell projection positions emojis along continuous affective dimensions.
The paper presents the projection as a means of analysis and visualization. Readers should keep the two outputs distinct — a dominant emotion in the ranking and a coordinate in valence–arousal space answer different questions about the same emoji, and the paper does not treat one as a substitute for the other.
Static polarity labels miss contextual variability
The paper frames its contribution against existing approaches that treat emojis as static sentiment markers, assigning fixed polarity labels or assuming stable emotional meaning. The authors argue that such simplifications overlook the contextual variability of emoji use and risk misinterpreting affective intent in informal, youth-dominated digital discourse.
The reported outcome is that emojis exhibit multidimensional and probabilistic emotional characteristics rather than fixed sentiment polarity, and that the constructed ranking dataset revealed clear dominant emotional tendencies for frequently used emojis. Note the asymmetry in that pair of claims: the multidimensional, probabilistic character is stated for emojis generally, while the clear dominant tendencies are reported specifically for frequently used emojis, where more contexts are available to aggregate.
Dominant tendencies are clearest for frequent emojis
Because the ranking is built by pooling contexts, an emoji's profile depends on how many contexts it appears in. The paper reports clear dominant emotional tendencies for frequently used emojis; it does not claim equally stable profiles for the long tail of the 1,393 distinct emojis in the corpus.
This is the practical threshold a reader should carry away. When the ranking is used to compare emojis, the comparison is best supported for emojis with substantial context counts. For rarely occurring emojis, the aggregated distribution rests on fewer messages, and the paper's own framing — probabilistic profiles rather than fixed meanings — applies with less stability.
English Twitter replies bound the conclusions
The framework integrates dataset construction, emoji–emotion ranking, and projection into the Russell valence–arousal space, and it was evaluated on the 100,000-reply English Twitter corpus. The authors state that the dataset is restricted to English-language Twitter replies and that this limits cross-cultural generalization.
They also note that the emotion detection was lexicon-based and that they plan extensive validation and comparison with other emotion-labeling approaches. That matters for how the ranking should be read: the emotion scores feeding the aggregation come from one labeling method, and the paper positions broader validation as future work rather than as a completed check. Future directions named in the paper include multilingual datasets, multiple social media networks, and the temporal dynamics of emoji usage.
Compare emojis by profile, not by a single label
The most decision-relevant shift this paper proposes is in what to compare. If emojis carry probabilistic, context-sensitive profiles, then ranking them by a single polarity discards the distribution that the paper treats as the actual signal. The five-emotion ranking and the valence–arousal projection are the two views the authors provide for that richer comparison.
The priority for a reader is therefore to check which output answers the question at hand: relative dominance among Happy, Angry, Sad, Fear, and Surprise for a given emoji, or its position in continuous affective space. And because the reported clarity of dominant tendencies is tied to frequently used emojis in an English-language reply corpus, comparisons are strongest where context counts are high and should be treated as provisional where they are not.
Sources used in this answer
Emoji–Emotion Ranking System Using Twitter Data
Emojis exhibit probabilistic, context-sensitive emotional profiles rather than fixed sentiment polarities, with clear dominant emotional tendencies reported for frequently used emojis.