Emoji-Emotion Ranking System Using Twitter Data

I. INTRODUCTION

• Most existing approaches treat emojis as static sentiment markers, overlooking contextual variability and risking misinterpretation of affective intent.The paper frames this as a challenge for sentiment and emotion analysis of informal digital discourse.Source: §I · p. 1
• The proposed framework constructs emoji–emotion distributions and rankings from real-world Twitter use, and embeds emojis in Russell’s continuous affective space.Source: §I · p. 1
• The study collects 100,000 emoji-containing Twitter replies and labels each tweet with weighted scores for five emotions.The emotions are Happy, Angry, Surprise, Sad, and Fear.Source: §I · p. 1

II. RELATED WORK

• Prior work uses emojis as cues, weak labels, polarity lexicons, or fused text–emoji features, with applications across languages and domains.Distant-supervision methods may assume stable emoji–emotion associations rather than modeling multidimensional distributions.Source: §II · pp. 1–2
• Open challenges include multifunctional emojis, domain adaptation, and explainability of deep models’ decisions in emoji–text interactions.Source: §II · p. 2

III. METHODS

• The corpus contains 100,000 publicly available Twitter replies collected from January 1, 2020, to November 1, 2025.Source: A. Data · p. 2
• The final corpus contains 1,393 distinct emojis, including emojis absent from the 614-item Emoji Sentiment Ranking v1.0 resource.It contains texts with 531 of that resource’s 614 emojis.Source: A. Data · p. 2
• Search queries targeted emojis; collection retained English replies without links, and two-stage deduplication removed repeated posts before merging records.Posts were deduplicated by URL during collection and again during final consolidation.Source: A. Data · p. 2
• 14,646 posts (14.6%) contain two or more emojis; posts average 11.94 tokens and 1.29 emojis.Source: A. Data · p. 2

B. Text Emotion Detection

• text2emotion maps each text to Happy, Angry, Surprise, Sad, and Fear scores using a lightweight lexicon-based approach.The approach requires neither manually labeled training data nor additional model fine-tuning.Source: B. Text Emotion Detection · pp. 2–3
• The scores estimate the emotional context around an emoji, not the emoji’s definitive or intrinsic meaning.The authors treat them as approximate corpus-level signals that become more informative after aggregation.Source: B. Text Emotion Detection · p. 3
• Normalized emotion scores for each post are associated with the emojis present and aggregated across posts to profile emoji use.Source: B. Text Emotion Detection · p. 3
• Lexicon-based detection may be less reliable for short, noisy posts containing sarcasm, irony, implicit affect, or non-standard language.Source: B. Text Emotion Detection · p. 3

1) Evaluation Criteria:

• The evaluation is descriptive and consistency-oriented, matching the goal of building corpus-level emoji–emotion profiles.Source: 1) Evaluation Criteria · p. 3
• Random preprocessed replies are inspected for plausible detected emotions and removal of obvious URL, mention, and hashtag artifacts.Source: 1) Evaluation Criteria · p. 3
• Emoji-profile stability is assessed by comparing emotion vectors and dominant-emotion agreement across random corpus halves.Source: 1) Evaluation Criteria · p. 3
• Valence–arousal coherence is checked against expected placements for positive, high-energy texts and clearly negative emojis.Source: 1) Evaluation Criteria · p. 3
• The study uses no external human annotations; results are exploratory and rely on aggregation over many posts.Source: 2) Limitations · p. 3

• The Russell circumplex represents emotion along valence and arousal dimensions.Source: Figure 2 · p. 3
• Short, noisy, sarcastic, or code-mixed texts can mislead a lexicon detector, and aggregation reduces but does not remove this effect.Source: 2) Limitations · p. 3
• Emoji scores reflect surrounding-text emotion rather than fixed universal meaning, and profiles may vary by domain and time.Source: 2) Limitations · p. 3
• The dataset covers only English Twitter/X replies from 2020–2025, limiting generalization to other languages and platforms.Source: 2) Limitations · p. 3
• Simple aggregation under-models blended or contrasting affect when a post contains multiple emojis.Source: 2) Limitations · p. 3
• A human-labeled subset and contextual neural baseline are proposed as follow-up work to strengthen validation.Source: 2) Limitations · p. 3

C. Russell Mapping for Emojis

• Valence (pleasant–unpleasant) and arousal (activation–deactivation) organize detected emotions in a two-dimensional space for comparing emoji characteristics.Source: C. Russell Mapping for Emojis · p. 3

D. Proposed Approach

• The proposed approach proceeds through sequential stages of collecting, preprocessing, and analyzing emoji-containing Twitter/X texts.Source: D. Proposed Approach · p. 3, Figure 3 · p. 4

1) Data Collection:

• Emoji-specific Twitter/X searches collected English-language replies from January 1, 2020, to November 1, 2025, excluding posts containing links.Source: 1) Data Collection · pp. 3–4

• The pipeline extracts text and metadata, removes URLs, mentions, hashtags, and non-textual artifacts, tokenizes words and emojis, and compiles the cleaned texts.Source: 2) Parsing and Preprocessing · p. 4
• Text2emotion estimates normalized relative-intensity scores for Happy, Angry, Surprise, Sad, and Fear in each text.Source: 4) Emoji–Emotion Analysis · p. 4

• Emotion intensity is linked to the emojis in each post and aggregated across posts containing the same emoji to build its emotional profile.Source: 4) Emoji–Emotion Analysis · p. 4

5) Data Reproducibility:

• The collection window, emoji-specific queries, reply and English filters, link exclusions, deduplication, and preprocessing steps are documented.These procedural details are presented as supporting reconstruction of the collection pipeline.Source: 5) Data Reproducibility · p. 4
• The resulting dataset’s main fields and final corpus size are described explicitly.Source: 5) Data Reproducibility · p. 4
• Exact reconstruction may depend on platform dynamics and data availability over time.Source: 5) Data Reproducibility · p. 4

IV. RESULTS

• Head emojis have emotion profiles estimated from more abundant context and are described as more stable; long-tail emojis warrant caution because of data sparsity.The text also notes that topic- and calendar-driven effects can shift distributions over time.Source: §IV · p. 4
• High conditionality in emoji pairs reveals thematic bundles and supports reasoning over pairs or short sequences rather than single-emoji priors.The text notes that some pair meanings are not simply the sum of their parts.Source: §IV · pp. 4–5, Table III · p. 5
• Laughter-related emojis skew toward Fear (0.2978 and 0.2999) rather than expected Happy (0.2039 and 0.1978), while another emoji skews Happy (0.3582) despite expected Sad (0.2131).These cases are presented as diagnostic examples of context sensitivity.Source: Table IV · p. 5, §IV · p. 5
• Fear has the highest median intensity (approximately 0.30) and widest distribution, while Angry has the lowest median (approximately 0.07) and narrowest.Source: §IV · pp. 5–6, Figure 5 · p. 6
• Positive, highly activating emojis cluster in the upper-right of valence–arousal space, while negatively valenced, low-arousal emojis tend toward the lower-left.Emoji coordinates are approximate and derived from aggregated emotion scores.Source: Figure 6 · p. 6, §IV · p. 6
• Figures 8 and 9 compare emoji emotion profiles across five emotions using radial intensity distances and heatmap cell colors.The radar axes represent Happy, Fear, Sad, Surprise, and Angry; heatmap rows represent emojis and columns represent emotions.Source: §IV · p. 6, Figure 8 · p. 7, Figure 9 · p. 7

• The framework uses weighted emotion scores across contexts in 100,000 replies to profile and rank 1,393 distinct emojis, then projects them into Russell valence–arousal space.Source: §V · p. 7
• Emojis show multidimensional, probabilistic emotional characteristics rather than fixed sentiment polarity, and frequently used emojis have clear dominant tendencies.Source: §V · p. 7
• Because the dataset contains only English-language Twitter replies, its cross-cultural generalization is limited.Source: §V · p. 7