The SuperEmotion dataset
Enric Junqué de Fortuny · 2025
Read The SuperEmotion dataset by Enric Junqué de Fortuny (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-15348.txt.
Introduction
Enric Junqué de Fortuny built the SuperEmotion dataset, set out in a 2025 report. Work on teaching machines to name feelings in text has lacked one large, shared base with a sound set of feeling types.
Older sets use mixed labels, hold few samples, or cover one narrow field. The new set joins text from many sources under one scheme. That scheme is drawn from a tested map of human feelings. Work in one field can thus be set next to work in the next. The report covers the second build of the set. It also gives first scores from a tuned modern text model, as a base mark for later work to beat.
Catalog record
| CREATOR | Enric Junqué de Fortuny |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 13 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-15348
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The SuperEmotion dataset Enric Junqué de Fortuny Abstract Despite the wide-scale usage and development of emotion classification datasets in NLP, the field lacks a standardized, large-scale resource that follows a psychologically grounded taxonomy. Existing datasets either use inconsistent emotion categories, suffer from limited sample size, or focus on specific domains. The SuperEmotion dataset addresses this gap by harmonizing diverse text sources into a unified framework based on Shaver’s empirically validated emotion taxonomy, enabling more consistent cross-domain emotion recognition research. 1 Introduction This report describes the SuperEmotion dataset, the world’s largest Shaver compliant emotion dataset for natural language processing. We developed the SuperEmotion dataset by aggregating multiple existing emotion datasets and remapping categories into Shaver’s primary emotion classes. The dataset encompasses 519,812 samples labeled across the primary emotions: joy, sadness, anger, fear, love , and surprise as well as a neutral category. The source datasets integrated into this collection include: • MELD ( Poria et al., 2019 ) : A multimodal dataset for emotion recognition in conversations. The text comes from scripts and transcribed dialogues from the TV show Friends , capturing multi-party spoken interactions with speaker context. • GoEmotions ( Demszky et al., 2020 ) : A large-scale dataset of carefully filtered English Reddit comments annotated with 27 fine-grained emotions plus neutrality. It is well-suited for studying subtle affect in user-generated social media content. • TwitterEmotion ( Saravia et al., 2018 ) : A Twitter-based dataset developed for context-aware emotion recognition, containing tweets labeled across multiple emotion categories. The short, informal text reflects real-world online communication. • ISEAR ( Scherer, 1997 ) : The International Survey on Emotion Antecedents and Reactions consists of structured questionnaire responses. Participants described personal experiences that triggered one of seven basic emotions, yielding formal first-person narratives. • SemEval ( Mohammad et al., 2018 ) : A dataset from the Semantic Evaluation series, particularly Task 1 on Affect in Tweets. It contains tweets annotated for emotional intensity and valence, providing rich insight into affective language in short-form, real-time social media. • CrowdFlower ( Van Pelt and Sorokin, 2012 ) : A crowdsourced dataset of tweets labeled with emotion categories such as sadness, joy, and anger. The data was used to evaluate how well laypeople agree on emotion labels in noisy, informal text. By consolidating these resources, the SuperEmotion dataset addresses class imbalances and provides a more diverse and extensive foundation for training robust emotion classification models following a well-understood taxonomy. Detailed statistics, data distribution, and guidelines for accessing the dataset are provided in this report. 2 Dataset Construction 2.1 Dataset Composition The SuperEmotion dataset is constructed by aggregating multiple publicly available emotion classification datasets. Table 1 summarizes the primary datasets integrated into this collection. Dataset Train Validation Test Classes MELD 9,989 1,109 2,610 7 TwitterEmotion 16,000 2,000 2,000 6 ISEAR 416,809 - - 6 GoEmotions 43,410 5,426 5,427 28 CrowdFlower 39,998 - - 13 SemEval 6,634 872 3,184 11 Total 532,840 9,407 13,221 SuperEmotions 441,478 55,088 58,902 7 Table 1: Overview of datasets aggregated in the SuperEmotion dataset, including the number of emotion classes in each source. 2.2 Preprocessing and Quality Control To ensure consistency across diverse source formats, we applied several preprocessing steps: 1. Text Normalization : We standardized text formatting by removing excessive whitespace, normalizing unicode characters, and ensuring consistent punctuation placement. 2. Deduplication : We identified and removed exact duplicate texts to prevent test set leakage and avoid biasing the dataset toward redundant patterns. 3. Data Splits : For datasets without predefined splits, we created stratified train/validation/test partitions (80%/10%/10%) to preserve label distribution across splits. 4. Metadata Preservation : While harmonizing the emotion taxonomy, we retained source information for each example, enabling analysis of domain-specific patterns and potential biases. 2.3 Shaver’s Taxonomy Emotion classification has been approached in a variety of ways in psychology, linguistics, and affective computing. Some models, such as Eckman’s framework ( Eckman, 1972 ) , define a small set of discrete universal emotions (e.g., anger, fear, joy, sadness, surprise, disgust), primarily grounded in facial expressions. Others, like Russell’s circumplex model ( Russell and Barrett, 1999 ) , represent emotions in a continuous space defined by valence and arousal dimensions. Plutchik’s wheel of emotions ( Plutchik, 1980 ; Plutchik, 2001 ) combines discrete emotions with intensity scaling and mixing of different primary emotions. After careful consideration, we adopt the taxonomy proposed by Shaver et al., 1987 , which provides a psychologically grounded, hierarchical classification of emotions based on empirical clustering of 135 emotion terms. Through free-listing, sorting, and similarity judgments, Shaver et al. identified six basic-level emotions ( love, joy, anger, sadness, fear, and surprise 1 1 1 Note: We include surprise even though Shaver gave it less important as it appears in all datasets and is clearly of interest to the NLP community. ) which serve as cognitively salient and linguistically frequent prototypes in English. We selected Shaver’s taxonomy because the taxonomy is lexically grounded—built from natural language emotion terms - which aligns well with the input modality of most NLP systems and simplifies annotation. Second, it balances granularity and coverage, capturing sufficient emotional nuance for robust modeling while remaining tractable for supervised learning and being robust for potential expansions of the dataset. 2.4 Label Harmonization To align heterogeneous label sets from different source datasets, we mapped related labels into these six core categories using Shaver’s original definitions and keywords ( Shaver et al., 1987 ) . When handling ambiguous cases during label harmonization, we prioritized semantic similarity to Shaver’s prototypical emotion terms. For instance, labels like ’optimism’ could reasonably map to either joy or be excluded as a distinct anticipatory state; we assigned it to joy based on its positive valence and association with pleasant feelings. Similarly, confusion was mapped to surprise rather than fear based on its cognitive rather than threat-oriented nature. Where a source dataset used idiosyncratic labels without clear mapping to Shaver’s categories (e.g., empty , curiosity ), we excluded these examples rather than forcing them into potentially inappropriate categories. A complete mapping is provided in Table 2 . Emotion Source Labels Anger anger , anger , anger , anger , anger , anger , annoyance , disapproval , disgust , disgust , disgust , hate Fear fear , fear , fear , fear , fear , nervousness , worry Joy amusement , enthusiasm , excitement , fun , happiness , joy , joy , joy , joy , joy , optimism , optimism , pride , relief , relief Love admiration , approval , caring , desire , gratitude , love , love , love , love , love , trust Sadness disappointment , embarrassment , grief , pessimism , remorse , sadness , sadness , sadness , sadness , sadness , sadness Surprise confusion , realization , surprise , surprise , surprise , surprise , surprise , surprise Neutral boredom , neutral , neutral , neutral Dropped anticipation , curiosity , empty Table 2: Emotion label breakdown mapped to Shaver’s categories. Dataset sources are color-coded as follows: GoEmotions , ISEAR , MELD , Crowdflower , SemEval , TwitterEmotion . We also introduce a Neutral category to capture instances with low or absent emotional valence. Labels that lacked a clear conceptual or empirical correspondence