EmoLex: The NRC Emotion Lexicon
- EmoLex is a crowdsourced lexicon that records word–sense associations with Plutchik’s eight basic emotions and polarity, enabling fine-grained affect analysis.
- It employs a sense-aware design from multiple sources and uses majority voting with ordinal scales, ensuring reliable annotation and distinguishing perceived emotion from denotative meaning.
- Practical applications include aggregate emotion tracking, embedding enrichment in NLP models, and serving as a benchmark in affective computing while highlighting ethical concerns over annotation bias.
EmoLex, commonly referring to the NRC Emotion Lexicon or NRC Word-Emotion Association Lexicon, is a high-coverage English word–emotion association lexicon introduced through crowdsourcing and later established as a central resource in affective NLP. In its original design, it records associations for word–sense pairs with Plutchik’s eight basic emotions—joy, sadness, anger, fear, trust, disgust, surprise, and anticipation—together with positive and negative polarity. Subsequent methodological discussion characterizes it as a manually curated, crowdsourced lexicon designed to capture how speakers of a language perceive word–emotion associations, rather than denotational meaning in the dictionary sense (Mohammad et al., 2013, Mohammad, 2020).
1. Definition and conceptual scope
EmoLex was created to address a limitation of earlier sentiment resources: polarity alone is often too coarse for emotion analysis. The original motivation was that negative language does not by itself distinguish anger, sadness, fear, or disgust, and positive language does not collapse cleanly into trust, joy, or anticipation. The lexicon therefore adopts Plutchik’s eight-emotion inventory and annotates polarity separately, yielding a resource that supports finer-grained lexical affect analysis than binary positive–negative sentiment (Mohammad et al., 2013).
A central conceptual point is that EmoLex encodes associations or connotations, not denotations. The later practical and ethical guidance states this explicitly: “party is associated with joy, but it does not mean (denote) joy.” This makes EmoLex a lexicon of perceived affective associations, not a dictionary of emotion words in the narrow sense. The same guidance also emphasizes that the annotation goal was to determine how speakers of a language perceive the emotion associations of words, and that this framing produced less variance in responses (higher inter-annotator agreement) than asking annotators about the emotions evoked in themselves (Mohammad, 2020).
This distinction also clarifies a common misconception. EmoLex is not restricted to terms that directly denote emotion. In the original study, only 826 terms, corresponding to 9.3% of the final target set, were judged to refer directly to emotions; the lexicon is broader because emotionally informative words often do not literally name an emotion (Mohammad et al., 2013).
2. Crowdsourced construction and annotation design
The original construction pipeline was explicitly sense-aware. The target inventory was drawn from Macquarie Thesaurus, WordNet Affect Lexicon, and General Inquirer, with annotation performed on term–sense pairs rather than bare word types. The union of all target sets comprised 10,170 term–sense pairs (Mohammad et al., 2013).
| Source | Selection rule | Size |
|---|---|---|
| Macquarie Thesaurus | frequent unigrams and bigrams by part of speech | 800 unigrams, 787 bigrams |
| WordNet Affect Lexicon | Ekman subset, at most two senses in Macquarie | 640 word–sense pairs |
| General Inquirer | at most three senses | 8,132 word–sense pairs |
Each annotation unit was implemented as a Mechanical Turk HIT containing 12 questions: 1 word-choice question, 2 polarity questions, 8 emotion questions, and 1 “is this an emotion?” question. The word-choice question asked which of four words was closest in meaning to the target, with one correct answer given by the head word of the relevant thesaurus category. This device served three purposes simultaneously: sense disambiguation, familiarity checking, and detection of random or malicious annotation (Mohammad et al., 2013).
The affect questions used a four-level ordinal scale: not, weakly, moderately, and strongly. Each HIT was assigned to 5 independent annotators. The paper’s pilot experiments compared two phrasings—whether a term evokes an emotion and whether it is associated with an emotion—and found that associated with yielded better agreement overall. At the binary emotive/non-emotive level, the micro-average proportion of terms with majority class size 5 increased from 55.8 under evokes to 61.0 under associated (Mohammad et al., 2013).
Quality control was extensive. Of the initial 50,850 assignments, 2,666 were discarded because of unanswered questions. The authors identified 1,045 terms with defective word-choice questions and discarded 5,225 assignments for those items while still paying workers. For the remaining valid HITs, more than 95% of assignments answered the word-choice item correctly. Annotators with less than 66.67% accuracy on word-choice checks had all their assignments rejected, and an additional 111 Turkers were removed by agreement-based outlier detection (Mohammad et al., 2013).
After filtering, the final master set contained 8,883 target terms, 38,726 retained assignments, and on average 4.45 valid annotations per term (Mohammad et al., 2013).
3. Representation, aggregation, and interpretive constraints
The original annotation was graded, but the released resource is often used in a binary form. For each term–emotion pair, the original aggregation rule selected the majority class among the four intensity labels, with ties broken in favor of the stronger intensity. For binary applications, the authors collapsed no + weak into non-emotive and moderate + strong into emotive; an analogous collapse was used for polarity into non-evaluative and evaluative (Mohammad et al., 2013).
Agreement was moderate for polarity and lower for the eight emotions, which the authors attribute to out-of-context annotation and the graded nature of lexical affect. At the binary emotion level, more than 60% of terms had unanimous agreement from all five annotators, and almost 85% had agreement from at least four annotators. The micro-average Fleiss’s over the eight emotions was 0.29, interpreted in the paper as fair agreement. For polarity, the micro-average Fleiss’s was 0.54, with 0.62 for negative and 0.45 for positive (Mohammad et al., 2013).
Later guidance adds an important interpretive caveat about the published word-level version. In the author’s description, “the original annotations are at word-sense level. We created the word-level lexicon by keeping all majority-voted emotion associations for each of the word’s senses.” This means that a published word-level entry may compress multiple sense-level majority associations into a single lexical item. The result can look broader or noisier than any single contextual usage, and it should not be mistaken for a context-sensitive semantic description (Mohammad, 2020).
The distribution of affective associations in the full set also illustrates the lexicon’s semantics. At the binary level, the proportions of terms associated with each emotion were 13% for anger, 12% for anticipation, 10% for disgust, 14% for fear, 16% for joy, 12% for sadness, 6% for surprise, and 16% for trust; 54% of terms were associated with at least one emotion. This suggests a lexicon whose signal is broad but sparse at the level of individual category assignments (Mohammad et al., 2013).
4. Practical use, multilingual translation, and ethical cautions
Practical guidance on EmoLex stresses that it is best used as a resource for aggregate, comparative analysis, not as a direct detector of internal mental state. The recommended reporting style is to make claims about emotion-word usage rather than speakers’ emotions. The example given is to prefer “the use of anger words grew by 20%” over “anger grew by 20%.” The same guidance explicitly warns: “Do not draw inferences about a single sentence or utterance from the emotion associations of its constituent words.” (Mohammad, 2020)
A second practical constraint is dominant-sense dependence. EmoLex entries mostly reflect the predominant senses of the words in general language. This is usually acceptable for common words, but domain-specific corpora may differ substantially. The strongest operational recommendation is therefore to inspect the most frequent terms in the target corpus manually and remove entries that are unsuitable because of sense mismatch, inappropriate human bias, or related issues (Mohammad, 2020).
The same paper treats multilingual availability cautiously. It notes that the NRC Emotion Lexicon and NRC VAD Lexicon are available as automatic translations in over 100 languages, but also stresses that automatic translation introduces errors and that emotion associations may vary culturally across languages. The balanced position is that translated resources can be useful, but should not be treated as equivalent to native, culturally validated lexica (Mohammad, 2020).
The ethical discussion is unusually explicit. Because EmoLex is crowdsourced, it reflects the perceptions of a specific annotator pool and may encode socio-cultural bias. Majority voting can suppress minority perceptions and create the false impression that the resulting norm is uniquely correct. The recommended response is not to treat majority-voted outputs as objective truth, but as the perceptions of the majority group among the annotators, and to use disaggregated annotations where available for more nuanced analysis. Identity-related terms and socially sensitive categories are singled out as requiring particular scrutiny before deployment (Mohammad, 2020).
5. Position within the affect-lexicon landscape
EmoLex occupies a distinctive place in the broader lexicon landscape because it combines manual / crowdsourced construction, binary associations, and Plutchik’s eight emotions. Comparative work on DepecheMood++ makes this contrast explicit: EmoLex is summarized there as a lexicon of about 14k lemmas with binary associations between words and Plutchik’s 8 emotions, whereas DepecheMood++ is automatically induced, bilingual, and real-valued. On the WASSA-2017 Tweet Emotion Intensity task, the paper reports an average of 0.26 for NRC-EmoLex and 0.35 for DM++ lemma, illustrating the empirical trade-off between curated interpretability and graded automatic induction (Araque et al., 2018).
EmoLex has also been treated as one component in lexicon-integration research. In Joint Emotion Label Space Modelling for Affect Lexica, it appears as “NRC Emotion”, a 14,182-word, binary, 8-label lexicon annotated with Plutchik’s emotions, and is merged with seven other English affect lexica by a multi-view VAE into a unified resource of 30,273 unique entries. The resulting study shows that EmoLex is a useful source lexicon, but that richer real-valued categorical resources, especially NRC Hashtag Emotion, can be more predictive in downstream tasks (Bruyne et al., 2019).
Work on multilingual lexicon induction has often positioned EmoLex as the archetype of a discrete/categorical resource. The 91-language lexicon-construction framework of Buechel, Ullrich, and Hahn distinguishes its own continuous VAD + BE5 representation from resources like EmoLex in three ways: continuous ratings rather than binary category membership, a different variable inventory, and a multilingual transfer method that creates 100k to 2M entries per language rather than stopping at manual curation or direct translation. In that comparison, EmoLex functions as the clearest reference point for a manually or semi-manually built categorical emotion lexicon (Buechel et al., 2020).
A related boundary appears in Emotion Representation Mapping. That work can map between VAD/VA(D) and BE5, and thereby create graded categorical lexicons for many languages, but it is not a full EmoLex replacement because it predicts only Joy, Anger, Sadness, Fear, Disgust, omitting Surprise, Trust, and Anticipation. This clarifies that EmoLex’s categorical inventory remains broader than several influential automatic alternatives (Buechel et al., 2018).
6. Applications and research influence
EmoLex has been used not only as a lookup lexicon but also as an experimental instrument for studying aggregate emotion, embedding structure, and affect-enriched modeling. In work on emotion arcs, the lexicon is one of the central resources in a systematic evaluation of lexicon-only versus ML-based temporal tracking. That study reports that, although lexicon-only methods are “markedly poor at instance level,” they are highly accurate at generating emotion arcs by aggregating information from hundreds of instances. For EmoLex-style analysis, the practical implication is that lexical aggregation becomes much more reliable as bin size increases, particularly in large-scale tracking settings (Teodorescu et al., 2022).
EmoLex has also been used as a probe of emotional geometry in embeddings. In one study, the eight EmoLex categories—Joy, Trust, Anticipation, Surprise, Fear, Anger, Disgust, Sadness—were used to examine the entanglement of affect in raw and transformed GloVe space. The paper reports, for example, that Anticipation and Trust had a raw-space similarity of , despite both being associated with positive valence, and that projected emotion space made such relations more coherent. It also reports a compositional result in which the sum of Joy and Trust vectors had similarity .62 with Optimism and -.19 with Remorse. This suggests that EmoLex can function as an external lexical probe for the latent affective organization of semantic embeddings (Wu et al., 2019).
In retrieval-augmented LLM systems for psychotherapy transcripts, the NRC Emotion Lexicon has been used as a list of English words associated with eight emotions for embedding enrichment. In that setting, segment embeddings were updated by adding a lexicon-derived affective component, and the reported effect was typically improved empathy together with reduced coherence. A plausible implication is that EmoLex is effective at making emotional cues more salient, while not by itself guaranteeing discourse-level consistency (Rasool et al., 2024).
Taken together, these applications show that EmoLex’s significance extends beyond direct word lookup. It serves as a manually grounded categorical lexicon, a benchmark against which continuous and multilingual lexica are contrasted, a component in unified affect-lexicon models, and a probe for emotional structure in contemporary representation learning.