
THE NRC WORDS OF WARMTH (WoW) LEXICON (aka Warmth and Competence Lexicon, aka WCST Lexicon)
-------------------------------------------------------------------------------------------

Version:	1
Released: 	June 2025
Copyright: 	2025 National Research Council Canada (NRC)
Created By: Saif M. Mohammad (Principal Research Scientist, National Research Council Canada)

Readme Last Updated: June 2025

Home Page: 	http://saifmohammad.com/WebPages/warmth.html

Contact: 	Dr. Saif M. Mohammad 
		 	saif.mohammad@nrc-cnrc.gc.ca 
		 	uvgotsaif@gmail.com

TABLE OF CONTENTS

            I.    General Description
            II.   What is Included: Forms Of The Lexicon, Files, And Format
            III.  Relevant Research Papers
            IV.   Python Code To Analyze Emotions In Text
            V.    WoW Lexicon In Various Languages 
            VI.   Using Polar Terms Only, Rescaling, Lemmatization, etc.
            VII.  Version Information And Change Log
            VIII. Other Emotion Lexicons
            IX.   Terms Of Use
            X.    Ethical Considerations


I. GENERAL DESCRIPTION
----------------------

The NRC Words of Warmth (WoW) Lexicon is a list of about 31,000 English terms (~26k
unigrams and ~5k multiword expressions (MWEs)) and real-valued scores indicating their
associations with warmth (W), sociability (S), and trust (T). The scores were obtained
from manual annotations through crowdsourcing. Since warmth analyses are often done along
with competence (C) (aka dominance) analyses, we also include in the lexicon the
competence scores for the terms (taken from the NRC VAD Lexicon v2 (Mohammad, 2025)). Thus
we also refer to this suite of lexicons as the Warmth–Competence Lexicons, or the WCST
Lexicons.

Who goes there: friend or foe?  This is a question human beings have asked from the
earliest of times to the present day.  A large body of social psychology research has
shown that warmth (W)  (friendliness, trustworthiness, and sociability) and competence (C)
(ability, power, dominance, and assertiveness) are core dimensions of social cognition and
stereotypes (Fiske et al., 2002; Bodenhausen et al., 2012; Fiske, 2018; Abele et al.,
2016; Koch et al., 2024).  That is, human beings quickly and subconsciously judge (assess)
other people, groups of people, and even their own selves along the dimensions of warmth
and competence---likely because of evolutionary pressures (MacDonald, 1992; Eisenbruch and
Krasnow, 2022).  Assessing W and C was central to early human survival (e.g., to
anticipate whether someone will help them build useful things or whether they might steal
their resources).

More recent work has started to explore how these dimensions develop, why they have
developed, and what they constitute. Of particular note, is the finding that warmth has
two distinct components: Trust (T) and Sociability (S).

Trust: honesty, fairness, dependability, reliability, morality, sincerity, honorableness, etc.

Sociability: sociableness, generosity, helpfulness, tolerance, understanding, thoughtfulness, etc.

We compiled sociability and trust association norms for over 31k English terms. The
lexicons were created by crowdsourcing and employing a slate of quality control
measures. We show that the resulting association scores have high reliability (repeating
the annotations leads to very similar scores and rankings). We created a third combined
lexicon for warmth by taking the union of the entries for the trust and sociability
lexicons.

Applications: The lexicons enable a wide variety of research and applications. Notably:

In Psychology and Social Cognition

-- What kind of trust assessments do children develop first? And what kinds are developed
later? (Trust can be of different kinds: care-based, character-based, consistency-based,
etc.) Similarly for sociability.

-- What are the mechanisms underpinning the development of WCST assessment capabilities
in children? How does exposure to different conditions impact these capabilities?

-- How different are the WCST capabilities of people in different cultures?

-- What role do differences in language play in the development of WCST capabilities?

In Computational Social Science, NLP

-- The lexicons can be used to study public discourse on topics of interest. For example,
how are the levels of warmth, competence, trust, and sociability in online discussions
about climate change or vaccines changing with time; how do these levels vary for
different stakeholders?; what sub-aspects of climate change (or vaccines or any topic of
interest) evoke the lowest amounts of warmth, competence, trust, and sociability?  etc.

-- How has the perceived WCST of a chosen target of interest (say government, banks,
immigrants, etc.) changed over the last 100 years?

In HCI and NLP

-- Understanding perceptions of WCST of people towards artificial agents.

In Digital Humanities and NLP

-- What role do warmth, trust, sociability, and competence play in developing compelling
characters and story arcs? How does this vary by genre and culture?

In Commerce

-- Tracking warmth, trust, sociability, and competence towards one’s product on social
media. This can help understand product branding, tracking user satisfaction, and
taking the appropriate remedial actions for product improvement and public-facing
communications.

-- Understanding how perceptions of warmth and competence of one’s product impact 
customer behavior.

See ethical considerations discussed further below and in the paper.

Companion lexicons:
- NRC Emotion Lexicon
- NRC Emotion Intensity Lexicon
- NRC VAD Lexicon
- WorryWords (the Word--Anxiety Association Lexicon)

Available here:
http://saifmohammad.com/WebPages/lexicons.html

This work was approved by the NRC Research Ethics Board (NRC-REB).  REB review seeks to
ensure that research projects involving humans as participants meet Canadian standards of
ethics.


II. WHAT IS INCLUDED: FORMS OF THE LEXICON, FILES, AND FORMAT
-------------------------------------------------------------

1. NRC-WCST-Lexicon-v1.0.txt: This is the main lexicon file with entries for ~31,000
English terms (words and multi-word expressions). It has five columns (separated by tabs):

- term: The English word for which W, C, S, and T scores are provided. The terms are listed in alphabetic order.
- warmth (W): score indicating association with warmth
- competence (C): score indicating association with competence
- sociability (S): score indicating association with sociability
- trust (T): score indicating association with trust 

All scores are in the range -1 to 1.

2. The directory 'OneFilePerDimension' has the same information as in
NRC-WCST-Lexicon-v1.0.txt, but in multiple files -- one for each dimension:

- warmth-NRC-WCST-Lexicon-v1.0.txt: Includes the W scores. The words are sorted in decreasing order of warmth.
- competence-NRC-WCST-Lexicon-v1.0.txt: Includes the C scores. The words are sorted in decreasing order of competence.
- sociability-NRC-WCST-Lexicon-v1.0.txt: Includes the S scores. The words are sorted in decreasing order of sociability.
- trust-NRC-WCST-Lexicon-v1.0.txt: Includes the T scores. The words are sorted in decreasing order of trust.

3. The directory 'OneFilePerDimension/PolarSubset' has a version of the lexicon that only
includes polar terms -- terms with scores <= -0.333 or >= 0.333.  (Other thresholds to
determine polar words may also be used.)

4. The directory 'CategoricalLabels' includes individual annotator labels, aggregated 
real-valued scores, standard deviation, categorical labels, etc. 

Two versions of the real-valued scores are included: 
(a) in the range -3 to 3: the average of the scores from the annotators
(b) in the range -1 to 1: a linear transformation of (a) to the range -1 to 1

OrdinalClass: A discrete class assigned to the term based on the real-valued score (a).
      The class is one of the following:
             3: very warm/trustworthy/sociable (when mean >= 2.5)
             2: moderately warm/trustworthy/sociable (when mean >= 1.5 and < 2.5)
             1: slightly warm/trustworthy/sociable (when mean >= 0.5 and < 1.5) 
             0: not associated with warm or cold/trustworthy or untrustworthy/sociable or unsociable (when mean > -0.5 and < 0.5) 
            -1: slightly cold/untrustworthy/unsociable (when mean > -1.5 and <= -0.5)
            -2: moderately cold/untrustworthy/unsociable (when mean > -2.5 and <= -1.5)
            -3: very cold/untrustworthy/unsociable (when mean <= -2.5)

MajorityLabel: This is the most common discrete label provided by the annotators.
      The label is one of the following:
			 3: very warm/trustworthy/sociable
			 2: moderately warm/trustworthy/sociable
			 1: slightly warm/trustworthy/sociable
			 0: not associated with warm or cold/trustworthy or untrustworthy/sociable or unsociable 
			-1: slightly cold/untrustworthy/unsociable
			-2: moderately cold/untrustworthy/unsociable
			-3: very cold/untrustworthy/unsociable
      We do not recommend using this label. It is provided just for information.
      Instead use either the mean real-valued score or ordinalClass.

MajorityLabelRatio: Ratio of the number of annotators that provided the majority label to
      the total number of annotators that provided a response for the term.

NumberAnnotations: the total number of annotators that provided a response for the term.

SetOfLabels: A comma-separated list of the individual responses provided by the annotators.

The directory also includes a sub-directory 'SupplementaryFiles' that has additional 
supplementary data files.
      It currently includes the files trust/sociability-highvariance-annotations.txt which has a 
	  list of terms not included in the main WCST file, because there was considerable variance in
      their annotations. 
      High variance can occur due to many reasons:
      -- not all words fit neatly in the trust/sociability dimensions;
      -- different senses of a word may be associated with very different degrees of trust/sociability;
      -- occasional poor annotation; etc.
      It is generally tricky to use these entries; and often the main WCST file should be 
      sufficient for most aggregate-level text analyses. However, entries from the highvariance
      files may also useful at times; best to exercise caution when using them.

5. The directory 'Unigrams' has a version of the lexicon that only includes unigrams
(single words).

6. The directory 'MWE' has a version of the lexicon that only includes multi-word
expressions.

PAPERS

P1. Paper-Words_of_Warmth.txt: Research paper describing the NRC WCST Lexicon.

P2. Paper-Practical-Ethical-Considerations-Lexicons.pdf: Research paper describing
practical and ethical considerations in the effective use of emotion and sentiment
lexicons.

P3. Paper-Ethics-Sheet-Emotion-Recognition.pdf: Research paper discussing ethical
considerations involved in automatic emotion recognition 'Ethics Sheet for Automatic
Emotion Recognition and Sentiment Analysis'.


III. RELEVANT RESEARCH PAPERS
-----------------------------

- Words of Warmth: Trust and Sociability Norms for over 26k English Words. Saif M.
Mohammad. In Proceedings of the 63rd Annual Meeting of the Association of Computational
Linguistics (ACL-2025), July 2025, Vienna, Austria.

If you use the lexicon in your work, then:

- Cite the paper:
		@inproceedings{wow-acl25,
    		title = {Words of Warmth: Trust and Sociability Norms for over 26k {E}nglish Words.}
    		author = {Mohammad, Saif M.},
    		booktitle = {Proceedings of the 63rd Annual Meeting of the Association of Computational
Linguistics (ACL-2025)},
    		year = {2025},
    		address = {Vienna, Austria},
    		publisher = {Association for Computational Linguistics}
		}

- Point to the lexicon homepage: 
		http://saifmohammad.com/WebPages/warmth.html

Other relevant papers:

- Best Practices in the Creation and Use of Emotion Lexicons.
  Saif M. Mohammad. Findings of the Association for Computational Linguistics: EACL 2023. 2023.

- Ethics Sheet for Automatic Emotion Recognition and Sentiment Analysis.
  Saif M. Mohammad. Computational Linguistics. 48 (2): 239–278. June 2022.


IV. PYTHON CODE TO ANALYZE EMOTIONS IN TEXT
-------------------------------------------

There are many third party software packages that can be used in conjunction with the NRC
Lexicons to analyze emotion word use in text. We recommend Emotion Dynamics:

	https://github.com/Priya22/EmotionDynamics

It is the primary package that we use to analyze text using the NRC Emotion Lexicon, the
NRC VAD Lexicon, and the Words of Warmth Lexicon.  It can be used to generate a csv file
with a number of emotion features pertaining to the text of interest, including metrics of
utterance emotion dynamics. 

See this paper for an example of the use of the lexicon to analyze emotions in text:

	Tweet Emotion Dynamics: Emotion Word Usage in Tweets from US and Canada. Krishnapriya
	Vishnubhotla and Saif M. Mohammad. In Proceedings of the 13th Language Resources and
	Evaluation Conference (LREC-2022), May 2022, Marseille, France.

	https://arxiv.org/pdf/2204.04862.pdf


V. NRC WORDS OF WARMTH LEXICON IN VARIOUS LANGUAGES 
---------------------------------------------------

The lexicon has WCST association scores for English words. Despite some cultural
differences, it has been shown that a majority of emotional norms are stable across
languages. Thus, we plan to provide versions of the lexicon in over 100 languages by
translating the English terms using Google Translate.

The NRC VAD Lexicon v1 already provides translations for a majority of the WoW words into 
these languages:

Afrikaans, Albanian, Amharic, Arabic, Armenian, Azerbaijani, Basque, Belarusian, Bengali,
Bosnian, Bulgarian, Burmese, Catalan, Cebuano, Chichewa, Corsican, Croatian, Czech,
Danish, Dutch, Esperanto, Estonian, Filipino, Finnish, French, Frisian, Gaelic, Galician,
Georgian, German, Greek, Gujarati, HaitianCreole, Hausa, Hawaiian, Hebrew, Hindi, Hmong,
Hungarian, Icelandic, Igbo, Indonesian, Irish, Italian, Japanese, Javanese, Kannada,
Kazakh, Khmer, Kinyarwanda, Korean, Kurmanji, Kyrgyz, Lao, Latin, Latvian, Lithuanian,
Luxembourgish, Macedonian, Malagasy, Malay, Malayalam, Maltese, Maori, Marathi, Mongolian,
Nepali, Norwegian, Odia, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian,
Samoan, Sanskrit, Serbian, Sesotho, Shona, Simplified, Sindhi, Sinhala, Slovak, Slovenian,
Somali, Spanish, Sundanese, Swahili, Swedish, Tajik, Tamil, Tatar, Telugu, Thai,
Traditional, Turkish, Turkmen, Ukranian, Urdu, Uyghur, Uzbek, Vietnamese, Welsh, Xhosa,
Yiddish, Yoruba, Zulu

Note some of the translations may be incorrect or they may simply be transliterations of 
the original English terms.


VI. LEMMATIZATION AND OTHER TECHNIQUES THAT MAY BE BENEFICIAL
-------------------------------------------------------------

The lexicon file can be used as is, but occasionally certain additional techniques can be
applied to make the most of it for one's specific application context.

1. LEMMATIZATION: The lexicon largely includes the base forms or lemmas of words. For
example, it may include an entry for 'attack', but not for 'attacks' or 'attacking'. In
many cases, such morphological variants are expected to have similar emotion scores. So
one can first apply a third-party lemmatizer on the target text to determine the base forms
before applying the lexicon. Note that lemmatization must be applied with care; for
example, while it it good to go from 'helplessness' to 'helpless' (they are expected to
have similar emotional connotations), 'helplessness' should not be lemmatized to 'help'
(they are expected to have markedly different emotional connotations). Further, various
factors such as tense and 'differing predominant senses for different morphological forms'
can impact emotionality. So benefits of lemmatization are limited, especially when
analyzing large pieces of text.

2. OTHER: Other techniques such discarding highly ambiguous terms (terms with many
meanings), or identifying most common terms in one's text and inspecting emotion entries
in the lexicon for those terms (and correcting entries where appropriate), etc. are also
good practice.


VII. VERSION INFORMATION AND CHANGE LOG
----------------------------------------

- Version 1 is the latest version (Released June 2025).  


VIII. OTHER EMOTION LEXICONS
----------------------------

- The NRC Emotion Intensity Lexicon is a list of English words (taken from the NRC Emotion
Lexicon and other sources) with real-valued scores of intensity for eight discrete emotions
(anger, anticipation, disgust, fear, joy, sadness, surprise, and trust). 

	Word Affect Intensities. Saif M. Mohammad. In Proceedings of the 11th Edition of the Language 
    Resources and Evaluation Conference (LREC-2018), May 2018, Miyazaki, Japan.
    
    http://saifmohammad.com/WebPages/AffectIntensity.htm

- The NRC Valence, Arousal, and Dominance (VAD) Lexicon includes a list of more than 20,000 English words and
their valence, arousal, and dominance scores.  For a given word and a dimension (V/A/D), the scores range from
0 (lowest V/A/D) to 1 (highest V/A/D).

    Obtaining Reliable Human Ratings of Valence, Arousal, and Dominance for 20,000 English Words.  Saif M.
    Mohammad. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics,
    Melbourne, Australia, July 2018.

	http://saifmohammad.com/WebPages/nrc-vad.html

Various other emotion lexicons can be found here:
http://saifmohammad.com/WebPages/lexicons.html

You may also be interested in some of the other resources and work we have done on the
analysis of emotions in text:

http://saifmohammad.com/WebPages/ResearchAreas.html
http://saifmohammad.com/WebPages/ResearchInterests.html#EmotionAnalysis


IX. TERMS OF USE 
----------------

1. Research Use: The lexicon mentioned in this page can be used freely for non-commercial
research and educational purposes.

2. Citation: Cite the papers associated with the lexicon in your research papers and
articles that make use of them. 

3. Media Mentions: In news articles and online posts on work using the lexicon, cite the
lexicon. For example: "We make use of the <resource name>, created by <author(s)> at the
National Research Council Canada." We would appreciate a hyperlink to the lexicon home
page and an email to the contact author (saif.mohammad@nrc-cnrc.gc.ca).  (Authors and
homepage information provided at the top of the README.)

4. Credit: If you use the lexicon in a product or application, then acknowledge this in
the 'About' page and other relevant documentation of the application by stating the name
of the resource, the authors, and NRC. For example: "This application/product/tool makes
use of the <resource name>, created by <author(s)> at the National Research Council
Canada." We would appreciate a hyperlink to the lexicon home page and an email to the
contact author (saif.mohammad@nrc-cnrc.gc.ca).

5. NO REDISTRIBUTION: Do not redistribute the data. Direct interested parties to the
lexicon home page.  You may not rent or license the use of the lexicon nor otherwise
permit third parties to use it. Do not upload the lexicon on a public website.

** Important **
Do not upload or store the lexicon in any location that can be scanned by companies or
other entities to create large language models. for example, do not upload to a public
website, Hugging Face, GitHub, etc. 

6. Proprietary Notice: You will ensure that any copyright notices, trademarks or other
proprietary right notices placed by NRC on the lexicon remains in evidence.

7. Title: All intellectual property rights in and to the lexicon shall remain the property
of NRC. All proprietary interests, rights, unencumbered titles, copyrights, or other
Intellectual Property Rights in the lexicon and all copies thereof remain at all times
with NRC.

8. Commercial License: If interested in commercial use of the lexicon, contact the author:
saif.mohammad@nrc-cnrc.gc.ca

9. Disclaimer: National Research Council Canada (NRC) disclaims any responsibility for the
use of the lexicon and does not provide technical support. NRC makes no representation and
gives no warranty of any kind with respect to the accuracy, usefulness, novelty,
validity, scope, or completeness of the lexicon and expressly disclaims any implied
warranty of merchantability or fitness for a particular purpose of the lexicon.  That
said, the contact listed above welcomes queries and clarifications.

10 Limitation of Liability: You will not make claims of any kind whatsoever upon or
against NRC or the creators of the lexicon, either on your own account or on behalf of any
third party, arising directly or indirectly out of your use of the lexicon. In no event
will NRC or the creators be liable on any theory of liability, whether in an action of
contract or strict liability (including negligence or otherwise), for any losses or
damages incurred by you, whether direct, indirect, incidental, special, exemplary or
consequential, including lost or anticipated profits, savings, interruption to business,
loss of business opportunities, loss of business information, the cost of recovering such
lost information, the cost of substitute intellectual property or any other pecuniary loss
arising from the use of, or the inability to use, the lexicon regardless of whether you
have advised NRC or NRC has advised you of the possibility of such damages. 

We will be happy to hear from you. For example,:
- telling us what you are using the lexicon for
- providing feedback regarding the lexicon;
- if you are interested in having us analyze your data for sentiment, emotion, and other affectual information;
- if you are interested in a collaborative research project. We regularly collaborate with graduate students,
post-docs, faculty, and research professional from Computer Science, Psychology, Digital Humanities,
Linguistics, Social Science, etc.

Email: Dr. Saif M. Mohammad (saif.mohammad@nrc-cnrc.gc.ca, uvgotsaif@gmail.com)


X. ETHICAL CONSIDERATIONS
-------------------------

Please see the papers below (included with the download) for ethical considerations
involved in automatic emotion detection and the use of emotion lexicons. (These also act
as the Ethics and Data Statements for the lexicon.)

- Words of Warmth: Trust and Sociability Norms for over 26k English Words. Saif M.
Mohammad. In Proceedings of the 63rd Annual Meeting of the Association of Computational
Linguistics (ACL-2025), July 2025, Vienna, Austria. 

- Ethics Sheet for Automatic Emotion Recognition and Sentiment Analysis.
Saif M. Mohammad. Computational Linguistics. 48 (2): 239–278. June 2022.

- Best Practices in the Creation and Use of Emotion Lexicons. Saif M. Mohammad. 
Findings of the Association for Computational Linguistics: EACL 2023. 2023.

Note that the labels for words are *associations* (and not denotations). As noted in the
paper above, they are limited by when the dataset was annotated, by the people that
annotated them, historical perceptions, and biases. (See bullets c through h in the
paper). It is especially worth noting that identity terms, such as those referring to
groups of people may be particularly prone to inappropriate biases. Further, marginalized
groups have historically faced more negative perceptions. 
