Research Offensiveness Scale

Offensiveness scale for the Fluenough adult deck: research synthesis

Source of truth: this wiki page.

The evidence points to a 4-band scale. Each band should say how a native speaker hears the word. Slurs, region and context should be separate fields on the card. No published norms exist for the app's Indian languages, so each language needs its own native-speaker rating round. No research was found for the two similarity matrices.

Only claims that passed the citation checks are used below. Where a check corrected a claim, the corrected wording is used. No primary PDF could be opened in full. Most confidence ratings are therefore Medium, based on abstracts and search excerpts.

1. The scales that exist

Research rating scales (psycholinguistics)

Source	Levels and anchors	How derived	Population	Confidence
Janschewitz, K. (2008). Taboo, emotionally valenced, and emotionally neutral word norms. Behavior Research Methods 40(4), 1065-1074. https://doi.org/10.3758/BRM.40.4.1065	1-9 for each dimension, 9 the highest. Tabooness question (quoted by a later paper): "How offensive the word is to people in general?" Offensiveness is how personally offensive or upsetting the rater finds the word. Other anchors not confirmed.	Seven dimensions. Personal reactions (personal use, offensiveness) are kept apart from perceived societal reactions (familiarity, tabooness). Familiarity and tabooness were rated higher than personal use and offensiveness, and the gap was largest for the taboo set.	78 US undergraduates, mostly English L1. 460 words, 92 of them taboo (secondary source). Only 52 of the 92 scored 4.5 or more on tabooness (citing paper).	Medium
Roest, S. A., Visser, T. A., & Zeelenberg, R. (2018). Dutch taboo norms. Behavior Research Methods 50(2), 630-641. https://doi.org/10.3758/s13428-017-0890-x	7-point Likert. Anchors not confirmed.	672 words (202 taboo), rated for general taboo, personal taboo, insult, valence and arousal. Split-half reliability was high. Many sex-related words scored high on taboo but low on insult.	87 Dutch psychology students, Erasmus University Rotterdam, Dutch L1	Medium
Beers Fägersten, K. (2007). A sociolinguistic analysis of swear word offensiveness. Saarland Working Papers in Linguistics 1, 14-37. https://publikationen.sulb.uni-saarland.de/bitstream/20.500.11880/23550/1/Beers_Faegersten.pdf	1-10: "'1' being 'Not Offensive', and '10' being 'Very Offensive'"	A word-list task plus six recorded campus dialogues. Sexual terms were rated above excretory and sacred terms. Sample means: n-word 8.5, cunt 6.6, motherfucker 5.9, bitch 5.0, hell 2.3, damn 2.3.	University of Florida undergraduates, English L1. N = 60 (33 male, 27 female).	Medium
Sulpizio, S., Günther, F., et al. (2024). Taboo language across the globe: A multi-lab study. Behavior Research Methods 56(4), 3794-3813. https://doi.org/10.3758/s13428-024-02376-6	Best-worst scaling (continuous scores, no Likert points). Paraphrased definitions: offensiveness is how inappropriate a person perceives the word to be; tabooness is how inappropriate the person believes society considers it.	Study 1 collected taboo words (N = 1046). Study 2 collected ratings on six dimensions (N = 455 per dimension). Split-half r > .80 for AoA, valence, tabooness and offensiveness in most samples.	L1 speakers of 13 languages in 17 countries, including Spain and Chile. No Indian language was seen.	Medium
Revised HurtLex. Tontodimamma, A., Fontanella, L., Anzani, S., & Basile, V. (2022; one listing 2023). An Italian lexical resource for incivility detection in online discourses. Quality & Quantity. https://doi.org/10.1007/s11135-022-01494-7. Companion paper: Tontodimamma, Anzani & Fontanella (2022), "Comparison of two annotation schemes to derive offensiveness scores in HurtLex".	Each headword has an offensiveness score. The README asks contributors for "a real value between 0 and 1".	5-point Likert rating per word, calibrated with a graded response (IRT) model and compared against best-worst scaling.	Italian, native-speaker annotators	Medium
Dewaele, J.-M. (2004). The emotional force of swearwords and taboo words in the speech of multilinguals. Journal of Multilingual and Multicultural Development 25(2/3), 204-222. https://doi.org/10.1080/01434630408666529	5-point Likert for "Do swear and taboo words in your different languages have the same emotional weight for you?" Anchors not confirmed.	Web questionnaire. Each respondent rated each of up to five languages.	1,039 adult multilinguals with many L1s	High (main finding), Medium (details)
Kapoor, H. (2016). Swears in context: The difference between casual and abusive swearing. Journal of Psycholinguistic Research. https://doi.org/10.1007/s10936-014-9345-z	Swears grouped as mild / moderate / severe. Participants rated how appropriate each was.	Casual vs abusive contexts (Study 1) and dialogue completion (Study 2).	254 participants (103 Indian, 151 non-Indian). The language of the stimuli was not confirmed and appears to be English.	Medium

Regulator and classification scales

Source	Levels and anchors	How derived	Population	Confidence
Ofcom / Ipsos MORI (2016). Attitudes to potentially offensive language and gestures on TV and radio, and its Quick Reference Guide. https://www.ofcom.org.uk/__data/assets/pdf_file/0023/91625/OfcomQRG-AOC.pdf	mild ("generally of little concern") / medium ("potentially unacceptable pre-watershed") / strong ("generally unacceptable pre-watershed") / strongest ("unacceptable pre-watershed"). Cunt and motherfucker are "strongest language, problematic for some even post-watershed".	Mixed methods with 248 participants. Built on the 2005 and 2010 studies.	UK adults, English L1, including minority-group participants	Medium
Ofcom / Ipsos MORI (2021). Public attitudes towards offensive language on TV and radio. https://www.ofcom.org.uk/__data/assets/pdf_file/0021/225336/offensive-language-summary-report.pdf	Acceptability of each word before and after the watershed. The 2021 tier labels are unresolved: one excerpt says "Mild, Moderate and Strong".	Survey run 22-26 Feb 2021: 368 respondents rated all 186 words. Plus 37 online groups and 25 depth interviews. Respondents rated "186 English words in isolation, which many found difficult".	UK adults	Medium
BBFC (2024). Classification Guidelines. https://www.bbfc.co.uk/education/issues/language	Roughly mild / moderate / strong / very strong. PG: "mild bad language only". 12A/12: "moderate bad language, and strong language may be permitted, depending on context and frequency". 15: "strong language, and very strong language may be permitted depending on the context and frequency". The 2024 wording comes from secondary summaries.	No fixed ranked list. Strength, context and tone are weighed, and aggressive, directed or sexual delivery can be rated higher. 2023 consultation of about 12,000 people. In force from 1 May 2024.	UK, English L1	Medium
Australian Government (2012). Guidelines for the Classification of Films 2012, F2012L02541. https://legislation.gov.au/F2012L02541/asmade/2012-12-19/text/original/pdf	"very mild - G, mild - PG, moderate - M, strong - MA 15+, high - R 18+, and very high - RC"	An impact scale for all classifiable elements. Impact may be lessened where a reference is verbal rather than visual.	Australia, English L1	Medium
ACMA TV descriptors (via MovieLabs). https://movielabs.com/md/ratings/v2.5/html/AU_ACMA_Ratings.html	PG: "Mild or socially offensive coarse language ... infrequently". M: "aggressive or strong coarse language should be infrequent".	Frequency and aggression change the level.	Australia, TV	Medium
MPA (via MovieLabs). https://movielabs.com/md/ratings/v2.4.4/html/US_MPAA_Ratings.html	One harsher sexually-derived word used as an expletive starts a film at PG-13. Two or more uses, or one in a sexual sense, mean R.	A count-and-sense rule.	USA	Medium
ESRB (secondary sources). https://www.animenewsnetwork.com/esrb-content-descriptors	Mild Language / Strong Language. The "Mild" label should be checked on esrb.org.	Descriptors	USA and Canada	Medium
FCC. https://fcc.gov/consumers/guides/obscene-indecent-and-profane-broadcasts	Binary. "Profane" means "grossly offensive" language, restricted from 6 a.m. to 10 p.m.	Law and Commission orders (FCC 04-43, 2004)	USA	Medium
India CBFC and the Cinematograph (Certification) Rules 2024	No graded language scale. Certificates are U / UA 7+ / UA 13+ / UA 16+ / A / S; the UA age markers are advisory.	Guidelines bar vulgarity and double-meaning words. A 28-word list was circulated on 12 Feb 2015 and then put on hold; it was later reported scrapped.	India, all film languages	Medium
India BCCC (set up in 2011)	None. It acts on complaints.	Advisories, for example on "crass, abusive, vulgar and double-meaning language with sexual overtones" (Dec 2012). First fine: Rs 2.5 lakh in 2013.	India, Hindi and regional TV	Medium

Category resources (no severity grade)

Source	What it gives	Languages	Confidence
HurtLex. Bassignana, Basile & Patti (2018), CLiC-it. https://github.com/valeriobasile/hurtlex	17 category codes (e.g. PS ethnic slurs, DDF/DDP disability, ASM/ASF genitalia, PR, OM, AN, CDS, RE) plus a stereotype flag. Its "conservative" vs "inclusive" field marks translation scope, not severity.	53 languages, including Hindi, Bengali and Spanish (machine-translated). No Telugu, Marathi, Kannada, Gujarati or Assamese.	High
SHARE. Plaza-del-Arco et al. (2022), LREC, pp. 1307-1316. https://aclanthology.org/2022.lrec-1.139	10,125 Spanish harmful terms and expressions, labelled by 5 annotators (kappa 78.8%). No severity grade found. Licence: research only, no redistribution.	Spanish	Medium
OffendES. Plaza-del-Arco et al. (2021), RANLP, pp. 1096-1108. https://aclanthology.org/2021.ranlp-1.123	Comment-level classes OFP / OFG / OFO / NOE / NO. NOE covers "rude words, blasphemes, or swearwords" used without aim to offend, "usually with a positive connotation".	Spanish (Spain)	Medium
HASOC 2019, 2021 and 2023 (FIRE)	HOF vs NOT. Hindi and English also have HATE / OFFN / PRFN. Marathi, Assamese, Bengali and Gujarati have binary labels only.	Hindi, Marathi, Bengali, Assamese, Gujarati	Medium
DravidianLangTech 2021 and HOLD-Telugu 2024	Offensive-target classes, no severity. Kannada: 7,772 comments.	Kannada, Telugu (plus Tamil, Malayalam)	Medium
BD-SHS. Romim et al. (2022), LREC. https://aclanthology.org/2022.lrec-1.552	Hate vs not, then target, then type (slander, religion, gender, call to violence). 50,281 comments.	Bengali (Bangladesh)	Medium
MACD. Gupta et al. (2022), NeurIPS Datasets and Benchmarks	About 150K ShareChat comments, binary abusive vs not (49% abusive)	Hindi, Tamil, Telugu, Malayalam, Kannada	Medium
MOLD. Gaikwad et al. (2021), RANLP, pp. 437-443. https://arxiv.org/abs/2109.03552	2,499 tweets collected by searching 22 Marathi curse words. Six native annotators labelled each OFF or NOT.	Marathi	Medium

2. What matters for this app

Learners hear these words differently from natives

• Confirmed (High). In Dewaele (2004), "the perceived emotional force of S-T words is highest in the L1 and gradually lower in languages learned subsequently".
• Classroom learners rated force lower than people who learned the language naturalistically (Medium).
• Nuance (Medium). In Dewaele (2016), LX (non-native) users of English rated most of 30 words as more offensive than L1 users did. They underrated only the single most offensive word. They were less sure of meanings. Dewaele links the overrating to a "red flag" that classroom warnings attach to these words.
• Dewaele, J.-M. (2016). Thirty shades of offensiveness. Journal of Pragmatics 94, 112-127. https://doi.org/10.1016/j.pragma.2016.01.009
• Consequence. Learners err in both directions: they overrate mid-level words and underrate the top one. The level on a card must state how a native speaker hears the word, not how strong it feels to the learner. The top band needs its own explicit warning.
• Also (Medium). In Jay & Janschewitz (2008), the offensiveness ratings of non-native speakers depended on their experience with English. Jay, T., & Janschewitz, K. (2008). The pragmatics of swearing. Journal of Politeness Research 4(2), 267-288. https://doi.org/10.1515/JPLR.2008.013

Regional and social variation

• Across countries (Medium). Sulpizio et al. (2024) report "a significant amount of cross-country variability in words' tabooness and offensiveness". One check found that maricón is similarly taboo in Spain and Chile but more offensive in Spain, and that puto is more offensive in Chile. A second check could not verify this (Low to Medium). Spanish cards need a region field.
• Context (Medium). Jay & Janschewitz (2008): the appropriateness of swearing depends on the speaker-listener relationship, the social-physical context and the particular word. Ofcom found rating words in isolation difficult. The BBFC raises the level for aggressive or directed delivery. MPA rules treat expletive and literal sexual use differently.
• Slurs move separately from swearing (Medium). Ofcom 2021: audiences are more relaxed about most swearing, but concern about discriminatory language is rising. In 2020, 1% of complaints were about swearing and 8% about racial discrimination. Most of the 2016 "strongest" tier was discriminatory terms (per press summaries).
• Indian specifics (Medium to Low). CBFC and BCCC both single out double-meaning language. A BCCC advisory on SC/ST (caste) language is reported only by a weak, undated source. Caste slurs are a class with no UK equivalent and must be covered.
• Content pattern (Medium). In Sulpizio et al. (2024), the most taboo and offensive words were almost exclusively about female genitalia or female family members. This matches common Hindi and Bengali abuse, but that comparison is my inference (Low).
• Gender (Medium). In Dewaele (2004), women rated emotional force higher; which languages this covers is unconfirmed. Kapoor (2016) found women rated swearing less appropriate. In Beers Fägersten (2007), the gender difference was not significant (p = 0.0707).
• Small core vocabulary (Medium). In Jay (2009), 10 words account for 80% of public swearing in US recordings. The counts from 1986, 1997 and 2006 showed the set stable over 20 years. A short deck can cover most of what a learner will hear. Jay, T. (2009). Perspectives on Psychological Science 4(2), 153-161. https://doi.org/10.1111/j.1745-6924.2009.01115.x

3. Options for the app's scale (best first)

#	Option	Main downside	Confidence
1	Four bands, mild / medium / strong / strongest, defined as "how a native speaker hears it in general" (Janschewitz's tabooness question). Separate fields on each card: type (swear, slur, sexual, family insult, using HurtLex-style categories), "can be friendly among peers" (the OffendES NOE idea), and a region note. Slurs against caste, religion, ethnicity, disability or gender always go in strongest.	The bands are coarse and one level hides context. Ofcom's 2016 tiers came from a small deliberative sample (248), and Ofcom may have relabelled the tiers in 2021.	Medium
2	Same as option 1, but with BBFC/Australian wording (mild / moderate / strong / very strong).	Same downsides as option 1. "Very strong" sounds less absolute than "strongest", which matters because learners underrate exactly the top word (Dewaele 2016).	Medium
3	Three bands, mild / moderate / severe (as in Kapoor 2016).	No top "never use" tier separate from strong words. Ofcom found some words "problematic for some even post-watershed".	Medium
4	A numeric score (1-9 or 0-1) from a rating study, shown as is.	Too fine to read on a card. It needs many raters per language, and no Indic norms exist to start from. Better kept as internal data behind option 1.	Medium
5	Two levels (ESRB style) or binary (FCC, CBFC, HASOC).	Too coarse to stop learners misjudging mid-level words.	High

How a reviewer per language would assign levels (Low: inference from the methods above)

1. Draft a candidate list from attested sources:
   • Hindi, Bengali and the other Indian languages: HASOC, MACD, BD-SHS and MOLD data.
   • Spanish: SHARE (licence permitting) and the Sulpizio et al. (2024) Spain and Chile items.
2. Have several native speakers per language rate each word, ideally from more than one region. Ask Janschewitz's question: "How offensive is this word to people in general?" Use a 1-9 or 5-point Likert scale. This follows the Revised HurtLex method.
3. Map the median rating to the four bands, with the cut points written down. Put slurs in strongest by rule.
4. Where raters from different regions disagree by a band or more, add a region note instead of averaging. For Spanish, check against the Sulpizio et al. Spain and Chile data.
5. Add a one-line context note (who says it to whom, friendly vs hostile). Record reviewer and date on the card.

4. Dropped claims and gaps

Dropped or changed after checking

• Dewaele (2004): the anchor labels "does not feel strong ... very strong" are unconfirmed and dropped.
• Dewaele (2016): "LX users avoid the most offensive words" is not in the abstract. It actually says LX users reported more use of less offensive words, and L1 users more use of the more taboo words.
• Sulpizio et al. (2024): dropped the exact tabooness wording ("not acceptable ... in most social situations"), the claim that disability words rank among the most offensive in most languages, the co-author names Badan, Basclain and Brysbaert, and the claim that scores are standardized within each language.
• Mexican Spanish ANEW: the claim of "regional differences in arousal" is unconfirmed and dropped.
• Jay (1992), Cursing in America: the page range 159-194 and the chapter DOI are unconfirmed. Cite the book DOI 10.1075/z.57.
• BBFC: the "infrequent use of strong language (e.g. fuck)" and "strongest terms (e.g. cunt)" thresholds are 2002-era wording, not the 2024 guidelines.
• FCC: the phrase "personally reviling epithets" is unconfirmed (it comes from Chaplinsky, 1942). The claim that the FCC treats slurs as reviling epithets is dropped.
• Australia: F2012L01934 is the computer-games instrument, not the films one. Corrected to F2012L02541.
• CBFC: the citation "Rajya Sabha PQ 154 (23 May 1990)" was not found and is dropped.
• Cinematograph Rules 2024: the language rules for the A certificate come only from Wikipedia and are dropped.
• Ofcom 2016: the placements of cow, crap and cock rest only on press summaries, and the exact quotes about discriminatory words and context were not seen verbatim.
• Beers Fägersten (2007): the finding that context lowers offensiveness is attributed to Beers Fägersten (2012) and Mormol (2018). The gender effect was not significant.

Gaps

• No primary full text was read. Scale anchors are unconfirmed for Janschewitz, Dewaele 2016, Sulpizio, Roest and Tontodimamma.
• The 2021 Ofcom tier labels are unresolved, and so is which South Asian terms Ofcom rates.
• There are no human-rated offensiveness norms for Hindi, Bengali, Telugu, Marathi, Kannada, Gujarati or Assamese (an absence finding, so Low). Kapoor (2016) appears to use English stimuli.
• The RAE "malsonante" and ASALE "vulg." / "desp." labels were not checked. Spain's regulator, CNMC, was not covered.
• Similarity matrices (the user's request): no research was found on words that sound like, or are spelled like, offensive words, in any of the app's languages. The only related idea is the CBFC and BCCC concern with double-meaning words. Both matrices will have to be built in-house. A possible method (Low): one list from phonetic distance over IPA transcriptions, one from spelling distance over native script and romanisation, both reviewed by native speakers. A separate literature search is still needed.

See also: Sources-and-Attributions · Home