Content uploaded by Jonas Sjöbergh
Author content
All content in this area was uploaded by Jonas Sjöbergh
Content may be subject to copyright.
A Multi-Lingual Dictionary of Dirty Words
Jonas Sj¨
obergh and Kenji Araki
Graduate School of Information Science and Technology
Hokkaido University
Sapporo, Japan
{js, araki}@media.eng.hokudai.ac.jp
Abstract
We present a multi-lingual dictionary of dirty words. We have collected about 3,200 dirty words in several languages and built a database
of these. The language with the most words in the database is English, though there are several hundred dirty words in for instance
Japanese too. Words are classified into their general meaning, such as what part of the human anatomy they refer to. Words can also
be assigned a nuance label to indicate if it is a cute word used when speaking to children, a very rude word, a clinical word etc. The
database is available online and will hopefully be enlarged over time. It has already been used in research on for instance automatic joke
generation and emotion detection.
1. Introduction
Dictionaries can be tremendously useful in many language
processing tasks, and are also useful sources of informa-
tion for human readers. One category of words that is often
not included in very large amounts in most dictionaries are
“dirty words”. By that we mean words that are generally
not used in polite company, referring to for instance sex-
ually related things, bodily functions, or cuss words and
insults.
We have collected dirty words in several languages and
built a multi-lingual dictionary linking words in different
languages with similar meanings. Other types of annota-
tion is also possible, many words are for instance annotated
with the nuance they carry, i.e. if a word is really rude or
perhaps a euphemism and so on. Information on whether a
word is unambiguously a dirty word or can also have non-
dirty meanings can also be added.
This dictionary can of course be used in translation appli-
cations, to find appropriate translation candidates for words
and phrases that are perhaps hard to find in other dictionar-
ies. A dictionary of dirty words is also useful in many other
ways, there have for example been quite a few products
launched world wide were the producer later found out that
in some of the markets the product name was a dirty word
or in a similar way gave bad impression. Such things could
be mitigated by having access to a large resource of dirty
words in different languages.
The dictionary can also be used in monolingual natural
language processing applications where information about
dirty words is useful. Three examples of areas where our
database has already been used are:
Humor recognition (Sj ¨
obergh and Araki, 2007b). In a
machine learning approach to classify texts as either
jokes or not, some features based on the presence of
dirty words in the text were used. A high presence of
dirty words was useful as an indication that the text
was a joke.
Humor generation (Sj ¨
obergh and Araki, 2007a; Sj¨
obergh
and Araki, 2008). Dirty words and euphemisms are
common in jokes, and are thus useful in automatic joke
generation. A system generating rather weak puns was
perceived as slightly funnier if the punch line of the
pun was a dirty word. Other joke generation meth-
ods were based on changing parts of idioms to similar
sounding euphemisms (dirty words) for sex etc.
Emotion recognition (Ptaszynski et al., 2007). In a simi-
lar way to the humor recognition case, sentences with
dirty words tended to carry emotive content in an ex-
periment on emotion detection in Japanese.
Other uses of dirty words, though not our database, in natu-
ral language processing include detecting if a message is
a flame (Spertus, 1997), and other machine learning ap-
proaches to humor recognition (Mihalcea and Strapparava,
2005).
2. Collecting the Dirty Words
We have collected dirty words and short phrases from sev-
eral different sources and in several different languages to
add to our database. These were then annotated manually
with various types of information. The original intended
use was humor generation and humor recognition in En-
glish and Japanese, so these two languages received the
most focus.
The single largest source of dirty words was a list collected
by George Carlin1, containing about 2,400 dirty word ex-
pressions in English. Most of these are euphemisms, tend-
ing towards joke like expressions, for example “trouser ana-
conda”.
For Japanese we extracted all words in the EDICT
dictionary (Breen, 1995) marked with the “vulgar” flag, and
also added various short lists of dirty words found on the
Internet. We also had several native speakers of Japanese
simply write down a lot of dirty words that they could come
up with by looking at the other words in the list.
We have also found useful information in the Alternative
Dictionaries2, the Swearsaurus3, and Wikicurse4, which are
1http://www.georgecarlin.com/dirty/2443.html
2http://www.notam02.no/˜
hcholm/altlang/
3http://www.insultmonger.com/swearing/
4http://www.wikicurse.com/
509
collections of “bad words” in many languages. There are
also many bad words in these resources in other languages
that we have not added to our database, mainly because of
a lack of native speakers to check if the words are really of
the kinds we want. These could of course be added later if
one so wishes.
After collecting the dirty words, they have been annotated
by hand with different types of information. Not all words
are annotated with all types of information yet. Annotation
regarding the meaning, nuance, and ambiguity of a word or
phrase is possible.
3. Structure
In the dictionary the words are annotated with the follow-
ing information: how to write the word, how to pronounce
the word, the meaning of the word, the nuance of the word,
whether the word is ambiguous in the sense that it has non-
dirty meanings too, what language the word comes from,
and the part of speech of the word. The dictionary also con-
tains many multi-word expressions, though they are treated
like one unit and we will refer to these too as “words” in this
paper except when talking specifically about the number of
words in the expressions.
The only information that is mandatory for a word is how
to write it. All other fields can be left unspecified, though
so far all words are also annotated with the language they
come from. Pronunciation is currently only provided for
the Japanese words, for which it can be non-trivial to fig-
ure out the reading of the ideographic characters used for
writing. The same ideographic character sequence can
have several different readings, some of which can be dirty
words while others are not.
The meanings are specified by links to special “interlin-
gua” like objects. These describe the general meaning of
a word using English (though adding explanations in other
languages too is of course also possible). Currently only
the general meaning is given, such as what part of the hu-
man anatomy a word refers too or that it is some form of
fornication. More detailed classifications can be done later
if it is found to be necessary for a specific application.
These interlingua meaning objects are also grouped into
three general groups: sex related, bodily functions, and in-
sults. The meanings of some words do not fit into any of
these three categories, in which case what group the mean-
ing belongs to is left unspecified. An example of an inter-
lingua object is “cuss word interjection” for things such as
“dammit”.
The nuances of words indicate if a word is a clinical word
used for instance in doctor patient conversations, if it is a
“cute” word used when speaking to children, if it is a eu-
phemism, or if it is an “extra bad” word (very rude), etc.
We have found use for this type of information ourselves in
other experiments, for instance in humor generation where
really bad words tended to offend rather than entertain, and
clinical words did not sound very funny either. This type
of information could also be useful for instance when se-
lecting from different translation candidates, so as to find a
translation with a similar nuance in the target language.
Nuance can of course be hard to determine for some words.
Words can be perceived as very rude by some people and
Language Words
English 2402
Japanese 397
Swedish 158
Bulgarian 147
Polish 125
Total 3229
Table 1: The number of words and expressions in different
languages currently in the database.
as fairly OK by others. The same word can also be very
rude in some contexts and not rude at all in other contexts.
Currently we have not made any efforts at more detailed
descriptions of nuances, but if there is interest in the future
it could be added later. Many words are still unproblematic
though, and can be fairly easily annotated with a simple
description of their nuance.
The ambiguity field indicates if a word has both dirty mean-
ings and non-dirty meanings. It is possible to just note that
both are possible (e.g. “pussy”) or if a word is always dirty
(e.g. “fuck”), and it is also possibly to specify in more de-
tail if the dirty meaning is much more common than any
non-dirty meanings (as perhaps “cock”), or if the word is
generally not dirty but can be in special contexts (e.g the
words “it” or “there” in many languages). Which meaning
is more common can of course in many cases be rather hard
to judge, in which case just noting that the word is ambigu-
ous is enough.
Part of speech is currently mostly not given, though the field
was added since this information was available in some of
the sources we used to build the dictionary and was used
in some of our text generation experiments using the data.
Other grammatical information could also be useful but is
currently not given. For instance, it could be useful to know
what forms the mutli-word expressions can take, if they
can have parts of the expression modified by adjectives etc.
without losing their dirty impression, etc.
The language field simply indicates what language a word
comes from. If the same string is a dirty word in several
languages, a separate entry is made in the database for each
language. The same is true if a word can mean several dif-
ferent dirty things in the same language.
The dictionary is stored in an SQL database. The database
has a primitive web interface that allows searching the
database, downloading the whole dictionary, adding new
words and meanings, and annotating existing words with
meanings, nuances, etc.
4. Statistics
Some statistics showing the contents of the dictionary can
be found in Tables 1 to 5. As can be seen in Table 1, the bulk
of the words are currently English words. That English has
the by far largest amount of words is probably caused by
English being the most widely used language both on the
Internet (a good source of dirty words) and in natural lan-
guage research (where one could perhaps expect such re-
sources to show up). Somewhere around 150 words seems
510
Category Words
Sex 2652
Bodily Functions 261
Insults 211
Unspecified 105
Table 2: The general grouping of the meanings of the words
in the database.
Nuance Words
Euphemism 1462
Fairly bad word 87
Used normally 70
Children’s speak 24
Very bad 22
Clinical 13
Unspecified 1551
Table 3: The nuances of the words.
to be the limit where people who collect dirty words mainly
for fun get tired and give up. The differences in number of
words between the languages in our database most likely
do not reflect any actual cultural differences in the amount
of dirty words. It is simply an effect of what purposes we
have used the data for and thus what languages we put the
most work into collecting dirty words for.
In Table 2, it can be seen that rather unsurprisingly the over-
whelming majority of the words are sex related words. This
varies a bit between languages though, some of the lists of
“dirty words” we have collected have contained mostly in-
sults to be hurled at other people to make them angry. Many
of these do have sexual connotations though.
Of the words that have so far been annotated with their nu-
ances, euphemisms are by far the most common, see Table
3, though only about half the words have been annotated so
far.
The annotation of the ambiguity of the words has only cov-
ered about a third of the data so far, see Table 4. Thus far,
about half the words are ambiguous, though a large part
of the remaining words are rather long euphemistic expres-
sions that are likely not very ambiguous.
In the final table, Table 5, statistics on the lengths of the
expressions is presented. About half the dictionary is made
Ambiguity Words
Ambiguous, can be either dirty or not 686
Always dirty 475
Ambiguous, not-dirty meaning most common 26
Ambiguous, dirty meaning most common 17
Unspecified 2025
Table 4: The ambiguity of the words.
Length in Words Expressions
1 1722
2 902
3 398
4 132
5 47
6 20
7 5
8 2
10 1
Multi-Word 1507
Table 5: The lengths of the expressions in the database.
Average length is 1.8 words.
up of multi-word expressions, though not many are made
up of four or more words. The longest expression so far is
“choke the sheriff and wait for the posse to come”, which
is an English expression for (male) masturbation.
This data gives a general idea of the contents of the
dictionary, but one should keep in mind that it is a bit com-
plicated to gather this type of information from such dif-
ferent languages. The Swedish part contains many quite
long compound words treated as only one word, while a
similar word in English would be a multi-word expression.
And Japanese has no space between words at all, so only a
quick cursory check of roughly how many “words” a phrase
contains was done for the Japanese part.
5. Availability
The dictionary is freely available on the web5, though the
web interface is still very primitive. It is possible to down-
load the whole dictionary, and also to add new words,
change or add more annotations to the words already in the
database etc.
We plan to extend the database ourselves, both by adding
more words to the languages already included but and by
adding more languages. Any volunteers are of course also
welcome to add more data too. We also plan to improve the
web interface.
6. Conclusions
We presented a dictionary of dirty words in several lan-
guages. The meanings of the words are linked, so it can be
used to find for instance translations of dirty words in other
languages. The nuances of the words (really rude, clinical,
euphemism, etc.) are also annotated, which can help in se-
lecting an appropriate translation. Words that are ambigu-
ous in the sense that they have other non-dirty meanings
too, can also be annotated with this information.
The dictionary contains about 3,000 words and expressions,
2,400 in English, 400 in Japanese, and slightly over 100
words each in Bulgarian, Polish, and Swedish. It is freely
available on the Internet, and it is also possible for volun-
teers to contribute new words to the dictionary.
5http://dr-hato.se/projects/dirtywords/
511
So far, the contents of the dictionary have mainly been used
in monolingual applications, for instance humor generation
and emotion classification.
Acknowledgements
This work was done as part of a project funded by the
Japanese Society for the Promotion of Science (JSPS). We
would like to thank some of the anonymous reviewers for
interesting suggestions for extending our work. We would
also like to thank the volunteers who have contributed dirty
words to the dictionary, especially Svetoslav Dankov who
also helped out with various practical things.
7. References
Jim Breen. 1995. Building an electronic Japanese-English
dictionary. In Japanese Studies Association of Australia
Conference, Brisbane, Australia.
Rada Mihalcea and Carlo Strapparava. 2005. Making
computers laugh: Investigations in automatic humor
recognition. In Proceedings of HLT/EMNLP, Vancou-
ver, Canada.
Michal Ptaszynski, Pawel Dybala, Wen Han Shi, Rafal
Rzepka, and Kenji Araki. 2007. Lexical analysis of
emotiveness in utterances for automatic joke generation.
ITE Technical Report, Vol. 31, No. 47, pages 39–42,
ME2007-204.
Jonas Sj¨
obergh and Kenji Araki. 2007a. Automatically
creating word-play jokes in japanese. In Proceedings of
NL-178, pages 91–95, Nagoya, Japan.
Jonas Sj¨
obergh and Kenji Araki. 2007b. Recognizing hu-
mor without recognizing meaning. In Francesco Ma-
sulli, Sushmita Mitra, and Gabriella Pasi, editors, Pro-
ceedings of WILF 2007, volume 4578 of Lecture Notes
in Computer Science, pages 469–476, Camogli, Italy.
Springer.
Jonas Sj¨
obergh and Kenji Araki. 2008. What is poorly said
is a little funny. In Proceedings of LREC-2008, Mar-
rakech, Morocco.
Ellen Spertus. 1997. Smokey: Automatic recognition of
hostile messages. In Innovative Applications of Artifi-
cial Intelligence (IAAI), pages 1058–1065, Providence,
Rhode Island.
512
