English. We describe the creation of HurtLex, a multilingual lexicon of hate words. The starting point is the Italian hate lexicon developed by the linguist Tullio De Mauro, organized in 17 categories. It has been expanded through the link to available synset-based computational lexical resources such as MultiWordNet and BabelNet, and evolved in a multi-lingual perspective by semiautomatic translation and expert annotation. A twofold evaluation of HurtLex as a resource for hate speech detection in social media is provided: a qualitative evaluation against an Italian annotated Twitter corpus of hate against immigrants, and an extrinsic evaluation in the context of the AMI@Ibereval2018 shared task, where the resource was exploited for extracting domain-specific lexicon-based features for the supervised classification of misogyny in English and Spanish tweets.
Italiano. L’articolo descrive lo sviluppo di Hurtlex, un lessico multilingue di parole per ferire. Il punto di partenza e` il lessico di parole d’odio italiane sviluppato dal linguista Tullio De Mauro, organizzato in 17 categorie. Il lessico e` stato espanso sfruttando risorse lessicali sviluppate dalla comunita` di Linguistica Computazionale come MultiWordNet e BabelNet e le sue controparti in altre lingue sono state generate semi-automaticamente con traduzione ed annotazione manuale di esperti. Viene presentata sia un’analisi qualitativa della nuova risorsa, mediante l’analisi di corpus di tweet italiani annotati per odio nei confronti dei migranti e una valutazione estrinseca, mediante l’uso della risorsa nell’ambito dello sviluppo di un sistema Automatic Misogyny Identification in tweet in spagnolo ed inglese.
Communication between people is rapidly
changing, in particular due to the exponential growth
of the use of social media. As a privileged place
for expressing opinions and feelings, social
media are also used to convey expressions of
hostility and hate speech, mirroring social and
political tensions. Social media enable a wide and viral
dissemination of hate messages. The extreme
expressions of verbal violence and their proliferation
in the network are progressively being configured
as unavoidable emergencies. Therefore, the
development of new linguistic resources and
computational techniques for the analysis of large amounts
of data becomes increasingly important, with
particular emphasis on the identification of hate in
language
The main objective of this work is the
development of a lexicon of hate words that can be used
as a resource to analyze and identify hate speech
in social media texts in a multilingual perspective.
The starting point is the lexicon ‘Le parole per
ferire’ developed by the Italian linguist Tullio De
Mauro for the “Joe Cox” Committee on
intolerance, xenophobia, racism and hate phenomena of
the Italian Chamber of Deputies. The lexicon
consists of more than 1,000 Italian hate words
organized along different semantic categories of hate
In this work, we present a computational version of the lexicon. The hate categories and lemmas have been represented in a machine-readable format and a semi-automatic extension and enrichment with additional information has been provided using lexical databases and ontologies. In particular we augmented the original Italian lexicon with translations in multiple languages.
HurtLex, the hate lexicon obtained with the method described in Section 3, has been tested with a corpus-based evaluation, through the analysis of a hate corpus of about 6,000 Italian tweets (Section 4.1), and through an extrinsic evaluation in the context of the shared task on Automatic Misogyny Identification at IberEval 2018, focusing on the identification of hate against women in Twitter in English and Spanish (Section 4.2).
The resource is available for download at http://hatespeech.di.unito.it/ resources.html 2
Lexical knowledge for the detection of hate
speech, and abusive language in general, has
received little attention in literature until recently.
Even for English, there are few publicly available
domain-independent resources — see for instance
the novel lexicon of abusive words recently
proposed by
Regarding Italian, apart from the lexicon of hate words developed by Tullio De Mauro described in Section 3, the literature is sparse, but it is worth mentioning at least the study by Pelosi et al. (2017) on mining offensive language on social media and the project reported in D’Errico et al. (2018) on distinguishing between pro-social and anti-social attitudes. Both the works rely on the use of corpora of Facebook posts. In particular, in Pelosi et al. (2017) the focus is on automatically annotating hate speech in a corpus of posts from the Facebook page “Sesso Droga e Pastorizia”, by exploiting a lexicon-based method using a dataset of Italian taboo expressions.
To conclude, let us mention that a new shared task on hate speech detection has been proposed in the context of the EVALITA 2018 evaluation campaign1, which provides a stimulating setting for discussion on the role of lexical knowledge in the detection of hate in language. 3
Our lexicon was created starting from preexisting lexical resources. In this section we give an overview of such resources and of the process we followed to create HurtLex.
We started from the lexicon of “words to hurt” Le
parole per ferire by the Italian linguist Tullio De
Mauro
plants (OR); animals (AN); male genitalia (ASM); female genitalia (ASF); words related to prostitution (PR); words related to homosexuality (OM). Other words and insults descriptive words with potential negative connotations (QAS); derogatory words (CDS); felonies and words related to crime and immoral behavior (RE); words related to the seven deadly sins of the Christian tradition (SVP). 3.2
WordNet
1http://www.di.unito.it/˜tutreeb/ haspeede-evalita18
WordNet is structured around synsets (sets of synonyms) and their 4 coarse-grained parts of speech: noun, verb, adjective and adverb.
MultiWordNet
BabelNet
The first step for the creation of our lexicon consisted in extracting every item from the lexicon Le parole per ferire. We obtain 1,138 items, but 1,082 unique items because several items were duplicated in multiple categories. We also removed 10 lemmas that belong to idiomatic multi-wordexpressions, e.g., “coccodrillo” (crocodile) in the expression “lacrime di coccodrillo” (crocodile tears), leaving us to 1,072 unique lemmas.
As a second step, we use MultiWordNet to augment the words with their part-of-speech tags. We use the Italian index of MultiWordNet, comprising, for each lemma, four fields containing the identifiers of the synsets in which the lemma is intended like a noun, an adjective, a verb and a pronoun. By joining this index with our lexicon, we obtain all the possible part-of-speech for 59,2 % of the lemmas, bringing the total number of lemmas from 1,072 to 1,156 to include duplicates with different part of speech. The remaining lemmas were annotated manually.
The third step consists of linking the lemmas of the lexicon with a definition. We use the BabelNet API to retrieve the definitions, aiming for high coverage. In total, we were able to retrieve a definition for 71,1% of the lemmas. Table 2 shows the Category PS RCI PA DDF DDP DMC IS OR AN
Percentage 2,76% 0,41% 5,38% 1,52% 8,55% 7,45% 1,38% 2,34% 10,07%
Category ASM ASF PR OM QAS CDS RE SVP distribution of the words not present in BabelNet across the HurtLex categories. All the information about the entries of HurtLex (lemma, part of speech, definition) and the hierarchy of categories is collected in one XML structured file for distribution in machine-readable format.
We leverage BabelNet to translate the lexicon into multiple languages, by querying the API2 to retrieve all the senses of all the words in the lexicon.
Next, we queried the BabelNet API again to retrieve all the lemmas in all the supported languages, thus creating a basis for a multilingual lexicon starting from an Italian resource.
Not surprisingly, some of the senses retrieved in the first step were unrelated to the offensive context, therefore their translation to other languages would generate unlikely candidates for a lexicon of hate words. For instance, BabelNet senses of named entities which are homograph to words in the input lexicon are extracted along with the other senses, but they are typically to exclude from a resource such as HurtLex.
Therefore, we performed a manual filtering of the senses prior to the automatic translation, with the aim of translating the original words only according to their offensive meaning. We manually annotated each pair lemma-sense according to one of three classes: Not offensive (used for senses that are totally unrelated to any offensive context), Neutral (senses that are not inherently offensive, but are linked to some offensive use of the word, for example by means of a semantic shift), and Offensive (senses that embody a crystallized offensive use of a word). To check the consistency 2https://babelnet.org/guide#java Definition Finocchio is a station of Line C of the Rome Metro.
Aromatic bulbous stem base eaten cooked or raw in salads.
Offensive term for an openly homosexual man.
Annotation Not offensive Neutral3 Offensive of the annotation, a subset of 200 senses were annotated by two experts, reporting an agreement on 87.6% of the items. Table 3 shows examples of the different annotation of senses of the same word.
After discussing the results of the pilot annotation, we decided to split the Neutral class into two additional classes. One of the new classes covers the cases where a sense is not literally pejorative, but it is used to insult by means of a semantic shift, e.g. metaphorically. The other additional class is for the senses which have a clear negative connotation, but not necessarily a direct derogatory use in a derogatory way, e.g., the main senses of “criminal”. Subsequently, the lexicon was annotated by two other experts reporting an agreement on 61% of the items. Most disagreement was concentrated in the distinctions Not offensive/Not literally pejorative (43% of the disagreement cases) and Negative connotation/Offensive (25% of the disagreement cases).
After the annotation, we discarded all the senses marked “not offensive”, and created two different versions of the multilingual lexicon in 53 languages: one containing only the translations of “offensive” senses (more conservative), and the other containing translations of “offensive”, “not literally pejorative” and “negative connotation” senses (more inclusive). 4
We evaluated the quality of the lexicon of hate words created with the method described in the previous section in two settings: by studying the occurrence of its words and their categories in a corpus of hate speech (Section 4.1), and by extracting features from HurtLex for supervised clas3The derogatory use of the word “finocchio” (fennel) in Italian is thought to originate from the middle ages, linking the fennel plant to the execution of gay men at the burning stake.
Category RE QAS CDS PS ASM OM AN PA DMC In order to gain insights on the composition of the HurtLex lexicon, we evaluated it against an annotated corpus of Hate Speech on social media, recently published by Sanguinetti et al. (2018b). The corpus consists of 6,008 tweets selected according to keywords related to immigration and ethnic minorities. Each tweet in the corpus is annotated following a rich schema, including hate speech (yes/no), aggressiveness (strong/weak/none), offensiveness (strong/weak/none), irony (yes/no) and stereotype (yes/no).
We searched the lemmas of HurtLex in the version of the hate speech corpus enriched with Universal Dependencies annotations4, by matching the pairs (lemma, POS-tag) in HurtLex with the morphosyntactic annotation of the corpus, and computed several statistics on the actual usage of such words in a specific abusive context of hate against immigrants. Table 4 shows the rate of messages in the corpus featuring words from each HurtLex category in the corpus.
For a more in-depth analysis, we also examined the relative frequency of single words in HurtLex with respect to the finer-grained annotation of the messages where they occur. Figures 1, 2, 3, 4 and 5 show examples of such analysis.
It can be noted how the relative frequency of words like “terrorismo” (terrorism), “ladro” (thief ) and “rubare” (stealing) decrease drastically as the tweets become more aggressive, offensive or with a higher level of hate speech (perhaps because, albeit negative, they are not swear words)), while 4The corpus of hate speech by Sanguinetti et al. (2018b) has been annotated with a method similar to that described in Sanguinetti et al. (2018a). words like “bastardo” (bastard) occur more as the tweets become more offensive (possibly also because they belong to the swearing sphere). Another class of words, like “zingaro” (gypsy), show a parabolic distribution. We hypothesize that this behavior is typical of words with an apparently neutral connotation that are sometimes used in abusive context with an offensive connotation. We plan to leverage this method of analysis for further studies on this line. 4.2
HurtLex was one of the resources used by the
Unito’s team to participate to the shared task
Automatic Misogyny Identification (AMI) at IberEval
2018
Valerio Basile and Viviana Patti were partially supported by Progetto di Ateneo/CSP 2016 (Immigrants, Hate and Prejudice in Social MediaIhatePrejudice, S1618 L2 BOSC 01).