English. In this paper we describe our submission to the shared task of Automatic Misogyny Identification in English and Italian Tweets (AMI) organized at EVALITA 2018. Our approach is based on SVM classifiers and enhanced by stylistic and lexical features. Additionally, we analyze the use of the novel HurtLex multilingual linguistic resource, developed by enriching in a computational and multilingual perspective of the hate words Italian lexicon by the linguist Tullio De Mauro, in order to investigate its impact in this task.
Hate Speech (HS) can be based on race, skin color,
ethnicity, gender, sexual orientation, nationality,
or religion, it incites to violence and
discrimination, abusive, insulting, intimidating, and
harassing. Hateful language is becoming a huge
problem in social media platforms such as Twitter and
Facebook
Misogyny can be linguistically manifested in
numerous ways, including social exclusion,
discrimination, hostility, threats of violence and
sexual objectification
The first shared task specifically aimed at
Automatic Misogyny Identification (AMI) took place
at IberEval 20182 within SEPLN 2018 considering
English and Spanish tweets
The aim of the AMI task is to detect
misogynous tweets written in English and Italian (Task
A)
The organizers provided the same amount of data for both languages: 4,000 tweets in the training set and 1,000 in the test set. The label distribution for Task A is balanced, while in Task B the distribution is highly unbalanced for both misogyny behaviors and targets. 3
Description of the System We used two Support Vector Machine (SVM) classifiers which exploit different kernels: linear and radial basis function (RBF) kernels.
SVM with Linear Kernel. Linear kernel was used to find the optimal hyperplane when SVM was firstly introduced in 1963 by Vapnik et al., long before Cortes and Vapnik (1995) proposed to use the kernel trick. Joachims (1998) recommends to use linear kernel for text classification, based on the observation that text representation features are frequently linearly separable. SVM with RBF Kernel. Choosing the kernel is usually a challenging task, because its performance will be dataset dependent. Therefore, we also experimenteed with a Radial Basis Function (RBF) kernel, which has been already proven as an effective classifier in text classification problems. The drawback of RBF kernels is that they are computationally expensive and obtain a worse performance in big and sparse feature matrices. 3.1
Features We employed several lexical features, performing a simple preprocessing step including tokenization and stemming, using the NLTK (Natural Language Toolkit) library4. A detailed description of the features employed by our model follows. Bag of Words (BoW). We used bags of words in order to build the tweets representation. Before producing the word vector, we changed all the characters from upper to lower case. Our vector space consists of the count of unigrams and 4https://www.nltk.org/ bigrams as a representation of the tweet. In addition, we also employed Bag of Hashtags (BoH) and Bag of Emojis (BoE) features, which are built by using the same technique as BoW, focusing on the presence of hashtags and emojis.
Swear Words. This feature takes into account the presence of a swear word and the number of its occurrences in the tweet. For English, we took a list of swear words from www.noswearing.com, while for Italian we gathered the swear word list from several sources5 including a translated version of www.noswearing.com’s list and a list of swear words from Capuano (2007).
Sexist Slurs. Beside swear words, we also considered sexist words, that are specifically targeting women. We used a small set of sexist slurs from previous work by Fasoli et al. (2015). We translated and expanded that list manually for our Italian systems. This feature has a binary value, 1 when at least one sexist slur presence on tweet and 0 when there is no sexist slur on tweet.
Women Words. We manually built a small set of
words containing synonyms and several words
related to word “woman" in English and “donna" in
Italian. Based on our previous work
Surface Features. We also considered several surface level features including: upper case character count, number of hashtags, number of URLs, and the length of the tweet counting the characters.
Hate Words Lexicon. HurtLex
5https://www.parolacce.org/2016/12/ 20/dati-frequenza-turpiloquio/ and https: //it.wikipedia.org/wiki/Turpiloquio_ nella_lingua_italiana Task A
English
Italian Misogynistic kernel (0:765 accuracy), while for Italian the best result has been obtained by runs #2 and #3 with the Linear kernel (0:893 accuracy). Different sets of categories from HurtLex were able to improve the classifier performance, depending on the language.
In order to classify the category and target of misogyny (Task B), we adopted the same set of features as Task A. Therefore, we did not build new systems specifically for Task B.
We experimented with different selections of categories from the HurtLex lexicon, and identified the most useful for the purpose of misogyny identification. As it can be seen in Table 3, the main categories are: physical disabilities and diversity (DDP), words related to prostitution (PR), words referring to male genitalia (ASM) and female genitalia (ASF). But also: derogatory words (CDS), words related to felonies and crime, and also immoral behavior (RE).
Language Systems Accuracy Bag of Word Bag of Hashtags Bag of Emojis S.W. Count S.W. Presence Sexist Slurs Woman Word Hashtag Link Presence Upper Case Count Text Length ASF Count PR Count OM Count DDF Count CDS Count DDP Count AN Count ASM Count DMC Count IS Count OR Count PA Count PS Count QAS Count RCI Count RE Count SVP Count Kernel
English run1 run2 0:765 0:72
Italian run3 run1 run2 run3 0:744 0:786 0:893 0:893 X X X X X X
X X X X X X X X X X X X X X X X X
X X X X X
X X X X X - X X X - - - X X X X X - - - - - - - - X X - X X X X - - - X X X - - - - - - X X - - - - - X X - - - - - - - - X X - - - - - - - - - - - - - - - - - - - - - X X - - - - -
RBF Linear RBF RBF Linear Linear
the test sets. Our best system in Task A ranked 3rd in Italian (0:839 in accuracy for run3) and 13th in English (0:621 in accuracy for run3). Interestingly, our best result on both languages were obtained by the best configuration submitted at the IberEval campaign. However, our English system performance was way worse compared to the result of IberEval (accuracy = 0:814). We will try to analyze this problem in the Section 6. Team himani.c.run3.tsv himani.c.run2.tsv AMI-BASELINE hateminers.c.run3 hateminers.c.run1 SB.c.run2.tsv himani.c.run1.tsv SB.c.run1.tsv hateminers.c.run2 SB.c.run3.tsv resham.c.run2 resham.c.run1 bakarov.c.run1 resham.c.run3 RCLN.c.run1 ITT.c.run2.tsv bakarov.c.run2 14-exlab.c.run1 bakarov.c.run3 14-exlab.c.run3 ITT.c.run1.tsv ITT.c.run3.tsv 14-exlab.c.run2
ENGLISH
Avg. 0.406 0.377 0.370 0.369 0.348 0.344 0.342 0.335 0.329 0.328 0.322 0.316 0.309 0.283 0.280 0.276 0.275 0.260 0.254 0.239 0.238 0.237 0.232 Cat. 0.361 0.323 0.342 0.302 0.264 0.282 0.280 0.282 0.229 0.269 0.246 0.235 0.260 0.214 0.165 0.173 0.176 0.124 0.151 0.107 0.140 0.138 0.205 Targ. 0.451 0.431 0.399 0.435 0.431 0.407 0.403 0.389 0.430 0.387 0.399 0.397 0.357 0.353 0.395 0.379 0.374 0.395 0.356 0.371 0.335 0.335 0.258 around 0:5 for Italian). Several under-represented classes such as DERAILING and DOMINANCE are very difficult to be detected in category classification (See Table 1 for details). Similarly, the label distribution was very unbalanced for target classification, where most of the misogynous tweets are attacking a specific target (ACTIVE).
Several features which focus on the use of offensive words were proven to be useful in English. For Italian, a simple tweet representation which involves Bag of Words, Bag of Hashtags, and Bag of Emojis already produced a better result than the baseline. Some of the HurtLex categories that were improving the system’s performance during training did not help the prediction on the test set CDS, ASM for Italian). However, similarly to the egories and discriminate whether the target is acSpanish case, the system configuration which utitive or passive. Both subtasks for both languages lized ASF, PR, and DDP obtained the best result have very low baselines (below 0:4 for English and in Italian.
We performed an error analysis on the gold standard test set, and analyzed 160 Italian tweets that our best system configuration mislabelled. The label “misogynistic” was wrongly assigned to 147 instances (false positives, 91.9% of the errors), while the contrary happened only 13 times (false negatives, 8.1% of the errors). The same situation happened in the English dataset, but with a less striking impact, with 228 false positives (60.2% of the errors), 151 false negatives (39.8% of the errors). In this section we conduct a qualitative error analysis, identifying and discussing several factors that contribute to the misclassification.
Presence of swear words. We encountered a lot
of “bad words” in the dataset of this shared task
for both English and Italian. In case of abusive
context, the presence of swear words can help to
spot abusive content such as misogyny. However,
they could also lead to false positives when the
swear word is used in a casual, not offensive
context
In Example 1, the swear word “bitch" is used
just to arouse interest/show off, thus not directly
insulting the other person. This is a case of
idiomatic swearing
Irony and world knowledge. In Example 3, the
sentence “Potrei morire per il dispiacere.”6 is
ironic. Humor is very hard to model for automatic
systems — sometimes, the presence of figurative
language even baffles human annotators.
Moreover, external world knowledge is often required
in order to infer whether an utterance is ironic
Preprocessing and tokenization. In computermediated communication, and specifically on Twitter, users often resort to a language type that is closer to speech, rather than written language. This is reflected in less-than-clean orthography, with forms and expressions that imitate the verbal face-to-face conversation.
4. @ XXXXXXXXX @ XXXXXXXXXX @ XXXXXXX @ XXXXXX x me glob prox2aa colpiran tutti incluso nemicinterno.. esterno colpopiúduro saràculogrande che bevetropvodka e inoltre x questiondisoldi progetta farmezzofallirsudfinitestampe: ciò nnvàben xrchèindebolis ! 4 me glob next2aa will hit everyone included internalenemy.. external harderhit willbebigass who drinkstoomuchvodka and also 4 mattersofmoney isplanning tomakethesouthfailwithprintings: dis notgood causeweaken In Example 4, preprocessing steps like tokenization and stemming are particularly hard to perform, because of the lack of spaces between one word and the other and the confused orthography. Consequently all the classification pipeline is compromised and error-prone.
Gender of the target. As defined in the Introduction, we know that misogyny is a specific type of hateful language, targeting women. However, detecting the gender of the target is a challenging task in itself, especially in Twitter datasets. 5. @realDonaldTrump shut the FUCK up you infected pussy fungus.
6. @TomiLahren You’re a fucking skank!
Both examples use bad words to abuse their targets. However, the first example is labeled as not misogyny since the target is Donald Trump (man), while the second example is labeled as misogyny with the target Tomi Lahren (woman).
6Translation: I could die for heartbreak.
Here we draw some considerations based on the results of our participation to the EVALITA 2018 AMI shared task. In order to test the multilingual potential of our model, one of the systems we submitted for Italian at EVALITA (run #3) was based on our best model for Spanish at IberEval. Based on the official results, this system performed well for Italian, consisting of features such as: BoW, BoE, BoH and several HurtLex categories specifically related to the hate against women. Concerning English, we obtained lower results in EVALITA in comparison to IberEval with the same system configuration. It is worth mentioning that even if the training set for the AMI EVALITA task was substantially bigger, in absolute terms all the AMI’s participants at EVALITA obtained worse scores than the ones obtained by the IberEval’s teams.
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).