We describe our participation to the Automatic Misogyny Identi cation (AMI) shared task at IberEval 2018. The task focused on the detection of misogyny in English and Spanish tweets and was articulated in two sub-tasks addressing the identi cation of misogyny at di erent levels of granularity. We describe the nal submitted systems for both languages and sub-tasks: Task A is a classical binary classi cation task to determine whether a tweet is misogynous or not, while Task B is a ner grained classi cation task devoted to distinguish di erent types of misogyny, where systems must predict (i) one out of ve categories of misogynistic behaviours and (ii) if the abusive content was purposely addressed to a speci c target or not. We propose an SVM-based architecture and explore the use of several sets of features, including a wide range of lexical features relying on the use of available and novel lexicons of abusive words, with a special focus on sexist slurs and abusive words targeting women in the two languages at issue. Our systems ranked rst in Task A for both English and Spanish (accuracy score of 0.913 for English; 0.815 for Spanish), outperforming the baselines and the other participant systems, and rst in Task B on Spanish.
In the era of mass online communication, more and more episodes of hateful
language and harassment against women occur in social media 3. Hate Speech
(HS) can be de ned as any type of communication that is abusive, insulting,
intimidating, harassing, and/or incites to violence or discrimination, and that
disparages a person or a group on the basis of some characteristics such as
race, color, ethnicity, gender, sexual orientation, nationality, religion, or other
characteristics [
Recently, an increasing number of scholars is focusing on the task of
automatic detection of abusive or hateful language online [
Detecting misogynist content and its author is still a di cult task for social
media platforms. For instance, the popular social network Facebook is still unable
to deal with this issue and it relies on its community to report misogynistic
content4. The work of Hewitt et al. [
In this paper, we describe the systems we submitted for detecting misogyny
in the context of the Automatic Misogyny Identi cation (AMI) shared task at
IberEval 2018 [
The list of lexical features includes: Bag of Words (BoW): sparse vector
encoding the occurrence of unigrams, bigrams and trigrams in a tweet. Swear
Word Count: this feature represents the number of swear words contained in
a tweet. We used the list of swear words from noswearing dictionary 5. Swear
Word Presence: this feature is a binary value representing the presence of swear
words. We used the same dictionary from noswearing. Sexist Slurs Presence:
4https://www.nytimes.com/2013/05/29/business/media/facebook-says-itfailed-to-stop-misogynous-pages.html
5https://www.noswearing.com/dictionary
we use a small set of sexist words aimed towards women from prior work [
The structural features employed by our systems include: Bag of Hashtags (BoH): similarly to BoW, we exploit the hashtags. Bag of Emojis (BoE): we also utilized the Emojis in the tweets as a feature. We used their CLDR short name 7 in our feature matrix. Therefore, we converted the emoji unicode to its CLDR short name by using PyPI library8. Hashtag Presence: this feature has a binary value 0 (if there is no hashtag in the tweet) or 1 (if there is at least one hashtag in the tweet). Link Presence: presence of URLs in the tweets as a binary value: 0 if there is no link, 1 if there is at least one link in the tweet. All the features are encoded as xed-size numerical or one-hot vector representations, allowing us to experiment extensively with their combination. 3
In this section, we report on the result of the evaluation of our system for misogyny detection according to the benchmark established by the AMI task.
6For the Spanish system development, we translated all the English word lists described here by using Google Translate: https://translate.google.com/. 7https://unicode.org/emoji/charts/full-emoji-list.html 8https://pypi.org/project/emoji/ 3.1
The organizers of AMI proposed an automatic detection task of misogynistic content on Twitter, in English (EN) and Spanish (SP). Two di erent tasks were proposed: Task A is a binary classi cation task, where every system should determine whether a tweet is misogynous or not misogynous. Task B is composed of two distinct classi cation tasks. First, participants were asked to classify the misogynous tweets into ve categories of misogynistic behavior including: \stereotype & objecti cation", \dominance", \derailing", \sexual harassment & threats of violence", and \discredit". Secondly, they were asked to classify the misogynous tweets based on their target, labeling whether it is active (i.e. referring to one woman in particular) or passive (i.e. referring to a group of women).
Task A is evaluated in terms of accuracy, while for Task B the evaluation consists in the macro-average of the F1-scores on the positive classes. Each participating team could submit a maximum of 5 runs, pertaining to two di erent scenarios: constrained and unconstrained.
Dataset As summarized in Table 1, the organizers provided 3,251 tweets for the English training set and 3,307 tweets for the Spanish training set. Each tweet, in both languages, was annotated at three levels: 1) presence of misogynous content, 2) categories of misogynistic behavior, as described in Section 3.1, and 3) target of misogyny (active or passive). The organizers provided a balanced label Task A
Task B
English Spanish Misogynistic
1,568 Not misogynistic
1,683 Total
Stereotype
Dominance 1,649 Derailing
Sexual Harassment Discredit Active
Passive 1,658 No class distribution for Task A (misogynous vs. not misogynous ), while the distribution of data for Task B was highly unbalanced, re ecting the natural distribution of misogynistic behaviours and targets in the corpus. 3.2
We built two variants of our system and trained them on the available training sets. We tuned the system on the basis of the results of a 10-fold cross validation, using accuracy as an evaluation metric for Task A. The system for English is based on SVM with Radial Basis Function (RBF) kernel, while the system built for Spanish is based on SVM with a linear kernel. Both systems were built by using scikit-learn Python library9. Additionally, we performed an ablation test on our feature sets to study the impact of the di erent features on the system performance. Table 2 shows the features selected for each of our submissions and accuracy scores from cross-validation on training sets for English and Spanish. For what concerns features based on HurtLex, in S4 (EN) and S3 (SP) we explored the impact of hate words belonging to categories speci cally related to gender-based hate (see Sec. 2). In addition, we tested the performance of the best-performing sets of features of one language applied to the other language, to gauge the multilingual potential of the best systems: the English submission 5 is based on the best-performing (in cross-validation) combination of features for Spanish, and the Spanish submission 5 is based on the best-performing combination of features for English. For Task B, we used exactly the same features as Task A in each submission. We only submitted constrained runs. 3.3
O
cial Results and Analysis Table 3 until Table 6 shows our submission ranking based on the competition o cial results 10. The submission name is based on the submission numbering on Table 2 (run 1 is result of S1 and so on). Our systems ranked rst in Subtask A for both English (accuracy 0:913 by run 1) and Spanish (accuracy 0:815 by run 3). Meanwhile, for Subtask B (Table 5 and Table 6), one of our systems was the best result on Spanish (average Macro F-measure 0:446 by run 2) and the 6th on English (average Macro F-measure 0:370 by run 5).
Our experiment in testing the multilingual setting proved to be a challenge. Not surprisingly, both submissions 5 were the worst-performing compared to other submissions. However, the English S5 shows a comparatively good performance in absolute terms. On Table 3, we can see that all of our submissions in English were above the competition baseline. However as we can see on Table 4, with the same system applied to the Spanish dataset, we obtained a very low accuracy score in Spanish (ranked 24th, accuracy 0:537). This asymmetry indicates that the combination of BoW, BoH and BoE is a better representation of tweets in a multilingual setting than more ad-hoc, task-speci c features. rank submissions accuracy 1 14-exlab.c.run1 0:913 2 14-exlab.c.run2 0:902 3 14-exlab.c.run4 0:898 4 14-exlab.c.run3 0:879 : : : : : : : : : 10 14-exlab.c.run5 0:824 : : : : : : : : : 15 ami-baseline 0:784 rank submissions accuracy 1 14-exlab.c.run3 0:815 4 14-exlab.c.run1 0:812 5 14-exlab.c.run2 0:812 6 14-exlab.c.run4 0:809 : : : : : : : : : 18 ami-baseline 0:767 : : : : : : : : : 24 14-exlab.c.run5 0:536
On Task B, most participants achieved relatively low results, showing the di culty of this task, especially in classifying misogynistic behavior categories. We found the datasets' unbalanced distribution of labels to be the main issue. Based on the detailed result provided by the organizers, we note that most of the submitted system are not able to detect the less represented classes including derailing (29), dominance (49), and stereotype & objecti cation (137). Also classifying the target of misogyny (active and passive) has not been an easy task, which can be seen looking at the F1-score of the result on o cial results. Features including Swear Words Count, Swear Words Presence, Hashtag Presence, Link Presence, Sexist Slurs, and Woman-related Words outperformed all 10https://amiibereval2018.wordpress.com/important-dates/results/ rank submissions F1-score : : : : : : : : : 6 14-exlab.c.run5 0:369 8 14-exlab.c.run3 0:351 10 14-exlab.c.run4 0:343 12 14-exlab.c.run2 0:342 15 14-exlab.c.run1 0:338 : : : : : : : : : 16 ami-baseline 0:337 rank submissions F1-score 1 14-exlab.c.run2 0:446 2 14-exlab.c.run3 0:445 3 14-exlab.c.run4 0:444 5 14-exlab.c.run1 0:441 : : : : : : : : : 14 ami-baseline 0:410 : : : : : : : : : 20 14-exlab.c.run5 0:279 other submissions in English. In Spanish, the use of terms from the HurtLex lexicon, which were selected as related to gender-based hate, improves system performance in submission 3. However, not all the lexicon categories have been shown useful on this task, as indicated by the result of submission 4. 4
In this paper we described the 14-ExLab@UniTO submission for the Automatic Misogyny Identi cation (AMI) shared task at IberEval 2018. Our approach based on lexical knowledge was successful and our systems turned out to be the bestperforming out of the ones participating in the Task A for both English and Spanish. We also introduced a novel hate-speci c lexical resource which helped to improve the performance on the misogyny identi cation task.
For what concerns Task B, it was hard for all systems to classify misogynous tweets into the 5 categories proposed. After a manual inspection of the data, it emerged that there is no clear demarcation line between one category and the other and that the high presence of swearing in categories such as dominance and/or discredit just depends on the focus (e.g. the agent (man) vs. the wounded part, the target (woman)). At the same time, stereotype & objecti cation is not so conceptually distant from the sexual harassment category, due to a strong use of language referring to sexual body parts or vulgar sexual practices. Some examples from the English and Spanish datasets: stereotype & objectification (EN): No girl is even capable of developing morals until they get the slut fucked out of them. Welcome to my generation dominance (EN): Bad girls get spankings derailing: Women want u to automatically believe women who scream rape they don't understand our position.... sexual harassment & threats of violence (EN): @ SynergyFinny hey bitch wassup bitch suck my dick bitch discredit (EN): @ Herbwilson1967 Fuck that money whore @HillaryClinton Too stupid to know consensual touching or grabbing is not assault. Only @ChelseaClinton is dumber stereotype & objectification (ES): Que cuza antes la calle, una mujer inteligente o una tortuga vieja? Una tortuga vieja porque las mujeres inteligentes no existen : : : dominance (ES): \Voy a ensen~arle a esta perra como se trata a un hombre" LMAO IN LOVE WITH EL TITI sexual harassment & threats of violence (ES): @ genesismys1985 Me gustar a abrirte las piernas y clavarte toda mi polla en tu culo.
discredit (ES): Porque ladra tanto mi perra? La puta madre callate un poco We are planning to participate to the upcoming AMI shared task at EVALITA 2018, in order to validate our approach also for the Italian language.
V. Basile and V. Patti were partially funded by Progetto di Ateneo/CSP 2016 (Immigrants, Hate and Prejudice in Social Media, S1618 L2 BOSC 01).