We describe our participation in the EVALITA 2020 (Basile et al., 2020) shared task on Automatic Misogyny Identification. We focus on task A -Misogyny and Aggressive Behaviour Identification- which aims at detecting whether a tweet in Italian is misogynous and, if so, whether it is aggressive. Rather than building two different models, one for misogyny and one for aggressiveness identification, we handle the problem as one single multi-label classification task, considering three classes: nonmisogynous, non-aggressive misogynous, and aggressive misogynous. Our threeclass supervised model, built on top of AlBERTo, obtains an overall F1 score of 0:7438 on the task test set (F1 = 0:8102 for the misogyny and F1 = 0:6774 for the aggressiveness task), which outperforms the top submitted model (F1 = 0:7406).1
In 2020, Twitter users in Italy amount to approximately 3.7 million and the number is expected to constantly increase by 2026.2 Although Twitter is conceived to express personal opinions, share today’s biggest news, follow people or simply communicate with friends, there has
Copyright ©2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
1Our official submission to the task obtained F1 = 0:6343 (F1 = 0:7263 for the misogyny and F1 = 0:5423 for the aggressiveness task). The reason behind this poor performance was the unintended use of a mistaken transformer. See Appendix A for further details.
2https://www.statista.com/forecasts/
1146708/twitter-users-in-italy; last visit: 6
November, 2020.
been an increasing number of users that misuse
the platform by engaging in trolling,
cyberbullying, or by posting aggressive and misogynous
content
However, two main issues exist. Since
Twitter mostly relies on the community subjective
perception of hate speech, many posts are not
subjected to report, review, and removal. Moreover,
the amount of abusive posts significantly
outnumbers the people that can manually control harmful
content. Therefore, there is a need to improve the
quality of algorithms to spot potential instances of
hate speech; in particular towards women, since
research shows that they are subjected to more
bullying, abuse, hateful language, and threats than
men on social media
AMI 2020 consists of two tasks
com/en/rules-and-policies/
hateful-conduct-policy
Unbiased Misogyny Identification— aims at
discriminating misogynistic contents from the
nonmisogynist ones, while guaranteeing the fairness
of the model (in terms of unintended bias) on
a synthetic dataset
The rest of the contribution is distributed as follows. Section 2 includes some background and a brief overview of research in automatic misogyny identification. Section 3 describes the employed dataset. Section 4 describes our model. Section 5 summarizes the experiments performed and discusses the obtained results. It includes an error analysis, in order to show the error trends of the model. Section 6 draws some conclusions and discusses further possible research lines. 2
Due to the subjective perception of misogyny and
aggressiveness, a definition of what can be
considered misogynous and aggressive is necessary:
Misogynous content expresses hating towards
women, in the form of insulting, sexual
harassment, male privilege, patriarchy, gender
discrimination, belittling of women, violence against
women, body shaming and sexual
objectification
From a computational point of view,
misogyny detection is a text classification task. Text
classification in Natural Language Processing has
been widely explored and it is typically addressed
by using supervised models
For what concerns misogyny identification, a
shared task took place at IberEval 2018,
focusing on English and Spanish tweets
As for misogyny identification in Italian, the
first edition of the AMI shared task took place
in 2018
As mentioned above, the aim of our model is to flag misogynous contents and aggressive attitudes towards women in Italian tweets. To address this task, a dataset was provided by the task organizers: 5; 000 tweets, manually labelled according to two classes, misogyny and aggressiveness. The first one defines whether a tweet has been flagged as misogynous (positive class) or not (negative class). If a tweet has been flagged as misogynous, it is further determined whether it is considered as aggressive (positive class) or not (negative class).
The training dataset is fairly balanced in terms of misogyny. It contains 2; 337 misogynous and 2; 663 non-misogynous instances. A total of 1; 783 of the former are also considered as aggressive, whereas only 554 are not. The test set was composed of 1; 000 tweets.
Since we opted for a constrained approach, we only used the data provided by the organizers. We randomly split the supervised data into training and validation sets: 4; 700 instances for the former and 300 for the latter. 4
Since the identification of aggressiveness is related to the identification of misogynous tweets, we opt for a 3-class setting, based on one single model. The three classes are hence non-misogynist, aggressive misogynist, and non-aggressive misogynist. The idea is to determine how well a multilabel classifier can perform when addressing these two related problems; handling aggressiveness as a consequential class of the misogyny one.
We decided to base our model on BERT
(Bidirectional Encoder Representations from
Transformers), a task-independent language
representation model based on the transformers
architecture
As for the pre-processing, we used the pretrained AlBERTo tokenizer for text tokenization, and then we encoded the data. We set the maximum length to 256 characters, since that was the length of the longest instance in the training material (even if Twitter allows up to 280 characters). team run constrained score UniBOa 2 yes 0.7438 jigsaw 2 no 0.7406 jigsaw 1 no 0.7380 fabsam 1 yes 0.7343 YNU OXZ 1 no 0.7314 fabsam 2 yes 0.7309 NoPlaceForHateSpeech 2 yes 0.7167 YNU OXZ 2 no 0.7015 fabsam 3 yes 0.6948 NoPlaceForHateSpeech 1 yes 0.6934 AMI the winner 2 yes 0.6869 MDD 3 no 0.6844 PoliTeam 3 yes 0.6835 MDD 1 yes 0.6820 PoliTeam 1 yes 0.6810 MDD 2 no 0.6679 AMI the winner 1 yes 0.6653 PoliTeam 2 yes 0.6473 UniBOb 1 yes 0.6343 AMI the winner 3 yes 0.6259 NoPlaceForHateSpeech 3 yes 0.4902 a Run submitted after the deadline.
b Buggy run submitted on the deadline (cf. Appendix A).
We used the Pytorch instance of AlBERTo-Base, Italian Twitter lower cased4 and fine-tuned it to the downstream task. We used a softmax output layer with three neurons to produce the classification.
In order to tune the network, we used the
AdamW optimizer, which decouples weight decay
from gradient computation, with a learning rate of
1e-5
We explored different batch sizes over an increasing number of learning epochs. Table 1 shows the performance evolution on the validation set. The best combination was to train the model over 8 epochs with a batch size of 16. This combination leads to an F1 score of 0:8491 on the three-class problem. It is worth noting that these scores are not comparable against those for the actual task, which consists of two independent binary decisions: whether a tweet is considered misogynist and, if the answer is yes, whether it is aggressive.6
AlBERTo-it,
5The implementation is available at https: //github.com/TinfFoil/unibo_ami2020/.
6Indeed, the official task score is computed as the average of the F1 measures for the misogyny and the aggressiveness classifications. tweet Type-1 errors 1.2 La cosa del filo del telefono attorno alla gola e` presa da una lite di
Stefanie e Brooke in Beautiful #chilhavisto 1.2 “Quanti schiaffi, #Malinconico. Li meriteresti tutti, perche´ anche questa volta sei riuscito a lasciare il segno.E no, non era facile. . . ” 1.3 Tutti quelli che ho conosciuto mi hanno sempre detto che a primo impatto sembro stronza e psicopatica ma in verita` sono una cucciola Type-2 errors 2.1 @Fedoc5 in ginocchio con te che comandi il ritmo del pompino e decidi quando sborrarmi in gola, perche` un pompino deve sempre finire con l’ingoio 2.2 @Nigagalsen @matteosalvinimi un follower dal 2016. . . e ti permetti di aprire quella lurida fogna di bocca che hai. 2.3 Antonio ti prenderei a schiaffi, come fa a dire mi manchi, quando con Ilaria fai tutto. Coglione 2.4 Posso volerlo vedere cagarsi in mano e prendersi a schiaffi finche´ non diventano dispari o sono nazista? Type-3 errors 3.1 La cena e` pronta. Lo squalo balena affonda nei banchi di pesce per saziare la sua immensa mole. Non e` un abitudine frequente visto che filtra dall’acqua i microorganismi come le balene. 3.2 Comunque le pringles piu` buone sono quelle alla panna acida e cipolla misogynist aggressive actual pred actual pred no no yes no no no no no no yes yes no yes yes no no yes yes no no no no no no no no no yes yes no yes no no no no no
Given these results, we trained a new model on the full trained and development sets during 8 epochs, using a batch size of 16, and predicted on the test set. Such model obtains F1 = 0:7438, resulting from 0:8102 on the misogyny task and 0:6774 from the aggressiveness one.
Table 2 shows the AMI shared task leaderboard.
It highlights both our official submission UniBO run 1 (cf. Appendix A) and our post-deadline submission UniBO run 2. Run 2 tops all the systems submitted to the shared task. Indeed, modelling the two tasks as one single multi-class problem (and using transformers for the right language) helps the algorithm significantly.
Error Analysis After the release of the gold labels, we performed an analysis of the classification errors. We analyzed 300 instances, taken randomly from the test set (100 at the beginning, 100 in the middle and 100 at the end). As observed from the reported performance, our model struggled mostly with the identification of aggressive instances. As a result, there are relatively few cases in which our model correctly labels nonaggressive misogynous instances. We noticed that most of the time, when our model labels an instance as misogynist, it also labels it as aggressive. On the contrary, the system performs very well in identifying non-misogynous instances and aggressive-misogynous instances. The most common mistakes are grouped into three categories: 1. The system identifies as aggressive the instances that contain verbs expressing an aggressive attitude.7 2. The system identifies as misogynous (and
most of the time also aggressive) instances 7One potential reason behind this confusion is that we suspect that there are aggressive tweets in the dataset which, not having been identified as misogynist in the first place, are mislabeled as non-aggressive. This hypothesis should be further explored. that are neither misogynous nor aggressive, but contain typical misogynous sentences. 3. The system identifies as misogynous instances that are neither misogynous nor aggressive, but they contain double-entendre words typically used to insult women.
Table 3 shows some examples for all three kinds of errors. Regarding the errors of type 1, in instance 1.1 the action of winding up a telephone cable around the neck was perceived as aggressive, despite the speaker did not express a misogynous or aggressive attitude towards a woman, and indeed she is just commenting on something watched on TV. In instance 1.2, the sentence meritare gli schiaffi (deserving slaps) denotes violence, but it is not addressed towards a woman.
This kind of mistake might be overcome by implementing a model trained on the misogynist partition of the data only. Finally, instance 1.3 represents the bias related to the subjectivity nature of what is perceived to be misogynous. According to the annotation guidelines, a tweet should be flagged as misogynous if it expresses hating towards women. In this case, the poster of the tweet is not expressing any misogynous attitude, but she is reporting what she is been told by males.
Therefore, our system flagged the instance as nonmisogynous and we could agree.
As for the errors of type 2, if we look at the text only, the instances could seem misogynous sentences. However, in the instances 2.1 and 2.2 the hashtag tells us that it is referred to a man and the system fails to understand that. On the contrary, the system performs well when a masculine name or a masculine pronoun is specified, instead of an hashtag, as we can observe in the instances 2.3 and 2.4. In these cases our system could understand that the aggressive actions, that usually tend to be classified as aggressive-misogynous, are not referred to a woman.
For the type 3 errors, in instance 3.1 balena (whale/fat woman) and in 3.2 acida (acid/peevish) could confuse the model causing it to flag such instances as misogynous. 6
In this paper we described our approach to the EVALITA 2020 task on misogyny and aggressiveness identification in Italian tweets —AMI. The purpose of our participation was to determine whether a multi-label classifier is a good way to address this two-step task. Although the task seems to be conceived to be addressed with two different models, one for the identification of misogyny and the other for aggressiveness, we decided to try a different approach and build a single model that could identify three cases: nonmisogynous, non-aggressive misogynous and aggressive misogynous tweets.
We built our model on top of AlBERTo, an Italian version of BERT, and we trained the model using only the dataset provided by the task organizers. We experimented by setting different batch sizes over an increasing number of epochs. The highest F1 score on the validation set was reached by a batch size of 16 during 8 epochs. When evaluated on the test set, our model obtained an overall F1 score of 0:7438; 0:8102 for the misogyny and 0:6744 for the aggressiveness task. We hypothesize that the model struggles to identify misogynist aggressive instances partly because it gets confused by non-misogynist aggressive tweets which are labeled simply as non-misogynous. The implementation is publicly available for research purposes.
For what concerns further experiments, we plan to build two separate models: one to detect misogyny and the other trained only on already-flagged misogynous tweets to identify instances of aggressiveness. Another step to undertake would be to use an unconstrained approach and increase the number of instances for the training set, so that the model will have more data to learn from.
A
Our official submission used a pre-trained BERT model trained only on the English language. The experimentation and tuning were identical to the one applied when using AlBERTo (cf. Section 5). Table 4 shows the tuning evolution. The best configuration of this model, derived from the English BERT, obtains an F1 score of 0:8222 on the validation set when dealing with our three-class problem. Nevertheless, the performance dropped to F1 = 0:6343 on the test set.