1 Introduction

1With awareness of violence against women growing in the public discourse and the spread of unfiltered and possibly anonymous communication on social media in our digital culture, the issue of misogyny online has become compelling. Violence against women has been described by the UN as a “Gender-based [..] form of discrimination that seriously inhibits women’s ability to enjoy rights and freedoms on a basis of equality with men”¹. On the web this often takes the form of female-discriminating attacks of different types, which undermine the women’s rights of freedom of expression and participation². Following erjavec 012you’s understanding of hate speech, reported in (Pamungkas et al., 2018) 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, colour, ethnicity, gender, sexual orientation, nationality, religion, or other characteristics”, we can define misogynist speech as any kind of aggressive discourse which targets women because they are women. Within the larger context of hate speech, online misogyny — or cybersexism — stands out as a large and complex phenomenon which reflects other forms of offline abuse on women (Poland, 2016). This holds true for the Italian case as well, where bouts of misogynistic tweets have been linked to episodes of femicides³. In recent years the NLP community has addressed the issue of automatic detection of hate speech in general (Schmidt and Wiegand, 2017) and misogyny in particular (Anzovino et al., 2018). This effort to detect and contain verbal violence on social media (or any kind of text) demonstrates how NLP tools can also be used for ethically beneficial purposes and should be considered in the newborn and crucial discourse on Ethics in NLP (Hovy and Spruit, 2016; Hovy et al., 2017; Alfano et al., 2018). We are therefore proud to take up the AMI challenge (Fersini et al., 2018) and present our contribution to the cause of stopping misogynistic speech on Twitter. In this paper we propose a simple linear model using n-grams: we show that such a simple setup can still yield good results. We decided to propose a simple model for three reasons: first, it has been shown that linear SVM can easily outperform more complex deep neural networks (Plank, 2017; Medvedeva et al., 2017); second, training and testing our model does not require expensive hardware but a common laptop is enough to replicate our experiments; third, we experiment with a transformation of the input (i.e. we extract abstract features) and a linear model allows for an easier interpretation of the contribution of this transformation.

2To summarise, the following are the contributions of this paper:

• We propose a simple and yet strong misogyny detection system for English and Italian (ranked first out of twelve systems for misogynistic category detection)

• We show how a single system can be trained to work across languages

• We release all the code⁴ and our trained systems for reproducibility and for a quick implementation of language technology systems that can help detect and mitigate cybersexism phenomena.

3Task Description The AMI task is combined binary and multi-label, short text classification task. Given a Tweet, we have to predict whether it contains or not misogyny (Task A) and if it does, we have to classify the misogynistic behaviour and predict who is the subject being targeted (Task B). The misogynistic behaviour’s space consists of five different labels:

• Stereotype & Objectification

• Dominance

• Derailing

• Sexual Harassment & Threats of Violence

• Discredit

4The target can be either Active when the message refers to a specific person or Passive when the message expresses generic misogyny. The setup is the same for both Italian and English.

2 Data

5We use only the data released by the task organisers: they consist of Italian and English Tweets. The organisers report that the corpus has been manually labelled by several annotators. We provide an overview of the data set in Table 1. As it can be seen from the table, the data for Task A is more or less balanced, while the data for Task B is highly skewed.

3 Experiments

6In this section we describe the feature extraction process and the model that we built.

3.1 Pre-processing

7We decide not to pre-process the data in any way, since we do not have linguistic (or non-linguistic) reasons for doing so. To tokenize the text we simply split at every white space.

3.2 Model and Features

8We built a sparse linear model for approaching this task.

9We use n-grams extracted at the word level as well as at the character level. We use 3-10 n-grams and binary tf-idf. We feed these features to a Support Vector Machine (SVM) model with a linear kernel; we use the implementation included in scikit-learn (Pedregosa et al., 2011). Furthermore, we experiment with feature abstraction: we follow the bleaching approach recently proposed by (van der Goot et al., 2018). First, we transform each word in a list of symbols that 1) represents the shape of the individual characters and 2) abstracts from meaning by still approximating the vowels and characters that compose the word; then, we compute the length of the word and its frequency (while taking care of padding the first one with a zero in order to avoid feature collision); finally, we use a Boolean label for explicitly distinguishing words from non-alphanumeric token (e.g. emojis). Table 2 shows an example of this feature abstraction process.

Table 2: An illustration of the bleaching process

	shape	freq	len	alpha
This	Ccvc	46	04	True
is	vc	650	02	True
an	vc	116	02	True
example	vcvcccv	1	07	True
.	.	60	01	False
		1	01	False

10(van der Goot et al., 2018) proposed this bleaching approach for modelling gender across languages, by leveraging the language-independent nature of these features: here, we try to re-use the technique for classifying misogynist text across languages. We slightly modify the representation proposed by (van der Goot et al., 2018) by merging the shape feature (e.g. Xxx) with the vowel-consonant approximation feature (e.g. CVC) into one single feature (e.g. Cvc).

11We propose three different multi-lingual experiments:

• TRAIN Italian  TEST English

• TRAIN English  TEST Italian

• TRAIN Ita. & Eng.  TEST Ita. & Eng.

12For the last experiment, we use half the data set for each language. We report scores obtained by training on the whole training set and testing on the official test set, using the gold labels released by the organisers after the evaluation period.

4 Evaluation and Results

13Since the data set labels for the sub-task B are not evenly distributed across the classes, we use f1-score to evaluate our model. First we report results obtained via a 10-fold cross-validation on the training set; then, we report results from the official test set, whose labels have been released. The official evaluation does not take into account the joint prediction of the labels, however here we report results considering the 0 label: since we train different models for the different label sets, we make sure that the models trained on Task B are able to detect if a message is misogynistic in the first place.

4.1 Development Results

14We report the development results obtained by using different text representations. Table 3 presents an overview of these results. We note that all four representations — words, characters, a combination of these two and the bleached representation — all yield comparable results. The combination of words and characters seems to be the best format. Overall, we note that the system performs better on the Italian corpus than on the English corpus.

4.1.1 Cross-lingual Results

15In Table 4 we present the results of our cross-lingual experiments. We train and test different systems using lexical and abstract features. We note that the abstract model trained on Italian outperforms the fully lexicalized model when tested on English, but the opposite is not true. The English data set seems particularly hard for both the abstract and the lexicalized model. Interestingly, the abstract model trained on both corpora shows good results.

4.2 Test Results

16In Table 5 we present official test results (Fersini et al., 2018). We submitted only one, constrained run; a run is considered constrained when only the data released by the organisers are used. We submitted the model using the combined representation with word- and character-ngrams, trained once on the English corpus and once on the Italian corpus. We achieve the top and the second position for the tasks B and A respectively on the Italian data set. On the English data set our system is ranked 15th and 4th on the tasks A and B respectively.

5 Discussion and Conclusions

17A warning to the reader: this section contains explicit language. In the attempt to understand better the big difference in performance between the English and Italian models, we show the importance of words as learned by the model: we print the ten most important words, ranked by their learned weights. The result is shown in Table 6. From the output we see that the model trained on Italian learned meaningful words for identifying a misogynist message, such as zitta [shut up], tua [your] and muori [die!]: these words stand out from the rest of the profanity for directly referring to someone, while the rest of the words and almost all the most important English words could be used as interjections or could be more generic insults.

18The results of the abstract system are satisfactory for eventually building a light, portable model that could be adapted to different language. In the future we will try training on English and Italian and testing on a third corpus (such as the Spanish version of the AMI data set).

19In this paper we described our participation to the AMI - Automatic Misogyny Identification for Italian and English. We proposed a very simple solution that can be implemented quickly and we scored a state-of-the-art result for classification of misogynistic behaviours in five classes.

Acknowledgements

20The authors would like to thank the two anonymous reviewers who helped improve the quality of this paper. The first author has conducted this research as he was still part of the Erasmus Mundus master in Language and Communication Technology, a shared master program between the University of Groningen (NL) and the University of Malta (MT).