1 Introduction

1During the last years, the phenomenon of hate against women increased exponentially especially in online environment such as microblogs (Hewitt et al., 2016; Poland, 2016). According to the Pew Research Center Online Harassment report (2017) (Duggan, 2017), we can highlight that 41% of people were personally targeted, whose 18% were subjected to serious kinds of harassment because of the gender (8%) and that women are more likely to be targeted than men (11% vs 5%). Misogyny, defined as the hate or prejudice against women, can be linguistically manifested in numerous ways, ranging from less aggressive behaviours like social exclusion and discrimination to more dangerous expressions related to threats of violence and sexual objectification (Anzovino et al., 2018). Given this relevant social problem, the Automatic Misogyny Identification (AMI) task has been proposed first at IberEval 2018 (Spanish and English) (Fersini et al., 2018) and later at Evalita 2018 (Italian and English) (Caselli et al., 2018). The main goal of AMI is to distinguish misogynous contents from non-misogynous ones, to categorize misogynistic behaviours and finally to classify the target of a tweet.

2 Task Description

2The AMI shared task is organized according to two main subtasks:

• Subtask A - Misogyny Identification: a system must discriminate misogynistic contents from the non-misogynistic ones. Examples of misogynous and non-misogynous tweets are reported in Table 1.

• Subtask B - Misogynistic Behaviour and Target Classification: a system must recognize the targets that can be either specific users or groups of women together with the identification of the type of misogyny against women.

3Regarding the misogynistic behaviour, a tweet must be classified as belonging to one of the following categories:

• Stereotype & Objectification: a widely held but fixed and oversimplified image or idea of a woman; description of women’s physical appeal and/or comparisons to narrow standards.

• Dominance: to assert the superiority of men over women to highlight gender inequality.

• Derailing: to justify woman abuse, rejecting male responsibility; an attempt to disrupt the conversation in order to redirect women’s conversations on something more comfortable for men.

• Sexual Harassment & Threats of Violence: to describe actions as sexual advances, requests for sexual favours, harassment of a sexual nature; intent to physically assert power over women through threats of violence.

• Discredit: slurring over women with no other larger intention.

4Examples of Misogynistic Behaviours are reported in Table 2.

5Concerning the target classification, the main goal is to classify each misogynous tweet as belonging to one of the following two target categories:

• Active (individual): the text includes offensive messages purposely sent to a specific target;

• Passive (generic): it refers to messages posted to many potential receivers (e.g. groups of women).

6Examples of targets of misogynous tweets are reported in Table 3.

3 Training and Testing Data

7In order to provide training and testing data both for Italian and English, three approaches were employed to collect misogynous text on Twitter:

• Streaming download using a set of manually defined representative keywords, e.g. bi**h, w**re, c*nt for English and pu****a, tr**a, f**a di legno for Italian;

• Monitoring of potential victims’ accounts, e.g. gamergate victims and public feminist women;

• Downloading the history of identified misogynist, i.e. explicitly declared hate against women on their Twitter profiles.

8Among all the collected tweets we selected a subset of text querying the database with the co-presence of keywords, originating two corpora initially composed of 10000 tweets for each language. In order to label both the Italian and English datasets, we involved a group of 6 experts exploiting the CrowdFlower¹ platform for internal use. At the end of the labelling phase, we provided one corpus for Italian and one corpus for English to all the participants. The inter-rater annotator agreement on the English dataset for the fields of “misogynous”, “misogyny_category” and “target” is 0.81, 0.45 and 0.49 respectively, while for the Italian dataset is 0.96, 0.68 and 0.76. Each corpus is distinguished in Training and Test datasets. Regarding the training data, both the Italian and English corpora are composed of 4000 tweets. Concerning the test data, we provided 1000 tweets for each language. The training data has been provided as tab-separated, according to the following fields:

• id denotes a unique identifier of the tweet.

• text represents the tweet text.

• misogynous defines if the tweet is misogynous or not misogynous; it takes values as 1 if the tweet is misogynous, 0 if the tweet is not misogynous.

• misogyny_category denotes the type of misogynistic behaviour; it takes value as:
  – stereotype: denotes the category “Stereotype & Objectification”;
  – dominance: denotes the category “Dominance”;
  – derailing: denotes the category “Derailing”;
  – sexual_harassment: denotes the category “Sexual Harassment & Threats of Violence”;
  – discredit: denotes the category “Discredit”;
  – 0 if the tweet is not misogynous.

• target denotes the subject of the misogynous tweet; it takes value as:
  – active: denotes a specific target (individual);
  – passive: denotes potential receivers (generic);
  – 0 if the tweet is not misogynous.

9Concerning the test data, only “id” and “text” have been provided to the participants. Examples of all possible allowed combinations are reported below. Additionally to the field “id”, we report all the combinations of labels to be predicted, i.e. “misogynous”, “misogyny_category” and “target”:

10The label distribution related to the Training and Test datasets is reported in Table 4. While the distribution of labels related to the field “misogynous” is almost balanced (for both languages), the classes related to the other fields are quite unbalanced. Regarding the “misogyny_category”, we can distinguish between the two considered languages. In particular, for the Italian language, the most frequent label is related to the category Stereotype & Objectification, while for English the most predominant one is Discredit. Concerning the “target”, the most predominant victims are specific users (active) with a strong imbalanced distribution on the Italian corpus, while it is almost balanced for the English training dataset and strongly imbalanced on the (active) targets for the corresponding test dataset.

4 Evaluation Measures and Baseline

11Considering the distribution of labels of the dataset, we have chosen different evaluation metrics. In particular, we distinguished as follows:

12Subtask A. Systems have been evaluated on the field “misogynous” using the standard accuracy measure, and ranked accordingly.

13Subtask B. Each field to be predicted has been evaluated independently on the other using a Macro F1-score. In particular, the Macro F1-score for the “misogyny_category” field has been computed as average of F1-scores obtained for each category (stereotype, dominance, derailing, sexual_harassment, discredit), estimating F₁(misogyny_category). Analogously, the Macro F1-score for the “target” field has been computed as average of F1-scores obtained for each category (active, passive), F₁(target). The final ranking of the systems participating to Subtask B was based on the Average Macro F1-score (F₁), computed as follows:

14 (1)

15In order to compare the submitted runs with a baseline model, we provided a benchmark (AMI-BASELINE) based on Support Vector Machine trained on a unigram representation of tweets. In particular, we created one training set for each field to be predicted, i.e. “misogynous”, “misogyny_category” and “target”, where each tweet has been represented as a bag-of-words (composed of 1000 terms) coupled with the corresponding label. Once the representations have been obtained, Support Vector Machines with linear kernel have been trained, and provided as AMI-BASELINE.

5 Participants and Results

16A total of 6 teams for Italian and 10 teams for English from 10 different countries participated in at least one of the two subtasks of AMI. Each team had the chance to submit up to three runs for English and three runs for Italian. Runs could be constrained, where only the provided training data and lexicons were admitted, and unconstrained, where additional data for training were allowed. Table 5 shows an overview of the teams² reporting their affiliation, their country, the number of submissions for each language and the subtasks they addressed.

5.1 Subtask A: Misogyny Identification

17Table 6 reports the results for the Misogyny Identification task, which received 13 submissions for Italian and 26 runs for English submitted respectively from 6 and 10 teams. The highest Accuracy has been achieved by bakarov at 0.844 for Italian and by hateminers at 0.704 for English, both in a constrained setting. Most of the systems have shown an improvement with respect to the AMI-BASELINE. While the bakarov team submitted only one run based on TF-IDF coupled with Singular Value Decomposition and Boosting classifier, hateminers achieved the highest performance with a run based on vector representation that concatenates sentence embedding, TF-IDF and average word embeddings coupled with a Logistic Regression model.

5.2 Subtask B: Misogynistic Behaviour and Target Classification

18Table 7 reports the results for the Misogynistic Behaviour and Target Classification task, which received 11 submissions by 5 teams for Italian and 23 submissions by 9 teams for English. The highest Average Macro F1-score has been achieved by bakarov at 0.501 for Italian (even if the amended run of CrotoneMilano achieved the highest effective performance) and by himani at 0.406 for English, both in a constrained setting. On the contrary of the previous task, most of the systems have shown lower performance compared to the AMI-BASELINE. It can be easily noted by looking at the Average Macro F1-score of all the approaches, that the problem of recognizing the misogyny category and the target is more difficult than the misogyny identification task.

19This is due to the fact that there can be a high overlapping between textual expressions of different misogyny categories, therefore it is highly subjective for an annotator (and consequently for a system) to select a category rather than another one. Regarding the target classification, systems can be easily misled by the presence of mentions that are not the target of the misogynous content.

20While for the bakarov team the system for Subtask B is the same one of Subtask A, himani achieved the highest performance on the English language with a run based on a Bag of N-Gram representation coupled with an Ensemble of 5 models for classifying the Misogynistic Behaviour and 2 models for Target Classification.

6 Discussion

21The submitted systems can be compared by taking into consideration the kind of input features that they have considered for representing tweets and the machine learning model that has been used as classification model.

22Textual Feature Representation. The systems submitted by the challenge participants’ consider various techniques for representing the tweet contents. Some teams have concentrated the effort on considering a single type of representation, i.e. the team ITT adopted the traditional TF-IDF representation, while bakarov and RCLN proposed systems considering only weighted n-grams at character level for better dealing with misspellings and capturing few stylistic aspects.

23Additionally to the traditional textual feature representation techniques (i.e. bag of words/characters, n-grams of words/characters eventually weighted with TF-IDF) several teams proposed specific lexical features for improving the input space and consequently the classification performances. The team of CrotoneMilano experimented feature abstraction following the bleaching approach proposed by Goot et al. (Goot et al., 2018) for modelling gender through the language. Specific lexicons for dealing with hate speech language have been included as features in the systems of SB, resham and 14-exlab. In particular, resham and 14-exlab made also use of environment-specific features, such as links, hashtags and emojis, and task-specific features, such as swear word, sexist slurs and women-related words.

24Differently from these approaches, StopPropagHate and hateminers teams proposed systems that consider the popular Embeddings techniques both at word and sentence level.

25Machine Learning Models. Concerning the machine learning models, we can distinguish between approaches that work with traditional Support Vector Machines and Logistic Regression, Ensemble Models and finally Deep Learning methods. Following, we report the models adopted by the systems that participated in the AMI shared task, according to the type of the machine learning model that has been adopted:

• Support Vector Machines have been exploited by 14-exlab by using both linear and RBF kernel, by SB investigating only a radial basis function kernel, and by CrotoneMilano by adopting again a simple linear kernel;

• Logistic Regression has been used by bakarov and hateminers;

• Ensemble Models have been adopted by three teams according to different settings, i.e. ITT and himani used a Simple Voting of different classifiers, resham induced a Simple Voting over different input features and RCLN used an Ensemble based on Random Forest;

• A Deep Learning classifier has been adopted by only one team, i.e StopPropagHate that trained a simple dense neural network.

26External Resources Several participants exploited external resources for providing task-specific lexical features.

27The lexicons for addressing AMI for Italian have been mostly obtained from lists available online. The team SB used an available specific Italian lexicon called “Le parole per ferire” built by Tullio De Mauro³. Starting from this lexicon provided by De Mauro, the HurtLex multilingual lexicon has been created (Bassignana et al., 2018). Beyond HurtLex, the team 14-exlab gathered a swear word list from several sources⁴ including a translated version of the noswearing dictionary⁵ and a list of swear words from (Capuano, 2007).

28Regarding the English language, both resham and 14-exlab used the list of swear words from noswearing dictionary and the sexist slur list provided by (Fasoli et al., 2015). The team resham further investigated the sentiment polarity retrieved from SentiWordNet (Baccianella et al., 2010). Differently, the team SB exploited a manually modeled lexicon for the misogyny detection task proposed in (Frenda et al., 2018a). The HurtLex lexicon has been used by the team 14-exlab also for the English task.

29Finally, pre-trained Word Embeddings have been considered by SB and hateminers teams, specifically GloVe (Pennington et al., 2014) for the English task and Word Embeddings built on the TWITA corpus for the Italian one (Basile and Novielli, 2014).

7 Conclusions and Future Work

30We presented here a new shared task about Automatic Misogyny Identification on Twitter for Italian and English. By analysing the runs submitted by the participants we can conclude that the problem of misogyny identification has been satisfactorily addressed by all the teams, while the misogynistic behaviour and target classification still remains a challenging problem. Concerning the future work, several issues should be considered to improve the quality of the collected data, especially for capturing those less frequent misogynistic behaviours such as Dominance and Derailing. The problem of hate speech against women will be further addressed in the HatEval shared task at SemEval in English and Spanish tweets⁶.

Acknowledgements

31The work of the third author was partially funded by the SomEMBED TIN2015-71147-C2-1-P research project (MINECO/FEDER). We thank Maria Anzovino for her initial help in collecting the tweets subsequently used for the labelling phase and the final creation of the Italian and English corpora used for the AMI shared task.