English. The automatic misogyny identification (AMI) task proposed at IberEval and EVALITA 2018 is an example of the active involvement of scientific Research to face up the online spread of hate contents against women. Considering the encouraging results obtained for Spanish and English in the precedent edition of AMI, in the EVALITA framework we tested the robustness of a similar approach based on topic and stylistic information on a new collection of Italian and English tweets. Moreover, to deal with the dynamism of the language on social platforms, we also propose an approach based on automatically-enriched lexica. Despite resources like the lexica prove to be useful for a specific domain like misogyny, the analysis of the results reveals the limitations of the proposed approaches.
Italiano. Il task AMI circa l’identificatione automatica della misoginia proposto a IberEval e a EVALITA 2018 e` un chiaro esempio dell’attivo coinvolgimento della Ricerca per fronteggiare la diffusione online di contenuti di odio contro le donne. Considerando i promettenti risultati ottenuti per spagnolo e inglese nella precedente edizione di AMI, nel contesto di EVALITA abbiamo testato la robustezza di un approccio simile, basato su informationi stilistiche e di dominio, su una nuova collezione di tweet in inglese e in italiano. Tenendo conto dei repentini cambiamenti del linguaggio nei social network, proponiamo anche un approccio basato su lessici automaticamente estesi. Nonostante risorse come i lessici risultano utili per domini specifici come quello della misoginia, analizzando i risultati emergono i limiti degli approcci proposti.
The anonymity and the interactivity, typical of
computer-mediated communication, facilitate the
spread of hate messages and the perpetuated
presence of hate contents online. As investigated by
Fox et al. (2015), these factors increase and
influence social misbehaviors also offline. In order
to foster scientific research to find optimal
solutions that could help to monitor the spread of hate
speech contents, different tasks have been
proposed in various campaigns of evaluation. An
example is the AMI shared task proposed at IberEval
20181 and later at EVALITA 20182. This task
focuses on the automatic identification of misogyny
in different languages. In particular, the first
edition focuses on Spanish and English languages,
and the second one on a new English corpus and
Italian language. The multilingual context
allows to observe the analogies and differences
between different languages. The AMI’s organizers
Following, Section 2 describes the studies that inspired our work. Section 3 explains the approaches employed in both languages. Section 4 discusses the obtained results and delineates some conclusions. 2
A first work about misogyny detection is
proposed in Anzovino et al. (2018). In this study, the
authors compared the performance of different
supervised approaches using word embeddings,
stylistic and syntactic features. In particular,
their results reveal that the best machine learning
approach for identification of misogyny is the
linear Support Vector Machine (SVM) classifier.
In general machine learning techniques are the
most used in hate speech detection
The AMI shared task proposed at EVALITA 2018 aims to detect misogyny in English and Italian collections of tweets. The organizers asked participants to detect misogynistic texts (Task A), and then, if the tweet is predicted as misogynistic, to distinguish the nature of target (individuals or groups labeled respectively “active” and “passive”), and identify the type of misogyny (Task B), according to the following classes proposed by Poland (2016): (a) stereotype and objectification, (b) dominance, (c) derailing, (d) sexual harassment and threats of violence, and (e) discredit.
Actually, these classes represent the different manifestations and the various aspects of this social misbehavior. Table 1 shows the composition of the datasets.
Considering the promising results obtained at the IberEval campaign, in this work we use two approaches mainly based on lexica. The first one (Section 3.1) is similar to the approach used in Frenda et al. (2018), based on topic, linguistic and stylistic information captured by means of modeled lexica and n-grams of characters and words.
The second one (Section 3.2) principally involves
the automatically extended versions of the
original lexica
Misogynistic
(d) The first proposed approach aims to capture topic, linguistic and stylistic information by means of manually-modeled lexica and n-grams of words and characters. Below the features description for each language.
English Features. For the detection of misogyny in English tweets, we employed the manuallymodeled lexica proposed in Frenda et al. (2018).
These lexica concerns sexuality, profanity, femininity and human body as described in Table 2.
These lexica contain also slang expressions.
Moreover, we take into account hashtags and abbreviations collected in Frenda et al. (2018): 40 misogynistic hashtags, such as: #ihatef emales or #bitchesstink; and a list of 50 negative abbreviations, such as wtf or stf u. Considering the most relevant n-grams of words, we employ the bigrams for the first task and the combination of unigrams, bigrams and trigrams (hence defined as UBT) for the second task. Moreover, the bag of characters (BoC) in a range from 1 to 7 grams is employed to manage misspellings and to capture stylistic aspects of digital writing. In order to perform the experiments, each tweet is represented as a vector. The presence of words in each lexicon is pondered with Information Gain, and character and word n-grams are weighted with Term Frequency-Inverse Document Frequency (TF-IDF) measure. In addition, considering the fact that in Frenda et al. (2018) several misclassified misogynistic tweets were ironic or sarcastic, we try to analyze the impact of irony in misogyny detection in English.
Indeed, Ford and Boxer (2011) reveal that sexist jokes that in general are considered innocent, truthfully they are experienced by women as sexual harassment. In particular, inspired by Barbieri and Saggion (2014), we calculate the imbalance of the sentiment polarities (positive and negative) in each tweet using SentiWordNet provided by Baccianella et al. (2010). For each degree of imbalance, we associate a weight used in the vectorial representation of the tweets. Despite our hypothesis is well funded, we obtained lower results for the runs that contain sentiment imbalance among the features (see Table 4).
Italian Features. For the Italian language, we selected some specific issue groups, described in Bassignana et al. (2018), from the Italian lexicon “Le parole per ferire” provided by Tullio De Mauro3. In particular, we consider the lists of words described in Table 3. Differently from English, the experiments reveal that: the UBT is useful for both tasks and the best range for BoC is from 3 to 5 grams4. Indeed, in a morphological complex language like Italian the desinences of the words (such as the extracted n-grams “tona” or “ana ”) contain relevant linguistic information. Diversely, in English, longer sequences of characters could help to capture multi-word expressions containing also pronouns, adjectives or prepositions, such as “ing at” or “ss bitc”.
To extract the features correctly, in order to train our models, we pre-process the data deleting emoticons, emojis and URLs. Indeed, from our experiments, the emoticons and emojis do not prove to be relevant for these tasks. In order to perform a correct match between the dictionaries of the corpora and the single lexicon, we use the lemmatizer provided by the Natural Language Toolkit (NLTK5) for English, and the Snowball Stemmer for Italian. Differently from English, the use of lemmatizer for Italian tweets hinders the match.
3http://www.internazionale.it/ opinione/tullio-de-mauro/2016/09/27/ razzismo-parole-ferire 4The experiments are carried out using the Grid Search. 5http://www.nltk.org/ Profanity Femininity Human body
Words 290
Definition contains words relative to sexual subject (orgasm, orgy, pussy) and especially male domination on women (rape, pimp, slave) is a collection of vulgar words such as motherf ucker, slut and scum is a list of terms used to identify the women as target. It contains personal pronouns or possessive adjectives (such as she, her, herself ), common words used to refer to women (girl, mother) and also offensive words towards women (such as barbie, hooker or non male) is a lexicon strongly connected with sexuality collecting words referred especially to feminine body also with negative connotations (such as holes, throat or boobs)
Definition collects words relative to animals, such as sanguisuga or pecora contains terms referred to female genitalia, such as f essa contains terms referred to male genitalia, such as verga is a list of derogatory words, such as bastardo or spazzatura contains words derived from plants but that are used as offensive words, such as f inocchio or rapa is a list of professions or jobs that have also a negative connotations, such as portinaia or impiegato contains terms about prostitution, such as bagascia or zoccolona is a list of words relative to stereotypes, such as negro or ostrogoto collects words that have in general negative connotations, such as parassita or dilettante contains terms relative to criminal acts or immoral actions, such as stupro or violento The second approach aims to deal with the dynamism of the informal language online trying to capture new words relative to contexts defined in each lexicon. Therefore, we use enriched versions of the original lexica (described above), and stylistic and linguistic information captured by means of n-grams of words and characters as in the first approach. The method for the expansion of a given lexicon shares the idea of identifying new words by considering their contextual similarity with known words, as defined by some pre-trained word embeddings. For its description, let assume that L = fl1; : : : ; lmg is the initial lexicon of m words, and W = f(w1; e(w1)); : : : ; (wn; e(wn))g is the set of pre-trained word embeddings, where each pair represents a word and its corresponding embedding vector. This method aims to enrich the lexicon with words strongly related to the context from the original lexicon without being necessarily associated to any particular word. Its idea is to search for words having similar contexts to the entire lexicon. This method has two main steps, described below.
Dictionary modeling. Firstly, we extract the embedding e(li) for each word li 2 L; then, we compute the average of these vectors to obtain a vector describing the entire lexicon, e(L). We name this vector the context embedding.
larity, we compare e(L) against the embedding e(wi) of each wi 2 W ; then, we extract the k most similar words to e(L), defining the set EL = (w1; : : : ; wk). Finally, we insert the extracted words into the original lexicon to build the new lexicon, i.e., LE = L [ EL.
Therefore, we carry out the experiments using
different pre-trained word embeddings for each
language: GloVe embeddings trained on 2
billion tweets
To carry out the experiments, a SVM classifier is employed with the radial basis function kernel (RBF) using the following parameters: C = 5 and = 0:1 for English and = 0:01 for Italian. Considering the complexity of the target classification for the Italian language due to imbalanced training set (see Table 1), we used a Random Forest (RF) classifier that aggregates the votes from different 6http://valeriobasile.github.io/twita/ about.html decision trees to decide the final class of the tweet.
The evaluation is performed using the test set provided by the organizers of the AMI shared task. For the competition, they use as evaluation measures the Accuracy for Task A and the average of F-score of both classes for Task B.
Table 4 and Table 5 show the results obtained in the competition compared with the baselines provided by the organizers for each task. Comparing the two approaches, in general AEL seems to work better than MML. However, the improvement of the results is very slight, especially for Italian language. This soft variation is unexpected considered the results obtained during the experiments employing 10-fold cross validations. In fact, AEL with enriched lexica using k equal 100 performed an Accuracy of 0.880. Moreover, looking at Table 4, reporting the official results of the AMI Task, only run 2 overcomes the baseline for the detection of misogyny in English, and for this run we used AEL approach excluding the sentiment imbalance as feature. About the identification of misogyny in Italian, the obtained results are lower than provided baselines as well as the values of F-score obtained in Task B for both languages (see Table 5). Despite the usefulness of lexica for a specific domain like misogyny, a lexicon-based approach proves to be insufficient for this task. Indeed, as the error analysis will confirm, misogyny, as well as general hate speech, involves linguistic devices such as humour, exclamations typical of orality and contextual information that completes the meaning transmitted by the tweet. Moreover, the low values obtained also in Task B suggest the necessity to implement dedicated approach for each misogynistic category.
This paper reports our participation in the AMI shared task. The organizers provide also the gold test set that helps us to understand better what are the misclassified cases and the aspects that should be considered in the next experiments. Carrying out the error analysis, we notice that in both datasets the content of URL affects the transmitted information in the tweet (such as Right! As they rape and butcher women and children !!!!!! https://t.co/maEhwuYQ8B). The swear words are often used also as exclamation without the aim to offend (such as Volevo dire alla Yamamay che tettona non sinonimo di curvy dato che di vita ha una 40, quindi confidence sta minchia.). Moreover, despite the actual English corpus does not contain several jokes, Italian misclassified tweets involve humourous utterances (such as @GrianneOhmsfor1 @BarbaraRaval A parte il fatto poi che culona inchiavabile” e` il miglior giudizio politico sentito sulla Merkel negli ultimi anni??”). In fact, in general, humour, irony and sarcasm hinder the correct classification of the texts, as we noticed in English and Spanish corpora provided in the IberEval framework. Participating in this shared task gave us the opportunity to analyze and compare multilingual datasets, and thus, to discover and infer general aspects typical of hate speech against women.
The work of Simona Frenda was partially funded by the Spanish research project SomEMBED TIN2015-71147-C2-1-P (MINECO/FEDER). We also thank the support of CONACYT-Mexico (projects FC-2410, CB-2015-01-257383). 7This run does not involve the sentiment imbalance 8This run involves the expansions of lexica with k = 100 Francesco Barbieri and Horacio Saggion. 2014. Modelling irony in twitter. In Proceedings of the StuEnglish Run baseline AMI run 2 run 1 run 3 Italian Run baseline AMI run 3 run 1 run 2 F-score 0.534 0.485 0.483 0.480
Target UBT+BoC UBT+BoC UBT+BoC Target UBT+BoC UBT+BoC UBT+BoC F-score 0.440 0.414 0.414 0.411 Total 0.487 0.449 0.448 0.446 dent Research Workshop at the 14th Conference of the European Chapter of the ACL.