English. We describe the systems the RuG Team developed in the context of the Hate Speech Detection Task in Italian Social Media at EVALITA 2018. We submitted a total of eight runs, participating in all four subtasks. The best macro-F1 score in all subtasks was obtained by a Linear SVM, using hate-rich embeddings. Our best system obtains competitive results, by ranking 6th (out of 14) in HaSpeeDe-FB, 3rd (out of 15) in HaSpeeDe-TW, 8th (out of 13) in Cross-HaSpeeDe FB, and 6th (out of 13) in Cross-HaSpeeDe TW.
Italiano. Illustriamo i dettagli dei due sistemi che il Team RuG ha sviluppato nell’ambito dell’esercizio di valutazione su riconoscimento di messagi d’odio in testi da Social Media per l’italiano. Abbiamo partecipato a tutti e quattro i sottotask, inviando un totale di otto predizioni. La migliore macro-F1, e` ottenuta da un SVM che usa embedding polarizzati, costruiti sfruttando contenuto ricco di odio. Il nostro miglior sistema ha ottenuto dei risultati competitivi, classificandosi 6 (su 14) in HaSpeeDe-FB, 3 (su 15) in HaSpeeDe-TW, 8 (su 13) nel Cross-HaSpeeDe FB, e 6 (su 13) in Cross-HaSpeeDe TW.
The use of “bad” words and “bad” language has
been the battleground for freedom of speech for
centuries. The spread of Social Media platforms,
and especially of micro-blog platforms (e.g.
Facebook and Twitter), has favoured the growth of
online hate speech. Social media sites and platforms
have been urged to deal with and remove
offensive and/or abusive content but the phenomenon is
so pervasive that developing systems that
automatically detect and classify offensive on-line content
has become a pressing need
The Natural Language Processing and
Computational Social Science communities have been
receptive to such urgency, and the automatic
detection of abusive and/or offensive language, trolling,
and cyberbulling
The haspeede task derives from the
harmonization process of originally separate annotation
efforts from two research groups, converging onto a
uniform label granularity
2https://sites.google.com/view/alw2018
3http://bit.ly/2EEC7Me
4http://bit.ly/2P7pTQ9
5http://di.unito.it/haspeedeevalita18
overview paper
The hate detection task is articulated in four binary (hate vs non-hate) sub-tasks, two in-domain, two cross-domain. The in-domain sub-tasks require training and test data to belong to the same text type, either Facebook (HaSpeeDe-FB) or Twitter (HaSpeeDe-TW), while the cross-domain sub-tasks require training on one text type and testing on the other: Facebook-Twitter (CrossHaSpeeDe FB) and Twitter-Facebook (CrossHaSpeeDe TW). 3
All of our runs for all subtasks are based on supervised approaches, where data (and features) play a major role for the final results of a system. Furthermore, our contribution adopted a closed-task setting, i.e. we did not include any training data beyond what was provided within the task. We did however build enhanced distributed representations of words exploiting additional data (see Section 3.2). This section illustrates the datasets and language resources used in our submissions. 3.1
The organizers provided a total of 6,000 labeled Italian messages for training, split as follows: 3,000 comments from Facebook, and 3,000 messages from Twitter. For test, they subsequently made available 1000 instances for each text type. Table 1 illustrates the distribution of the classes in the different text types both in training and test data. Note that the distribution of labels in the test data is unknown at developing time. Although the task organisers have balanced the datasets with respect to size, and have adopted the same annotation granularity (hate vs. non-hate), the two datasets are very different both in terms of class distribution (i.e. 46.06% of messages labelled as hateful in Facebook vs. 32.40% in Twitter in training) and with regard to their contents. For instance, the Facebook data is concerned with general topics that may contain hateful messages such as immigration, religion, politics, gender issues, while the Twitter dataset is focused on specific targets, i.e., categories or groups of individuals who are likely to become victims of hate speech (migrants, Muslims, and Roma6). It is also interesting to note that the label distribution in the Facebook test data is flipped compared to training, with a strong majority of hateful comments. 3.2
We addressed the task by adopting a closed-task setting. However, as a strategy to potentially increase the generalization capabilities of our systems and tune them towards better recognition of hate content, we developed hate- and offensesensitive word embeddings.
To do so, we scraped comments from a list of
selected Facebook pages likely to contain
offensive and/or hateful content in the form of
comments to posts, extracting over 1M comments. We
built word embeddings over the acquired data with
the word2vec tool skip-gram model
We detail in this section our final submissions.
The models have been developed in parallel to
our participating systems at the GermEval 2018
Shared Task
Our first model is a Linear Support Vector
Machine (SVM), built using the LinearSVC scikit
learn implementation
We performed minimal pre-processing by removing stop words using the Python module stop-words7, and lowercasing the tokens.
6The Romani, Romany, or Roma are an ethnic group of traditionally itinerant people who originated in northern India and are nowadays subject to ethnic discrimination. 7https://pypi.org/project/stop-words/ We used two groups of surface features, namely: i.) word n-grams in the range 1–3; and ii.) character n-grams in the range 2–4. The sparse vector representation of each (training) instance is then concatenated with its dense vector representation, as follows: for every word w in an instance i, we derived a 300 dimension representation, w~ , by means of a look-up in the hate-rich embeddings. We performed max pooling over these word embeddings, w~ , to obtain a 300 dimension representation of the full instance, ~i. Words not covered in the hate-oriented embeddings are ignored. Finally, class weights are balanced and SVM parameters use default values (C = 1). 4.2
Our second submission uses a binary ensemble model, which combines a Convolutional Neural Network (CNN) system and the linear SVM (Section 4.1), with a logistic regression meta-classifier on top. Predictions on training data are obtained via ten-fold cross-validation.
In the ensemble model, each input instance to the meta-classifier is represented by the concatenation of four features: a) the class predictions for that instance made by the SVM, b) the predictions of the CNN, and c) two additional surfacelevel features: the instance’s length in terms of characters and the percentage of offensive terms in the instance. This latter feature is obtained via a look-up in a list of offensive terms in Italian obtained from the article Le Parole per ferire by Tullio De Mauro8 and the “bad words” category in the Italian Wiktionary. The feature is expressed by the ratio between the frequency of any of the instance’s tokens comprised in the list and the instance’s length in terms of tokens. Figure 1 shows the features fed to the ensemble meta-classifier.
The CNN is an adaptation of available
architectures for sentence classification
The max-pooling layer output is flattened, concatenated, and fed to the fully-connected layer composed of 50 hidden-units with the ReLU activation function. The final output layer with the
sigmoid activation function computes the distribution of the two labels. (Other network hyperparameters: Number of filters: 6; Filter sizes: 3, 5, 8; Strides: 1). We used binary cross-entropy as loss function and Adam as optimiser. In training, we set a batch size of 64 and ran it for 10 epochs. We also applied two dropouts: 0.6 between the embeddings and the convolutional layer, and 0.8 between the max-pooling and the fully-connected layer. 5
Table 2 reports the results and ranking for our runs for all four subtasks. We also include the scores of the CNN (not submitted to the official competition), marked with a .9 The SVM models obtain, by far, better results than the Ensemble models. It is likely that the Ensemble systems suffer from the lower performances of 9Being allowed to submit a maximum of two runs per subtask, we based our choice of models on the results of a 10-fold cross validation of the three architectures on the training data.
10The SVM correposnds to run id 1 and the Ensemble model to run id 3 in the official submitted runs - see Submissions-Haspeede in the GitHub repository https: //github.com/tommasoc80/evalita2018-rug/ tree/master/Submissions-Haspeede
Subtask HaSpeeDe-FB HaSpeeDe-TW Cross-HaSpeeDe FB Cross-HaSpeeDe TW
SVM Ensemble CNN SVM Ensemble CNN SVM Ensemble CNN SVM Ensemble CNN 6/14 9/14 n/a 3/15 9/15 n/a 8/13 9/13 n/a 6/13 7/13 n/a 0.7751 0.7428 0.7138 0.7934 0.7530 0.7363 0.5409 0.4845 0.4692 0.6021 0.5545 0.6093 the CNN. We also observe differences in performance on the two datasets across the subtasks.
In-domain, in absolute terms, we do better on Twitter (.7934) than on Facebook (.7751), and this is even truer in relative terms, as performance overall in the competition is better on Facebook (best: 0.8288) than on Twitter (best: 0.7993). Our high score on HaSpeeDe-TW comes from high precision and recall on non-hate, while for HaSpeeDe-FB, we do well on the hate class. This can be due to label distribution (hate is always minority class, but more balanced in Facebook), but also to the fact that we use Facebook-based haterich embeddings, which might push towards better hate detection.
Cross-domain, results are globally lower, as
expected, with best scores on Cross-HaSpeeDe FB
and Cross-HaSpeeDe TW of 0.6541 and 0.6985,
respectively
The drop in performance in the cross-domain settings is likely due to topics, and data collection strategies (general topics on Facebook, specific targets on Twitter). In other words, despite the use of hate-rich embeddings as a strategy to make the systems generalize better, our models remain too sensitive to training data, which is strongly represented as word and character n-grams.
The impact of the hate-rich embeddings is most strongly seen in HaSpeeDe-FB and CrossHaSpeeDe FB, with recall for the hate class being substantially higher than for the non-hate class. This could be due to the fact that the hate-rich embeddings have been generated from comments in Facebook pages, that is, the same text type as the training data in the two tasks, so that possibly some jargon and topics are shared. While this has a positive effect when training and testing on Facebook (HaSpeeDe-FB), it has instead a detrimental effect when testing on Twittter (CrossHaSpeeDe FB), since this dataset has a large majority of non-hate instances, and we tend to overpredict the hate class (see Table 3).
In HaSpeeDe-TW and Cross-HaSpeeDe TW (training on Twitter) the impact of the hate-rich embeddings is a lot less clear. Indeed, recall for the hate class is always lower than non-hate, with the large majority of errors (more than 50% in all runs) being hate messages wrongly classified as non-hateful, thus seemingly just following the class imbalance of the Twitter trainset.
In both datasets, hate content is expressed either in a direct way, by means of “bad words” or direct insults to the target(s), or more implicitly and subtly. This latter type of hate messages is definitely the main source of errors for our systems in all subtasks. Finally, we observe that in some cases the annotation of messages as hateful is subject to disagreement and debate. For instance, all messages containing the word rivoluzione [revolution] are marked as hateful, even though there is a lack of linguistic evidence. 7
Developing our systems for the Hate Speech
Detection in Italian Social Media task at
EVALITA 2018, we focused on the generation of
distributed representations of text that could not
only enhance the generalisation power of the
models, but also better capture the meaning of words
in hate-rich contexts of use. We did so exploiting
Facebook on-line communities to generate
haterich embeddings
A Linear SVM system outperformed a metaclassifer that used predictions from the SVM itself, and a CNN, due to the low performance of the CNN component. Major errors of the systems are due to implicit hate messages, where even the hate-rich embeddings fail. A further aspect to consider in this task is the difference in text type and class balance of the two datasets. Both of these aspects have a major impact on system performance in the cross-genre settings.
Finally, to better generalize to unseen data and
genres, future work will focus on developing
systems able to further abstract from the actual
lexical content of the messages by capturing general
writing patterns of haters. One avenue to explore
in this respect is “bleaching” text