1. Introduction and Motivations

1From a NLP perspective, much attention has been paid to the automatic detection of Hate Speech (HS) and related phenomena (e.g., offensive or abusive language among others) and behaviors (e.g., harassment and cyberbullying). This has led to the recent proliferation of contributions on this topic (Nobata et al. 2016; Waseem et al. 2017; Fortuna et al. 2019), corpora and lexica1, dedicated workshops2, and shared tasks within national3 and international4 evaluation campaigns.

2As for Italian, the first edition of HaSpeeDe (Bosco et al. 2018), a task specifically focused on HS detection, was proposed at EVALITA 2018 (Caselli et al. 2018). The task consisted of the binary classification (HS vs not-HS) of texts from Twitter and Facebook. For each social media platform, training and test data were provided. Furthermore, two cross-platform sub-tasks were introduced to test the systems’ ability to generalize across platforms.

3The ultimate goal of HaSpeeDe 2 at EVALITA 2020 (Basile et al. 2020) is to take a step further in state-of-the-art HS detection for Italian. By doing this, we also intend to explore other side phenomena and see the extent to which they can be automatically distinguished from HS.

4We propose a single training set made of tweets, but two separate test sets within two different domains: tweets and news headlines. While social media are still one of the main channels used to spread hateful content online (Alkiviadou 2019; Wodak 2018), an important role in this respect is also played by traditional media, and newspapers in particular.

5Furthermore, we chose to include another HS-related phenomenon, namely the presence of stereotypes referring to one of the targets identified within our dataset (i.e., muslims, Roma and immigrants). With the term stereotype we mean any explicit or implicit reference to typical beliefs and attitudes about a given target (Sanguinetti et al. 2018). An error analysis of the main systems on the HaSpeeDe 2018 dataset itself (Francesconi et al. 2019) showed that the occurrence of these elements constitutes a common source of error in HS identification.

6Finally, it has been observed that in social media and newspapers’ headlines, the most hateful parts are often verbless sentences or a verbless fragments, also known as Nominal Utterances (NUs) (Comandini, Speranza, and Magnini 2018). The relevant presence of NUs has been investigated in the POP-HS-IT corpus (Comandini and Patti 2019). In order to have a better understanding of the syntactic strategies used in HS, we include the recognition of NUs in hateful tweets and news headlines.

2. Task Description

7HaSpeeDe 25 consists of a Main task and two Pilot tasks and is based on two datasets, one containing messages from a social media platform, namely Twitter, and the other one news headlines. The three tasks are shortly described as follows:

• Task A - Hate Speech Detection (Main Task): binary classification task aimed at determining the presence or the absence of hateful content in the text towards a given target (among immigrants, Muslims and Roma)

• Task B - Stereotype Detection (Pilot Task 1): binary classification task aimed at determining the presence or the absence of a stereotype towards the same targets as Task A

• Task C - Identification of Nominal Utterances (Pilot Task 2): sequence labeling task aimed at recognizing NUs in data previously labeled as hateful.

8This edition of the task presents several distinguishing features with respect to the first one. Besides including new and more-richly annotated data, news headlines were introduced as cross-domain test data. Furthermore, two additional tasks are proposed. Finally, the Twitter test set intentionally contains tweets published in a different time frame than those in the training set to verify the systems’ ability to detect HS forms independently of biases. These biases result from context-related features, such as events – regarding one of our HS targets – that can be controversial or be subject to heated and polarized debates.

3. Datasets and Formats

9In this section we describe the datasets and formats used in the three tasks.

3.1 Twitter Dataset

• Task A: The Twitter portion of the data of HaSpeeDe 2018 was included in the training set (4,000 tweets posted from October 2016 to April 2017). Moreover, new Twitter data were included for this competition, a subset of the data gathered for the Italian hate speech monitoring project “Contro l’Odio” (Capozzi et al. 2019). The data were retrieved using the Twitter Stream API and filtered using the set of keywords described in . The newly annotated tweets were posted between September 2018 and May 2019 and were annotated by Figure Eight (now Appen) contributors for hate speech and by the task organizers for the stereotype category. In particular, only data posted between January and May 2019 were included in the test set.

• Task B: The HaSpeeDe Twitter corpus – used in the first edition of the task – was already annotated for stereotype since it was part of the Italian Hate Speech corpus described in Sanguinetti et al. . We then used the same guidelines to enrich the new data from “Contro l’Odio” with this annotation layer. The annotation was carried out by the task organizers.

• Task C: The HaSpeeDe Twitter corpus was also annotated for the presence of Nominal Utterances (NUs) within a side project (Comandini and Patti 2019). We used an updated version of its guidelines (available in the task repository) to enrich the new hateful data introduced in the campaign. Similarly to the stereotype level, the annotation of NUs was carried out by the task organizers specifically for this task’s purposes.

3.2 News Dataset

• Task A: For task A a new test corpus composed of newspapers’ headlines about immigrants was made available. The data were retrieved between October 2017 and February 2018 from online newspapers (La Stampa, La Repubblica, Il Giornale, Liberoquotidiano) and annotated within the context of a Master’s degree thesis discussed in 2018 at the Department of Foreign Languages at the University of Turin. Data annotation includes the same categories annotated in the Twitter corpus.

• Task B: The News corpus also includes stereotype annotation, performed according to the same guidelines used for developing the Twitter corpus.

• Task C: Similarly to the Twitter dataset, the third annotation level was added in the News corpus from scratch and specifically for the present task.

10Tables 1, 2 and 3 show the data distribution for each task.

11The whole dataset consists of 8,012 tweets and 500 news headlines for Task A and B, and 3,388 tweets and 181 news (i.e., the sub-set with hateful data only) for Task C.

12In Task A and B, HS and stereotype represent the 41.8% and 44.6%, respectively, of the Twitter dataset. In contrast, in the News dataset, the portion of hateful content and stereotype lowers to 36% and 35%.

13Table 3 shows statistics about the total number of texts with or without NUs in Task C. The percentage of hateful tweets featuring at least one NU is 57.4%; the percentage of news headlines having at least one NU is 83.4%. This distribution is in line with the one found in Comandini and Patti .

3.3 Formats

• Task A and B: For both tasks A and B data are provided in a tab-separated values (TSV) file including ID, text, HS and stereotype class (0 or 1). Mentions and URLs were replaced with @user and URL placeholders. Table 4 shows some annotation examples.

• Task C: The dataset provided for Task C was annotated using WebAnno and converted into a IOB (Inside-Outside-Beginning) format. The resulting IOB2 alphabet consists of I-NU-CGA, O and B-NU-CGA.
  The annotation includes the ID, followed by an hyphen to mark the token number, the token, and the IOB2 annotation of the NUs.
  Below an example taken from the training set.6

14To prevent participants from cheating, the released test set for Task C also contains non-hateful messages. However, the evaluation of the systems is conducted only on the hateful messages since we are interested in investigating the relationship between these two phenomena.

4. Evaluation

15For each task, participants were allowed to submit up to 2 runs. A separate official ranking was provided, and the evaluation was performed according to the standard metrics, i.e, Precision, Recall and F-score.

16For Task A and Task B, the scores were computed for each class separately, and finally the F-score was macro-averaged to get the overall results.

17For Task C, token-wise scores were computed, and a NU was considered correct only in case of exact match, i.e., if all tokens that compose it were correctly identified.

18Different baseline systems were built according to the task type:

• For Task A and B, besides a typical classifier based on the most frequent class (Baseline_MFC in Tables 5–8), a Linear SVM with TF-IDF of unigrams and 2–5 char-grams was used (Baseline_SVC).

• For Task C, the baseline replicates the one presented for the COSMIANU corpus (Comandini, Speranza, and Magnini 2018), which identifies as correct in the test the NUs that appear in the training set (memory-based approach); baseline results in Table 9.

5. Task Overview: Participation and Results

5.1 Participants

19A total amount of 14 teams participated in the Main task on HS detection, 6 teams also submitted their results for the Pilot task 1 (i.e. Task B) on stereotype detection, while we did not receive any submission for the Pilot task 2 (i.e. Task C) on NUs identification. Except for one case, all teams submitted 2 runs for their tasks. Furthermore, 4 teams used the same systems to participate in other (and partly related) tasks within the EVALITA 2020 campaign: YNU_OXZ and Jigsaw participated in the task on Automatic Misogyny Identification (AMI) (Fersini, Nozza, and Rosso 2020), while TextWiller and Venses also participated in the task on Stance Detection in Italian Tweets (SardiStance) (Cignarella et al. 2020). It is worth pointing out that in this second edition we registered a higher participation of non-Italian and non-academic teams, and that HaSpeeDe 2 has been one of the most participated EVALITA 2020 tasks.

5.2 Systems Overview

Approaches

20The participating models are characterized by different architectures that exploit principally BERT-based models and linguistic features. Transformers are a popular choice in this edition. Jigsaw (Lees, Sorensen, and Kivlichan 2020), Svandiela (Klaus, Bartle, and Rossmann 2020), DH-FBK (Leonardelli, Menini, and Tonelli 2020), TheNorth (Lavergne et al. 2020) fine-tuned BERT, AlBERTo7 and UmBERTo8 language models for both runs. YNU_OXZ (Ou and Li 2020) exploited the pre-trained XLM-RoBERTa9 multi-language model as input of Neural Networks architecture. Fontana-Unipi (Fontana and Attardi 2020) developed a model that is an ensemble of fixed number of instances of two principal transformers (AlBERTo and DBMDZ10) and a combination of DBMDZ input and a dense layer. The DBMDZ is used also by By1510 (Deng, Bai, and Dai 2020) in a transfer learning approach. UO team (Rodriguez Cisnero and Ortega Bueno 2020), on the other hand, used a Bi-LSTM with the addition of linguistic features in the first run, while using the pre-trained DBMDZ model in the second one. CHILab (Gambino and Pirrone 2020) experimented transformer encoders in the first run and depth-wise Separable Convolution techniques in the second one. Moreover, some teams explored classical machine learning approaches such as No Place For Hate Speech (S. R. da Silva and T. Roman 2020), TextWiller (Ferraccioli et al. 2020), UR_NLP (Hoffmann and Kruschwitz 2020) and Montanti (Bisconti and Montagnani 2020). Finally, Venses (Delmonte 2020), based on the parser for Italian ItGetaruns, applied six different rule-based classifiers.

Features and Lexical Resources

21Various features are tested and explored by participants. Morphosyntactic features are exploited by CHILab, using Part-of-Speech tags as additional input. To adapt the POS tagging model provided by Python’s spaCy library to social media language, they added emoticons, emojis, hashtags and URLs to vocabulary. In addition, to preprocess the texts, they used sentiment lexicon for replacing emoticons with appropriate labels about the expressed sentiment. Semantic and lexical features are exploited by Venses and UO teams. In particular, UO team used WordNet to catch lexical ambiguity, syntactic patterns and similarity among words; calculated information gain to capture the most relevant words; used lexicons such as HurtLex (Bassignana, Basile, and Patti 2018) and SenticNet11 to feature words with hateful categories and sentiment information. Finally, different types of representation of tweets are tested by Montanti: TF-IDF, DistilBert12 and GloVe (Pennington, Socher, and Manning 2014) vectors as well as their combination.

Additional data

22Some teams preferred to use additional data to improve the knowledge of their classifiers. To extend the provided training set, YNU_OXZ exploited Facebook data provided in the first edition of HaSpeeDe and DH-FBK used a set of Italian tweets that covers similar topics. Jigsaw, for one of the submissions, used additional user-generated comments to fine-tune their model. CHILab used additional tweets taken from TWITA 201813 by means of some keywords extracted from the provided training set to extend the embedding layer of their model. Finally, the SENTIPOLC 2016 dataset was exploited by UO team.

Interaction between Task A and B

23Except for TheNorth team, most of the participants did not consider the interaction between Task A and B. Taking into account the possible correlation between texts containing hate speech and texts expressing stereotyped ideas about targets, TheNorth tested the performance of multitasking approach for both tasks (second run) against a fine-tuned UmBERTo model (first run). In particular, observing competition results we can notice the efficacy of multitasking in hate speech identification and not in stereotype detection.

5.3 Results

24In Table 4, 5, 6 and 7, we report the official results of HaSpeeDe 2 for Task A and B, ranked by the macro-F1 score. In case of multiple runs, a suffix has been appended to each team name, in order to distinguish the run ID of the submitted file.

25As a general remark, we can observe that the in-domain Main task registered better results (macro-F1=0.8088) both compared to the cross-domain counter-part (0.7744) and the Pilot task 1; in turn, better results were obtained in the latter with the in-domain data compared to the News set (0.7744 and 0.7203, respectively). The best performances overall provided by the systems used for Task A on Twitter data is also reflected in the average value of the macro-F1 scores of each ranking: 0.6899 for the latter, 0.6306 for Task B on Twitter data, 0.6144 for Task A on News data and 0.5972 for Task B on News data.

26We also considered the overall results achieved by all participating teams and observed that, as regards Task A, 12 and 13 teams (in the Twitter and News test set, respectively) obtained higher scores than the SVM-based baseline with at least one of the submitted runs, and 13 teams, on both domains, outperformed the one based on the most frequent class. For Task B, and with respect to the SVM baseline, the same is true for 4 teams out of 6 in the Twitter set and for 3 teams in the News set, while all teams beat the majority-class baseline with at least one run.

27Regarding Task C, since the training set is composed of tweets, we first investigated the macro F-score value on a validation set created by splitting the training set in 80%-20%. We then tested the memory-based baseline described in Section 4 on the two test sets released for the task. Table 8 shows the macro-F1 values obtained in the validation set, in the Twitter test set as well as in the News test set. As mentioned earlier, no submissions were made for this task, but the baselines’ values for both domains are reported in this overview as reference points for further works.

6. Discussion

28A discussion of results, especially those regarding the Main task, necessarily involves a preliminary comparison with the ones obtained in the first edition of HaSpeeDe, in particular in the two tasks where Twitter data were used for training, i.e. HaSpeeDe_TW and Cross-HaSpeeDe_TW. The best systems attained macro-F1=0.7993 in the former task and 0.6985 in the latter. While these results are in line with those reported for Task A on the in-domain data, the results obtained in this edition on News data are better than the part cross-domain task, where the test set was made up of Facebook comments. We hypothesize that the homogeneity of hate target in News and Twitter corpora (immigrants) has meant more than the similar linguistic features in Twitter and Facebook data, stemming from the fact that they are both social media texts.

29Participants achieved promising results in the detection of stereotypes, a new pilot task proposed at HaSpeeDe this year for the first time. In our view, stereotype and HS are meant as orthogonal dimensions of abusive language, which do not necessarily coexist. This influenced the design of HaSpeeDe 2, where we proposed two independent tasks for the detection of such categories. However, a first analysis of systems participating in both tasks suggests that most teams did not design a dedicated system for stereotype recognition, but focused on developing a HS detection model, adapting the same model to stereotype recognition, reducing de facto stereotypes to characteristics of HS. We hypothesize that this could be one of the factors that led the systems to not generalize well when applied to the stereotype detection task, especially in texts that are not hateful but contain stereotypes. This hypothesis is confirmed by the high percentage of false negatives (21% in tweets and 35% in news headlines) of the stereotype class in non-hateful texts, with respect to false negatives (5% in tweets and 28% in news headlines) in hateful ones. It is possible to notice the same increase also in false positives in hateful texts. These values suggest that stereotype appears as a more subtle phenomenon that could not give rise to hurtful message. The percentages have been computed taking into account the set of common incorrect predictions of the three best runs in Task B, and calculated in relation to the actual distribution of HS and stereotype in the test set. Analyzing the predictions of the three best runs in Task A, similar influence of stereotype is observed in false negative and positive, but to a minor extent. These results are in line with the observations about emerged from the error analysis of HaSpeeDe 2018 (Francesconi et al. 2019).

30To conclude the discussion on this edition’s results, we comment on the baseline scores obtained for Task C. As it can be noticed from Table 8, the value obtained on the validation set is higher than the ones obtained on both test sets. This variation can be explained by the main characteristics of the data at hand: on the Twitter side, this is due to the different time frames of tweet’s publication included in training and test set, while on the News side, such low value is expected by virtue of the different text domain. Since this baseline uses a memory-based approach, such a low performance is to be expected in datasets from different time frames, since the discussion topics are different and Twitter users change their hashtags and slogans, which are the main repeated items.

7. Conclusions

31In its second edition, the HaSpeeDe task proposed the detection of hateful content in Italian, by challenging systems along two dimensions, time and domain, and taking into account also the category of stereotype, which often co-occurs with HS. This paves the way for further investigations also about the relationships linking stereotype and HS.

32In order to take a step further in state-of-the-art HS detection, the task provided novel benchmarks for exploring different facets of the phenomenon and laying the foundations for deeper studies about the impact of bias, topic and text domain. In this line, also a pilot task about recognition of NUs was proposed, devoted to study this kind of linguistic form in hateful messages in tweets and newspaper headlines, as it has been proved that both headlines in journalistic writings (Mortara Garavelli 1971) and social media texts (Ferrari 2011; Comandini, Speranza, and Magnini 2018) are a fertile ground for NUs. Even though we did not receive any submission for Pilot task 2, our hope is that the fine-grained annotation of hateful data concerning these aspects can be the subject of deeper studies to shed light on the syntax of hate, a topic still understudied.