In this paper we describe the Hate and Morality (HaMor) submission for the Profiling Hate Speech Spreaders on Twitter task at PAN 2021. We ranked as the 19th position - over 66 participating teams according to the averaged accuracy value of 73% reached by our proposed models over the two languages. We obtained the 43th higher accuracy for English (62%) and the 2nd higher accuracy for Spanish (84%). We proposed four types of features for inferring users attitudes just from the text in their messages: HS detection, users morality, named entities, and communicative behaviour. The results of our experiments are promising and will lead to future investigations of these features in a finer grained perspective.
The Profiling Hate Speech (HS) Spreaders on Twitter is an Author Profiling task organized at
PAN [
The Hate and Morality (HaMor) team participates to the task with a system that combines
HS, and moral values detection [
Four types of features for inferring users attitudes just from the text in their messages have been selected to train our model: HS detection (Section 2.1), users morality (Section 2.2), Named Entities (Section 2.3), Communicative behavior (Section 2.4).
Hate Speech Detection (HSD) is the automated task of detecting whether a piece of text contains
hate speech. Several shared tasks on HSD have taken place and large annotated corpora are
available in different languages. For example, the HatEval dataset for hate speech detection
against immigrants and women in Spanish and English tweets have been released to be used at
the Task 5 of the SemEval-2019 workshop [
SemEvalSVM (SESVM): 1-dimensional feature that counts - for each user - the number of hateful tweets predicted by a linear SVM trained using a text 1-3grams bag-of-words representation.
Atalaya (ATA) [12]: 1-dimensional feature that counts - for each user - the number of hateful tweets predicted by a linear-kernel SVM trained on a text representation composed of bag-of-words, bag-of-characters and tweet embeddings, computed from fastText word vectors. We have taken inspiration from the system proposed by the Atalaya team that obtained the best result at the HatEval Spanish sub-task.
Fermi (FER) [13]: a 1-dimensional feature that counts - for each user - the number of hateful tweets predicted by SVM with the RBF kernel trained on tweet embeddings from Universal Sentence Encoder. We based the system on that proposed by the Fermi team that obtained the best result at the HatEval English sub-task.
Furthermore, the growing interest on this topic leads the research community (and not only) to develop some lexica of hateful words such as HurtLex [14], NoSwearing1, and The Racial Slur Database2. HurtLex is a lexicon of offensive, aggressive, and hateful words in over 50 1https://www.noswearing.com/ 2http://www.rsdb.org/full languages (including English and Spanish). The words are divided into 17 different categories. Then, NoSwearing is a list of English swear words, bad words, and curse words. The Spanish translation was made by Pamungkas et. al [15]. Finally, the Racial Slur Database is a list of words that could be used against someone - of a specific race, sexuality, gender etc. - divided into more then 150 categories. The list is only available in English, we thus computed the Spanish translation using Babelnet’s API [16]. We also take advantage of spaCy3 models en_core_web_lg, and es_core_news_lg for expanding the three lexica. Indeed, we used the tok2vec embedding representation for including in the three lists the 10 most similar tokens of each word in order to also capture inflection forms and synonyms. We have thus proposed the following features: − −
HurtLex (HL): a 18-dimensional feature that evaluates the number of hateful words used by each user, the mean of hateful words in each tweet, and the standard deviation. We exploited the following 6 categories: negative stereotypes ethnic slurs, moral and behavioral defects, words related to prostitution, words related to homosexuality, words related to the seven deadly sins of the Christian tradition, felonies and words related to crime and immoral behavior (we exclusively considered the conservative level).
No Swearing (NoS): a 3-dimensional feature that evaluates the number of swearing words used by each user, the mean of swearing words in each tweet, and the standard deviation. The Racial Slur Database (RSdb): a 27-dimensional feature that evaluates the number of swearing words used by each user, the mean of swearing words in each tweet, and the standard deviation for each of the following 9 categories: Asians, Arabs, Black people, Chinese, Hispanics, Jews, Mexicans, Mixed Races, Muslims.
According to many scholars, morality is a pluralistic rather than an universal concept.
Several configuration of values are possible, and some of them are in conflict, such as autonomy
versus community [
In building our model, we considered binding moral dyads as a potential feature characterizing
a HS spreader. More specifically, we claimed that users who rely on loyalty/betrayal and
authority/subversion might be inclined to post hatred contents online. Hence, we referred to
two existing resources: the extended Moral Foundations Dictionary (eMFD) [18], and the Moral
Foundations Twitter Corpus (MFTC) [
The eMFD is a dictionary of terms categorized by a specific moral foundation. We chose all those related to loyalty/betrayal and authority/subversion moral concerns, and translated them in Spanish with the BabelNet’s API. Finally, we expanded the words list using the same methodology explained in Section 2.1. The result is the following feature: − extended Moral Foundations Dictionary (eMFD): a 12-dimensional feature that comprises the mean, the standard deviation, and the total amount of terms occurring in her/his tweets for the four categories loyalty/betrayal and authority/subversion.
The MFTC is a collection of 35, 000 tweets annotated for their moral domains, and organized in 7 subcorpora, each focusing on a specific discourse domain (eg: the Black Lives Matters, and #metoo movements, and the US 2016 presidential elections). Using transfer learning as a label assignment method, we converted the original multi-label annotation schema in a binary-label one: 9, 000 texts annotated as loyalty, betrayal, authority or subversion were considered as potentially correlated with HS (true), while the other not (false). Using the resulting corpora as training set, we thus proposed the following feature.
−
Moral Foundations Twitter Corpus (MFTC): a 1-dimensional feature that counts - for each user - the number of hateful tweets predicted by a linear SVM trained using a text 1-3grams bag-of-words representation.
In a message, the mention of a person belonging to a group vulnerable to discrimination might be seen as a signal of hatred contents, since the clear presence of a target in this kind of expressions allows discriminating between what is HS and what is not. Thereby, we implemented a feature aimed at detecting the presence of a potential HS target within a tweet.
We first collected all the entities of type PERSON in the whole training set detected by the transition-based named entity recognition component of spaCy. Then, we searched the retrieved entities on Wikipedia through the Opensearch API4. The example below shows the Wikipedia pages returned by the Opensearch API when the entity Kamala is requested.
[‘Kamala’,‘Kamala Harris’,‘Kamal Haasan’, ‘Kamala (wrestler)’,‘Kamala Khan’,‘Kamala Surayya’, ‘Kamala Harris 2020 presidential campaign’, ‘Kamaladevi Chattopadhyay’,‘Kamala Mills fire’, ‘Kamalani Dung’]
However, this operation is revealed to be not accurate. In fact, it does not return a unique result for each entity detected by spaCy, but a set of 10 potential candidates. Therefore, we decided to create two lists - one for each language - of HS targets including only persons that belong to categories that could be subject to discrimination.
With the aim of detecting the relevant categories, we scraped the category box from the Wikipedia pages of all entities of type PEOPLE detected by spaCy (3, 996 English, and 5, 089 Spanish). The result is a list of Wikipedia’s categories per language, which needed to be filtered to avoid not relevant results.
4https://www.mediawiki.org/wiki/API:Opensearch
The Figure 1 shows a partial selection of Kamala Harris category box, which contains several references to unnecessary information, such as ‘1964 births’, or ‘Writers from Oakland, California’, but also usefully ones, such us ‘African-American candidates for President of the United States’ or ‘Women vice presidents’.
After a manual analysis of the two lists, we thus narrowed them by a regex filtering, in order to obtain only a set of relevant categories: 279 for English, and 415 for Spanish. Finally, we collected all the individuals who are their members. As final result, we obtained two gazetteers of potential HS targets (7, 5890 entities for English, and 31, 235 for Spanish) in the following format.
{Margaret Skirving Gibb : Scottish feminists,
Melih Abdulhayog˘lu : Turkish emigrants to the USA,
James Adomian : LGBT people from Nebraska [...]} We thus proposed a feature that counts the mentions towards persons belonging to a group vulnerable to discrimination.
−
Named Entity Recognition of HS target (NER): a 5-dimensional feature expressing the total number of potential HS targets mentioned in her/his tweets, the mean, the standard deviation, and the ratio between the number of HS target, and all the HS targets mentioned by the user.
Under the label ‘Communicative Behavior’ a set of features related to the structure of the tweet and to the user’s style has been grouped. The total number, the mean, and the standard deviation have been computed for each feature over all users feeds.
− −
Uppercase Words (UpW): this feature refers to the amount of words starting with a capital letter and the number of words containing at least two uppercase characters.
Punctuation Marks (PM): a 6-dimensional feature that includes the frequency of exclamation marks, question marks, periods, commas, semicolons, and finally the sum of all the punctuation marks mentioned before.
Length (Len): 3 different features were considered to build a vector: number of words, number of characters, and the average of the length of the words in each tweet. Communicative Styles (CoSty): a 3-dimensional feature that computes the fraction of retweets, of replies, and of original tweets over all user’s feed.
Emoji Profile (EPro): this feature tries to distinguish some user’s traits from the emoji her/his used. We implemented a one-hot encoding representation of the modifiers used in the emoji ZWJ sequences (e.g. man: medium skin tone, beard) that includes the 5 different skin tone modifiers and the gender modifiers, in addition to the religious emojis (e.g. Christian Cross) and the national flags.
The organizers provided a train dataset of 400 Twitter’s feeds - 200 written in English and 200 in Spanish - binary labelled with HS Spreader. The distribution is perfectly balanced among the true and false labels. In order to assess the performance of the participating systems, a test set of 200 unlabelled Twitter’s feeds - 100 for each language - was provided. The accuracy was used for evaluating the performance in terms of HS Spreader identification for each language. The final ranking will be the average of the accuracy values per language.
In our experiments, we addressed HS Spreader identification employing a 5-fold validation over the train set with the aim of maximizing the predictive accuracy. The code is available on GitHub for further exploration and for allowing reproducibility of our experiments5. We carried out several experiments by combining all the features introduced in Section 2 for training a linear Support Vector Machine (SVM). Table 1 shows some of the performances we obtained in this experimental setting for each languages. The values for the accuracy are the average over three 5-fold validations using different random shufefls of the train set. The runs we submitted have been highlighted in bold print.
The first row contains the higher result that we obtained in the experimental setting for the two languages. Despite this, we have chosen to submit a Spanish run that reached a lower result. We experimented in other shared tasks that the results obtained in the experimental setting is often much lower than the ofcfiial ones [ 19]. It was precisely for this reason that we chose features - that reached good results, though not the best - based on external sources such as other annotated corpora (FER, ATA, HatEval), lexica (HL, eMFD, NoS), and semi-structured contents (NER). Indeed, if, on one hand, the submitted model for English does not reached the 5https://github.com/mirkolai/PAN2021_HaMor higher level of accuracy reached in the experimental setting, on the other hand, the submitted model for Spanish reached a better result than experimental ones. The omission of BoW and BoE features adversely affects the experimental runs for Spanish, but not for English. However, our choice of not include the features BoW and BoE in the submitted run for English involves the creation of a little feature-space representation of each instance (vectors length of 52 variables). This representation could indeed include too few variables for effectively detecting a so complex phenomena such as HS. Then, we did not investigated cross-cultural differences among English and Spanish speaking countries although it is well known that historical and political factors are relevant on the orientation of moral values [20]. Maybe also for that reason, the features based on the moral values we investigated performs better in Spanish then in English tweets. However, the introduction of the external-sources-based features positively affects the runs for both languages.
Our models obtained 84% (2nd higher result) and 62% (43th higher result) in terms of accuracy on HS Spreader identification respectively for Spanish and English. The final score, used in determining the the final ranking, is the averaged accuracy values per language which corresponds to 73% (19th position - over 66 participating teams).
In this paper we presented an overview of the HAMOR submission for the Profiling Hate Speech Spreaders on Twitter task at PAN-2021. We participated by submitting one run for both tweets written in English and Spanish. Our approach, chiefly based on external resources such as other annotated corpora, lexica, and semi-structured content, proved to be highly successful concerning the task of HS Spreader identification in both languages. The results show that the use of external resources preserves stable values of accuracy between the experimental setting and the prevision of the test set on Spanish sub-task. The proposed lexica gave a considerable contribution for obtaining these results and the use of named entity recognition for detection potential target of HS looks promising. In the future, we plan to employ the features discarded from the submitted run for a prediction on the test set. We also aim to explore a finer grained approach to MFT detection features, considering different combination of moral values, and analyzing how moral attitudes may vary across different countries. Finally, the Named Entity Recognition feature needs to be improved through testing different NER tools, and referring to other semantic resources, in addition to Wikipedia. [12] J. M. Pérez, F. M. Luque, Atalaya at SemEval 2019 task 5: Robust embeddings for tweet classification, in: Proceedings of the 13th International Workshop on Semantic Evaluation, Association for Computational Linguistics, 2019, pp. 64–69. [13] V. Indurthi, B. Syed, M. Shrivastava, N. Chakravartula, M. Gupta, V. Varma, FERMI at SemEval-2019 task 5: Using sentence embeddings to identify hate speech against immigrants and women in Twitter, in: Proceedings of the 13th International Workshop on Semantic Evaluation, Association for Computational Linguistics, 2019, pp. 70–74. [14] E. Bassignana, V. Basile, V. Patti, Hurtlex: A multilingual lexicon of words to hurt, in: 5th Italian Conference on Computational Linguistics, CLiC-it 2018, volume 2253, CEUR-WS, 2018, pp. 1–6. [15] E. W. Pamungkas, A. T. Cignarella, V. Basile, V. Patti, et al., 14-exlab@ unito for ami at ibereval2018: Exploiting lexical knowledge for detecting misogyny in english and spanish tweets, in: 3rd Workshop on Evaluation of Human Language Technologies for Iberian Languages, IberEval 2018, volume 2150, CEUR-WS, 2018, pp. 234–241. [16] R. Navigli, S. P. Ponzetto, Babelnet: The automatic construction, evaluation and application of a wide-coverage multilingual semantic network, Artificial intelligence 193 (2012) 217–250. [17] J. Graham, J. Haidt, B. A. Nosek, Liberals and conservatives rely on different sets of moral foundations., Journal of personality and social psychology 96 (2009) 1029. [18] F. R. Hopp, J. T. Fisher, D. Cornell, R. Huskey, R. Weber, The extended moral foundations dictionary (eMFD): Development and applications of a crowd-sourced approach to extracting moral intuitions from text, Behavior Research Methods 53 (2021) 232–246. [19] M. Lai, A. T. Cignarella, D. I. Hernández Farías, iTACOS at IberEval2017: Detecting stance in Catalan and Spanish tweets, in: Proceedings of the Second Workshop on Evaluation of Human Language Technologies for Iberian Languages (IberEval 2017) co-located with 33th Conference of the Spanish Society for Natural Language Processing (SEPLN 2017), volume 1881 of CEUR Workshop Proceedings, CEUR-WS.org, 2017, pp. 185–192. [20] L. Myyry, K. Helkama, M. Silfver-Kuhalampi, K. Petkova, J. P. Valentim, K. Liik, Explorations in reported moral behaviors, values, and moral emotions in four countries, Frontiers in Psychology 12 (2021) 1468.