In recent years, tasks revolving around hate speech detection have experienced a growing interest in the field of Natural Language Processing. Two main trends stand out in the context of sexism recognition: the focus on overt forms of sexism such as misogyny on social media and tackling the problem as a text classification task. The main objective of this work is to introduce a new approach to tackle sexism recognition as a sequence labelling task, operating on the token level rather than the document level. To achieve this goal, we introduce (i) the FGSDI (Fine-Grained Sexism Detection in Italian) corpus, containing Italian newspaper articles annotated with fine-grained linguistic markers of sexism, and (ii) a two-step pipeline that sequentially performs sexism detection on the sentence level and sexism classification on the token one. Our primary ifndings include that (i) tackling the task of sexism recognition as a sequence labelling task is possible, however, a large amount of labelled data is needed; (ii) leveraging few-shot learning for sexism detection proves to be an efective solution in scenarios where only a limited amount of data is available; (iii) the proposed pipeline approach allows for better results compared to the baseline by doubling the overall precision and achieving a better F1-score.
media platforms such as Twitter and Facebook.
The main contributions of this paper are as follows. First,
According to the Sapir-Whorf hypothesis [
1Code available at: https://github.com/fede-m/Fine-grained-sexismdetection-in-Italian-Newspapers
The interest in the role of language in reflecting and per- We concentrated our analysis on newspaper articles,
petuating societal gender inequalities emerged during which represent an underexplored text type in automatic
the so-called second-wave feminism. In Italy, the main sexism recognition and provide the opportunity to
invescontributor was Alma Sabatini, whose works focused on tigate the presence of linguistic sexism in a more formal
analysing the language used in mass media and educa- context (and covert style) than social media. In particular,
tional publishing, identifying discriminatory patterns [
Most relevant works on this topic focus on sexism de- 3.2. Label Definition and Annotation
tection and categorization tackled as classification tasks,
which assess whether a sentence exhibits sexist content Since we decided to approach the problem of
recognizand which type of sexism it contains. As shown by [6], ing diferent linguistic markers of sexism as a sequence
the most significant shift in this field was the advent of labelling task, the first step was to define the labels to
transformers [
Regarding the Italian language, the main focus so far has 3.2.1. Label Definition
been on misogyny and hate speech detection. In
particular, [16] and [17] studied misogyny and aggressiveness As a baseline, we referred to the work of Alma Sabatini
in Twitter posts. Although not leveraging classification [
it hard for the model to focus on tokens associated with the other labels relevant to our analysis.
During hyperparameter tuning, better results were achieved when training for 5 epochs with a learning rate of 6e-5, and weight decay of 0.004. Training for more epochs caused the model to overfit.
After having defined the FGSDI corpus, we next evaluate on how well we can automatically detect sexist markers. In particular, we decided to tackle the problem as a token classification task, comparing two approaches: ifne-tuning a [5] model and a two-step pipeline inspired by [6]. For each approach, we exper- 4.2. Pipeline imented with diferent models and settings. All experiments were conducted on Google Colab using a T4 GPU. The second approach is a modular two-step pipeline illustrated in Figure 1, which leverages both sequence and 4.1. Baseline token classification sequentially. The main diference from the baseline was the introduction of a preliminary The first approach, which we used as baseline, involved ifltering step, modelled as a binary sentence classificaifne-tuning a model on the token classi- tion task, whose goal was to reduce the total number of ifcation task. We chose RoBERTa since it achieves better non-sexist tokens passed on to the second step which results in many diferent tasks compared to other models performs token classification. from the same BERT family [5]. In particular, we com- We changed our input to sentences instead of entire docpared the performance of two models, namely XLM-R uments to fit the binary sentence classification task. In [23] and Hugging Face’s Italian.5 order to prevent an information loss deriving from havWe used whole documents as input to maximize the con- ing less context for the model to make the prediction, we text provided to the model to make the prediction. As modified the original text by applying coreference resolupre-processing steps, we truncated and padded the texts tion. In particular, we first extracted all coreference heads to fit the 512-token limit of the RoBERTa tokenizer. and respective clusters from the full text using crosslinFor training, we used Cross Entropy with cost-sensitive gual coreference.6 Then, for each sentence, we looked at learning techniques [24] to assign a higher penalty to the whether it contained a coreference and, if so, we added model when it misclassified one of the minority classes the corresponding coreference head at the beginning of (i.e. all the classes marking signs of sexism). The best the sentence in square brackets. Finally, we adjusted the results were achieved by initializing the weight of the "O" labels by assigning label 1 to all sentences containing at label to 0.05 and all the others to 2. This intervention was least one sexist marker, and 0 otherwise. necessary since, due to the sparsity of the data, the ma- In the first step of the pipeline, we applied binary senjority of tokens were classified with the "O" label, making tence classification to filter out sentences that did not 5https://huggingface.co/osiria/roberta-base-italian 6https://pypi.org/project/crosslingual-coreference/ contain markers of sexism (i.e. were assigned label 0 from the model). In performing this task, we compared two diferent transfer-learning methods, namely fine-tuning and few-shot learning and selected the one producing the best results.
For the former approach, we employed the pre-trained model fine-tuned on Italian that we used for the baseline, this time trained on the binary classification task. The model was trained for 14 epochs using a learning rate of 2e-5. No cost-sensitive learning techniques were applied for this task since the labels were more balanced compared to the token classification setting.
For few-shot learning, we employed the prompt-free SetFit (Sentence Transformer Fine Tuning) framework [25] which is composed of two steps. Firstly, it leverages pretrained Sentence Transformer models [26] to generate semantically meaningful embeddings for the provided labelled examples. Then a classification head assigns a class to the embeddings generated by the first step. After experimenting with diferent models, we picked
distilusebase-multilingual-cased-v17 as transformer and the default logistic regression model for the predictions. As additional parameters, we used 10 iterations i.e. number of sentence pairs to generate for contrastive learning (see [25] for more information), 1 epoch with batch size 16 and Cosine Similarity to calculate the distance between embeddings in the learned vector space.
Unlike LLMs and other few-shot learning methods [27] [28], SetFit ofers the advantages of not relying on prompt engineering and of providing outputs in the form of vectors directly containing predictions that do not need additional formatting. Moreover, using few-shot learning allowed us to re-distribute the presence of the labels so that each class was equally represented. In particular, for label 1, we randomly sampled 30 sentences for each phenomenon, whereas for label 0 we sampled 45 sentences.
Finally, the sentences that were assigned label 1 from the ifltering step were used to train the for Italian on the token classification task.
The metrics we considered for evaluation are precision, recall, and F1-score. Given the unbalanced distribution of labels in the dataset in favour of non-sexist tokens, we excluded accuracy, since models could achieve high accuracy by predicting the majority class for all tokens. To assess the token classification results, we used the seqeval [29] framework, which is specifically suited for measuring models’ performance on sequence labelling tasks providing both overall and per-class metrics. For the pipeline, we additionally incorporated the results 7https://huggingface.co/sentence-transformers/distiluse-basemultilingual-cased-v1 Models Performance Metrics. For each metric mean and standard deviation are reported.
Model
and pipeline approaches. For the baseline, we consider the Italian, and for the pipeline, the combination of SetFit for the binary classification and Italian for the token classification.
The pipeline method almost doubled the value for precision compared to the baseline, despite achieving a worse recall. This result was expected since the goal of the ifltering step was to reduce the number of non-sexist sentences to pass on to the next step, therefore lowering the overall recall, to achieve higher precision in the token classification of the remaining sentences. This shows the importance of the filtering step in reducing the imbalance between majority and minority labels, allowing the model to concentrate on more subtle relationships between tokens.
Another aspect to consider is that the baseline is applied to whole documents, whereas the pipeline is based on single sentences. Despite using coreference resolution, this could prevent the model from considering certain relationships between tokens that could help better classify them.
With a higher F1-score, the pipeline approach had overall better results, although both approaches only reached modest values. Nevertheless, this result was expected due to the high imbalance of the dataset, the complexity of the task, and the fact that most minority labels did not have suficiently many examples for the model to learn from. However, we hypothesize that increasing the amount of relevant data could lead to a greater performance gain.
This hypothesis is backed by the error annotation we
conducted to acquire a more detailed overview of which
phenomena were better and which worse recognized by
the model. In the analysis, we focused on the results
of the pipeline only, since it achieved higher precision
showing, therefore, a more fine-grained understanding
of the labels at hand. Moreover, we only consider the
classes with a precision higher than 0.25 since the
remaining ones are characterized by too few instances for grained Sexism Detection in Italian) corpus for which we,
an in-depth examination. Overall, the best results were importantly, provided new in-depth annotation
guideachieved for labels Feminine article before surname and lines. They are based on foundational linguistic work by
Identification through man , followed by Masculine of pro- [
We refer to appendix C for more comprehensive per-label results and error analysis.
By looking at the discrepancies between annotations and model predictions, we could not only shed light on which specific phenomena within a class needed more examples to improve results but also test the robustness of the annotation. In some cases, legitimate doubts arose, highlighting the dificulty of the task and the need for additional annotators to increase the confidence level of the annotation itself.
This work aimed to bridge a gap in the research area of sexism detection and classification in Italian by the following contributions. First, we proposed the FGSDI (Fine[5] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, for Computational Linguistics: Human Language Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Technologies, Volume 1 (Long and Short Papers), Luke Zettlemoyer, Veselin Stoyanov, RoBERTa: A Association for Computational Linguistics, Minrobustly optimized bert pretraining approach, arXiv neapolis, Minnesota, 2019, pp. 4171–4186. URL: (2019). doi:10.48550/arXiv.1907.11692. https://aclanthology.org/N19-1000. [6] Angel F. M. de Paula, Roberto F. da Silva, Detec- [16] Arianna Muti, Francesco Fernicola, Alberto Barróntion and classification of sexism on social media Cedeño, Misogyny and aggressiveness tend to come using multiple languages, transformers, and ensem- together and together we address them, in: Proble models, in: Proceedings of the Iberian Lan- ceedings of the Thirteenth Language Resources guages Evaluation Forum (IberLEF 2022), CEUR, and Evaluation Conference, European Language Spain, 2022. Resources Association, Marseille, France, 2022, pp. [7] Alma Sabatini, Raccomandazioni per un uso non 4142–4148. URL: https://aclanthology.org/2022.lrecsessista della lingua italiana per la scuola e per 1.0. l’editoria scolastica, Presidenza del consiglio dei [17] Samuel Fabrizi, fabsam @ AMI: A Convolutional ministri-Direzione generale delle informazioni della Neural Network Approach, EVALITA Evaluation of editoria e della proprietà letteraria artistica e scien- NLP and Speech Tools for Italian - December 17th, tifica, Rome, 1986. . ed., Accademia University Press, 2020, pp. 35–39. [8] Dagmar Stahlberg, Friederike Braun, Lisa Irmen, doi:10.4000/books.aaccademia.6782.
Sabine Sczesny, Representation of the sexes in [18] Alessandra T. Cignarella, Mirko Lai, Andrea Marra, language, Social Communication, 1st ed., A. W. Manuela Sanguinetti, ’La ministro é incinta’ : A Kruglanski and J.P. Forgas, New York, 2007, pp. 163– Twitter account of women’s job titles in italian, in: 187. doi:10.4324/9780203837702. Proceedings of the Eighth Italian Conference on [9] Irene Biemmi, Il sessismo nella lingua e nei libri Computational Linguistics CliC-it 2021, Torino: Acdi testo: Una rassegna della letteratura pubbli- cademia University Press, Milan, Italy, 2022, pp. 85– cata in italia In: Educazione sessista: Stereotipi di 91. doi:10.4000/books.aaccademia.10525. genere nei libri delle elementari, Educazione Ses- [19] Pierluigi Cassotti, Andrea Iovine, Pierpaolo Basile, sista. Stereotipi di genere nei libri delle elemen- Marco de Gemmis, Giovanni Semeraro, Emerging tari, Rosenberg & Sellier, Torino, 2017, pp. 19–60. trends in gender-specific occupational titles in italdoi:10.4000/books.res.4696. ian newspapers, in: Proceedings of the Eighth Ital[10] Jennifer Prewitt-Freilino, T. Andrew Caswell, Emmi ian Conference on Computational Linguistics CliCLaakso, The gendering of language: A comparison it 2021, Torino: Accademia University Press, Miof gender equality in countries with gendered, nat- lan, Italy, 2022, pp. 369–374. doi:10.4000/books. ural gender, and genderless languages, Sex Roles aaccademia.10907. 66 (2012) 268—-281. doi:10.1007/s11199-011- [20] Elisa Merkel, Anne Maass, Laura Frommelt, Shield0083-5. ing women against status loss. the masculine form [11] Gianna Marcato, Eva-Maria Thüne, Italian. Gender and its alternatives in the italian language, Jourand female visibility in Italian, Gender Across Lan- nal of Language and Social Psychology 31 (2012) guages, Torino, 2002, pp. 187–217. doi:10.1075/ 311–320. doi:10.1177/0261927X12446599. impact.10.14mar. [21] Anna Lepschy, Giulio Lepschy, Helena Sanson, Lin[12] Cecilia Robustelli, Linee guida per l’uso del genere gua italiana e femminile, Quaderns d’Italià 6 (2001) nel linguaggio amministrativo, Progetto Accademia 9–18. doi:10.5565/rev/qdi.51. della Crusca e Comune di Firenze Comune di [22] Elisabeth Burr, Agentivi e sessi in un corpus di Firenze, Firenze, 2012. giornali italiani., in: Dialettologia al femminile. [13] Federica Formato, Linguistic markers of sexism Atti del Convegno Internazionale di Studi, Padova: in the italian media: A case study of ministra and CLUEB, Sappada/Plodn (Belluno), 1995, pp. 349– ministro, Corpora 11 (2016) 371–399. doi:10.3366/ 365.
cor.2016.0100. [23] Alexis Conneau, Kartikay Khandelwal, Naman [14] Fabiana Fusco, Stereotipo e genere : il punto di vista Goyal, Vishrav Chaudhary, Guillaume Wenzek, della lessicografia, Linguistica 49 (2009) 205–225. Francisco Guzmán, Edouard Grave, Myle Ott, Luke doi:10.4312/linguistica.49.1.205-225. Zettlemoyer, Veselin Stoyanov, Unsupervised cross[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, lingual representation learning at scale, in: ProceedKristina Toutanova, BERT: Pre-training of deep ings of the 58th Annual Meeting of the Associabidirectional transformers for language understand- tion for Computational Linguistics, Association for ing, in: Proceedings of the 2019 Conference of Computational Linguistics, Online, 2020, pp. 8440– the North American Chapter of the Association 8451. doi:10.18653/v1/2020.acl-main.747. [24] Victoria López, Alberto Fernández, Salvador Gar- Language (2016) 17.
cía, Vasile Palade, Francisco Herrera, An insight [34] Gilda Sensales, Alessandra Areni, Alessandra Dal into classification with imbalanced data: Empirical Secco, Linguistic sexism in the news coverage of results and current trends on using data intrinsic women ministers from four italian governments: characteristics, Information Sciences 250 (2013) An analysis from a social-psychological perspec113–141. doi:10.1016/j.ins.2013.07.007. tive, Journal of Language and Social Psychology 35 [25] Lewis Tunstall, Nils Reimers, Unso E. S. Jo, Luke (2016) 1–9. doi:10.1177/0261927X16629787.
Bates, Daniel Korat, Moshe Wasserblat, Oren Pereg, [35] Daniel Jurafsky, Universal tendencies in the seEficient few-shot learning without prompts, arXiv mantics of the diminutive, volume 72, De Gruyter (2022). doi:10.48550/arXiv.2209.11055. Mouton, Berlin, Boston, 1996, pp. 533–578. doi:10. [26] Nils Reimers, Iryna Gurevych, Sentence-BERT: Sen- 2307/416278.
tence embeddings using siamese BERT-networks, [36] Robin Lakof, Language and woman’s place, volin: Proceedings of the 2019 Conference on Empiri- ume 2, Cambridge University Press, 1973, pp. 45—cal Methods in Natural Language Processing, Asso- 79. doi:10.1017/S0047404500000051. ciation for Computational Linguistics, Hong Kong, [37] Federica Formato, ‘Ci sono troie in giro in ParlaChina, 2019, pp. 3982—-3992. doi:10.18653/v1/ mento che farebbero di tutto’: Italian female politiD19-1410. cians seen through a sexual lens, Gender and Lan[27] Tom Brown, Benjamin Mann, Nick Ryder, Melanie guage 11 (2016) 389–414.
Subbiah, Jared D. Kaplan, Prafulla Dhariwal, [38] Istituto della Enciclopedia Italiana fondata da GioArvind Neelakantan, Pranav Shyam, Girish Sastry, vanni Treccani, 2012.
Amanda Askell, Sandhini Agarwal, Ariel Herbert- [39] Caitlin Hines, Let me call you sweetheart: The Voss, Gretchen Krueger, Tom Henighan, Rewon WOMAN AS DESSERT metaphor, in: Cultural perChild, Aditya Ramesh, Daniel Ziegler, Jefrey Wu, formances, Proceedings of the Third Women and Clemens Winter, Chris Hesse, Mark Chen, Eric Language Conference, April 8-10, 1994, Berkeley Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Women and Language Group, University of CaliforJack Clark, Christopher Berner, Sam McCandlish, nia, Berkeley, California, 1994, pp. 295–303. Alec Radford, Ilya Sutskever, Dario Amodei, Lan- [40] Caitlin Hines, What’s So Easy about Pie?: The Lexguage models are few-shot learners, arXiv (2020). icalization of a Metaphor, CSLI Publications, Standoi:10.48550/arXiv.2005.14165. ford, California, 1996, pp. 189–200. [28] Haokun Liu , Derek Tam, Mohammed Muqeeth, Jay [41] Caitlin Hines, She-wolves, tigresses, and morphoseMohta, Tenghao Huang, Mohit Bansal, Colin Rafel, mantics, in: Gender and Belief Systems. ProceedFew-shot parameter-eficient fine-tuning is better ings of the Fourth Berkeley Women and Language and cheaper than in-context learning, arXiv (2022). Conference, April 19-21, 1996, Berkeley Women and doi:10.48550/arXiv.2205.05638. Language Group, University of California, Berkeley, [29] Hiroki Nakayama, seqeval: A python framework California, 1996, pp. 303–311. for sequence labeling evaluation, 2018. URL: https: [42] Caitlin Hines, Foxy chicks and Playboy bunnies: //github.com/chakki-works/seqeval. A case study in metaphorical lexicalization, Ben[30] Jeanette Silveira, Generic masculine words and jamins, Amsterdam, 1999, pp. 9–23. doi:10.1075/ thinking, Women’s Studies International Quarterly cilt.152.04hin. 3 (1980) 165–178. doi:10.1016/S0148-0685(80) [43] Caitlin Hines, Rebaking the Pie: The ‘WOMAN 92113-2. AS DESSERT’ Metaphor, Oxford University Press, [31] Peter V. Hegarty, Sandra Mollin, Rob Foels, Bi- New-York and Oxford, 1999, pp. 145–162. doi:10. nomial word order and social status, in: Ad- 1093/oso/9780195126297.001.0001. vances in intergroup communication, Peter Lang [44] Sara Mills, Feminist Stylistics, Routledge, 1995. Publishing, 2016, pp. 119—-135. URL: https://api. doi:10.4324/9780203408735.
semanticscholar.org/CorpusID:151450125. [45] Daniel Gutzmann, Erik Stei, How quotation marks [32] Marlis Hellinger, Hadumod Bußmann, Gender what people do with words, Journal of PragmatAcross Languages - The linguistic representation of ics 43 (2011) 2650–2663. doi:10.1016/j.pragma. women and men., volume 1, John Benjamins Pub- 2011.03.010.
lishing Company, 2001. doi:10.1075/impact.9. [33] Gilda Sensales, Alessandra Areni, Alessandra Dal
Secco, Italian political communication and gender A. Annotation Guidelines bias: Press representations of men/women presidents of the houses of parliament (1979, 1994, and We present the annotation guidelines where for each 2013), International Journal of Society, Culture & label we provide a general description of the phenomena falling within that label, relative examples, and, where deemed necessary, an explanation of the specific example.
Additionally, we provide a translation into English made by us of the Italian examples.
This phenomenon encompasses the usage of the masculine form of substantives as "neutral" to address people of all genders. For this class, the label we used corresponds to the same level of granularity as the one proposed by Sabatini. The only diference is that we decided to follow [12] and exclude the generic masculine used to refer to indefinite groups or individuals. For example, we decided not to include the following cases:
contro il ddl recentemente approvato alla Camera [..] 8 English: [...] the mobilisation of Italian journalists against the recently approved bill [...] Italian: Ma cosa prevede la legge e quali sono le tappe in caso di dimissioni di un sindaco?9 English: So, what does the law stipulate and what are the steps to follow in case a mayor resigns? Explanation: In both examples, using techniques such as splitting [11] (dei giornalisti e delle giornaliste italiane and un sindaco o una sindaca) might hurt the readability of the article, especially if this technique is employed in all cases featuring this type of generic masculine, which is the most common and frequent one.
It follows a list of phenomena the annotator should include in the category Generic Masculine. For each specific phenomenon, we provide examples, an eventual explanation of the example and motivation for considering it in the analysis. a) Usage of words uomo/uomini (man/men) with generic meaning, instead of using more inclusive words such as esseri umani (human beings) or persone (people).
Examples: 8http://www.repubblica.it/cultura/2015/10/08/news/appello_ contro_la_nuova_legge_bavaglio_primo_firmatario_rodota_124630230/?rss 9http://roma.repubblica.it/cronaca/2015/10/08/news/dimissioni_ del_sindaco_ecco_l_iter_che_ne_consegue_secondo_la_legge124599210/?rss Italian: [...] e dei barconi utilizzati dagli scafisti e dai mercanti di uomini. 10 English: [...] and the boats used by men smugglers and traders.
Italian: Insieme avevano deciso di tenere in piedi il sindaco fino alla fine del Giubileo per votare nel 2017, con una sostanziale sovrapposizione del partito e dei suoi uomini nella gestione del Campidoglio. 11 English: Together, they decided to hold up the mayor until the end of the Jubilee to vote in 2017 with a substantial overlap of the party and its men in the management of the Capitol.
Italian: Sono un anarchico io, sono per il libero pensiero però come diceva Lucrezio metto l’uomo al centro della natura, noi siamo ifgli del De Rerum Natura. 12 English: I am an anarchist, I stand for free thinking but as Lucrezio used to say, I put the man at the centre of nature, we are children of the De Rerum Natura.
Italian: Mi avrebbe fatto piacere se avesse parlato [Silvio Berlusconi], ma ha scelto di non intervenire in attesa di un risarcimento, il pronunciamento della Corte europea dei diritti dell’uomo. 13 English: I would have liked him [Silvio Berlusconi] to talk, but he decided not to speak pending compensation, the pronouncement of the European Court of
Men Rights.
Italian: [...] hanno firmato la delega afidata agli uomini del nucleo di polizia giudiziaria [...]. 14 English: [...] they signed the proxy entrusted to the men of the judicial police [...]. 10https://www.repubblica.it/politica/2015/10/11/news/i_ protagonisti_sono_tre_obama_putin_e_francesco-124804296/ 11https://www.repubblica.it/politica/2015/10/09/news/renzi_ha_ gia_deciso_niente_primarie_il_nome_lo_scelgo_io_-124662736/ 12https://firenze.repubblica.it/cronaca/2015/10/26/news/cecchini_ gli_allarmismi_oramai_sono_di_moda_-125935450/ 13http://www.corriere.it/politica/15_ottobre_22/no-stop-twittergasparri-il-selfie-orban-stimo-8e46ae4e-78f6-11e5-95d8a1e2a86e0e17.shtml 14http://roma.corriere.it/notizie/cronaca/15_ottobre_12/rischi-ilgiubileo-roma-piedi-oltre-duemila-anni-305f7aa4-70bd-11e5a92c-8007bcdc6c35.shtml#post-0 The use of the generic "man" contributes to making women invisible and it reinforces the idea of women as someone who deviates from the norm. Diferently from the generic masculine used to refer to indefinite groups or individuals and as underlined by Sabatini, there are good alternatives that can be used to avoid the word "man" in this generic meaning. Moreover, [30] also argues that it is more dificult for a woman to feel included in the concept of "man" or "he". b) Usage of plural masculine with names where at least one of the names is masculine, even if there are more females than males in the group.
Italian: In questi giorni infatti diversi attori, artisti e cantanti come Christiane Filangeri, Claudia Zanella, Claudio Corinaldesi, Daniela Poggi, Elena Santarelli, Fabio Troiano, Filippo Timi, Francesca Inaudi, Giulia Bevilacqua, Jasmine Trinca, Libero De Rienzo, Lillo Petrolo, Lorenza Indovina, Lorenzo Lavia, Luca Argentero, Lucia Ocone, Ludovico Fremont, Maria Rosaria Omaggio, Maya Sansa, Michele Riondino, Sonia Bergamasco, Susanna Tamaro, Valentina Lodovini, Vinicio Marchioni, Remo Girone [...]. 15 English: In the past days, a large number of actors, artists and singers such as Christiane Filangeri, Claudia Zanella, Claudio Corinaldesi, Daniela Poggi, Elena Santarelli, Fabio Troiano, Filippo Timi, Francesca Inaudi, Giulia Bevilacqua, Jasmine Trinca, Libero De Rienzo, Lillo Petrolo, Lorenza Indovina, Lorenzo Lavia, Luca Argentero, Lucia Ocone, Ludovico Fremont, Maria Rosaria Omaggio, Maya Sansa, Michele Riondino, Sonia Bergamasco, Susanna Tamaro, Valentina Lodovini,
Vinicio Marchioni, Remo Girone [...].
Explanation: In the rather long list of names, we can notice that the majority of names are female (14 women and 11 men). Nevertheless, the substantives attori, artisti, and cantanti are only declined in the masculine form. Also, we can see that a long list of names is provided, so the space required to add the words attrici (actress) and artiste 15https://roma.repubblica.it/cronaca/2015/09/30/news/_no_alla_ privatizzazione_dei_canili_comunali_di_roma_il_presidio_dei_ lavoratori_all_ex_cinodromo-123987648/ (artists feminine) would have had a minimal impact on the readability of the article.
Italian: [...] per permettere il soccorso dei due feriti (due donne alla guida delle utilitarie: non sarebbero gravi).16 English: [...] to allow the two injured (two women driving an economy car, none of them seem to be in danger) to be rescued.
Explanation: This last example shows how even in circumstances where the definite referents are all females, the generic masculine is still employed. Note how the author had to add a parenthesis to specify that the two injured were both women, highlighting how the generic masculine alone was not enough to correctly include them.
In this case, we are not referring to an indefinite group of people but to a definite one, in which both women and men are present. Therefore, the specification made at the beginning about leaving out of the analysis the generic masculine when referred to an indefinite group does not hold. c) Usage of the male form for word pairs where female and masculine have diferent lexical roots: fratello (brother), padre (father), fratellanza (brotherhood).
Italian: Io, che nel tempo vengo da lontano quando usava il buon costume, la fratellanza e la gente viveva felice senza tante pretese [...]. 17 English: I come from a past time when good manners were used, there was brotherhood and people lived happily without many pretensions [...].
Explanation: The word fratellanza (brotherhood) comes from the word fratello (brother). The symmetric feminine would be sorellanza (sisterhood). Note that in
Italian there is no word like the English 16http://www.lastampa.it/2015/10/20/edizioni/cuneo/incidente-infrazione-s-benigno-due-feriti-GbEo1nAQWh0SQFVYZwGSxM/ pagina.html 17http://www.corriere.it/cronache/15_ottobre_15/compie-99-annichiede-eutanasia-decido-io-quando-ora-morire-0c8ca15c-730a11e5-8fc1-d31255f25c65.shtml siblings or the German Geschwister to indicate the generic brother and sister relationship.
Italian: [...] Ma il Papa che c’entra? «È venuto un anno fa a Redipuglia, ha fatto un gran discorso sull’amore fraterno, la comunista si è infatuata e ha montato la tendopoli davanti alla scuola [...]. 18 English: [...] What does the Pope have to do with this? «He came last year to Redipuglia and gave a great speech about fraternal love, the communist got a crush and put together the tent city in front of the school [...].
Explanation: Like the word fratellanza (brotherhood), also the word fraterno (fraternal) comes from fratello (brother). Note that in this case, no symmetric equivalent of fraterno is commonly used in Italian, although some proposals such as sorerno or sorellesco have been made.19 Interestingly, the pair materno (maternal) paterno (paternal), similar in meaning and relationship to each other, do preserve the symmetry. Also, as noted by [11] in the case of mother and father, there is a tendency to explicitly include both genders in complex expressions, a practice that seems to be instead considered a "stretch" in basically any other situation. A possible interpretation could be that the realm of motherhood and mother is the only space and role in society which is considered suitable for women and in which women are at least as important as men.
All the words included in this sub-category belong to the class of nouns in Italian in which gender is expressed by using diferent lexical roots rather than adding sufixes. As for the generic "man", also these words tend to hide women’s presence and make them invisible. d) Masculine precedence in male/female oppositional couples. 18http://www.corriere.it/cronache/15_ottobre_11/goriziamigranti-quel-bivacco-parco-caduti-be33ec74-6fe7-11e5a08a-e76f18e62e8d.shtml#post-0 19https://accademiadellacrusca.it/it/consulenza/concorrenti-alfemminile-di-fraterno-scendono-in-gara-sororale-sororiosorellevole-e-sorellesco/10082
Italian: E ricordare che "la pari dignità fra uomo e donna [...] all’insegna della sola differenza che tenta di allontanare le identità uomo-donna". 20 English: And remember that "the equal dignity between men and women [...] under the sign of the only diference that tries to keep man-woman identities apart".
Italian: [...] italiani come noi vogliono una buona legge sui diritti civili ma non vogliono che si tolga il diritto ad un bambino di avere un papà ed una mamma.
21 English: [...] Italians like us want a good civil right law but don’t want to take away the right of children to have a dad and a mum.
Italian: "Sono convinto che la maggioranza degli italiani ritenga che la famiglia naturale sia quella formata da un uomo e una donna." 22 English: "I believe that the majority of Italians considers a natural family the one of a man and a woman." Italian: Il progetto presentato dalla società prevede spazi dedicati alla vendita di abbigliamento maschile e femminile e accessori[...]. 23 English: The project presented by the company includes spaces dedicated to the sale of men and women clothing.
Italian: Arrivano in piazza del Campidoglio in piccoli gruppetti, marito e moglie, tre amiche [...]24 20http://www.repubblica.it/vaticano/2015/10/09/news/sinodo_ emendamenti_italiani_contro_il_gender_e_per_la_famiglia_ uomo-donna-124706948 21http://www.lastampa.it/2015/10/14/italia/politica/unioni-civili-ilsenato-boccia-lo-stop-2JF3S0Cf01foHuw9kzCecM/pagina.html 22https://www.repubblica.it/cronaca/2015/10/01/news/_il_rifiuto_ della_diversita_dietro_queste_mistificazioni_-124088213/ 23https://milano.repubblica.it/cronaca/2015/09/30/news/milano_ hugo_boss_galleria-124024653/ 24http://www.lastampa.it/2015/10/26/italia/politica/lapiazza-spontanea-di-marino-adesso-imbarazza-il-pdvxCVzqw4AWLnQM4LlgSb3I/pagina.html English: They arrive in Campidoglio square in small group, husbands and wives, three friends [...] The underlying idea for this class is that word order can be used as a syntactic means to express existing hierarchies in society. We refer to [31] for an in-depth overview of this phenomenon. e) Usage of donne (women) to indicate a separate category (as if they would not be a part of the other mentioned categories).
Italian: Arrestati cinque cittadini marocchini e due italiani, tra cui una donna, per rapina aggravata in concorso. 25 English: Five Moroccan and two Italian citizens, one of which a woman were arrested for aggravated robbery in complicity.
Italian: Gorizia è città di frontiera, siamo abituati ad accogliere, quando scoppiò la guerra in Jugoslavia arrivarono 17 mila profughi; ma c’erano anche donne e bambini. 26 English: Gorizia is a border city, we are used to hosting, when the war in Yugoslavia broke out, 17 thousand refugees came; but at the time there were also women and children.
Italian: Tre giovani, di 17, 22 e 23 anni, e una ragazza di 22 anni [...]. 27 English: Three young people aged 17, 22 and 23, and a 22-year-old girl [...].
The fact that women are appointed as a separate category where a group of individuals are mentioned has two main efects. On the one hand, this validates the fact that the generic masculine 25https://milano.repubblica.it/cronaca/2015/10/17/news/milano_ rapine_sui_treni-125266311/ 26http://www.corriere.it/cronache/15_ottobre_11/goriziamigranti-quel-bivacco-parco-caduti-be33ec74-6fe7-11e5a08a-e76f18e62e8d.shtml#post-0 27http://www.corriere.it/cronache/15_ottobre_07/catania-scontromoto-quattro-giovani-muoiono-carbonizzati-3e3ade90-6d3711e5-8dcf-ce34181ab04a.shtml#post-0 is not really neutral, as pointed out by the first and last examples. On the other hand, women are perceived as a whole homogenous category, as if their gender would already attribute certain characteristics to them. f) Use of masculine forms for specific female subjects (also for personifications). Italian: Dopo 20 anni di gestione animalista e no profit, vincitore è risultata una impresa barese, proprietaria di un mega canile da 1200 posti a Bari assai fatiscente e gestore di stabulari per animali da laboratorio per l’università di Bari. 28 English: Thanks to 20 years of animal welfare and non-profit management, the winner was a company from Bari, owner of a big, very run-down 1200-seat dog shelter in Bari, which manages a facility for lab animals for the University of Bari.
Explanation: Here we can see that the subject of the sentence is una impresa (a company), which is grammatically feminine. Despite that, both vincitore (winner) and gestore (manager, supervisor) are declinated in the masculine form. It is interesting that proprietaria (owner) is instead correctly feminine.
Italian: Questo vuole Putin, che sa tuttavia di dover stipulare un accordo con gli Usa e con Obama in particolare perché chi tra un anno gli succederà non è detto che conceda alla Russia il ruolo di comprimario che Obama, pur cercando di limitarlo, è comunque disposto a riconoscergli. 29 English: This is what Putin wants, however, he knows that he will have to make a deal with the US and Obama in particular because whoever will succeed him might grant Russia the supporting role that Obama, although trying to limit it, is still willing to recognise it.
Explanation: Russia is feminine in Italian, nevertheless both the word comprimario and the clitic gli are in the masculine form. As 28https://roma.repubblica.it/cronaca/2015/09/30/news/_no_alla_ privatizzazione_dei_canili_comunali_di_roma_il_presidio_dei_ lavoratori_all_ex_cinodromo-123987648/ 29https://www.repubblica.it/politica/2015/10/11/news/i_ protagonisti_sono_tre_obama_putin_e_francesco-124804296/ noted in [11], clitic gli often replaces the feminine le even in contexts where the referent is clearly feminine.
Italian: Il ministro Maria Elena Boschi [...]. 31 English: Minister Maria Elena Boschi [...]. This is in line with the tendency of the generic masculine already observed before. Also, in all these cases the choice could additionally be biased by the fact that the roles expressed by these terms are usually associated with men.
We added this label as sub-category of Generic masculine to identify cases where the "rule" of generic masculine was not applied if the profession or role indicated by the substantive was stereotypically occupied by women. This phenomenon is related to the concept of social gender described by [32], which refers to the tendency to use female pronouns or nouns when referring to professions which are lower status and usually occupied by women and male ones in all other cases. We found however only one example for this class.
This class corresponds to Sabatini’s label Asymmetries Italian: [...] racconta il suo avvocato Erika Galati [...].
32 English: [...] says her lawyer Erika Galati [...]. Italian: [...] candidare una donna premier?33
Italian: Henriette Reker, la candidata che sabato è stata vittima di un’aggressione xenofoba per il suo impegno a favore dei migranti, è stata eletta sindaco di Colonia. 34 English: Henriette Reker, the candidate who was the victim of xenophobic aggression on Saturday due to her commitment to immigrants, was elected mayor of Cologne.
Italian: [...] candidata sindaco di Colonia alle elezioni
in programma domani [...] 35 English: [...] mayor candidate of Cologne in tomor
row’s elections [...] Italian: Per questo, Salvini dopo l’endorsement al leader di Fratelli d’Italia Giorgia Meloni come possibile candidato sindaco del centrodestra a
Roma [...]. 36 English: For this reason, Salvini, after the endorsement of Giorgia Meloni, leader of Fratelli d’Italia, as a possible centre-right mayor candidate in Rome [...]. in the usage of agentives and analyses two phenomena. 31http://www.corriere.it/politica/15_ottobre_13/senato-riformaOne of them is the usage of the masculine form of traguardo-opposizioni-non-voteranno-ae4eb4fc-716c-11e5professional titles (especially for high-status ones) to b015-f1d3b8f071aa.shtml refer to specific female referents. The other is the use of 32http://www.corriere.it/cronache/15_ottobre_08/funeralidonna (woman) as a modifier attached to the masculine 1ca1tet5o-l8icaie-cla-3-m6da7d8rfe2-dfca6t0im4.ash-jtimhal#dpisotsat--i0taliana-7bdc0eb6-6de3form of the profession. Sabatini also included in this 33http://www.lastampa.it/2015/10/20/italia/politica/la-grandecategory the creation of the agentive forms through the tentazione-di-casaleggio-in-campo-direttamente-lui-oppuresufix -essa, for which we decided to create a separate una-donna-O6NTs3Vtws2OYuRIPoIl8J/pagina.html class. 34http://www.corriere.it/esteri/15_ottobre_18/colonia-candidatavittima-aggressione-stata-eletta-sindaco-cfb286da-75bf-11e5a6b0-84415fd3d85.shtml Examples: 35http://www.lastampa.it/2015/10/17/esteri/agguato-acolonia-ferita-a-coltellate-candidata-sindaco-indipendente30http://www.repubblica.it/scuola/2015/10/20/news/l_ora_di_ sDXbWLi9YLLuwPujF7TQpO/pagina.html religione_in_aule_semivuote_ma_e_vietato_unire_le_classi_- 36https://milano.repubblica.it/cronaca/2015/10/09/news/salvini_ 125463096 maroni-124724471/ Explanation: Although having used the neutral form A.5. Asymmetric usage of names, leader to refer to the politician Meloni, the author surnames, and titles of the article still endorses the masculine form by using the male preposition al (to) instead of the Sabatini includes in this class instances where female correct female one, alla. The same can be noted referents are only referred to by their first name, also for the compound candidato sindaco, where asymmetries in the usage of the word signora (which both nouns are declined in the masculine forms, translates to both lady and Mrs), and the usage of the although they refer to a woman. This compound feminine article before surnames. We primarly focused occurred often in the corpus, also in the form on the first phenomenon i.e. the asymmetry of the candidata sindaco (first noun in the feminine and usage of the first-name-only to refer to women, and the second in the masculine form) but never in the latter, to which, given the high frequency with which it whole feminine form candidata sindaca. occurred, we dedicated a separate class.
One can argue that this last form might sound In [33], the authors note how using first-names-only has incorrect, but it is asymmetric with respect to a trivializing and degrading function since first names similar constructions such as candidata maestra are commonly used to refer either to children, people vs candidata maestro (teacher candidate), where belonging to the personal sphere, or those deemed probably the first option would sound more ap- occupying an inferior position in the social hierarchy propriate than the second one, although both are scale. Additionally, while appearing in the news provides rare in usage. visibility, this is ofset by the impossibility of obtaining more information about the referenced people, as it is not feasible to search for somebody only by their first A.4. Usage of "-essa" sufix name (e.g. in a search engine). In general, we noted the co-occurrence of this phenomenon almost exclusively with female referents.
Among the diferent sufixes that the Italian language uses to derive the feminine form from the masculine one, the -essa sufix seems to be consistently considered in the literature as bearing a negative connotation (see [20], [21] and [22]). This is also evident from the fact that there exist alternative forms for nearly all substantives that make use of this sufix. In this regard, we must make a distinction between words which are nowadays commonly used in Italian and which have therefore lost the negative connotation, such as professoressa (professor), and more recent neologisms such as avvocatessa (lawyer) for which using the form avvocata is to be preferred.
Explanation: [20] analyses the perception of people towards diferent professional titles used to refer to women. At the time of the analysis, avvocata (which is the grammatical feminine derivation of avvocato) was still considered to be agrammatical.
Nevertheless, participants in the study attributed a higher degree of competence to female referents designated with this title, than with the more spread avvocatessa. 37http://www.corriere.it/esteri/15_ottobre_15/oscar-pistoriusandra-domiciliari-partire-20-ottobre-dbda2808-7337-11e5-b97329d2e1846622.shtml
Italian: La vita (social) di una moderna eremita Rachel Denton, 52 anni, è una carmelitana cattolica. [...] Ma a diferenza degli eremiti del passato, Rachel non vive in una grotta [...] Rachel ha comunque deciso di continuare a vivere in solitudine. 38 English: The social media life of a modern hermit.
Rachel Denton, 52 years old, is a Carmelite catholic. [...] However, diferently from the hermits of the past, Rachel does not live in a cave [...] Rachel still decided to keep living in solitude.
o Sonya ma indosseranno il reggiseno «Elena», o quello «Sofia», il modello «Gioia» oppure «Francesca». 39
Fiona or Sonya, but they will wear the «Elena» or «Sofia» bra, or the «Gioia» or «Francesca» model. 38http://www.corriere.it/foto-gallery/esteri/15_ottobre_12/vitasocial-una-moderna-eremita-ba59d8c2-70da-11e5-a92c8007bcdc6c35.shtml#post-0 39http://www.corriere.it/moda/news/15_ottobre_12/miriambarbara-marsica-modelle-sono-ragazze-normali-c5300ad4-710f11e5-a92c-8007bcdc6c35.shtml Explanation: In the first example, Rachel Denton is introduced with both name and surname only at the beginning of the text. Instead of referring to her by surname, as we noticed to be the norm in analogous cases where men were subjects, the author keeps calling her by first-nameonly throughout the whole article. In the second example, women’s surnames were not mentioned even at the beginning of the article.
Explanation: See here the dissymmetric use of the article la in front of the surnames Taverna and Boschi, but not in front of Castaldi.
Italian: Berlusconi chiede alla Merkel un aiuto [...] 42 English: Berlusconi asks the Merkel for help [...]
A.6. Feminine article before surname This category is part of Sabatini’s Asymmetries in the usage of adjectives, substantives, diminutives, and verbs, We decided to dedicate a separate class to this phe- though we decided to address each of these phenomena nomenon due to its high frequency. The asymmetric separately. The decision was mainly motivated by the usage of the feminine article la followed by the surname low frequency of the single categories, whose specific of a woman, also defined as dissymmetric feminine in nuances were easier to identify using smaller and less [34], is widely spread in the Italian language. Being not ambiguous labels. used for men, the functionality of this marker is mainly The adjectives that we considered in the analysis refer to make the gender of the person visible, attaching even mainly to three semantic areas that perpetuate the to proper names, as noted in [11], the gender bias that gender bias of seeing women as small, silent, and perceives women as the exception to the norm. uniquely identified through physical characteristics (which reinforces the idea of women as sex objects).
Examples: Additionally, we included other adjectives that we noticed being used asymmetrically for men and women.
Italian: La Eva Longo... che lo sai, no? è grande amica Following the approach used in [13], we double-checked di Nicola Cosentino, Nick o’ mericano, a sua volta each potentially asymmetric adjective on Word Sketch43, amico dei Casalesi... beh, la Longo s’aspetta di a tool that shows in which contexts a word typically diventare presidente della commissione Infras- appears and to which other words it is generally trutture... Poi c’è [...]. 40 associated.
English: The Eva Longo... who, you know right?, is a good friend of Nicola Cosentino, Nick the Ameri- Examples: can, who is in turn a friend of the Casalesi family... Italian: Lei è stata per decenni la nostra vivacissima, well, the Longo expects to become president of intelligentissima ’spalla’. [...] era una persona inthe Infrastructure Commission... Then there is tellettualmente vivace [...] con quel suo musetto [...]. dolce e furbo [...] sproporzionata rispetto al corpo esile [...]. Quel lavoro silenzioso [...]. 44 Explanation: Note how here even the full name of
Eva Longo is preceded by the feminine article la, asymmetrical to Nicola Cosentino’s name which has no article.
Italian: Anche quando la Taverna chiama prostituta la Boschi, o quando Castaldi mi dà del parassita sociale. 41 English: Also when the Taverna calls the Boschi a prostitute, or when Castaldi calls me a social parasite. 40http://www.corriere.it/politica/15_ottobre_02/accuse-dollarifalsi-veleni-verdiniani-resa-ce-chiudono-3a2b9798-68c5-11e5a7ad-17c7443382c3.shtml 41https://www.corriere.it/politica/15_ottobre_05/non-devoscusarmi-quel-gestaccio-l-ha-fatto-lezzi-io-l-ho-mimato9194e8f4-6b4a-11e5-9423-d78dd1862fd7.shtml
English: She has been for decades our very lively, very clever ’sidekick’. [...] she was an intellectually lively person [...] with her lovely, astute little face [...] disproportionate to the slight body [...].
Her silent work [...].
Explanation: Noteworthy is here the usage of the word vivace (lively). This term is usually used to refer to children, for example in the expression è un bambino vivace (he is a lively child). This is also backed by Word Sketch, where the only 42http://www.lastampa.it/2015/10/23/italia/politica/berlusconichiede-alla-merkel-un-aiuto-per-tornare-in-sella
AYibqnhAmZhvGxdgRHlJaJ/pagina.html 43https://www.sketchengine.eu/guide/word-sketch-collocations
and-word-combinations/ 44http://www.repubblica.it/cultura/2015/10/14/news/daniela_
bellingeri_lutto-125034845 human referents for the adjective are namely the substantives bambino, bimbo, bambina.
The expression is not typically used to refer to adult men. The underlying idea is to draw a parallelism between women and children [35]. Also, the adjectives dolce (sweet) and esile (slender) are rarely used for men since they do not adhere to their stereotypical gender roles.
Both mostly refer to inanimate subjects and their only human subject indicated on Word Sketch is femmina (female). Additionally, [36] mentions the adjectives lovely and sweet (both translated as dolce in Italian) as being typically feminine. As for the association between women and silence, here the silent work conveys precisely the idea of knowing one’s place, highlighted also by the use of ’sidekick’ to describe the referent’s attitude to do her job in the shadows, without seeking due recognition.
ista nizzarda Solenne Païdassi approda domani sera alla Verdi, sull’onda di una notorietà ormai internazionale. 45 English: Gutsy and quite resolute, the violinist from Nice Solenne Païdassi will land tomorrow evening at the Verdi theatre, on the wake of an at this point international notoriety.
Explanation: The word grintoso (gutsy) has femminilità grintosa (gutsy femininity) among its noun modifiers on Word Sketch. Note that in [14], in the initial examples that refer to the Signorino Buonasera, we find the word grinta (grit), sarcastically used to refer to a man. It is noteworthy, that the word determinata (resolute) would itself be more stereotypically masculine, therefore the author uses quite to smooth its meaning.
mali». Perchè loro, le splendide «modelle per caso» di Intimissimi [...] rendendo protagoniste le personalità di donne reali[...]. 46 English: Simple and charming names for «normal» women. Because they, the splendid «models by chance» from Intimissimi [...] featuring real women [...]. 45https://milano.repubblica.it/cronaca/2015/10/08/news/solenne_ pai_dassi_il_mio_stravinskij_brioso_e_ardente_vi_emozionera_124622485/ 46http://www.corriere.it/moda/news/15_ottobre_12/miriambarbara-marsica-modelle-sono-ragazze-normali-c5300ad4-710f11e5-a92c-8007bcdc6c35.shtml Explanation: It is interesting to see how the words normal and real are here used to refer to and comment on the bodies and the physical appearance of these women. This intention is made evident by the fact that the referenced women are called "models by chance", which explicitly draws a parallel between models’ physical appearance and the one of "normal" women. We can note the asymmetric usage of the adjectives by changing the referent to a male one, since the expression real man is more related to moral and behavioural attitudes. The same can be noticed also for the expression normal man, where normal also refers more to the mental/psychological sphere rather than the physical one. Splendida, which has in Italian a connotation similar to amazing in English, and charming are also mentioned in [36] as typically feminine adjectives.
Italian: [...] ci sono scatti di Sebastiano F. assieme a una showgirl bionda che gli cinge la vita, a una mora altrettanto famosa e procace.47
This class is part of Sabatini’s Asymmetries in the usage of adjectives, substantives, diminutives, and verbs and exhibits asymmetry in two key aspects.
The first is the presence of words exclusively associated with women, for which a corresponding male form does not exist. These words mostly come from semantic domains such as sexuality, physical appearance, and marital status, which describe societal realms in which women are often relegated. This phenomenon can be exemplified by the absence of a masculine form for the word prostituta (prostitute). As noted in [11], this is not in line with the trend in Italian of creating a masculine term when men start occupying professions traditionally occupied by women only (see the word ostetrico, obstetrician).
The second phenomenon we want to investigate in this 47http://roma.corriere.it/notizie/cronaca/15_ottobre_23/scandalogay-gigolo-collaborava-la-onlus-fondata-un-cardinale6defd1b8-7903-11e5-95d8-a1e2a86e0e17.shtml class is word pairs that, despite having a denotatively equivalent male version, carry a negative connotation when used in their feminine form. In this context as well, the semantic loading (see [11]) attributed to the female version of these words often has a sexual undertone.
One example of this phenomenon is the asymmetric usage of zitella (spinster) and scapolo (bachelor). Italian: Il video della campagna - che ha come testimo
nial la showgirl Filippa Lagerbäck - [...]. 48 English: The campaign video - which has showgirl
Filippa Lagerbäck as testimonial - [...] Explanation: The term showgirl does not have a male equivalent, since *showboy does not exist. Rather, the word presentatore (host, presenter) would be used for men. Also, the original meaning of showgirl in English was a young woman regarded as an object of display49, which gives a sexual connotation to the term, moving the attention to the outer appearance of women rather than to their profession or talent and reinforcing the idea of women as objects.
Italian: Anche quando la Taverna chiama prostituta la Boschi, o quando Castaldi mi dà del parassita sociale. 50 English: Also when the Taverna calls the Boschi a pros
titute, or when Castaldi calls me a social parasite.
Explanation: Prostituta is asymmetric in that there exists no male equivalent, both grammatically (*prostituto) and semantically (the use of gigolò does not have the same negative connotation).
Italian: Si tutela il diritto del fanciullo alla continuità afettiva e si rendono entrambi i partner titolari di diritti e doveri verso di esso. 51 English: The right of the child to emotional continuity is protected, and both partners are appointed rights and duties towards it. 48http://www.repubblica.it/ambiente/2015/10/02/news/
salvaciclisti_limite_auto_citta_-124169246 49https://www.oed.com/dictionary/showgirl_n?tab=meaning_and_
use 50https://www.corriere.it/politica/15_ottobre_05/non-devoscusarmi-quel-gestaccio-l-ha-fatto-lezzi-io-l-ho-mimato9194e8f4-6b4a-11e5-9423-d78dd1862fd7.shtml 51http://www.lastampa.it/2015/10/14/italia/politica/unioni-civili-ilsenato-boccia-lo-stop-2JF3S0Cf01foHuw9kzCecM/pagina.html
Explanation: In this case, the word fanciullo is used to indicate children in general. However, the asymmetry here lies in the fact that, while the male form is used as a synonym for children, the feminine fanciulla is employed also for young women. The definition of fanciulla is namely young woman or non-married woman of any age or young woman with whom one makes love52.
This last definition shows how the term can also be loaded with sexual connotations (not carried by the word bambina, which better defines a girl-child). Therefore, we can argue that the word bambino is to be preferred in this context.
Italian: La signora, assunta con un contratto a tempo
determinato di cinque mesi [...]. 53 English: The lady, employed on a fixed-term contract
for five months [...].
Italian: [...] Lui in due anni e mezzo ha fatto quello che “questi qua” non hanno fatto in 40 anni», protesta una signora. 54 English: [...] In two years and a half, he managed to do what "that others" did not manage in 40 years», complains a lady.
Explanation: In the last two examples, we can see the asymmetric usage of the word signora, in the meaning of lady. In the examined corpus, all instances of signore were always followed by the last and/or first name of the referent, which suggests its usage mainly as a title. In contrast, signora, exactly like lady in English, can be used as a synonym for woman and appears in contexts, like the ones in the examples, where for the masculine the word man would be employed.
Italian: Ma che sarebbe solo la piccola parte scoperta di una imponente rete sommersa, bracconieri e commercianti che farebbe capo proprio alla signora
Yang Feng Glan [...]. 55 52https://www.treccani.it/vocabolario/fanciulla_%28Sinonimi-e
Contrari%29/ 53http://roma.corriere.it/notizie/cronaca/15_ottobre_09/buzzivince-tribunale-ma-solo-contro-l-ex-amante-c55a683e-6e8911e5-aad2-b4771ca274f3.shtml#post-0 54http://roma.corriere.it/notizie/cronaca/15_ottobre_12/rischi-ilgiubileo-roma-piedi-oltre-duemila-anni-305f7aa4-70bd-11e5a92c-8007bcdc6c35.shtml 55http://www.corriere.it/cronache/15_ottobre_08/tanzania-larresto-regina-dell-avorio-007-italiani-ca0e9016-6df4-11e58aec-36d78f2dc604.shtml#post-0 Explanation: Conversely, we decided not to consider cases like this last example, where signora was followed by the name and/or surname of the person. This choice was motivated by the fact that, at least in the corpus examined, we did not find strong asymmetries with the masculine counterpart.
English: However, this would be only the small un- Explanation: The asymmetry lies here in the reference covered part of a huge underground network of to give kisses, which is a verb that belongs to poachers and traders under the control of Mrs. the private sphere and is here used instead in Yang Feng Glan [...]. a public context. This is in line with what was noted by [37] about the overlapping of the private and public spheres which permeates the Italian political scene and becomes even more evident in connection with women.
Italian: [...] Matteo Salvini che considera «pazzesco» che venga indagato e «sputtanato» un «leghista onesto e concreto». 58
Since Sabatini provided only some very specific examples for this category, we tried to identify and assess possible asymmetries based on the examples found in the corpus and what was examined for the other categories. Through this analysis, we identified two main trends. The first pertained to the usage of verbs derived from the same or similar semantic areas stereotypically associated with women that were pointed out in the previous classes. The second focuses on the roles assumed by female and male actors in the use of certain verbal constructions. In particular, we limited our analysis to verbs in which both men and women referents were included in the action, but only men had the agentive roles, leaving women the role of passive objects. Italian: [...] il compagno musicista, la portava in campagna. 56 English: [...] the partner, who’s a musician, took her to the countryside.
Explanation: In the construction "male subject + take + female object + to do something", men and women do not participate together in the action. Rather, the man takes on an agentive role and the woman the passive role of being the one "taken somewhere to do something".
Italian: [...] e poi alla Boschi passerà la voglia di ridere, di dare baci e inizierà a sudare freddo. 57 English: [...] and then, the Boschi will get over the urge to laugh, give kisses, and she will break out in a cold sweat. 56http://www.repubblica.it/cultura/2015/10/14/news/daniela_ bellingeri_lutto-125034845 57http://www.corriere.it/politica/15_ottobre_05/ddl-boschisenato-articolo-6-voto-segreto-846e6dae-6b80-11e5-9423d78dd1862fd7.shtml#post-0 English: [...] Matteo Salvini, who considers «insane» that an «honest and authentic member of the Lega party» will be investigated and «fucked up». Explanation: The verb sputtanare (to fuck up) comes from the root of puttana (slut). As for prostituta, puttana does not have a male equivalent, which makes the word itself and all its derivations asymmetric.
Diminutives are the last aspect taken into consideration in Asymmetries in the usage of adjectives, substantives, diminutives, and verbs. In [35], the author draws a detailed picture of the semantic meanings associated with the diminutive. In particular, he identifies a link between diminutives and the female gender across all languages, based on the conceptual metaphor of women as children and "small things" in general. This conceptualization derives from the opposition between female/male, which sees women as smaller than men, both on a physical and power level. It is interesting to note, that this parallel between women and children could also explain the asymmetry in first-name references to women and men.
Italian: Con il sorriso, con quel suo musetto dolce e furbo, gli occhialetti [...]. 59 English: With her smile, her lovely astute little face, the small glasses [...].
Explanation: In Italian, diminutives are formed using sufixes -etto, -ino, -ello, and -uccio [38] as modiifers of the lexical root to which they are attached. Note that the article from which both examples are taken refers to a woman in her 50s, although 58http://www.lastampa.it/2015/10/13/italia/politica/berlusconimantovani-corretto-sono-stupito-MeStPSkDhe5HPxfSAV3iyH/ pagina.html 59http://www.repubblica.it/cultura/2015/10/14/news/daniela_ bellingeri_lutto-125034845 the use of diminutives associates her more with a child than with an adult woman. Moreover, the word musetto, diminutive of muso (face, snout), contributes to the metaphor of women as small animals (see A.11).
This label corresponds to the same level of granularity as Sabatini’s Asymmetries in the usage of images and tone. Concerning the tropes, we focused mainly on the use of metaphors, metonymy and synecdoche since they are more common, but other types of tropes should also be considered in this category if instances of them are present in the corpus. The methaphors we focused on are based on [39], [40], [41], [42], and [43], and are: • Women as small animals: echoes back to the idea of women as prey in the "sex-is-hunting" metaphor • Women as femmes fatales: compares women, usually occupying positions of power, to either felines (tigers, lionesses, cats), to underline their slyness and charm, or insects known to have power over their male counterparts (lucciola, firefly) • Women as flowers : suggests the idea of the fragility and powerlessness of women.
Another trope that seems to be widely used in this context is metonymy, and more specifically synecdoche, in which women are presented by only referring to their single body parts. This has the result (and aim) of objectifying the woman referent by presenting her as a mere anatomical fragment, only there for the male gaze to be pleased [44].
As for the asymmetric usage of tone, we limited our analysis to a single phenomenon which seemed to co-occur frequently with women referents in the corpus, namely the use of scare quotes [45]. This decision was motivated by the high level of interpretability of what to consider a "sexist tone" and the dificulty (already for human beings, let alone for models) to assess it. We also included in this class idioms and proverbs that have a misogynistic and sexist undertone. Italian: [...] uno scricciolo di donna. 60 English: [...] a little slip of a woman. 60http://www.corriere.it/cronache/15_ottobre_23/ciao-vera-fattamercurio-elegante-irrequieta-dificile-non-averti-qui-23ca7324796f-11e5-a624-46f9df231ebf.shtml Explanation: Here we can see the usage of the woman as small animal metaphor. In Italian, scricciolo literary means Winter Wren, a bird characterized by its small dimensions. Moreover, the definition provided by Treccani 61 attests to its usage to refer specifically to children, which makes the whole metaphor also in line with the parallel woman-child.
Italian: E quello che “rinuncia a 42 milioni di euro mentre gli altri hanno approvato la Legge Boccadutri (o bocca di rosa) con tempi da speedy gonzales”. 62 English: And the one who ’gives up 42 millions euros, while the other approved the Boccadutri Law (or bocca di rosa) at speedy gonzales speed.’ Explanation: The expression bocca di rosa (mouth of rose) is particularly interesting. On the one hand, it represents the metaphor women as flower due to the reference to the rose, which is rich in symbolism in Western cultures. On the other, bocca di rosa is the title of a song by Fabrizio De Andrè, a famous Italian singer-songwriter. The song narrates the story of a sex worker, who is referred to namely as bocca di rosa, and the term has therefore become a synonym for prostitute in Italian. Thus, in this example, the dimension of fragility and that of sex intertwine in a single oxymoronic metaphor.
Italian: Dall’altro il pragmatismo di Casaleggio che fa capire con chiarezza chi porta - e continuerà a portare per un po’ - i pantaloni in casa Movimento 5 Stelle [...]63 English: On the other side, we have the pragmatism of Casaleggio which shows who wears - and will keep wearing for a while - the trousers in the house of Movimento 5 Stelle.
Explanation: This example refers to sexism in idioms.
Trousers were in the past a piece of cloth worn only by men so that the expression has the same meaning as to be the man of the house. This refers to the clear patriarchal hierarchy that sees men as the ones who decide and rule within the 61https://www.treccani.it/vocabolario/scricciolo/ 62http://www.corriere.it/politica/15_ottobre_17/grillo-bis-sognotogliere-mio-nome-logo-maio-candidato-premier-non-certoabbiamo-regole-5faea604-750e-11e5-a7e5-eb91e72d7db2.shtml 63http://www.lastampa.it/2015/10/19/italia/politica/ casaleggio-stoppa-di-maio-non-passiamo-il-testimoneVHrr5YruY7MTPfLZtITs2I/pagina.html domestic walls. This idea is reinforced by the juxtaposition of the word casa (home), which indicates a private space, and the name of the political party, which is instead public [37].
Italian: [...] ci sono scatti di Sebastiano F. assieme a una showgirl bionda che gli cinge la vita, a una mora altrettanto famosa e procace.64 English: [...] there is a photo shoot with Sebastiano F. together with a blonde showgirl encircling his waist, and an equally famous and provocative brunette.
Explanation: This is an example of synecdoche. Note how the information provided to identify the subjects varies across the sentence. First, the only man among them is presented by his first name (and the initial of the surname, probably for privacy reasons). Then, the first woman is described by her hair colour and her professional title (asymmetric as we noted in A.8 ). Finally, the last one is only denoted by a fragment of her body, namely her hair colour, and her attitude, which additionally carries a clear sexual undertone.
cinese di successo, traficante di zanne nel tempo libero. 65 English: The «queen of ivory» is a successful Chinese entrepreneur, who trafics ivory fangs in her free time.
Italian: [...] sarebbe diventata un «capo» assoluto [...].
66 English: [...] she would have become an absolute «boss» [...].
Italian: [...] la presidente nazionale della Fiab Giulietta Pagliaccio si è "armata" di vernice bianca e pennello [...]. 67 64http://roma.corriere.it/notizie/cronaca/15_ottobre_23/scandalogay-gigolo-collaborava-la-onlus-fondata-un-cardinale6defd1b8-7903-11e5-95d8-a1e2a86e0e17.shtml 65http://www.corriere.it/cronache/15_ottobre_08/tanzania-larresto-regina-dell-avorio-007-italiani-ca0e9016-6df4-11e58aec-36d78f2dc604.shtml#post-0 66http://www.corriere.it/cronache/15_ottobre_23/ciao-vera-fattamercurio-elegante-irrequieta-dificile-non-averti-qui-23ca7324796f-11e5-a624-46f9df231ebf.shtml 67http://www.repubblica.it/ambiente/2015/10/02/news/ salvaciclisti_limite_auto_citta_-124169246 English: [...] the national president of Fiab Giulietta Pagliaccio "armed" herself with white paint and brush [...].
Italian: Senza mezzi termini le ’ha cantate’ su Facebook a un’agenzia di modelle che le aveva chiesto di dimagrire [...]. 68 English: Bluntly, she ’gave it’ on Facebook to a modelling agency who asked her to lose weight [...].
Explanation: In many texts, we detected the usage of quotation marks to attenuate the meaning of verbs or substantives usually associated with masculinity when used to refer to female subjects. The first two cases exemplifying this phenomenon are the words capo (boss) and regina (queen) in quotation marks. Regarding the ifrst, there is no contextual reason that suggests such use of scary quotes, since being a boss should not be something extreme for women. For the second, one can argue that the intention was to mark the whole expression queen of ivory as a nickname for the woman. If that is the case, this would attribute a sense of paternalism and trivialization to the story, which is nevertheless to be considered an instance of sexism in the use of tone and therefore classified under this category.
The remaining examples employ scare quotes to attenuate verbs. In the first case, the verb armarsi (to arm oneself), clearly echoes images of war and violence. This must have seemed too strong to be associated with a woman, and therefore the author preferred to attenuate its meaning by adding quotation marks. As for the second, the choice of the verb cantarle is already attributing a note of attenuation and trivialization to the narration, even without the usage of scare quotes. Italian: [...] intelligentissima ’spalla’, l’anima dell’archivio [...]. Lei era la nostra ’complice’ [...] le piaceva ’regalare’ le sue capacità [...] molti di noi hanno continuato a ’saccheggiare’ la disponibilità e cultura di Daniela [...]. 69 English: [...] very clever ’sidekick’, the life of the archive [...]. She was our ’accomplice’ [...] she 68http://www.corriere.it/salute/nutrizione/15_ottobre_16/modelladice-basta-andate-fare-c-non-posso-tagliarmi-ossa-6e4a9b5a7400-11e5-846d-a354bc1c3c5e.shtml 69http://www.repubblica.it/cultura/2015/10/14/news/daniela_ bellingeri_lutto-125034845 liked ’giving away’ her abilities [...] many of us continued to ’plunder’ Daniela’s willingness and knowledge [...].
Explanation: Diferently from the previous examples, we can see here the apologetic usage of quotation marks (see [45]) to express detachment from the arguably not-quite-correct attitude of Daniela’s colleagues towards her. The picture which this description evokes is a woman with many capabilities (she is elsewhere in the text defined as "very intelligent", "well-read" and "educated"), but who nonetheless has a marginal role and whose knowledge is exploited by others (here saccheggiare is in quote marks to achieve some sort of attenuation of the behaviour, although the term exactly describes the attitude of the colleagues towards her).
We decided to split Sabatini’s Asymmetries in the usage of identification of women through men, age, profession and role into two categories, namely this one and the one in the following section A.13. Also, we did not include in the analysis the variables of age and profession. On the one hand, this choice was motivated by the fact that Sabatini herself did not provide any examples for these categories. On the other, both profession and age were variables already analysed in other classes in the current study.
In general, this class refers to instances where women are presented in texts through their relationship to a man in expressions such as daughter of, wife of or girlfriend of. Italian: Sergio e la moglie erano finiti in carcere
nell’ambito dell’inchiesta del procuratore [...]. 70 English: Sergio and his wife were imprisoned as a re
sult of an investigation by the prosecutor [...].
Explanation: Here Sergio’s wife has no name and she is just identified through the relationship to her husband.
Italian: La prima vittoria in un’aula di tribunale Salvatore Buzzi l’ha ottenuta con la sua ex amante. 71http://roma.corriere.it/notizie/cronaca/15_ottobre_09/buzziKatia Cipolla, con cui [...]. Buzzi aveva denun- vince-tribunale-ma-solo-contro-l-ex-amante-c55a683e-6e89ciato la ex [...]. Dietro la richiesta, la minaccia 11e5-aad2-b4771ca274f3.shtml#post-0 velata di rivelare la relazione alla moglie. [...] 72shttatplk:/i/nwg-wdwal.llaa-setaxm-mpao.gitl/i2e0-c1i5n/q10u/a1n7t/eendnizei-oinmi/pimerpieesreia-/aagclci-uasrarteos-tdi-idomiciliari-whZpW2q8UJlOinsRECEJvL/pagina.html 70https://milano.repubblica.it/cronaca/2015/10/05/news/milano_ 73http://firenze.repubblica.it/cronaca/2015/10/11/news/ scarcerato_dopo_tre_mesi_i_genitori_di_fatima_la_foreing_ perseguitava_l_ex_fidanzata_arresti_domciliari_per_un_25enneifghter_dell_is-124402650/ 124838732 Ma al processo non si è costituito parte civile contro l’ex amante, per la quale il pm aveva chiesto l’assoluzione. 71 English: Salvatore Buzzi achieved his first win in court against his ex-lover. Katia Cipolla, with whom [...]. Buzzi pressed charges against the ex [...]. Behind the request, there was the threat of revealing the afair to the wife. [...] But at the trial, he did not bring a civil action against the ex-lover, of whom the public prosecutor asked for acquittal.
Italian: [...] per evitare che le cose potessero degenerare in atti di violenza nei confronti della ex moglie e del figlioletto. 72 English: [...] to avoid that the situation could degenerate in violence against the ex-wife and the little child.
Italian: Il giovane, tra giugno 2011 e aprile 2012, aveva più volte perseguitato e minacciato l’ex fidanzata. 73 English: The young man had harassed and threatened the ex-girlfriend multiple times between June 2011 and April 2012.
Explanation: Sabatini highlights as particularly ofensive the expression ex-girlfriend/lover/wife, which implies that a woman continues to be identified by her male partner, even after the relationship has ended.
Note that the two last examples refer to situations of possible domestic violence. This makes even more problematic the usage of the terms ex-wife and ex-girlfriend respectively because it suggests the identification of possible victims through their oppressors.
In this section, our primary objective is to highlight instances where women are portrayed in texts through their role as mothers. Note that we excluded instances of mother of from the previous category, as one can be the mother of individuals of any gender, rendering it incongruent with the description Indentification through men.
In this context, the asymmetry arises from the societal expectation that becoming a mother constitutes a defining and comprehensive experience for women, while the same expectation does not apply to men. We evaluate this phenomenon in two aspects. Firstly, when information about being a mother is mentioned out of context, diverting attention from other aspects of the referent’s life. Secondly, when being a mother is relevant to the context, but no additional information is provided about the woman in question, suggesting that being qualified as a mother alone sufices for identification. Furthermore, we will consider cases where women are identified by their gender rather than their profession, particularly in situations where the latter holds significance. Italian: Molti esponenti politici si sono detti scandalizzati, ma la reazione più eficace è stata quella di Caroline Boudet, mamma di Louise [...]. 74 English: Many politicians said to be shocked, but the most impressive reaction was the one by Caroline Boudet, mother of Louise [...].
Explanation: Caroline Boudet is a journalist. Although in this specific context, the fact that she was the mother of Louise was relevant, it was not the only main focus of the story. Nonetheless, this is the only title used to qualify her in the whole article. We argue that the contrast is here made even more evident by the contraposition with the word politicians, who are described exclusively by their professional role and not by that of parents (since it is highly likely that most of them are parents themselves).
Italian: Finisce così la storia di Assunta, la madre di
Fatima, la jihadista italiana [...]. 75 English: Thus ends the story of Assunta, mother of
Fatima, the Italian jihadist [...].
Explanation: Similarly to the previous example, we have no further information about Assunta 74http://www.corriere.it/esteri/15_ottobre_08/vignetta-choccharlie-hebdo-cita-de-gaulle-ma-ofende-down-4bc6e35c-6df911e5-8aec-36d78f2dc604.shtml 75http://www.corriere.it/cronache/15_ottobre_08/funeralicattolici-la-madre-fatima-jihadista-italiana-7bdc0eb6-6de311e5-8aec-36d78f2dc604.shtml#post-0 except that she is the mother of someone. Additionally, note how both women referents (namely Assunta and Fatima) are here only introduced by their first names. Notably, in other parts of the article, the father of Fatima is not presented only through his relationship with the daughter.
Italian: Le mamme sono preoccupate [...]. 76 English: The mothers are worried [...].
Explanation: Here, women are considered as a separate homogeneous category, where members are uniquely characterized by the fact of being mothers.
Italian: E così, le ragazze si allenano tutto l’anno con sessioni di training speciali tra yoga, pilates e boxe. [...] E poi, diciamocelo, una ragazza farebbe qualsiasi cosa per non perdere il posto su quella passerella [...]. 77 English: Thus, the girls work out the whole year with special training sessions involving yoga, pilates, and boxing. [...] And let’s be honest, a girl would do anything not to lose her spot on that catwalk [...].
Explanation: Here, ragazze is used as a synonym of models, which is the profession occupied by the subjects of this article. The suggested efect is of trivialization of the profession, probably because mainly associated with women and based on outer appearance, which is one of the few aspects considered important for women.
This category was not directly included in Sabatini’s work. Nevertheless, we wanted to gather in one class all instances in which women were depicted through their physical appearance and that could not be resolved in one of the previous categories. Here, we are not delving into specific word classes as we did for the asymmetries in the usage of substantives, adjectives, and verbs. Instead, our focus lies on the organization of information and the decision to emphasize aspects of women’s outer appearance rather than other facets. 76 77http://www.corriere.it/moda/news/15_ottobre_08/soltanto-4litri-d-acqua-angeli-victoria-s-secret-dieta-1b54aab8-6de711e5-8aec-36d78f2dc604.shtml#post-0
Explanation: In all these examples, but particularly in the last two, references to the outer appearance of these women are completely out of context. Note that we excluded from this category references to women’s bodies in cases where it could be considered relevant for the profession, for example in the case of models. Although this choice can be considered arguable, we explicitly wanted to consider only cases where the inappropriateness of these comments was obvious. 78http://www.repubblica.it/spettacoli/cinema/2015/10/24/news/ morta_maureen_o_hara_stella_di_john_ford-125818160 79https://milano.repubblica.it/cronaca/2015/10/08/news/solenne_ pai_dassi_il_mio_stravinskij_brioso_e_ardente_vi_emozionera_124622485/ 80http://www.corriere.it/moda/news/15_ottobre_12/miriambarbara-marsica-modelle-sono-ragazze-normali-c5300ad4-710f11e5-a92c-8007bcdc6c35.shtml Table 2 shows the labels’s distribution in the dataset. The results are cumulative of all newspapers included in the analysis, namely Repubblica, Il Corriere della Sera and La Stampa.
We present the error analysis and the concrete results achieved by both pipeline and baseline for the labels Generic masculine, Masculine of professions, Asymmetric usage of names, surnames and titles, Feminine article before surname and Identification through man .
The error annotation was done manually by first extrapolating all misclassified sentences for each label, splitting false positives and false negatives. Then, we collected and clustered similar error patterns in the misclassified instances and analysed the possible reasons that led to diferent error types.
This was the most diverse class among those considered in this analysis. Table 3 shows the results obtained for this label. Overall, the model was able to understand the main features of the phenomena falling into this category, although not always classifying them correctly.
With a higher number of false positives than false negatives, the model tended to classify more instances than the annotated ones, sometimes showing only a superficial understanding of the phenomena, and other times posing legitimate doubts about the annotation itself.
In particular, the precedence of masculine in female/male oppositional couples was the marker that was better recognized by the model, which even pointed out cases that were not correctly included in the annotation. These are the only instances misclassified by the model: Italian: «[il Pd] Caro Pd siamo pronti a difendere il
diritto dei bambini ad avere mamma e papà».81 Translation: «[Democratic Party] Dear Pd, we are ready to defend the right of children to have a mother and a father».
Explanation: Non-sexist oppositional couple, where the feminine precedes the masculine, which should therefore not be classified as a member of this class.
Italian: [...] Arrivano in piazza del Campidoglio in piccoli gruppetti, marito e moglie, tre amiche, due compagni di sezione del Pd [...]. 82 Translation: [...] They come to Campidoglio Square in small groups, husband and wife, three friends, two fellow members of the Pd [...].
Explanation: This instance was misclassified as
Identification through man , probably for the occurrence of the words husband and wife that are common for this class.
Italian: Invece in Italia ci sono voluti circa quindici anni, e un lavoro di mediazione certosina, perché mani, anche ai genitori dell’afido di " concorrere" all’adozione del ragazzino e della ragazzina dei quali, di fatto, sono già figure fondamentali. 83
Translation: It took Italy instead about fifteen years, and painstaking mediation work, to come to a law that, starting tomorrow, allows also foster parents to "compete" for the adoption of the boy and girl to whom they actually already are fundamental ifgures.
Explanation: In this example, both genders are made explicit by using splitting (i.e. both the male and female forms occurred). Although this results in the masculine form preceding the feminine one, during the annotation process, we decided not to classify it as Generic masculine because by using splitting the authors intended to precisely avoid the use of generic masculine, and we did not want to penalize this choice. However, the model correctly identified the precedence of the masculine form in this case. Therefore, the annotation should probably be revisited to make it more strict in this regard and less ambiguous.
Additionally, the model was able to link the presence of the substantives uomo/uomini (man/men) with this class. However, it seemed to limit itself to identifying and marking all occurrences of these words, rather than showing an actual understanding of the phenomenon.
For example, in many cases, the model wrongly classified instances of the word uomo when referring to one or more explicit male referents.
Italian: Moravia, un uomo che amava le donne [...]. 84 Translation: Moravia, a man who loved women [...].
Italian: Mentre sono partite le indagini continua la cac
cia ai due uomini. 85 Translation: While investigations have started, the
hunt for the two men continues.
Finally, the model struggled to recognize sexist markers where women were treated as a separate category and the disagreement in gender between a subject and its nominal predicate. While both did not present enough had the additional obstacle of being more abstract and less ascribable to the occurrence of specific words.
Italian: Arrestati cinque cittadini marocchini e due italiani, tra cui una donna, per rapina aggravata in concorso. 86 si arrivasse ad una legge che permetterà, da do- examples for the model to properly learn from, the latter 84https://www.corriere.it//cultura//15_ottobre_26//creare-poiafido_puo_diventare_adozione_la_legge_sulla_continuita_ 124842364//?rss 85https://firenze.repubblica.it//cronaca//2015//10//11//news/ 86https://milano.repubblica.it//cronaca//2015//10//17//news/ /milano_rapine_sui_treni-125266311//?rss Translation: Five Moroccan and two Italian citizens, Italian: Il ministro dell’Economia Pier Carlo Padoan one of which a woman were arrested for aggravated robbery in complicity.
Translation: The promoters are thirteen organiza- professional titles, some of which still struggle to permeItalian: I promotori sono tredici organizzazioni di varie nazioni [...]. 87 tions from diferent countries [...].
As already pointed out, the model shows some understanding of which phenomena belong to this class and hardly ever misclassifies it with other labels. However, the diversity of the markers included in Generic Masculine has the detrimental efect of making it dificult for the model to focus more specifically on single phenomena, especially in our setting, where only a scarce number of examples per label is provided. Hence, a possible solution could be to split this class into smaller classes, each identifying a more specific marker.
Albeit being the class with the most samples and describing a less complex phenomenon compared to other classes, the model presented some dificulties in correctly assessing this sexist marker (see table 4). Pipeline results for label "Masculine of professions" Even though it seemed to recognize the link of this label with high-status professions such as minister, lawyer or mayor, it was unable to identify the key aspect considered in this class, lying in the usage of the masculine form also for women. Rather, it marked all instances of these titles, regardless of the gender of the referent. Italian: «Ma possiamo ancora migliorare», ammette il direttore sportivo Carlo Deslex. 88 Translation: «But we can still do better», acknowledges the sports director Carlo Deslex. 124200861//1//?rss 1.35216208/ 18/news/basket-la-poli-oppisti-cipir-vince-al-debutto-in-casa
Translation: The Minister of Economy Pier Carlo This behaviour could be caused by the absence of "positive" examples of the correct feminine forms for these ate and become part of the Italian language. In this case, eforts in providing more such examples could help the model focus on the key aspect of this class and thereby achieve better performance.
This was the second class in terms of the number of samples after Masculine of Professions, and as shown by table 5 it obtained comparable results.
Pipeline results for label "Asymmetric usage of names, surclass, we can notice that the model can correctly link the class to the presence of female names. Notably, it seems even more strict than the annotator in classifying instances where women are referenced only by name.
The reason could be that the model struggles to identify contexts in which using only names might be appropriate.
This is made worse by the fact that the pipeline takes into account single sentences so that only a limited context is provided to the model for the prediction.
Italian: [DANIELA Bellingeri] Daniela era una persona intellettualmente vivace, colta, amava la musica e la poesia. 90 Translation: [DANIELA Bellingeri] Daniela was an intellectually lively, well-read, loved music and poetry.
Explanation: In this case, context mattered for the annotation since the author of the article was writing about a person they knew, therefore 89https://www.corriere.it//economia//15_ottobre_23//padoan90https://www.repubblica.it//cultura//2015//10//14//news//daniela_ referencing her only by first name. However, this context was not provided to the model, which was therefore correctly pointing out the use of the name only.
Italian: Fatima. Sono stati scarcerati dopo 3 mesi di detenzione Sergio Sergio e la moglie Assunta, i genitori di Maria Giulia ’Fatima’ Sergio , la presunta jihadista italiana convertita all’Islam e partita per la Siria per combattere nelle fila del Califato.
Translation: Fatima. After 3 months in prison, Sergio
Sergio and his wife Assunta, parents of Maria Giulia ’Fatima’ Sergio, the alleged Italian jihadist who converted to Islam and went to Siria to fight for the Caliphate, have been released.
Explanation: Here, the name Fatima could be correctly considered a member of this class. We decided not to annotate it since it was used as a nickname, but this decision can give rise to interpretations.
The last two examples show the dificulty of the annotation process and the interpretability of single phenomena.
A possible solution could be to be more strict in the annotation or expose the model to more fine-grained examples where the usage of names can be appropriate. The trade- Table 6 of between the two should be considered with respect to the specific use case where the model is employed.
Additionally, in the second example, the model classifies Maria as asymmetric, although the name does contain her surname. This points out a possible inability of the model to distinguish cases where either multiple first and last names are present or some nicknames are introduced in the middle of the name. Similarly, potential errors derive from not correctly distinguishing names from surnames or not recognizing names as such, especially when the referent does not have an Italian name.
Italian: [Amazon] [Global] [Jay Carney] [all’ inchiesta del New York Times] In un post su Medium dal titolo Quello che il New York Times non ti ha metodo di lavoro dei due giornalisti che hanno curato l’inchiesta. 92 Translation: [Amazon] [Global] [Jay Carney] [investigation of New York Times] In a Medium post titled What the New York Times did not say, Carney harshly attacked the working method of the two journalists that curated the investigation. /milano_scarcerato_dopo_tre_mesi_i_genitori_di_fatima_la_ foreing_fighter_dell_is-124402650//?rss ribatte-al-new-york-times-la-vostra-inchiesta-non-rispetta-icriteri-giornalistici-yIvf1nQCNzl8AFWRirtIrJ//pagina.html raccontato, Carney ha attaccato duramente il Italian: Anche quando la Taverna chiama prostituta la Boschi, o quando Castaldi mi dà del parassita 93https://palermo.repubblica.it//cronaca//2015//10//02//news/ 94https://www.corriere.it/politica/15_ottobre_05/non-devoscusarmi-quel-gestaccio-l-ha-fatto-lezzi-io-l-ho-mimato9194e8f4-6b4a-11e5-9423-d78dd1862fd7.shtml Explanation: In this case, the model interpreted
Carney as a female name and misclassiefid it as member of this class.
Italian: Lo ha detto Piera Maggio, la madre di Denise
Pipitone, subito dopo la sentenza di assoluzione per Jessica Pulizzi, la sorellastra di Denise accusata di sequestro di persona. 93 91
Translation: This is what Piera Maggio, mother of
Denise Pipitone, said right after the verdict of acquittal for Jessica Pulizzi, Denise’s step-sister, accused of kidnapping.
Explanation: On the contrary, here Denise was not recognized as a female name and therefore not correctly classified by the model.
Thanks to the limited variability and high repetitiveness of the phenomenon which made it easier for the model to recognize, this was the class that achieved the best overall results (see Table 6).
Pipeline results for label "Feminine article before surname"
Baseline Pipeline
Precision
Recall However, we can point out some examples where the model was unable to identify the label, mainly because it did not correctly assess the presence of a surname following the article. In some cases, surnames were interpreted as nouns either because they also function as nouns in Italian or because they have a structure that recalls the one of an Italian noun. This is the case in the following two examples.
sociale. 94 parasite.
Translation: Also when the Taverna calls the Boschi a prostitute, or when Castaldi calls me a social Italian: A motivare le pressioni sul protagonista di svolto dalla Cipolla nell’estate del 2011 [...]. 95 Translation: Pressures on the lead of Mafia Capitale were motivated by the loss of her job as a bartender, which the Cipolla did during the summer of 2011 [...]. Moreover, the model struggled with some foreign sur- figlia , fidanzata names or surnames with a particular structure such as O’Hara in the following example, which the model did not recognise as a surname.
Italian: [...] la O’Hara aveva ricevuto nel febbraio
scorso l’Oscar alla carriera. 96 Translation: [...] the O’Hara received last February an false positives.
Oscar to her career.
Finally, there were a few instances where the model was misled by the surrounding context, resulting in errors where names of other entities, like bands (first example) or cars (second example), were mistakenly identified as surnames: Italian: San Siro, la Banda Bassotti e la Champions
sfumata: gli striscioni sfottò anti-Juve.
vanished Champions League: the mocking banners against Juve. 97 Italian: Un nuovo diesel per la Opel Meriva Opel torna alle cabrio la Cascada a 29.400 euro [...]. 98 Translation: A new diesel for Opel Meriva Opel reverts to convertibles the Cascada for 29.400 euros [...].
only preceded by Feminine article before surname, the highest precision (see Table 7).
The model correctly identifies a link between this class and the presence of female substantives such as moglie, 125818160//?rss This was the class that achieved the highest recall and, both phenomena in a single class. foreing_fighter_dell_is-124402650//?rss 98https://www.repubblica.it//motori//sezioni//prodotto//2015//10/ /legnano_dalla_nonna_ai_cugini_sgominata_un_intera_ famiglia_di_ricettatori_piu_una_complice-124904073//?rss Pipeline results for label "Identification through man" or compagna. In many cases, model predictions raised legitimate doubts about the annotation, which sometimes had to be reconsidered. Nevertheless, as we also noted for the occurrence of uomo/uomini in the Generic Masculine class, the model tends to classify any instance of such words in the text without lingering on more subtle analysis. However, diferently from uomo/uomini, this poses fewer problems, as it causes fewer One of the most common errors in the model’s predictions is neglecting whether the relationship is actually with a man. For example, in the following sentence, the relationship sorella di (sister of) is referred to a woman, Fatima, and was therefore not included in the annotation.
However, one can argue that the phenomenon can be extended to all cases where someone is presented by their relationship with someone else, independently of gender.
The annotation could therefore be revisited to include also these cases.
Italian: La sorella di Fatima è ancora detenuta. 99 Translation: Fatima’s sister is still in custody.
Another interesting factor to consider is that the model classifies instances of type
mother of as members of this class, where we had instead set up a separate class to include them, namely Identification through gender/ role .
This could lead to two possible solutions. Either introducing more instances of the latter class, so that the model can correctly learn to distinguish between the two cases.
Or we could restore the original class by [
Furthermore, by analysing the errors, we noticed that the word compagna could potentially pose a problem since it can mean both partner in a romantic relationship and mate in a sports team. Hence, more focused examples on this aspect might be needed to teach the model to distinguish between these two usages.
Italian: [...] Nadia Fanchini, solo undicesima al traguardo dello slalom gigante a 3 secondi e un decimo dalla compagna di squadra. 101 Translation: [...] Nadia Fanchini, who finished only eleventh in the giant slalom, 3 and one-tenth seconds after the teammate.
Finally, the model correctly identified some instances of this class in the part added by the coreference resolution at the beginning of the sentence, that had however not been annotated. This can be solved by adding the annotation also for the coreference part or creating ad hoc examples to teach the model not to consider the text in that part of the sentence for the annotation. However, this does not have any negative efect on the performance of the model and can therefore be overlooked.