Recently, social networks have become the primary means of communication for many people, leading computational linguistics researchers to focus on the language used on these platforms. As online interactions grow, recognizing and preventing ofensive messages targeting various groups has become urgent. However, finding a balance between detecting hate speech and preserving free expression while promoting inclusive language is challenging. Previous studies have highlighted the risks of automated analysis misinterpreting context, which can lead to the censorship of marginalized groups. Our study is the first to explore the reappropriative use of slurs in Italian by leveraging Large Language Models (LLMs) with a zero-shot approach. We revised annotations of an existing Italian homotransphobic dataset, developed new guidelines, and designed various prompts to address the LLMs task. Our findings illustrate the dificulty of this challenge and provide preliminary results on using LLMs for such a language specific task. Warning: This paper contains examples of explicitly ofensive content. Our positionality: This paper is situated in Italy in 2024 and is authored by researchers specializing in Natural Language Processing (NLP). Beyond our academic work, we are sensitive to anti-hate speech issues. Our backgrounds fields are theoretical linguistics, computer science and NLP.
CEUR ceur-ws.org
In recent years, social networks have become the primary means of communication for most people. With the daily growth of online interactions, it has become urgent to recognize and prevent the spread of ofensive messages against diferent target groups based on gender, sex, sexual orientation, race, religion, language, or political orientation. Moreover, categorizing hate speech with clear-cut boundaries is overly simplistic, as it includes various forms of abusive language that imply disrespect and hostility. A recent challenge is finding a balance between detecting hate speech and preserving the free spread of ideas and opinions on the web, while promoting inclusive and fair language. Thiago et al. (2021) [1] highlighted how automated analysis can misinterpret context, risking the censorship of marginalized groups other study by Pamungkas and colleagues (2020) [2, 3] emphasized the importance of considering context in Natand solidarity within the group members [4]. Although community visibility and the use of specific slang have been approached for years, to our knowledge only some hate speech studies specifically addressed slurs, and few focused on slurs semantic reappropriation [5]. Nowadays, recognizing this kind of semantic shift through
need to take a step ahead from the existing abusive language detection models. Having in mind the capability of LLMs in classification task, we leveraged a LLM with a zero-shot approach in order to recognize the presence of reappropriative uses in our dataset.
This study makes the following contributions: • We partially revised the original annotation previously conducted on the HODI dataset (Homolanguages, such as those of the LGBT+ community. An- reappropriative use of slurs in Italian, highlighting the transphobic Dataset in Italian)1[8], by developing new annotation guidelines. • We used a LLM specifically fine-tuned on Italian
language by leveraging prompt engineering. • From a linguistic point of view, we showed why certain features of the Italian language make this task particularly challenging.
review the most significant related work on hate speech detection and zero-shot approaches leveraging LLMs. In the Section 3 we describe our methodology for the dataset creation and the implementation of zero-shot tasks. In Sections 3 and 5 we respectively report results, analysis and main limitations of this work. Finally, in the last Section 6 we draw conclusions of the current research.
(RHSD), the first hate speech dataset dedicated at investigating the use of reclaimed slur terms, and by fine-tuning a baseline model which resulted in the Reclaimed Hate Speech (RHS) model.
As far as the Italian language is concerned, slurs recently became a significant topic from a linguistic and philosophical point of view, but there are not studies focusing on slurs reappropriation detection task. Philosophy of language studies highlighted that a key area of interest is slurs echoic uses, where target communities reappropriated derogatory terms to express pride, solidarity, or use them as tools for political and social activism [13, 4]. Nossem (2019) [14] observed a productive role in creating localized versions of queer by reappropriating and redefining existing local alternative terms, specifically frocio and frocia, femminiell@, and ricchione. At this point, it should be noted that Italian, diferently from English language, lacks terms like queer, which bring with them such a long socio-cultural and historical background. The semantic requalification process of homotransphobic slurs is at its first steps and consists of a challenging task that has not yet been investigated in computational domains with LLMs.
to magnitude of the phenomenon and the dificulties of defining clear boundaries. Some recent developments in AI underlined the challenge of building corpora and models to automatically detect the abusive (or not abusive) nature of slurs in social media texts. Pamungkas’ et al. 3. Methodology (2020) [2] research focused on the use of swear words in English and aimed at diferentiate between ofensive and 3.1. Dataset creation non-ofensive occurrences of slurs. A Twitter English corpus, SWAD (Swear Words Abusiveness Dataset), was To our knowledge, there are no available annotated developed by manually annotating the abusive charge at datasets in the Italian language focusing on the phethe word level and models were trained to automatically nomenon of slurs semantic reappropriation. To address predict abusiveness. the issue of limited data, in this preliminary research we
Over the last decade, most studies approached the hate utilized the HODI dataset [8], which contains 6000 Italian speech detection in terms of binary classification [ 9]. Twitter messages collected by using a set of 21 keywords For instance, Plaza et al. (2023) [10] examines this task (i.e., gay, pride, lesbica, frocio). The dataset is a collection by comparing the performances of an encoder-decoder of sentences directed against LGBT+ community who model with several BERT-based models in both zero-shot are target of homotransphobia. Our argument is that learning and fine-tuning scenarios. The findings show in such a corpus it is possible to find slurs used in both that BERT-based models perform poorly in zero-shot abusive and reappropriative contexts. With the aim of learning, while the others, even without additional train- collecting messages suitable for our study, we filtered the ing, achieves results comparable to fine-tuned models. HODI dataset by selecting tweets that contain at least
Nowadays, research indicates that hate speech changes one denigratory term, by adopting a two-fold strategy. depending on the target groups [9]. Detecting homo- To select the homotransphobic swear words, we used the transphobic hate speech (i.e. a specific abusive language HurtLex lexicon2 [15], a multilingual lexicon containing addressed to LGBT+ community) has emerged as a critical an organized list of denigratory terms divided into 17 research area, with various scholars proposing solutions categories (i.e. negative stereotypes, ethnic slurs, moral in diferent languages such as English [ 11] and Italian and behavioral defects, words related to homosexuality). [8, 12]. From HurtLex, we selected only the words categorized as
However, only few studies focused on the detection homotransphobic, then we further narrowed the list to of slurs that have undergone a semantic reappropriation those that satisfy the slur definition 3 provided by Bianchi process. Zsisku and colleagues (2024) [5] approached the task by collecting the Reclaimed Hate Speech Dataset 2https://github.com/valeriobasile/hurtlex 3Bianchi (2014) [4] defines slurs ”derogatory terms -such as ‘nigger’ and ‘faggot’-targeting individuals and groups of individuals on the basis of race, nationality, religion, gender or sexual orientation. According to most scholars, slurs generally have a neutral counterpart,
fying homotransphobia in Italian tweets.
all’oratorio, alle medie, al liceo e tutta la vita. E mi . seclehmooeln,tinarhyigshchsochool,oal,tatnhde aylolumtyh lcifeen. tAern,dinI’mmiodkdaley va bene così, c’è più colore in questo mondo
with that, it adds more color to this world. i.e. a non-derogatory correlate: ‘Boche’ and ‘German’, ‘nigger’ and ‘African-American’ or ‘black’, ‘faggot’ and ‘homosexual’”.
After obtaining the described subset, we utilized zeroshot Learning (ZSL) with prompting to assess the model’s • Mamma mia raga come mi ha messa di buon umore ability to determine whether the target words are used il #LiguriaPride non mi sentivo così da un sacco in abusive or non-abusive context. Specifically, we emgrazie energia frocia ployed the Qwen model [17], a multilingual decoded-only [English translation: Mamma mia guys how LLM pre-trained on Italian. the #LiguaPride has put me in such a good mood We define the temperature of the model to be 1, a fair I haven’t felt this way in a long time thanks FRO- trade-of between randomness and determinism in the CIA energy] results, and a maximum sequence length of 2024. For inference, an A100 GPU provided by Google Colab was (2014,2015) [4, 16]. We chose to exclude words such as gay, omosessuale, omofilo, pederasta , and diverso because they are not strictly derogatory terms, hypothesizing that if words are not perceived as abusive, they cannot undergo a process of semantic reappropriation. After obtaining a list of 17 words, we filtered the HODI dataset by selecting only the tweets that contained at least one of the following target words: anomalo, chiappa, frocio, invertito, travestiti, checca, deviato, culattone, finocchio, ifnocchi, finocchietto, sesso anale, frocia, ricchione, trans, troia, stesso sesso. The resulting subset is a collection of 1742 tweets (see two examples in table 1). 3.2. Annotation guidelines Establishing guidelines for such a subjective and previously unexplored topic has been challenging. Since the phenomenon lacks clear boundaries, we aimed to describe the task as clearly as possible. With this in mind, we based our guidelines on previous works in the field of the philosophy of language [ 4, 16, 13]. We asked three expert annotators to decide whether the target words in each tweet are used in a reappropriative context or not. Building on previously cited works, we defined reappropriation as the use of derogatory epithets by members of the target groups in a manner that is generally considered non-ofensive. To better define the phenomenon we highlighted diferent contexts in which this linguistic behaviour could occur:
the derogatory terms in a non-ofensive way in informal contexts.
Political reappropriation contexts – target groups reclaim the use of derogatory epithets as a tool to emphasize a conscious and common political struggle.
• Happy #PrideMonth e ricordatevi che l’orgoglio si celebra non solo quando andate a ballare nelle discoteche gay, ma anche quando si tratta di metterci la faccia e combattere per la causa perché altrimenti il ricchione lo state facendo solo col culo degli altri e non è carino [English translation: Happy #PrideMonth and remember that pride is celebrated non only when you go dancing in gay discos, but also when it comes to put your face out there and to fight for the cause because otherwise you are just being RICCHIONE on other people’s ass and it is not nice]
Artistic contexts – artists reclaim derogatory epithets to subvert the dominant socio-cultural norms.
• Poca gente che li guarda, c’è una checca che fa il tifo Se #LucioDalla avesse scritto #AnnaEMarco nel 2022 sarebbe stato accusato di omofobia, lui. Invece ha scritto una canzone immensa [English translation: Few people look at them, there is a CHECCA cheering if #LucioDalla had written #AnnaEMarco in 2022 he would have been accused of homophobia. Instead he wrote a great song] 3.3. Zero-shot learning approach Table 2 on tweets labeled diferently (some examples in Appendix Inter-annotator agreement metrics B). We observed that out of a total of 217 tweets with annotation disagreement, 67 (30.88%) contained the word Fleiss’ Kappa 0.57 ”frocia”. This word likely caused confusion due to its Annotators Cohen’s kappa unique history: unlike the other target words ”frocia”, feminine form of ”frocio”, originated in an already reapAnnotator 1 vs Annotator 2 0.559 propriative context 5 [14]. In some cases, due to a lack of AAnnnnoottaattoorr 12 vvss AAnnnnoottaattoorr 33 00..562187 context, it was very dificult to understand the real communicative intent of tweets (i.e., Sono ricchione. (senso andiamo) - ”I’m gay. (like, let’s go)”). In other instances, used. The code is available on the following GitHub it was challenging to determine whether the person who page4. wrote the message is part of the LGBT+ community or
As previously discussed, collecting a large-scale corpus not (Oggi il mondo mi sta urlando contro che sono un ricfor reappropriated language detection is challenging. To chione colossale senza speranza ed io gli sto dando ragione address the lack of data, we used a ZSL approach, prompt- - ”Today the world is shouting at me that I’m a colossal ing the model to recognize the presence of semantic re- hopeless queer, and I’m agreeing with it”), assuming that qualification without providing additional information. only members of target community can use slurs in reapThis method evaluates the model’s ability to generalize propriative sense. Finally, we also identified some noisy efectively with no training data, taking into account only data in which target words have diferent meanings. For information acquired during the LLM training phase. example, in the sentence Il 4 è l’onomastico di checca fren
Diferent studies [ 18, 19] showed that ZSL results are zis ci ubriachiamo (”On the 4th, it’s Checca Frenzis’ name significantly influenced by the appropriateness and pre- day, so we’re getting drunk”) the term ”checca”6 is likely cision of the prompts used. Additionally, multiple re- used as a diminutive of the Italian name ”Francesca”. searchers [19] proposed diferent methods to improve We also noticed that in some cases tweets labelled as performances. Plaza-del-Arco et al. (2022) [18] demon- reappropriative were also labelled as homotransphobic strated that one of the most critical factors is ensuring in the original annotation of HODI dataset. Due to this that the prompt fits well with the utilized corpus. Taking apparent contradiction, we conducted a qualitative linthis into account, we designed four diferent prompts guistic analysis on this data. We realized that in four using the HODI sub-corpus with the reappropriation an- examples (Oggi avrò di che parlare coi colleghi..un etero notation as the gold standard, each including specific analfabeta che conquista l’attenzione di una checca alfadetails about the task and the corpora. The first one is betizzata , mi raccomando vai a fare la quarta dose the most general - explaining only the task in few words - che forse ti aiuta a dimenticarmi. Ciao - ”Today I’ll have while the fourth is as precise as possible providing full list something to talk about with my colleagues... an illiterof target words (full prompts are provided in Appendix ate straight guy who captures the attention of a literate A). queer. Make sure to get your fourth dose, maybe it’ll help you forget about me. Bye”), it is unclear whether the writer is part of the LGBT+ community or not. In 4. Results other words, it is uncertain if the users were using slurs to refer to themselves with reappropriative intent or to 4.1. Annotation statistics other persons in abusive term. In addition, in some of these examples, target words were used as part of figures We calculated the annotator agreement firstly by using of speech, mostly similes (Fare come una checca - ”Behave Fleiss’ Kappa, obtaining 0.57, secondly through Cohen’s as a faggot”). These expressions, highly lexicalized in Kappa between pairs of annotators (all metrics are dis- Italian and often used as abusive idiomatic phrases, likely played in table 2). The moderate agreement and metrics increased the dificulty in recognizing the correct usages. variability highlighted the task’s dificulty and subjectivity. Despite the three annotators being experts on the topic, they encountered challenges in distinguishing the use of slurs.
The majority annotation indicates that out of a total of 1742 examples, only 168 were annotated as reappropriated.
To better understand annotators disagreements and collect challenging examples, we conducted an analysis 5Nossem (2019) considers ”frocia” as a calque of the English “queer” or ”Alternatively, we could see it as a new concept which is specific to the Italian linguistic and cultural context, rather than an adaption or appropriation of the English “queer”, i.e. some sort of a territorialised post-queer” [14]. 6”Checca” as well as being a diminutive form of the Italian name ”Francesca” is a colloquial and somewhat derogatory term in Italian used to refer to a gay man
process; therefore, a new scalar annotation scheme is probably required. Furthermore, the fact that only experienced young researches sensitive to LGBT+ issues were involved in the annotation task may have led to bias in the results.
As future work we plan to: • create a new dataset and annotating it by following a perspectivist approach 7[20], i.e. by collecting diferent points of view from various social media, involving annotators with diferent backgrounds, in terms of age, origin, education, in/out target groups, and providing more context information during the annotation phase in order to better understand slurs’ meanings and intents. • through diferent LLMs, investigate which approach has better performances in recognising diferent uses of slurs, for instance by using ZSL approach between pairs of examples or defining few-shot with new suitable data. • regarding ethical considerations, it is crucial to directly and actively involve the LGBT+ community. Gathering viewpoints and suggestions from those who experience daily oppression and denigration is essential not only to strengthen the research methodology but also to ensure its relevance and sensitivity to their lived experiences.
4.2. LLM classification results
Notably, performances change among the prompts. The fourth prompt, which is the most specific, achieves the highest performance as it specifies all the target words considered during dataset construction. In contrast, the third one, focusing specifically on detecting homotransphobia by asking if the text intends to ofend on the basis of sexual orientation or gender identity, has low performances. Among the four prompts, the first one (”Determine if the sentence contains semantic reappropriation; respond ’True’ if it does and ’False’ otherwise.”) has the worst performances, likely due to the ambiguity of the expression ”semantic reappropriation” for the model. Additionally, the model struggled to recognize the minority class (semantic requalification) because it is very complex for the model to recognize the context of the use of a slur, whether it is used to ofend or not.
This requires a deep understanding of the context and social dynamics, and it can also be a challenging task for humans.
To address this issue, balancing the information in the prompt by providing more details about semantic requalification could improve the model’s overall performances.
Therefore, we did not achieve very good performances, highlighting the importance of collecting new data and reviewing the computational approach.
This paper presents the first attempt to specifically address the detection of slur reappropriation in the Italian language. One of the reasons that motivated us to undertake this task is the need to ensure a safe linguistic environment on social networks without risking the censorship of individual freedom of expression. Since there was no existing dataset to explore homophobic slurs in the Italian language, we filtered a pre-existing homotransphobic dataset to build a subset containing only tweets with slurs occurrences, used both abusively and non5. Limitations and future works abusively. We then designed precise new guidelines and annotated the filtered subset, focusing on the presence of The semantic requalification of slurs turned out to be a slur semantic reappropriation. With the newly annotated complex and time-consuming process in several aspects. dataset, we approached a classification task using LLMs Although the study has taken its first steps, some limi- with zero-shot techniques. Leveraging the Qwen model tations must be acknowledged. Firstly, we realized that [17], we proposed four diferent prompts. As suggested the HODI dataset [8] was not completely suitable for our by previous literature, more specific prompts and those purposes. Tweets had been collected for the homotrans- better suited to the dataset yielded better performance. In phobia detection aim and the diference of research goals this work, we proposed an important and under-explored did not provide us the right data to investigate the seman- task through a two-fold contribution. On one hand, we tic requalification process of slurs. Secondly, a binary highlighted the lack of data in the Italian language dealannotation proved to be limiting due to the dificulty of ing with this phenomenon and the necessity of building the task. The subjective evaluation of the annotators does not allow the problem to be simplified in terms of the presence or absence of semantic requalification 7https://pdai.info/ an up-to-date corpus that comprehensively includes mul- ings of the Eighth Evaluation Campaign of Natutiple sources and semantic contexts. On the other hand, ral Language Processing and Speech Tools for Italwe demonstrated a possible approach by leveraging new ian. Final Workshop (EVALITA 2023), Parma, Italy, state-of-the-art LLMs. Finally, it is important to have in September 7th-8th, 2023, volume 3473 of CEUR mind that compared to English, Italian has a diferent Workshop Proceedings, CEUR-WS.org, 2023. URL: history and cultural background, resulting in a much https://ceur-ws.org/Vol-3473/paper26.pdf. slower linguistic evolution. This makes establishing pre- [9] A. Ollagnier, E. Cabrio, S. Villata, Unsupervised cise characteristics of this topic a challenging task due ifne-grained hate speech target community detecto the lack of solid foundational knowledge. In conclu- tion and characterisation on social media, Social sion, we believe that bringing attention to the issue will Network Analysis and Mining 13 (2023) 58. lead to anti-discrimination activities, the creation of safer [10] F. M. Plaza-del arco, D. Nozza, D. Hovy, Respectspaces in online communication, and the inclusion and ful or toxic? using zero-shot learning with lanacceptance of LGBT+ communities. guage models to detect hate speech, in: Y.-l. Chung, P. Röttger, D. Nozza, Z. Talat, A. Mostafazadeh Davani (Eds.), The 7th Workshop on Online Abuse References and Harms (WOAH), Association for Computational Linguistics, Toronto, Canada, 2023, pp. [1] D. O. Thiago, A. D. Marcelo, A. Gomes, Fighting 60–68. URL: https://aclanthology.org/2023.woah-1. hate speech, silencing drag queens? artificial intelligence in content moderation and risks to lgbtq [11] 6S.. dKoui:m10a.r,18A6.5N3/avg1a/r,2A02.3K.uwomaahr-, 1A..6.Singh, Hate voices online, Sexuality & culture 25 (2021) 700–732. speech detection: A survey, in: 2022 4th Interna[2] E. W. Pamungkas, V. Basile, V. Patti, Do you really tional Conference on Advances in Computing, Comwant to hurt me? predicting abusive swearing in munication Control and Networking (ICAC3N), social media, in: N. Calzolari, F. Béchet, P. Blache, IEEE, 2022, pp. 171–176.
K. Choukri, C. Cieri, T. Declerck, S. Goggi, H. Isa- [12] D. Locatelli, G. Damo, D. Nozza, A cross-lingual hara, B. Maegaard, J. Mariani, H. Mazo, A. Moreno, study of homotransphobia on twitter, in: ProceedJ. Odijk, S. Piperidis (Eds.), Proceedings of the ings of the First Workshop on Cross-Cultural ConTwelfth Language Resources and Evaluation Con- siderations in NLP (C3NLP), 2023, pp. 16–24. ference, European Language Resources Associa- [13] B. Cepollaro, et al., Linguaggio d’odio, in: Pragtion, Marseille, France, 2020, pp. 6237–6246. URL: matica Sperimentale, Società Editrice il Mulino spa, https://aclanthology.org/2020.lrec-1.765. 2022, pp. 145–156. [3] E. W. Pamungkas, V. Basile, V. Patti, In- [14] E. Nossem, Queer, frocia, femminielle, ricchione et vestigating the role of swear words in abu- al. - localizing ’queer’ in the italian context, GSI: sive language detection tasks, Lang. Re- Gender, Sexuality, Italy 6 (2019) 1–27. sour. Evaluation 57 (2023) 155–188. URL: https:// [15] E. Bassignana, V. Basile, V. Patti, Hurtlex: A muldoi.org/10.1007/s10579-022-09582-8. doi:10.1007/ tilingual lexicon of words to hurt, in: E. Cabrio, S10579- 022- 09582- 8. A. Mazzei, F. Tamburini (Eds.), Proceedings of the [4] C. Bianchi, Slurs and appropriation: An echoic Fifth Italian Conference on Computational Linguisaccount, Journal of Pragmatics 66 (2014) 35–44. tics (CLiC-it 2018), Torino, Italy, December 10[5] E. Zsisku, A. Zubiaga, H. Dubossarsky, Hate speech 12, 2018, volume 2253 of CEUR Workshop Proceeddetection and reclaimed language: Mitigating false ings, CEUR-WS.org, 2018. URL: https://ceur-ws.org/ positives and compounded discrimination, in: Pro- Vol-2253/paper49.pdf. ceedings of the 16th ACM Web Science Conference, [16] C. Bianchi, Il lato oscuro delle parole: epiteti deni2024, pp. 241–249. gratori e riappropriazione, Sistemi intelligenti 27 [6] T. Gillespie, Custodians of the Internet: Platforms, (2015) 285–302.
content moderation, and the hidden decisions that [17] J. Bai, S. Bai, Y. Chu, Z. Cui, K. Dang, X. Deng, shape social media, Yale University Press, 2018. Y. Fan, W. Ge, Y. Han, F. Huang, B. Hui, L. Ji, M. Li, [7] N. Strossen, Hate: Why we should resist it with free J. Lin, R. Lin, D. Liu, G. Liu, C. Lu, K. Lu, J. Ma, speech, not censorship, Oxford University Press, R. Men, X. Ren, X. Ren, C. Tan, S. Tan, J. Tu, P. Wang, 2018. S. Wang, W. Wang, S. Wu, B. Xu, J. Xu, A. Yang, [8] D. Nozza, A. T. Cignarella, G. Damo, T. Caselli, H. Yang, J. Yang, S. Yang, Y. Yao, B. Yu, H. Yuan, V. Patti, HODI at EVALITA 2023: Overview of Z. Yuan, J. Zhang, X. Zhang, Y. Zhang, Z. Zhang, the first shared task on homotransphobia detec- C. Zhou, J. Zhou, X. Zhou, T. Zhu, Qwen technical tion in italian, in: M. Lai, S. Menini, M. Polignano, report, CoRR abs/2309.16609 (2023). URL: https:// V. Russo, R. Sprugnoli, G. Venturi (Eds.), Proceed- doi.org/10.48550/arXiv.2309.16609. doi:10.48550/ ARXIV.2309.16609. arXiv:2309.16609. [18] F. M. Plaza-del Arco, M.-T. Martín-Valdivia,
R. Klinger, Natural language inference prompts for zero-shot emotion classification in text across corpora, in: N. Calzolari, C.-R. Huang, H. Kim, J. Pustejovsky, L. Wanner, K.-S. Choi, P.-M. Ryu, H.H. Chen, L. Donatelli, H. Ji, S. Kurohashi, P. Paggio, N. Xue, S. Kim, Y. Hahm, Z. He, T. K. Lee, E. Santus, F. Bond, S.-H. Na (Eds.), Proceedings of the 29th International Conference on Computational Linguistics, International Committee on Computational Linguistics, Gyeongju, Republic of Korea, 2022, pp. 6805–6817. URL: https://aclanthology.org/ 2022.coling-1.592. [19] P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, G. Neubig, Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, ACM Computing Surveys 55 (2023) 1–35. [20] S. Frenda, G. Abercrombie, V. Basile, A. Pedrani, R. Panizzon, A. T. Cignarella, C. Marco, D. Bernardi, Perspectivist approaches to natural language processing: A survey, Language Resources and Evaluation Journal (2024). doi:10.
1007/s10579-024-09766-4.
4
Stabilisci se la frase contiene riappropriazione semantica, rispondi ’Vero’ se lo contiene e ’Falso’ altrimenti.
propriation; respond ’True’ if it contains it and ’False’ otherwise.
Stabilisci se la frase contiene un linguaggio che non ha intenzione di ofendere, Rispondi ’Vero’ se lo contiene e ’Falso’ altrimenti.
has not abusive intent. Respond ’True’ if it does and ’False’ otherwise Stabilisci se la frase contiene un linguaggio che intende ofendere delle persone per il loro orientamento sessuale e le loro identità di genere, rispondi ’Vero’ se lo contiene e ’Falso’ altrimenti.
Stabilisci se nelle frasi proposte le seguenti parole “frocio, invertito, travestit*, checca, deviato, culattone, finocchio, finocchi, omosex, finocchietto, omosessuali, frocia, ricchione, trans, troia” sono utilizzate per ofendere le persone per il loro orientamento sessuale e/o identità di genere. Rispondi “Vero” se c’è un intento ofensivo, altrimenti “Falso”.
Determine if the following words in the proposed sentences—’frocio, invertito, travestit*, checca, deviato, culattone, finocchio, finocchi, omosex, finocchietto, omosessuali, frocia, ricchione, trans, troia’—are used to ofend people based on their sexual orientation and/or gender identity. Respond ’True’ if there is an ofensive intent, otherwise respond ’False’.
Translation
myself.
I’m at university and I just can’t stop being so gay today, help! How gay is she, I love her, she wants a paper map to explore the gardens.
half gay..
Today the world is screaming at me that I am a colossal hopeless fag, and I’m agreeing with it.
I’m about to tweet something very gay.