The increasing prevalence of sexist and misogynistic statements on social media poses a serious challenge. Due to the high volume of content, manual moderation is not feasible; therefore, automated detection systems are urgently needed. The EXIST task within CLEF 2025 is dedicated to the automatic identification of sexist content in social networks and the determination of the intention and subtype of sexism. This paper describes the COMFOR team's contribution to the first task of the competition, which focuses on tweets. A support vector machine (SVM) based on a comprehensive feature representation, including embeddings and lexical features, was used. For the third subtask, this classifier was used as the basis for a classifier chain. Additionally, the results of the ifrst subtask were compared with those of a large language model (LLM) with an assigned persona. Our best models achieved an Information Contrast Measure (ICM) of 0.4928 (Subtask 1.1), −0.2203 (Subtask 1.2), and −0.4635 (Subtask 1.3) in the hard evaluation of the English test data.
Social networks are becoming increasingly important in today’s society as a source of information,
entertainment and even for exchanging opinions across geographical borders. Furthermore, a key
advantage of online communication is the anonymity that allows people to express their opinions freely.
However, this anonymity also has its downsides, as it can encourage the sharing of condescending or
ofensive content more frequently [
A particularly critical phenomenon is so-called hate speech, i.e. statements that attack individuals
or groups based on characteristics attributed to the groups [
To counteract the problem of online hate, including sexism, content moderators are tasked with
monitoring communication and promptly removing harmful content if necessary [
The sEXism Identification in Social neTworks (EXIST) competition addresses this challenge by
calling for the identification of content in text, image, and video data that either expresses critical or
contemptuous views towards women or refers to such events [
To accomplish these tasks, two diferent systems were compared with each other: On the one hand, a Support Vector Machine (SVM) was used, which was based on features such as embeddings, emojis and the frequency of sexist words. On the other hand, large language models (LLMs) were used as part of a few-shot prompting approach, in which the classification was carried out from the perspective of an assigned role (e.g., a male or female person). However, due to the long computing times of the LLMs, results could only be submitted for Subtask 1.1 in the competition using this approach.
The rest of this paper is organised as follows: After discussing the current state of the literature in section 2, the datasets used for this work and the methodology are briefly described in section 3. The results are then discussed in section 4. Finally, we conclude in section 5.
Since this paper uses both traditional classification methods and advanced LLMs, the following section provides an overview of recent approaches to sexism detection, both in the field of traditional methods and modern language models, as well as related areas of application.
The automatic detection of sexist tweets and the determination of their intention and subtype have
already been addressed in two previous shared tasks [
SVM was chosen as the classification algorithm. As demonstrated by the survey conducted by Abdollah Zadeh et al. [15], SVM is robust against outliers and particularly well-suited to high-dimensional data, a property that is particularly relevant in text classification due to the large number of potential features. However, the disadvantage of classic SVM is that it cannot handle overlapping classes. For Subtask 1.2, in which multiple classes exist, this problem can be solved by using separate binary classifiers for each class.
Subtask 1.3, however, represents a genuine multi-label problem that requires more complex solutions. Asti et al. [16] investigated various methods for multi-label classification, including the combination of classification algorithms such as SVM, Multinomial Naive Bayes (MNB) and Random Forest (RF) with transformation methods such as the binary relevance method [17], classifier chains [ 18] and the label powerset transformation [19]. The combination of SVM and classifier chains, as introduced by Read et al. [20] in particular, showed superior performance with identical features [16]. That means that SVM can be used not only as a binary classifier for Subtask 1.1, but also, with appropriate extensions, for the more complex Subtasks 1.2 and 1.3.
A key element in modelling is the selection of suitable features. Fasoli et al. [21], for example, compiled a list of sexist swear words and their social acceptability. Pamungkas et al. [22] also utilised swear words from both formal and informal language, drawing on data from the NoSwearing website [23], among other sources.
In addition, binary, dictionary-based features can be defined, for example, by checking whether a tweet contains terms associated with the word “woman” [21]. A hate word lexicon based on Bassignana et al. [24] was also used, which is divided into 17 weightable subcategories.
In the field of lexical features, classic representations such as Bag of Words (BoW), Bag of Hashtags and Bag of Emojis were used [22]. In addition, newer approaches demonstrate that semantic representation using word embeddings can yield better results [25]. Asudani et al. [25] provided a comprehensive overview of various methods, including Word2Vec [26] and its further development Global Vectors for Word Representation (GloVe) [27]. GloVe not only takes local contexts into account but also global co-occurrence statistics, and is particularly well suited for this application thanks to pre-trained models on Twitter data.
In summary, the overview of related research indicates that SVM, in conjunction with a classifier chain and features such as embeddings, presents an intriguing approach and can contribute to the diversification of methods within shared tasks.
In contrast to traditional approaches, participants in the two previous EXIST competitions primarily
employed smaller and larger language models [
Instead of BERT-based and other so-called encoder-only models, larger, mostly decoder-only models,
for instance from the Generative Pre-trained Transformer (GPT) [
The work of Samani et al. [
One possible approach for improvement is to combine few-shot prompting with the specification of
a persona (i.e., role) or a specific perspective from which the LLM should perform the classification
[
In particular, Tian et al. [
Despite their potential, it should not be forgotten that LLMs also have weaknesses, such as dificulties
in recognising implicit sexism [
To address the problem of sexism detection in social media, SVM and LLM-based approaches were employed for Subtask 1.1, while the SVM system was utilised for Subtasks 1.2 and 1.3. We submitted a total of six runs. An overview of the approaches is provided in Table 1. Concerning Subtask 1.3, run COMFOR_1 and COMFOR_2 difer only in that hard labels were provided in the first run and soft labels in the second. The following paragraphs provide a detailed description of the individual methods.
The Twitter dataset for all subtasks of Task 1 addressed in this paper was initially created for the previous
EXIST 2023 edition and described in detail by Plaza et al. [
A special characteristic of the annotations provided for the training data is that the “learning with
disagreement” paradigm [
Therefore, as is customary in the literature [
Subtask 1.3: In the third subtask, the multi-label classification of the targeted facets of a woman’s life, only the sexist tweets (majority label YES) were given a label. Due to the design of this task as a multi-label classification task, a slightly modified procedure was necessary here: a label was assigned as soon as at least two annotators agreed on an assignment. If two annotators did not agree on at least one label for a tweet, this tweet was removed from the dataset.
The resulting class distribution of the training dataset, enriched with the translated tweets, is presented
in Table 2, where NO indicates non-sexist tweets. For descriptions of the other classes, please refer to
Plaza et al. [
For all three subtasks, runs were submitted that utilised a traditional classification approach employing an SVM. The implementation of this system is described in more detail in the following subsections.
Several cleanup steps were performed. Initially, URLs and brackets were removed, as those were not
relevant for classification. In addition, mentions beginning with an @ sign were replaced with the
placeholder “person”, as it was irrelevant who exactly was being addressed since no information about
users was available. However, for the second sub-task, which aims to distinguish whether the tweet
reports on sexism, describes it or is itself sexist [
Since hashtags can also contain sexist references [
Finally, all tweets were converted to lowercase for normalisation and lemmatised using the English
UDPipe language model by Straka and Straková [
To generate the feature representation of the tweets for all three subtasks, various types of features were extracted and concatenated into a single feature vector for each instance. Those included a sparse term vector representation based on term frequency–inverse document frequency (TF-IDF) weighting, a dense embedding of the tweet, and additional numerical features such as the number of dictionary words and the frequency of specific token types, including emojis. Each feature type is described in more detail subsequently.
TF-IDF vectors Following the BoW approach, a TF-IDF weighted document-term matrix was generated, with each row containing the weighted term vector of a tweet.
GloVe-Embeddings In addition, the GloVe model by Pennington et al. [27] was used, which was trained on a large tweet corpus and is therefore particularly well suited for this data. Subsequently, to obtain a vector representation for a tweet, the arithmetic mean of the embeddings of the words occurring in that tweet was calculated.
Emojis The semantic information of the emojis was taken into account by including their word descriptions in the BoW representation and the embedding representation after preprocessing (see subsubsection 3.2.1). Their word descriptions were included in the bag-of-words representation and the embedding representation. To also take into account quantitative information about the use of emojis, the number of emojis per tweet was added as a numerical feature.
Hashtags The original hashtags served as the basis for two additional features: First, following the approach of Pamungkas et al. [22], a TF-IDF weighted bag-of-hashtags representation was created, analogous to the standard bag-of-words model. For each tweet, this approach yielded a sparse vector whose elements corresponded to the TF-IDF values of the hashtags appearing in that tweet. Second, the total number of hashtags per tweet served as an additional numerical feature.
Swear words Words with an ofensive or hurtful character play a central role in the detection of sexist language. Since traditional dictionaries often do not include such slang expressions, a list of swear words from the website NoSwearing.com [23] was used. This source, which is continually expanded, was particularly well-suited for identifying these terms. The number of swear words per tweet was integrated as a numerical feature.
Sexist words With a special focus on sexist language, a separate word list was created. To this end, the 100 words with the highest Pearson correlation to the YES label (i.e. sexism) in the training data were extracted. The frequency of these presumably sexist words served as a numerical feature.
After feature extraction, feature selection was carried out using the approach proposed by Kuhn [
Using the extracted features, SVM models were trained for each subtask. A sigmoid kernel was chosen
because Srivastava and Bhambhu [
The diferent types of classification tasks required the use of diferent strategies for using the SVM algorithm: Subtask 1.1: Binary classification SVM could be applied directly to this task.
Subtask 1.2: Multi-class classification Following the one-vs-one approach [
Subtask 1.3: Multi-label classification To take into account possible correlations between the
categories, an ensemble classifier chain was trained using the approach of [ 20]. For each category
(i.e. each facet), a binary classifier was created and linked in such a way that the predictions of
previous classifiers were incorporated into the following classifiers as additional features. Since
there is no natural order of the categories, the order of the classifiers in the chain was chosen
arbitrarily. The classifier chain provided the probability of tweets belonging to the categories.
These soft labels represented the COMFOR_2 run of the subtask for the soft-soft evaluation, as
described by Plaza et al. [
Subtasks 1.2 and 1.3 were treated as a hierarchical classification problem: The systems were trained only on sexist annotated training data and applied exclusively to tweets that were predicted as sexist in Subtask 1.1. This approach met the requirement to only classify sexist tweets further.
For Subtask 1.1, the results of the traditional approach, based on SVM, were compared with those of
LLMs using persona-based prompting in combination with in-context learning. The focus was on
opensource LLMs rather than paid APIs such as GPT-4 [
The classification task was presented to all models using the same system and user prompt. System prompt The system consists of an assignment to a persona or sociodemographic characteristic, a task description, and a definition of sexism. The following roles (personas) were each assigned to an LLM, which then annotated the entire English validation data set or test dataset: • a male person • a female person • Alice Schwarzer, a well-known feminist
The selection of the first two roles – a male person and a female person – was based on the work
of Tian et al. [
System Prompt You are <persona>. Your task is to decide whether a tweet expresses ideas related to sexism: Consider the following definition: Ideas about sexism can be expressed in the following three forms: the tweet is sexist itself, the tweet describes a sexist situation in which discrimination towards women occurs, or the tweet criticizes a sexist behaviour. Remember that you should annotate the tweet from the point of view of <persona>.
In addition, the LLM was provided with both the specific task and a definition of sexism via the
system prompt—an approach also suggested, for example, by Reuver et al. [
User prompt The user prompt consisted of the following parts: a repetition of the task and the
assigned role, instructions on the desired output format, and examples of the classification task for
in-context learning. Specifically, the one-class shot strategy recommended by Assis et al. [
User Prompt Determine whether the input tweet contains sexism. You should only reply with “YES” or “NO”. Do not provide explanations or notes. Respond with a single word. Respond “YES” if the tweet contains sexism and “NO” if it does not. Respond using JSON. Always remember that you should annotate the tweets from the perspective of <role>. Examples of classification are: • People really try to convince women with little to no ass that they should go out and buy a body. Like bih, I don’t need a fat ass to get a man. Never have. Label: YES • @messyworldorder it’s honestly so embarrassing to watch and they’ll be like “not all white women are like that Label: NO
The following section compares the models using the validation data and discusses the final results of the competition based on the test data.
First, the performance of the LLMs was compared on the validation dataset, with the Information
Contrast Measure (ICM) [
The results presented in Table 3 reveal apparent performance diferences between the models.
Regardless of the assigned persona, Gemma 3-27B consistently achieves the best scores, while
LLaMA3-8B performs the worst. For example, Gemma 3-27B achieves approximately three times the ICM value
of LLaMA-3-8B for male personas. This result supports the widely held assumption that model size is a
key performance factor in LLMs, as explained by Bender et al. [
Furthermore, there was no tendency for the models to favour a particular gender for better
classification results. While Gemma 3-27B and LLaMA-3-8B performed better when assigned the male
persona, Qwen2.5-32B and Mixtral 8x22B achieved the highest scores in the role of Alice Schwarzer.
Remarkably, none of the models achieved the best results with the unspecified female persona. One
possible explanation for the diferent performance of the models depending on the assigned persona
lies in the variation of the pre-training datasets [
For the larger models, the assigned persona appears to have less influence on performance, especially in the Gemma 3-27B model, which raises the question of whether the assignment of personas makes a remarkable contribution to the classification result. For clarification, the classification task was subsequently repeated with all models, this time without persona assignment. The system prompt and user prompt remained identical to those in paragraph 3.3 and paragraph 3.3; only the text elements marked in red, which defined the persona, were omitted.
As Table 3 (results highlighted in grey) demonstrates, persona assignment led to a slight improvement
in performance for all models except Gemma 3-27B. One possible explanation for the fact that the
best-performing model did not show any improvement is that specifying a specific perspective is
particularly helpful when a model is less confident in performing the task. However, it should also be
noted that the diference in performance between prompting with and without persona was minor,
especially for the larger models. This observation is also consistent with the findings of Civelli et al.
[
All systems used for submitting the runs were first evaluated on the English validation data (see Table 4). For Subtask 1.1, the SVM-based system (COMFOR_1) and the two most powerful LLM-based approaches were used. As shown in Table 3, the latter are Gemma 3-27B with an assigned male persona (COMFOR_2) and a female persona (COMFOR_3). The runs listed in Table 4 for Subtask 1.2 and Subtask 1.3 are based on the SVM system described in subsection 3.2.
It should also be noted that, for technical reasons, the ICM-Hard and ICM-Hard Norm scores could not
be determined for the run COMFOR_1 for Subtask 1.3 based on the validation data. The run COMFOR_2
for Subtask 1.3 yielded so-called soft labels (see subsubsection 3.2.4). Accordingly, the metrics ICM-Soft
and ICM-Soft Norm were calculated for this run, as described by Plaza et al. [
Subtask 1.1 When comparing the results of Subtask 1.1, it is noteworthy that the two LLM-based runs - COMFOR_2 (Gemma 3-27B with male persona) and COMFOR_3 (Gemma 3-27B with female persona) outperformed the SVM baseline (COMFOR_1) in terms of the ICM hard score, the primary evaluation metric selected by the organizers, with a relative diference of 62.5%. All other calculated metrics also yielded lower scores for the SVM-based system compared to the LLM-based approaches. One possible reason for the lower performance of the SVM is the significant class imbalance in the training data COMFOR_1 COMFOR_1 COMFOR_2 (see subsection 3.1). This problem did not afect the LLM, as no further training or fine-tuning was performed on the training dataset.
Another limitation of the SVM approach arises from the large number of concatenated features. Although attempts were made to identify redundancy through correlation analyses, these experiments should be repeated with alternative correlation measures such as XICOR [74], which also capture non-linear relationships.
Furthermore, it is questionable whether the combination of diferent lexical feature types – such as GloVe embeddings and the BoW representation – ofers added value. A systematic investigation of the respective contribution of individual feature types and their combinations to model performance is therefore necessary. More targeted feature engineering could further improve the results of SVM and may bring them closer to those of LLM-based approaches.
Concerning the results of the LLM approaches, it is remarkable that the macro precision, macro recall and macro F1 values hardly difer when assigning a male (COMFOR_2) or female persona (COMFOR_3). This result confirms the previous impression that the selected persona in Gemma 3-27B has little influence on the classification results.
Subtask 1.2 Regarding Subtask 1.2, it is noticeable that a significantly lower 1 measure of only 0.47 and a negative ICM score were achieved. One possible reason for this result is that only tweets that were classified as sexist (YES) by the SVM in Subtask 1.1 were passed on to the SVM for fine-grained classification. Misclassifications by the SVM for Subtask 1.1, therefore, have a direct impact on the performance of the SVM for Subtask 1.2.
An alternative approach would be to address the task not as a hierarchical problem, but to directly classify the tweets into four categories, including the NO category, i.e., no sexism. However, this approach would exacerbate the problem of imbalance, as the non-sexist tweets strongly outnumber the tweets assigned to the various categories of sexist tweets. The fact that the imbalance is already problematic in the three categories of author intent is evident from the 1 scores achieved for these individual categories: • DIRECT: 0.76 • JUDGEMENTAL: 0.08 • REPORTED: 0.27
The 1 measure for the overrepresented class DIRECT clearly exceeds that of the two
underrepresented classes. In addition to class imbalance, the performance diferences could also be due to the fact
that feature types such as BoW or dictionary-based features are particularly well suited for detecting
explicitly sexist tweets of the DIRECT category. Tweets in the JUDGEMENTAL category, however,
describe (social) circumstances perceived as sexist [
Subtask 1.3 Both the problem of imbalance and the dependence on the classifier’s performance from Subtask 1.1 also afect Subtask 1.3. Nevertheless, the macro 1 measure for this task was higher than for Subtask 1.2. This result was achieved through comparatively high macro precision, but at the expense of lower macro recall. The decisive factor in this trade-of is primarily the threshold value at which a tweet is assigned a label, which is why systematic testing of diferent threshold values could lead to performance improvements.
The results of the soft evaluation were considerably weaker, with a negative ICM soft score of −10.89
and an ICM soft norm of 0.0. A possible reason is the training on an aggregated gold standard (see
subsection 3.1), where agreement between two annotators was suficient for label assignment. Thus,
annotation uncertainties were not taken into account, which may have afected the prediction of the
actual proportion of annotators who assigned a label [
The final evaluation by the competition organisers resulted in several rankings: on the one hand, the
systems were compared only with the gold standard of one language, and on the other hand, with
the gold standard for both languages [
For all subtasks, even the first one, the submitted runs are listed at the bottom of the rankings. However, these results can mainly be explained by the fact that we only submitted results for English tweets. Therefore, the results and ranking of our runs, based solely on the English test data, which can be found in Table 6, are more interesting.
Table 6 clearly shows that LLM-based models continue to outperform SVM on the test data for
Subtask 1.1. In contrast to the results of the validation data, COMFOR_3, i.e. Gemma 3-27B with a female
persona, achieved a slightly higher ICM hard score than COMFOR_2. Ranked 77th out of 158, this run
is roughly in the middle of the submitted results. Accordingly, there is also room for improvement
for the LLM-based system. In addition to the aspects already discussed in subsection 4.2, it should
also be investigated whether, as emphasised by Jiang et al. [
Moreover, the SVM-based systems achieved poor results in Subtasks 1.2 and 1.3 with negative ICM
hard scores. However, the fact that over 20 % of the teams in Subtask 1.2 and 10 % in Subtask 1.3
achieved an ICM hard score of 0.0 [75] underscores the challenge of reliably predicting aggregated
labels in these classification tasks, considered subjective [
Finally, the results for the soft-soft evaluation of Subtask 1.3 are shown in Table 7, both for all test data and for English data only.
What is particularly notable here is that the ICM Soft norm is zero in both cases. However, when comparing these results with those of the other participants, it becomes apparent that over 20 participants achieved a score of zero in this measure, which highlights that predicting the proportion of annotators who assigned a label also poses challenges. In this context, the organisers of the Shared Task themselves emphasised that label ambiguity and the lack of agreement between annotators make this task particularly dificult [75].
The aim of the first task of the EXIST-2025 task was to detect sexism in tweets and then to classify sexist content in more detail. In this paper, we presented our approach to these subtasks specifically for English-language tweets. For binary sexism detection (Subtask 1.1), two approaches were compared: On the one hand, LLMs were used in combination with persona-based prompting and in-context learning. On the other hand, a traditional machine learning approach was used, in which an SVM was trained on a dataset enriched with tweets translated from Spanish into English. Features based on a GloVe model pre-trained specifically for Twitter, as well as additional lexical and statistical features, were used for the SVM. The SVM was also used for the other subtasks: to classify the author’s intention (Subtask 1.2) and to detect the facet of a woman’s life afected in the tweet (Subtask 1.3). For multi-label classification in Subtask 1.3, the SVM served as the basis for the classifier chain method.
For Subtask 1.1, our best model, the LLM Gemma 3-27B with an assigned male persona, achieved an ICM score of 0.4928 on the English test data, outperforming the SVM approach, which achieved an ICM score of 0.2033. In the other two subtasks, however, negative ICM values were achieved in each case, a result primarily due to the strong class imbalance in the training data, which negatively impacted the performance of the SVM.
One possible approach to mitigate this problem is to follow the example of Lee et al. [76] and utilise LLMs to generate additional training instances for underrepresented classes artificially. Furthermore, since LLMs delivered better results than classic models in Subtask 1.1, it makes sense to extend their use to Subtasks 1.2 and 1.3. To date, it has been demonstrated that assigning a gender-specific persona has had only a minor impact on the classification results. Therefore, further potential can be tapped by assigning alternative perspectives, such as a specific cultural background.
Additionally, it has not yet been taken into account that the intense subjectivity of the task can lead to disagreements among annotators. A possible solution here could be a combination of LLMs with classic methods such as SVM. In this case, the LLM would not act as the sole classifier, but as a kind of ‘artificial additional annotator’ that takes over a decision-making function in the event of label inconsistencies and can thus contribute to the consistency of the training dataset.
During the preparation of this work, the authors used ChatGPT, Grammarly in order to: Grammar and spelling check, Paraphrase and reword, Improve writing style. Further, the authors used DeepL in order to: Text Translation. After using these tool/services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [12] S. Gross, J. Petrak, L. Venhof, B. Krenn, GermEval2024 Shared Task: GerMS-Detect – Sexism Detection in German Online News Fora, in: B. Krenn, J. Petrak, S. Gross (Eds.), Proceedings of GermEval 2024 Task 1 GerMS-detect Workshop on Sexism Detection in German Online News Fora (GerMS-detect 2024), Association for Computational Lingustics, Vienna, Austria, 2024, pp. 1–9. [13] P. Donabauer, Pd2904 at GermEval2024 (Shared Task 1: GerMS-Detect): Exploring the Efectiveness of Multi-Task Transformers vs. Traditional Models for Sexism Detection (Closed Tracks of Subtasks 1 and 2), in: B. Krenn, J. Petrak, S. Gross (Eds.), Proceedings of GermEval 2024 Task 1 GerMS-detect Workshop on Sexism Detection in German Online News Fora (GerMS-detect 2024), Association for Computational Lingustics, Vienna, Austria, 2024, pp. 39–47. [14] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, in: J. Burstein, C. Doran, T. Solorio (Eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1, Association for Computational Linguistics, Minneapolis, Minnesota, 2019, pp. 4171–4186. doi:10.18653/v1/N19-1423. [15] A. Abdollah Zadeh, J. Leevy, T. Khoshgoftaar, A Survey on the Choice Between Binary Classification and One-Class Classification, in: Proceedings of the 27th ISSAT International Conference on Reliability and Quality in Design, International Society of Science and Applied Technologies, Virtual Event, 2022. [16] A. D. Asti, I. Budi, M. O. Ibrohim, Multi-label Classification for Hate Speech and Abusive Language in Indonesian-Local Languages, in: Proceedings of the 2021 International Conference on Advanced Computer Science and Information Systems (ICACSIS), IEEE, Depok, Indonesia, 2021, pp. 1–6. doi:10.1109/ICACSIS53237.2021.9631316. [17] E. Montañes, R. Senge, J. Barranquero, J. Ramón Quevedo, J. José Del Coz, E. Hüllermeier, Dependent binary relevance models for multi-label classification, Pattern Recognition 47 (2014) 1494–1508. doi:10.1016/j.patcog.2013.09.029. [18] E. Alvares-Cherman, J. Metz, M. C. Monard, Incorporating label dependency into the binary relevance framework for multi-label classification, Expert Systems with Applications 39 (2012) 1647–1655. doi:10.1016/j.eswa.2011.06.056. [19] M. R. Boutell, J. Luo, X. Shen, C. M. Brown, Learning multi-label scene classification, Pattern
Recognition 37 (2004) 1757–1771. doi:10.1016/j.patcog.2004.03.009. [20] J. Read, B. Pfahringer, G. Holmes, E. Frank, Classifier Chains for Multi-label Classification, in: W. Buntine, M. Grobelnik, D. Mladenić, J. Shawe-Taylor (Eds.), Machine Learning and Knowledge Discovery in Databases, Springer Berlin Heidelberg, Berlin, Heidelberg, 2009, pp. 254–269. [21] F. Fasoli, A. Carnaghi, M. P. Paladino, Social acceptability of sexist derogatory and sexist objectifying slurs across contexts, Language Sciences 52 (2015) 98–107. doi:10.1016/j.langsci.2015. 03.003. [22] E. W. Pamungkas, V. Basile, V. Patti, Misogyny Detection in Twitter: A Multilingual and CrossDomain Study, Information Processing & Management 57 (2020) 102360. doi:10.1016/j.ipm. 2020.102360. [23] All Slang Network, NoSwearing.com: Swear Word List, Dictionary, Filter, and API, https://www.noswearing.com/, 2025. [24] E. Bassignana, V. Basile, V. Patti, Hurtlex: A Multilingual Lexicon of Words to Hurt, in: Proceedings of the 5th Italian Conference on Computational Linguistics (CLiC-it), volume 2253, CEUR-WS, Turin, Italy, 2018, pp. 51–56. doi:10.4000/books.aaccademia.3085. [25] D. S. Asudani, N. K. Nagwani, P. Singh, Impact of word embedding models on text analytics in deep learning environment: A review, Artificial Intelligence Review 56 (2023) 10345–10425. doi:10.1007/s10462-023-10419-1. [26] T. Mikolov, I. Sutskever, K. Chen, G. Corrado, J. Dean, Distributed Representations of Words and Phrases and their Compositionality, Advances in Neural Information Processing Systems 26 (2013) 1–8. [27] J. Pennington, R. Socher, C. D. Manning, GloVe: Global Vectors for Word Representation, in: Proceedings of the 2014 Conference on Empirical Methods in Natural Language ProcessCan Language Models Be Too Big?, in: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, ACM, Virtual Event Canada, 2021, pp. 610–623. doi:10.1145/ 3442188.3445922. [73] C. Zhou, P. Liu, P. Xu, S. Iyer, J. Sun, Y. Mao, X. Ma, A. Efrat, P. Yu, L. Yu, S. Zhang, G. Ghosh, M. Lewis, L. Zettlemoyer, O. Levy, LIMA: Less Is More for Alignment, in: Proceedings of the 37th International Conference on Neural Information Processing Systems, Nips ’23, Curran Associates Inc., Red Hook, NY, USA, 2024, pp. 55006–55021. doi:10.48550/arXiv.2305.11206. [74] S. Chatterjee, A New Coeficient of Correlation, Journal of the American Statistical Association 116 (2021) 2009–2022. doi:10.1080/01621459.2020.1758115. [75] L. Plaza, J. Carrillo-de-Albornoz, I. Arcos, P. Rosso, D. Spina, E. Amigó, J. Gonzalo, R. Morante, Overview of EXIST 2025: Learning with Disagreement for Sexism Identification and Characterization in Tweets, Memes, and TikTok Videos, in: J. Carrillo-de-Albornoz, A. G. S. de Herrera, J. Gonzalo, L. Plaza, J. Mothe, F. Piroi, P. Rosso, D. Spina, G. Faggioli, N. Ferro (Eds.), Experimental IR Meets Multilinguality, Multimodality, and Interaction. Proceedings of the Sixteenth International Conference of the CLEF Association (CLEF 2025), CEUR-WS, Madrid, Spain, 2025. [76] D.-H. Lee, J. Pujara, M. Sewak, R. White, S. Jauhar, Making large language models better data creators, in: H. Bouamor, J. Pino, K. Bali (Eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Singapore, 2023, pp. 15349–15360. doi:10.18653/v1/2023.emnlp-main.948.