Warning: This paper contains examples of language and images that may be ofensive. This paper presents a probabilistic approach to identifying the disagreement-related elements in misogynistic memes by considering both modalities that compose a meme (i.e., visual and textual sources). Several methodologies to exploit such elements in the identification of disagreement among annotators have been investigated and evaluated on the Multimedia Automatic Misogyny Identification (MAMI) [ 1] dataset. The proposed unsupervised approach reaches comparable performances, and in some cases even better, with state-of-the-art approaches, but with a reduced number of parameters to be estimated. The source code of our approaches is publicly available†.
dataset [
Detecting abusive language has become an increasingly important task. The challenges introduced by the new modes of representation, which require a multimodal analysis, are further compounded when considering the subjectivity of the task. The subjectivity of the task derives from the fact that individuals’ perception of what characterizes a message of hate varies widely. Such diversification is reflected in the labeling phase in the form of disagreement among annotators. Identifying elements within the sample that can lead to disagreement is of paramount importance for several reasons. For content that can lead to disagreement, specific annotation policies might be introduced, and the number of annotators might be enlarged to capture multiple perspectives [3, 4, 5].
In this work, we propose a methodology to identify the disagreement-related elements in multimodal samples by exploring both visual and textual elements in the Many natural language tasks, such as hate speech detection, humor detection, and sentiment analysis, involve subjectivity since they require an interpretation based on human judgment, cultural context, or personal opinion [6]. Such phenomenon is reflected in the dataset through multiple labels from diferent annotators or via the inclusion of a confidence level to ground truth labels. Labels derived from diferent interpretations are therefore able to capture multiple perspectives and understandings [6].
Information about annotators’ disagreement has primarily been exploited as a means to improve data quality by excluding controversial instances [7, 8]. Alternatively, aiming at improving model performances, different strategies have been developed to exploit disagreement information in the training phase. For instance, in [9], the authors assign weights to instances to prioritize the ones with higher confidence levels. Another commonly adopted strategy [6, 10] aims at directly learning from disagreement without considering any aggregated label. While a considerable amount of research has been conducted to understand the reasons behind annotators’ disagreement [11, 12, 8] and to leverage disagreement when training classification models agreement has been proposed by [12]. Such taxonomy to a multimodal scenario. In particular, [23] introduces articulates four macro categories of reasons behind dis- a methodology to identify disagreement related conagreement: sloppy annotations, ambiguity, missing infor- stituents that, however, is limited to textual content. The mation, and subjectivity. Moreover, the authors evaluate approach includes a strategy to identify disagreementthe impact on classification performance of the diferent related textual constituents and an approach for gentypes. eralization towards unseen textual constituents. Both
Only recently, works have focused on the task of ex- methods have been extended to a multimodal scenario plaining disagreement [20, 21, 22, 23]. In [21], the au- in order to identify disagreement related elements both thors propose exploratory text visualization techniques in textual and visual sources that compose a meme. as a method for analyzing diferent perspectives from Given an element , a corresponding Element Disagreeannotated data. In [22], the authors identify textual con- ment Score ( EDS(e)) has been computed according to the stituents that contribute to hateful message explanation following equation: by exploiting integrated gradients within a filtering strategy. A more recent approach [23] proposes a probabilistic () = (|) − (¬|) (1) semantic approach for the identification of disagreementrelated constituents (e.g. textual elements) in hateful content. Overall, the findings indicate that, while LLM can yield promising results, comparable outcomes can be attained with less complex strategies and fewer computational resources. While previous research has concentrated on the analysis of textual disagreement, this study represents, to the best of our knowledge, a first insight into the explanation of multimodal disagreement.
In particular, we have revised and extended to the multimodal environment the methodology proposed in [23] in order to consider not only textual elements but also visual ones.
where (|) represents the conditional probability that there is agreement on a meme given that the meme contains the element . Analogously, (¬|) denotes the conditional probability that there is no agreement on a meme given that, that meme, contains the element . Given that EDS represents a difference between two complementary probabilities, it is bounded within the range of -1 to +1. A higher positive score indicates stronger agreement between annotators, whereas a lower negative score suggests disagreement.
The score can be estimated on the training data and exploited to identify additional disagreement-related elements on unseen memes.
3.1. Identification of Once the Element Disagreement Scores have been
estiDisagreement-Related Elements mated for each visual and textual element in the training
dataset, they can be exploited to qualify the level of
disThe first phase of the proposed approach aims to evalu- agreement on unseen samples. Analogously to what
ate the relationship between elements (both visual and was carried out in [23], diferent aggregation strategies
textual) that compose a meme and annotators’ disagree- have been investigated, relying on the hypothesis that
ment. Preliminary preprocessing operations have been the identified elements can be exploited for identifying
performed before identifying disagreement-related ele- the disagreement thanks to their diferent distribution in
ments. For what concerns the textual components, pre- samples with and without an agreement.
processing operations have been performed (i.e., tok- For each meme in the test set, the corresponding list
enization, lemmatization, lower casing and stop word of elements and the corresponding Elements
Disagreeremoval) to identify a valid set of tokens1 that might be ment Score estimated on the training data have been
related to disagreement. Considering the image com- extracted. In particular, for each meme, the textual and
ponent, the set of 14 human readable concepts (tags) visual elements have been identified and paired with the
identified by [
In order to measure the relationship among each ele- A qualitative evaluation, comprehensive of a compariment in the memes and the disagreement among annota- son with the specific misogynistic terminology and an tors, the approach proposed in [23] has been extended evaluation of the keyword included in the dataset creation phase, has been performed to assess the quality of the EDS, while both the F1-score for the two considered
10 times in the dataset have been removed. classes (agreement (+) and disagreement (-)) and a global F1-score have been computed to validate the MDS.
The score estimation is strongly based on what is observed in the training data, resulting in the lack of scores for any elements that do not appear in the training samples. This is particularly relevant for textual components rather than visual ones. In fact, while we can assume an open-word vocabulary (where a few terms on unseen data can not appear in the training set) for the textual source, we limited the visual tags to closed-word settings (only 14 tags can be considered both in training and unseen memes). Since we need to generalize only on unseen textual constituents, for each (unseen) textual element ˆ, an approximated EDS score has been computed as follows: of the training lexicon:
∑︀ [(, ˆ) · ()] (ˆ) = ∈ (3) ∑︀ (, ˆ) ∈ • Multimodal Disagreement Score with unseen constituents: All the above-proposed strategies for MDS estimation have been extended to also include elements that do not belong to the training lexicon and for which the EDS score has been estimated. In particular, given a multimodal sample , the aggregation functions presented in Section 3.2 will in this case consider the values of both seen (by considering the ()) and unseen (by considering the (ˆ)) elements. Such generalized aggregation functions will be later referred to through the prefix − .
• Embeddings of the training lexicon: the
contextualized embedding representation of each
textual element has been obtained via mBert [
textualized embedding representation has been Table 1 reports the top-10 highest positive and
highcomputed via mBert [
where (⃗x, ⃗v^) is the cosine similarity between the average contextualized embedding representation of element and ˆ, and is a grid search estimated threshold. • Unseen terms score: the EDS score for an unseen textual element ˆ is computed as the weighted average of the most similar constituents
training dataset is adopted, as the proposed work is focused on the analysis and prediction of disagreement and the test dataset is constructed to include only samples with complete agreement.
The training dataset, instead, is characterized by 65% of data with
complete agreement. Therefore, it has been divided in order to
isolate the 90% for token estimation and the remaining 10% for the
evaluation.
3Additionally, a boolean disagreement label has been derived to
represent complete agreement among annotators. In particular, this
last label is set to 1 if all the annotators have indicated the same
label, to 0 otherwise.
Table 1 subject of debate, particularly in relation to its
intersecTerms with the highest positive and lowest negative scores tion with misogynistic ideologies [
Table 2 reports the top-5 highest positive and highest sexual desirability. Achieved results suggest the ability negative disagreement scores derived for the visual com- of the approach to detect such variety in interpretations ponent. It is easy to notice how all the scores are positive and reflect them within the EDS scores. and achieve small values, denoting a tendency of such Figure 2 reports two memes that share the same text tags to be weakly related to the agreement label. and a diferent image. Despite such commonalities, the
Figure 1 reports an example of a meme with disagree- memes have been labeled diferently: while the first ment along with the visual representation of the EDS of meme has been labeled as misogynous by 2 annotators its textual and visual elements. Moreover, as highlighted out of 3, the second one has been unanimously labeled with a grey bar, some of the reported scores have been es- as non-misogynous. Since such memes share a common timated. Such scores correspond, in fact, to constituents textual representation, the derived textual elements and that are not present in the training dataset and for which textual-EDS are also equal, resulting in an indistinguishit was not possible to calculate the ESD score. The visual able representation that is inefective for disagreement representation of the scores related to such elements cor- identification. Moreover, although the memes difer in responds to the score obtained through the estimation the visual content, resulting in diferent tags and, therestrategy. Overall, it is easy to notice the presence of ele- fore, diferent textual-EDS, as previously mentioned, such ments strongly related to disagreement (i.e., sexual and a component alone is not suficient for disagreement premarket), highlighted in pink. diction.
The concept of the "sexual marketplace" is often the The findings demonstrate the necessity of joint considera
of the scores derived from all the elements (i.e., terms and tags). Results achieved on the textual component only highlight G-Mean as the most performing approach.
tionships for unseen terms. Furthermore, BERT [
Finally, Table 5 reports the performances of the different approaches for disagreement identification jointly considering both modalities. Furthermore, for a better comparison of the performance achieved by the proposed 4BERT has been implemented and finetuned using the hugging-face method) up to 8% for the mean-based approach. However, the best performances are achieved by the minimum and G-minimum approaches, for which the estimation methodology is not efective. Such behavior may be attributed to the imbalance in the dataset. The larger the number of samples with agreement, the greater the num5CLIP has been implemented and finetuned using the huggingface framework with default hyperparameters. In particular, we used framework with default hyperparameters. We adopted "bert-basethe version available at https://huggingface.co/openai/clip-vit-l cased" available at https://huggingface.co/google-bert/bert-base-c arge-patch14 to which we concatenated a linear layer for binary ased.
classification. 0.34 0.48 0.49 0.49 0.52 0.45 0.40 0.40
disagreement-related elements in multimodal content. The proposed approach allows for the identification of elements that could be used as a proxy to identify samples that might be perceived diferently by the annotators, and therefore, that could lead to disagreement. Achieved results highlight the dificulty of the task, denoting the need for a more advanced approach. Future work will include diferent strategies for image analysis in order to provide a better description of the image itself in all the elements that compose it. Furthermore, a study of the compositionality might be carried out to better represent the relationship among such elements inside the meme. The sense of a meme is often derived from the meanings of its individual parts (i.e. the image and text) and the way they are combined. By analyzing how diferent elements interact and contribute to the overall message, it is possible to gain a deeper understanding of how the meaning is represented within the diferent modalities. This will help in identifying complex patterns and improve the accuracy of classification models.
in the framework of the FairTransNLP-Stereotypes research project (PID2021-124361OB-C31) funded by MCIN/AEI/10.13039/501100011033 and by ERDF, EU A way of making Europe. todimamma, How do we study misogyny in the digital age? a systematic literature review using a computational linguistic approach, Humanities and
[3] P. Kralj Novak, T. Scantamburlo, A. Pelicon,
ment in hate speech modelling, in: International
ment of Uncertainty in Knowledge-Based Systems, Springer, 2022, pp. 681–695. [4] C. van Son, T. Caselli, A. Fokkens, I. Maks,
layered annotation scheme for perspectives, in:
A. Moreno, J. Odijk, S. Piperidis (Eds.), Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), European Language Resources Association (ELRA), Portorož, Slovenia, 2016, pp. 1177–1184. URL: https: (SemEval-2023), Association for Computational Lin//aclanthology.org/L16-1187. guistics, Toronto, Canada, 2023, pp. 171–176. URL: [5] S. Frenda, G. Abercrombie, V. Basile, A. Pedrani, https://aclanthology.org/2023.semeval- 1.24.
R. Panizzon, A. T. Cignarella, C. Marco, D. Bernardi, doi:10.18653/v1/2023.semeval-1.24. Perspectivist approaches to natural language pro- [15] M. Sullivan, M. Yasin, C. L. Jacobs, University at cessing: a survey, Language Resources and Evalua- bufalo at semeval-2023 task 11: Masda–modelling tion (2024) 1–28. annotator sensibilities through disaggregation, in: [6] A. Uma, T. Fornaciari, D. Hovy, S. Paun, B. Plank, Proceedings of the 17th International Workshop M. Poesio, Learning from disagreement: A survey, on Semantic Evaluation (SemEval-2023), 2023, pp. Journal of Artificial Intelligence Research 72 (2021) 978–985.
1385–1470. [16] A. de Paula, G. Rizzi, E. Fersini, D. Spina, et al., [7] B. Beigman Klebanov, E. Beigman, From annotator Ai-upv at exist 2023–sexism characterization usagreement to noise models, Computational Linguis- ing large language models under the learning with tics 35 (2009) 495–503. disagreements regime, in: CEUR WORKSHOP PRO[8] Y. Sang, J. Stanton, The origin and value of dis- CEEDINGS, volume 3497, CEUR-WS, 2023, pp. 985– agreement among data labelers: A case study of 999. individual diferences in hate speech annotation, in: [17] J. Erbani, E. Egyed-Zsigmond, D. Nurbakova, P.Information for a Better World: Shaping the Global E. Portier, When multiple perspectives and an Future: 17th International Conference, iConference optimization process lead to better performance, 2022, Virtual Event, February 28–March 4, 2022, an automatic sexism identification on social media Proceedings, Part I, Springer, 2022, pp. 425–444. with pretrained transformers in a soft label context, [9] A. Dumitrache, F. Mediagroep, L. Aroyo, C. Welty, Working Notes of CLEF (2023).
A crowdsourced frame disambiguation corpus with [18] M. E. Vallecillo-Rodríguez, F. del Arco, L. A. Ureñaambiguity, in: Proceedings of NAACL-HLT, 2019, López, M. T. Martín-Valdivia, A. Montejo-Ráez, Intepp. 2164–2170. grating annotator information in transformer fine[10] T. Fornaciari, A. Uma, S. Paun, B. Plank, D. Hovy, tuning for sexism detection, Working Notes of M. Poesio, et al., Beyond black & white: Leveraging CLEF (2023). annotator disagreement via soft-label multi-task [19] G. Rizzi, M. Fontana, E. Fersini, Perspectives on learning, in: Proceedings of the 2021 Conference hate: General vs. domain-specific models, in: of the North American Chapter of the Association Proceedings of the 3rd Workshop on Perspecfor Computational Linguistics: Human Language tivist Approaches to NLP (NLPerspectives)@ LRECTechnologies, Association for Computational Lin- COLING 2024, 2024, pp. 78–83.
guistics, 2021. [20] M. Michele, V. Basile, F. M. Zanzotto, et al., Change [11] L. Han, E. Maddalena, A. Checco, C. Sarasua, my mind: How syntax-based hate speech recogU. Gadiraju, K. Roitero, G. Demartini, Crowd nizer can uncover hidden motivations based on difworker strategies in relevance judgment tasks, in: ferent viewpoints, in: 1st Workshop on PerspecProceedings of the 13th international conference tivist Approaches to Disagreement in NLP, NLPeron web search and data mining, 2020, pp. 241–249. spectives 2022 as part of Language Resources and [12] M. Sandri, E. Leonardelli, S. Tonelli, E. Ježek, Why Evaluation Conference, LREC 2022 Workshop, Eudon’t you do it right? analysing annotators’ dis- ropean Language Resources Association (ELRA), agreement in subjective tasks, in: Proceedings of 2022, pp. 117–125. the 17th Conference of the European Chapter of the [21] L. Havens, B. Bach, M. Terras, B. Alex, Beyond exAssociation for Computational Linguistics, 2023, pp. planation: A case for exploratory text visualizations 2428–2441. of non-aggregated, annotated datasets, in: G. Aber[13] S. Shahriar, T. Solorio, Safewebuh at semeval- crombie, V. Basile, S. Tonelli, V. Rieser, A. Uma 2023 task 11: Learning annotator disagreement (Eds.), Proceedings of the 1st Workshop on Perspecin derogatory text: Comparison of direct training tivist Approaches to NLP @LREC2022, European vs aggregation, arXiv preprint arXiv:2305.01050 Language Resources Association, Marseille, France, (2023). 2022, pp. 73–82. URL: https://aclanthology.org/202 [14] E. Gajewska, eevvgg at SemEval-2023 task 11: 2.nlperspectives-1.10.
Ofensive language classification with rater-based [22] A. Astorino, G. Rizzi, E. Fersini, Integrated gradiinformation, in: A. K. Ojha, A. S. Doğruöz, ents as proxy of disagreement in hateful content, in: G. Da San Martino, H. Tayyar Madabushi, R. Ku- CEUR WORKSHOP PROCEEDINGS, volume 3596, mar, E. Sartori (Eds.), Proceedings of the 17th CEUR-WS. org, 2023.
International Workshop on Semantic Evaluation [23] G. Rizzi, A. Astorino, P. Rosso, E. Fersini, Unrav