=Paper=
{{Paper
|id=Vol-3432/paper33
|storemode=property
|title=Explainable Classification of Internet Memes
|pdfUrl=https://ceur-ws.org/Vol-3432/paper33.pdf
|volume=Vol-3432
|authors=Abhinav Kumar Thakur,Filip Ilievski,Hông-Ân Sandlin,Zhivar Sourati,Luca Luceri,Riccardo Tommasini,Alain Mermoud
|dblpUrl=https://dblp.org/rec/conf/nesy/ThakurISSL0M23
}}
==Explainable Classification of Internet Memes==
https://ceur-ws.org/Vol-3432/paper33.pdf
Explainable Classification of Internet Memes
Abhinav Kumar Thakur1 , Filip Ilievski1,∗ , Hông-Ân Sandlin2 , Zhivar Sourati1 ,
Luca Luceri1 , Riccardo Tommasini3 and Alain Mermoud2
1
University of Southern California, Information Sciences Institute, USA
2
Cyber-Defence Campus, armasuisse Science and Technology, Switzerland
3
Institut National des Sciences Appliquées, France
Abstract
Nowadays, the integrity of online conversations is faced with a variety of threats, ranging from hateful
content to manufactured media. In such a context, Internet Memes make the scalable automation of
moderation interventions increasingly more challenging, given their inherently complex and multimodal
nature. Existing work on Internet Meme classification has focused on black-box methods that do not
explicitly consider the semantics of the memes or the context of their creation. This paper proposes a
modular and explainable architecture for Internet Meme classification and understanding. We design
and implement multimodal classification methods that perform example- and prototype-based reasoning
over training cases, while leveraging both textual and visual SOTA models to represent the individual
cases. We study the relevance of our modular and explainable models in detecting harmful memes on
two existing tasks: Hate Speech Detection and Misogyny Classification. We compare the performance
between example- and prototype-based methods, and between text, vision, and multimodal models,
across different categories of harmfulness (e.g., stereotype and objectification). We devise a user-friendly
interface that facilitates the comparative analysis of examples retrieved by all of our models for any
given meme, informing the community about the strengths and limitations of these explainable methods.
Keywords
explainability, neuro-symbolic integration, case-based reasoning, internet memes
1. Introduction
In the Internet era, the real and the virtual worlds are becoming increasingly closer, and nearly
every person, event, and idea have a Web counterpart. A particularly unique information
medium that arises in this context is the Internet Meme (IM): “a piece of culture, typically
a joke, which gains influence through online transmission” [1]. An IM is based on a medium,
typically an image representing a well-understood reference to a prototypical situation within a
certain community. IMs have been extremely popular: according to a recent survey by Facebook,
75% of people between 13 and 36 share Internet Memes (IMs), and 30% do it daily.1 IMs can
be viral: following another study [2], 121,605 different variants of one particular meme were
NeSy 2023, 17th International Workshop on Neural-Symbolic Learning and Reasoning, Certosa di Pontignano, Siena,
Italy
∗
Corresponding author.
Envelope-Open akthakur@isi.edu (A. K. Thakur); ilievski@isi.edu (F. Ilievski); hongan.sandlin@ar.admin.ch (H. Sandlin);
souratih@isi.edu (Z. Sourati); lluceri@isi.edu (L. Luceri); riccardo.tommasini@insa-lyon.fr (R. Tommasini);
alain.mermoud@ar.admin.ch (A. Mermoud)
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
CEUR Workshop Proceedings (CEUR-WS.org)
Proceedings
http://ceur-ws.org
ISSN 1613-0073
1
https://www.facebook.com/notes/10158928003998415, accessed 17/12/2022.
posted across 1.14 million status updates.2 As multimodal (they combine visual and language
information creatively), relatable (and, thus, dependent on community and virtual context),
succinct (they spread complex messages with a minimal information unit that connects the
virtual circumstances to the real ones), and fluid (subject to variations and alterations), IMs
provide essential data to study the flux of ideas on the Web.
Conversely, as online platforms have been already weaponized in a variety of geo-political
events and social issues [4, 5, 6, 7], the scale and complexity of IMs make content moderation
even more difficult. The inaccurate classification of memes can lead to inadequate moderation
interventions (removal, flagging, demotion, etc.) that, combined with the lack of tracking
mechanisms across platforms, has the potential to further decrease public trust in social media
platforms and related moderation policies. Content moderation policies, or the lack thereof,
can have serious implications on individuals, groups, and society as a whole. On the one hand,
content moderators may react late, inconsistently, or unfairly, thus angering users [8], as well
as contributing to reinforcing and exacerbating conspiratorial narratives [9, 6]. On the other
hand, minimal content moderation may permit coordinated influence operations [10] or enable
the spontaneous formation of toxic and dangerous communities, e.g., the study by Mamié
et al. demonstrates how “the Manosphere”, a conglomerate of men-centred online communities,
may serve as a gateway to far-right movements. At this scale, moderation of IMs requires
machine-augmented methods for tracking and hate speech classification of IMs.
Existing computational work on IMs has recognized the need to assist content moderators.
Much of this work has focused on tracking their temporal spread over time (i.e., virality) [12, 13,
14]. The recent introduction of two tasks on the topic of hate speech, namely Hateful Memes [15]
and Misogyny identification [16], has inspired methods that classify IMs based on unimodal
or multimodal features, largely dominated by perceptual information. While these methods
are a step in the right direction, they are typically modeled as black boxes and optimized for
accuracy. Considering the complex interplay of text, vision, and background knowledge in
IMs, the decisions reached by such models cannot be trusted by human stakeholders. To assist
human moderators and social scientists to understand the semantics and the pragmatics of IMs
at scale, these methods must be designed with explainability as a requirement.
In this paper, we explore explainable multimodal methods for IM classification. We rely on the
general idea of Case-Based Reasoning (CBR), where a prediction can be traced back to similar
memes that the method has observed at training time. Considering the complex nature of IMs,
we opt for CBR because it can provide transparent insights into the model reasoning, while still
leveraging the representation learning ability of state-of-the-art (SOTA) models. Based on these
premises, this paper provides three key contributions:
1. We devise an explainable framework for IM classification, consisting of explainable
methods that perform CBR over features that represent individual memes. We adopt
representative CBR methods based on example- and prototype-based reasoning over text,
vision, and multimodal feature extractors.
2. We evaluate the accuracy and the explainability of our framework on two tasks:
Hate Speech Detection and Misogyny Classification, and across different categories of
2
While the term meme usually refers to biological memes [3], in this paper, we treat it as synonymous to Internet
Meme (IM).
harmfulness (e.g., stereotype and objectification). We perform ablation experiments to
understand the impact of different modalities and modeling choices.
3. We devise a user-friendly interface that facilitates the comparative analysis of examples
retrieved by all of our models for any given IM. We apply the user interface to understand
the ability of different explainable models to retrieve useful instances for CBR and inform
future work about the strengths and limitations of these methods.
We make our code available to facilitate future research on explainable IM classification.3
2. Related Work
Most prior works on Internet Memes in AI have focused on understanding their virality and
spread on social media over time [12, 13, 14]. Another popular direction has been detecting
forms of hate speech in memes. The Hateful Memes Challenge and Dataset [15] is a competition
accompanied by an open-source dataset with over 10 thousand examples, where the goal is to
leverage vision and language understanding to identify memes with hateful content. Kirk et
al. [17] compare memes in this challenge to memes in the ‘wild’, observing that extraction of
captions is an open challenge, and that open-world memes are more diverse than memes in
curated benchmarks. The Multimedia Automatic Misogyny Identification (MAMI) [16] challenge
asks systems to identify misogynous memes, based on both text and images in the targeted
input memes. Methods for these challenges typically employ Transformer-based models that
incorporate vision and language, like ViLBERT [18], UNITER [19], and CLIP [20]. For a more
comprehensive overview of methods for detecting hate speech in memes, we refer the reader
to the recent review by Hermida and Santos [21]. Sheratt [22] aims to organize memes into
a genealogy, with the goal of building a comprehensive knowledge base going forward. The
combination of efforts to explain IMs with explicit knowledge and the generalization power of
large visual, textual, and multimodal models holds a promise to advance the SOTA of meme
understanding and classification. However, to our knowledge, no prior work has focused on
such multi-faceted and explainable methods for understanding IMs. To bridge this gap, we
design a modular architecture that integrates visual and textual models with prototype- and
example-based reasoning methods. Our framework thus balances the goals of obtaining SOTA
performance and providing transparent access to the model reasoning.
There has been a surge in using example-based explanations to enhance people’s compre-
hension of black-box deep learning models’ behavior and acquired knowledge. [23] propose
and evaluate two kinds of example-based explanations in the visual domain. The extracted
similar training data points help the end-users understand and recognize the capabilities of
the model better. Although [24] come to the same conclusion as [23] confirming the effect of
examples to boost the comprehension of the model by end-users, they do not see any evidence
supporting the same effect about the trust of end-users when presented with example-based
explanations. Similarly, methods for prototype-based classification have been developed for
visual tasks in the past, such as xDNN [25]. However, to our knowledge, we are the first work to
employ example-based and prototype-based methods for downstream tasks of IM classification.
3
https://github.com/usc-isi-i2/meme-understanding
Figure 1: Classification and feature extraction model within the example-based explanation model.
3. Explainable Framework for IM Classification
An IM classification model that detects offensive or inappropriate IMs can be easily trained.
However, the black-box nature of ML models makes it difficult to interpret why a meme is
flagged [26], especially when misclassified. To address this limitation, we adopt a neuro-symbolic
framework with explainable example- and prototype-based predictions for IM classification
tasks. Both methods utilize a frozen pre-trained model to extract features from a meme in a
transfer learning setup with a separate downstream classification model, which leverages the
features to make a final decision. The modularity of the framework enables an easy comparison
over the combination of the feature extraction model and the explanation method used.
3.1. Explainable methods
Example-Based Meme Classification [27] categorizes IMs and explains them based on similar
memes found in the training data. Example-based explanation works by retrieving training
examples that have a similar representation to the test example from the model’s point of view
to act as a proxy to understand the model’s behavior. This approach provides a direct insight
into the model reasoning, enabling users to analyze errors and detect latent biases in the dataset
[28]. Figure 1 depicts the IM classification model, which applies a classification head (L1-L3 and
Predictions layer) over the features extracted from a frozen pre-trained model for prediction.
The last hidden state (output of L3) of the trained classifier is used for calculating the similarity
between IMs based on cosine similarity. Then, for an unlabeled meme, we predict the labels
using this classification model. The features extracted from a pre-trained model can be fed
into a query engine to select similar images from a database that stores pre-computed features
computed by the same pre-trained model for the training memes. To display the retrieved
similar memes in a user-friendly way, we develop a visualization tool to display the model-wise
predictions and similar memes from the training dataset, thus supporting the predictions with
example-based explanations.
Prototype-Based Meme Classification is based on the prototype theory [29] of categorization
in psychology and cognitive linguistics. The prototype theory dictates that any given concept in
any given language has a real-world example that best represents this concept, i.e., its prototype.
For other concepts, there is a graded degree of belonging to a conceptual category, and some
members are closer to the prototype than others. The prototype-based classification relies
on learning label-wise prototypes from the training dataset followed by a rule-based decision
algorithm for the classification, which makes these models inherently interpretable. We adopt
the method of Explainable Deep Neural Networks (xDNN) [25] (Figure 5 in the appendix), which
combines statistical learning and reasoning to make an explainable prediction. xDNN uses
a neural backend to automate feature extraction followed by multiple statistical layers that
dynamically learn from the underlying data distribution. The learning objective of xDNN is to
create classwise prototypes, formalized as the local peaks for class distribution representing the
data cloud in the prototype’s vicinity. At inference time, given a new sample, xDNN computes
similarities over the features from the feature extraction layer against all the prototypes. Based
on these similarities, the new sample is classified following a rule-based decision approach.
The decision-making consists of two steps: (i) Local (per class) Decision Making: finding the
classwise prototype with the highest similarity, (ii) Global Decision Making: comparing the best
prototypes across all the classes and choosing the one with the highest similarity to decide the
final label. While the explanations of the example-based method are extracted post hoc, those
in the prototype-based method are part of the decision-making model.
Both example-based and prototype-based classifiers are instances of case-based reasoning,
and there has been some controversy over the superiority of one over the other. There are both
claims about the superiority of prototypical examples over normal examples [30], as well as
their counterparts [31] who state that a context theory of classification, which derives concepts
purely from exemplars, works better than a class of theories that included prototype theory.
3.2. Pretrained Models for Feature Extraction (FE)
We chose the following pre-trained models for feature extraction to analyze the information
captured by models trained over different modalities and pretraining strategies.
Textual Models. We use BERT𝑏𝑎𝑠𝑒 [32], trained on BooksCorpus (800M words) and English
Wikipedia (2,500M words) using two unsupervised tasks of Masked LM and Next Sentence
Prediction. We expect that BERT would help analyze explainability for general-purpose formal
language. Expecting that command of slang in social media is essential for meme understanding,
we use the BERTweet model [33] having the same architecture as BERT𝑏𝑎𝑠𝑒 and trained using
the RoBERTa [34] pretraining procedure over 80 GB corpus of 850M English tweets. We
expect BERTweet to be a better fit for encoding meme text as tweets have short text length and
generally contain informal grammar with irregular slang vocabulary, similar to IMs.
Vision Models. To capture visual information, we used the CLIP (Contrastive Language-Image
Pre-training) model [35]. CLIP is trained with Natural Language Supervision over 400 million
(image, text) pairs collected from the Internet with the contrastive objective of creating similar
features for an image and text pair. Because of the variety of training data and unrestricted
Table 1
MAMI dataset characteristics.
Sets Total Misogynous Shaming Stereotype Objectification Violence
Training 10,000 5,000 1,274 2,810 2,202 953
Test 1,000 500 146 350 348 153
text supervision, CLIP reaches SOTA-comparable zero-shot performance over various tasks
like fine-grained object classification and action recognition in videos. CLIP is robust to the
distribution shift between various datasets and shows better domain generalization across
datasets.
Mixed Models. As IMs are based on a complex and creative interplay between textual and
visual information [36, 37], we apply mixed FEs to capture both graphical and textual information
simultaneously. To do so, we concatenate features from both BERTweet and CLIP together.
4. Experimental Setup
We experiment with meme classification tasks over two existing datasets: MAMI and Hateful
Memes.
SemEval-2022 Task 5: Multimedia Automatic Misogyny Identification (MAMI) [16]
consists of two sub-tasks of misogyny detection and its type classification. Sub-task A: Misogyny
Detection Task focuses on detecting whether a meme is misogynous. The inter-annotator Fleiss-k
Agreement for sub-task A is 0.5767. Sub-task B: Misogyny Type Classification Task is a multi-class
task that categorizes a meme into one or more misogynous types, namely, shaming, stereotype,
objectification, and violence. A more formal description of these categories can be found in
[16]. The inter-annotator Fleiss-k Agreement for sub-task B is 0.3373. The inter-annotator
Fleiss-k Agreement clearly shows that sub-task B is comparatively more difficult than sub-task
A. Data statistics for both sub-tasks of the MAMI dataset are presented in Table 1.
Hateful Memes [38] consists of a single task of meme hate detection. The dataset consists of
10K memes equally divided into hateful and not-hateful classes; the dev and test set consist of
5% and 10% of the dataset, respectively. The average human accuracy on this task is 84.70%.
Evaluation. We keep our evaluation of classification performance consistent with the original
paper about the MAMI dataset. Sub-task A is evaluated using macro-average F1 measure for
each class label (misogynous and not misogynous). Likewise, Sub-task B is evaluated using
weighted-average F1 measure, weighted by the true label count for each label. For the Hateful
Meme dataset, we also follow the original work and evaluate the models based on classification
accuracy. In addition, we manually evaluate the example-based explanation approach using the
visualization tool by analyzing the prediction and similar memes from the training dataset. We
evaluate the prototype-based explanation method (xDNN) by its classification performance and
manually investigating the prototypes identified from the training dataset.
Table 2
Classification results for MAMI and Hateful Memes. We report results for our four FEs and two methods.
To contextualize these results, we provide statistics about the six baselines of the MAMI task and the
eleven baselines of the Hateful Memes task, as well as human accuracy when available.
Method Model Misogyny Misogyny type Hateful
detection classification Memes
Min 0.481 0.467 0.520
Task
Mean 0.680 0.663 0.594
Baselines
Max 0.834 0.731 0.647
Prototype BERT𝐵𝑎𝑠𝑒 0.537 0.524 0.496
-based BERTweet 0.543 0.534 0.470
method CLIP 0.642 0.629 0.552
(xDNN) CLIP + BERTweet 0.648 0.626 0.554
BERT𝐵𝑎𝑠𝑒 0.602 0.589 0.558
Example
BERTweet 0.600 0.594 0.546
-based
CLIP 0.685 0.686 0.593
method
CLIP + BERTweet 0.701 0.688 0.609
Human - - - 0.847
5. Analysis
5.1. Accuracy Analysis
Table 2 shows the performance of individual models within our framework on the MAMI and
Hateful memes datasets. The table compares our two methods over different feature extractor
models. For context, we show the min, max, and mean scores from the tasks’ papers [16, 15].
MAMI v/s Hateful Meme. Between MAMI sub-task A (misogyny detection) and Hate
detection over the Hateful Meme dataset, all models perform better on the misogyny detection
task. This is intuitive, as the presence of misogyny directly relates to the mention of women
(or related terminology), while hate is a more open-ended and multifaceted problem. For both
tasks and methods, we also observe that BERTweet, which is trained on Twitter data, performs
better than the BERT-based models for the MAMI dataset, though the difference between the
two models is relatively small. Thus, exposure to slang on social media data has a positive, yet
limited, impact on models for meme content classification. However, for the Hateful meme task,
BERT performs better than the BERTweet model due to the shift in distribution between the
two datasets. Finally, we note that the result of our best model is consistently better than the
average baseline result, but worse than the best result. The performance of our methods could
be further improved by fine-tuning the FEs on the downstream tasks.
Prototype-Based v/s Example-Based Explanation. For both datasets and different pre-
training models, the example-based method, which uses a neural classification head, performs
better than the prototype-based (xDNN) on the same pre-trained model. The prototype-based
models rely entirely on the pre-trained features and might lose performance on learning complex
patterns, which the deep learning model can learn. However, xDNN is much faster to train than
training the neural classifier head, as it needs just a single pass over the training data.
Modality Performance Analysis (Text v/s Image v/s Mixed). For each meme dataset
Figure 2: Explanatory interface for our example-based classification method. We cover highly explicit
or offensive content with white boxes.
and method combination, the CLIP-based image model performs comparatively better than
BERT-based text models. The combined model using CLIP and BERTweet features outperforms
all models, including those using CLIP alone. However, the improvement of the joint model
over CLIP is relatively low (0.5-1.5 points) compared to the improvement over BERT models
(6-10 points) for both explanation strategies. This means that either visual information is more
important than text, the CLIP model can also capture the textual information in the IM, or both.
For misogyny classification, the combination of CLIP and BERTweet also performs the best,
with the CLIP-only model performing very closely as the second best.
5.2. Explanability Analysis
Example-Based Classification Figure 2 shows the most similar IMs retrieved by the example-
based classification for one test image, visualized by our custom-made user interface. The
interface displays the model-wise predictions for BERTTweet, CLIP, and their combination,
together with similar memes from the training dataset for explainability. The test image is
misogynous, portraying a stereotype about women. The predictions from each model are
correct with high confidence about the misogyny detection and stereotype classification, which
is explainable to some degree by looking at similar examples from the training dataset. Focusing
on the most similar images per model, we observe that the combined model retrieves three
images that also depict misogyny (red background) and stereotyping. The interplay between
the text and the vision components is consistent across these IMs, two of which refer to the
relationship between the kitchen and women within the context of feminist discourse. This
example shows that the instances retrieved by the combined multimodal model are the most
reliable, which also correlates with its best performance. The examples by CLIP and BERTTweet
Figure 3: Prototypes for the running example IM according to the xDNN method with BERTTweet
extractor (left) and CLIP/CLIP+BERTTweet (right).
are partially useful, with CLIP retrieving more relevant images than BERTTweet in most cases.
In both the image-only and text-only settings, the retrieved memes relate to the same overall
topic e.g., feminism, but their intended meaning is somewhat different than the input IM.
This confirms that, although CLIP is retrieving images that are related to the test image and
can encode text to some extent, it is still beneficial to combine the features of CLIP with a
dedicated language model that understands slang, like BERTTweet. Focusing on the text-setting
only, although the model is performing well, we can observe that none of the memes are
accurately about the subject of discussion in the test image (women in the kitchen). This
confirms our expectation that IM classification should be evaluated both in terms of accuracy
and explainability simultaneously, as models might predict correctly for the wrong reasons.
Prototype-Based Classification To our surprise, for both datasets, xDNN creates prototypes
equal to the training supports for the class, i.e., the number of prototypes coincides with the
number of IMs in the training data. The memes, even though belonging to the same category, can
have very different textual/visual information content and representation. To understand better
the predictions of the prototype-based classifier, we investigate the most similar prototypes
for the same input meme presented in Figure 2. We show the nearest prototype according
to BERTTweet on the left of Figure 3, and the prototypical image according to CLIP and the
combined model on the right of this figure. The prototype according to BERTTweet is an image
that is not misogynous and seemingly unrelated to the input meme. However, the prototypical
IM according to CLIP and the mixed feature extractor is more relevant, as it is misogynous
and expresses stereotype and objectification, similarly to the input meme. As this IM depicts a
cartoon character, its surface similarity to the input is low, but CLIP is seemingly able to connect
it to the input meme based on abstract similarity relating to objectification and sexuality. Yet, we
note that the retrieved prototypes with CLIP are less relevant compared to the example-based
retrieval.
6. Conclusions and Outlook
In this work, we implemented and analyzed example- and prototype-based approaches for
explainable Misogyny Identification and Hate Speech Detection in IMs. Our experiments
revealed that methods for IMs classification can balance the goals of explainability and good
accuracy. While the example-based method is simpler, it achieved higher performance and its
explanatory power was more intuitive, as demonstrated through our tool-supported analysis.
Among the feature extractors, we observed that vision models were more effective than language
models, and the combination achieved the best performance.
A key future direction is integrating background knowledge and figures of speech. The
example in Figure 2 includes a test IM which depicts a woman in a kitchen. The stereotype and
misogyny, in this case, are most likely linked to assumed background knowledge, such as the
women’s social status in the 1960s, the second wave of feminism, and the more expansive link
between housewives and the kitchen. While the text and the image separately already hint at
misogyny, it is their combination that exhibits non-ambiguous misogyny. This is also apparent
from the examples extracted by models. Particularly, the most similar meme in terms of content
and references, the central image in the bottom row has explicit references to World War 3 and
the discussion revolving around two opposing opinions: (1) many Gen Z and Millennial women
worried about being drafted, and (2) women wanting equality only until they have to be a part of
the draft and joining the military. Moreover, by focusing on the other retrieved IMs, we observe
references to various sources. For instance, in the center IM in the last row, we observe the
title of Back to the Future movie that substitutes “future” with “kitchen”, implying the relation
between these two terms. In the center IM in the middle row, we see the Annoyed Picard, the
Star Trek character that has long been associated with implying irritability or disappointment
that is also extended here towards women. While these cues are captured to some degree by the
combined model, still, we see a gap that should be filled by background commonsense and factual
knowledge, as well as internet folklore to build robust and explainable meme classification
methods. This lack of knowledge is also apparent in the remaining IMs, which seem relevant
on the surface but are in fact dissimilar to the input meme.
We intend to explore methods for injecting background knowledge and figures of speech,
for instance, by integrating the framework with our Internet Meme Knowledge Graph to
tap into rich background knowledge [39]. We also plan to integrate a wider set of models,
e.g., VilBERT [18], explore the role of large language model prompting [40], and improve our
explanations to be more informative for users, e.g., by using features generated by CLIP in
combination with the misogynous type classification predictions. As test data for internet
memes is limited, we intend to work towards creating larger and more diverse data for the
identification and explanation of hate speech in memes. We make our code available in the
hope that the community will help us in pursuing these challenges together.
7. Acknowledgments
The first two authors have been supported by armasuisse Science and Technology, Switzerland
under contract No. 8003532866.
References
[1] P. Davison, 9. the language of internet memes, in: The social media reader, New York
University Press, 2012, pp. 120–134.
[2] E. A. L. Adamic, T. Lento, P. Ng, The evolution of memes on facebook, Facebook Data
Science (2014).
[3] R. Dawkins, The Selfish Gene, Oxford University Press, 1976.
[4] F. Pierri, L. Luceri, E. Ferrara, How does twitter account moderation work? dynamics
of account creation and suspension during major geopolitical events, arXiv preprint
arXiv:2209.07614 (2022).
[5] G. Nogara, P. S. Vishnuprasad, F. Cardoso, O. Ayoub, S. Giordano, L. Luceri, The disinfor-
mation dozen: An exploratory analysis of covid-19 disinformation proliferation on twitter,
in: 14th ACM Web Science Conference 2022, 2022, pp. 348–358.
[6] E. Chen, J. Jiang, H.-C. H. Chang, G. Muric, E. Ferrara, et al., Charting the information
and misinformation landscape to characterize misinfodemics on social media: Covid-19
infodemiology study at a planetary scale, Jmir Infodemiology 2 (2022) e32378.
[7] F. Pierri, B. L. Perry, M. R. DeVerna, K.-C. Yang, A. Flammini, F. Menczer, J. Bryden, Online
misinformation is linked to early covid-19 vaccination hesitancy and refusal, Scientific
reports 12 (2022) 1–7.
[8] H. Habib, R. Nithyanand, Exploring the magnitude and effects of media influence on
reddit moderation, Proceedings of the International AAAI Conference on Web and Social
Media 16 (2022) 275–286. URL: https://ojs.aaai.org/index.php/ICWSM/article/view/19291.
doi:1 0 . 1 6 0 9 / i c w s m . v 1 6 i 1 . 1 9 2 9 1 .
[9] L. Luceri, S. Cresci, S. Giordano, Social media against society, The Internet and the 2020
Campaign (2021) 1.
[10] R. DiResta, S. Grossman, Potemkin pages & personas: Assessing gru online oper-
ations, 2014-2019, White Paper https://fsi-live. s3. us-west-1. amazonaws. com/s3fs-
public/potemkin-pagespersonas-sio-wp. pdf (2019).
[11] R. Mamié, M. Horta Ribeiro, R. West, Are anti-feminist communities gateways to the far
right? evidence from reddit and youtube, in: 13th ACM Web Science Conference 2021,
2021, pp. 139–147.
[12] G. Marino, Semiotics of spreadability: A systematic approach to internet memes and
virality (2015).
[13] V. Taecharungroj, P. Nueangjamnong, The effect of humour on virality: The study of
internet memes on social media, in: 7th International Forum on Public Relations and
Advertising Media Impacts on Culture and Social Communication. Bangkok, August, 2014.
[14] C. Ling, I. AbuHilal, J. Blackburn, E. De Cristofaro, S. Zannettou, G. Stringhini, Dissecting
the meme magic: Understanding indicators of virality in image memes, Proceedings of
the ACM on Human-Computer Interaction 5 (2021) 1–24.
[15] D. Kiela, H. Firooz, A. Mohan, V. Goswami, A. Singh, P. Ringshia, D. Testuggine, The hateful
memes challenge: Detecting hate speech in multimodal memes, 2021. a r X i v : 2 0 0 5 . 0 4 7 9 0 .
[16] E. Fersini, F. Gasparini, G. Rizzi, A. Saibene, B. Chulvi, P. Rosso, A. Lees, J. Sorensen,
Semeval-2022 task 5: Multimedia automatic misogyny identification, in: Proceedings
of the 16th International Workshop on Semantic Evaluation (SemEval-2022), 2022, pp.
533–549.
[17] H. R. Kirk, Y. Jun, P. Rauba, G. Wachtel, R. Li, X. Bai, N. Broestl, M. Doff-Sotta, A. Shtedritski,
Y. M. Asano, Memes in the wild: Assessing the generalizability of the hateful memes
challenge dataset, arXiv preprint arXiv:2107.04313 (2021).
[18] J. Lu, D. Batra, D. Parikh, S. Lee, Vilbert: Pretraining task-agnostic visiolinguistic represen-
tations for vision-and-language tasks, Advances in neural information processing systems
32 (2019).
[19] Y.-C. Chen, L. Li, L. Yu, A. El Kholy, F. Ahmed, Z. Gan, Y. Cheng, J. Liu, Uniter: Universal
image-text representation learning, in: European conference on computer vision, Springer,
2020, pp. 104–120.
[20] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell,
P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language su-
pervision, in: International Conference on Machine Learning, PMLR, 2021, pp. 8748–8763.
[21] P. C. d. Q. Hermida, E. M. d. Santos, Detecting hate speech in memes: a review, Artificial
Intelligence Review (2023) 1–19.
[22] V. Sherratt, Towards contextually sensitive analysis of memes: Meme genealogy and
knowledge base (2022).
[23] C. J. Cai, J. Jongejan, J. Holbrook, The effects of example-based explanations in a machine
learning interface, in: Proceedings of the 24th International Conference on Intelligent
User Interfaces, IUI ’19, Association for Computing Machinery, New York, NY, USA, 2019,
p. 258–262. URL: https://doi.org/10.1145/3301275.3302289. doi:1 0 . 1 1 4 5 / 3 3 0 1 2 7 5 . 3 3 0 2 2 8 9 .
[24] C. Ford, E. M. Kenny, M. T. Keane, Play mnist for me! user studies on the effects of post-
hoc, example-based explanations &; error rates on debugging a deep learning, black-box
classifier (2020). URL: https://arxiv.org/abs/2009.06349. doi:1 0 . 4 8 5 5 0 / A R X I V . 2 0 0 9 . 0 6 3 4 9 .
[25] P. Angelov, E. Soares, Towards explainable deep neural networks (xdnn), 2019. URL:
https://arxiv.org/abs/1912.02523. doi:1 0 . 4 8 5 5 0 / A R X I V . 1 9 1 2 . 0 2 5 2 3 .
[26] R. Andrews, J. Diederich, A. B. Tickle, Survey and critique of techniques for extracting
rules from trained artificial neural networks, Knowledge-Based Systems 8 (1995) 373–389.
URL: https://www.sciencedirect.com/science/article/pii/0950705196819204. doi:h t t p s : / /
d o i . o r g / 1 0 . 1 0 1 6 / 0 9 5 0 - 7 0 5 1 ( 9 6 ) 8 1 9 2 0 - 4 , knowledge-based neural networks.
[27] A. Renkl, Toward an instructionally oriented theory of example-based
learning, Cognitive Science 38 (2014) 1–37. URL: https://onlinelibrary.
wiley.com/doi/abs/10.1111/cogs.12086. doi:h t t p s : / / d o i . o r g / 1 0 . 1 1 1 1 / c o g s . 1 2 0 8 6 .
arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/cogs.12086.
[28] I. Sigler, Example-based explanations to build better ai/ml models,
2022. URL: https://cloud.google.com/blog/products/ai-machine-learning/
example-based-explanations-to-build-better-aiml-models.
[29] E. H. Rosch, Natural categories, Cognitive Psychology 4 (1973) 328–350. URL: https:
//www.sciencedirect.com/science/article/pii/0010028573900170. doi:h t t p s : / / d o i . o r g / 1 0 .
1016/0010- 0285(73)90017- 0.
[30] M. K. Johansen, J. K. Kruschke, Category representation for classification and feature
inference, J. Exp. Psychol. Learn. Mem. Cogn. 31 (2005) 1433–1458.
[31] D. L. Medin, M. M. Schaffer, Context theory of classification learning, Psychol. Rev. 85
(1978) 207–238.
[32] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of deep bidirectional
transformers for language understanding, in: Proceedings of the 2019 Conference of
the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, Volume 1 (Long and Short Papers), Association for Computational
Linguistics, Minneapolis, Minnesota, 2019, pp. 4171–4186. URL: https://aclanthology.org/
N19-1423. doi:1 0 . 1 8 6 5 3 / v 1 / N 1 9 - 1 4 2 3 .
[33] D. Q. Nguyen, T. Vu, A. Tuan Nguyen, BERTweet: A pre-trained language model for
English tweets, in: Proceedings of the 2020 Conference on Empirical Methods in Natural
Language Processing: System Demonstrations, Association for Computational Linguistics,
Online, 2020, pp. 9–14. URL: https://aclanthology.org/2020.emnlp-demos.2. doi:1 0 . 1 8 6 5 3 /
v1/2020.emnlp- demos.2.
[34] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer,
V. Stoyanov, Roberta: A robustly optimized bert pretraining approach, 2019. URL: https:
//arxiv.org/abs/1907.11692. doi:1 0 . 4 8 5 5 0 / A R X I V . 1 9 0 7 . 1 1 6 9 2 .
[35] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell,
P. Mishkin, J. Clark, G. Krueger, I. Sutskever, Learning transferable visual models from
natural language supervision, 2021. URL: https://arxiv.org/abs/2103.00020. doi:1 0 . 4 8 5 5 0 /
ARXIV.2103.00020.
[36] E. Zenner, D. Geeraerts, One does not simply process memes: Image macros as multi-
modal constructions, Cultures and traditions of wordplay and wordplay research 6 (2018)
9783110586374–008.
[37] B. Dancygier, L. Vandelanotte, Internet memes as multimodal constructions, Cognitive
Linguistics 28 (2017) 565–598.
[38] D. Kiela, H. Firooz, A. Mohan, V. Goswami, A. Singh, P. Ringshia, D. Testuggine, The
hateful memes challenge: Detecting hate speech in multimodal memes, 2020. URL: https:
//arxiv.org/abs/2005.04790. doi:1 0 . 4 8 5 5 0 / A R X I V . 2 0 0 5 . 0 4 7 9 0 .
[39] R. Tommasini, F. Ilievski, T. Wijesiriwardene, IMKG: The Internet Meme Knowledge
Graph, in: Extended Semantic Web Conference, 2023.
[40] R. Cao, R. K.-W. Lee, W.-H. Chong, J. Jiang, Prompting for multimodal hateful meme
classification, arXiv preprint arXiv:2302.04156 (2023).
8. Appendices
8.1. Example-based Classification Process
The example-based classification method process at training and inference time is illustrated in
Figure 4.
8.2. Architecture of xDNN
Our architecture for prototype-based explainable classification, called Explainable Deep Neural
Networks (xDNN) is shown in Figure 5.
Figure 4: Example-based explanation based on similarity-based meme search. The Train meme-features
database contains pre-computed features using the Feature Extractor module.
Figure 5: Our architecture for prototype-based explainable classification, called Explainable Deep
Neural Networks (xDNN). Figure reused from the original paper [25].
8.3. Model Training Details
The classification model (Figure 1) used in the example-based explanation setup applies a
trainable neural head over frozen pre-trained models, which is trained with the Binary Cross
Entropy Loss using the Adam optimizer with a learning rate of 10−4 . Table 3 describes each
layer of the classification head, and the hidden state of L3 is used for feature extraction for
similar example searches over the training dataset.
xDNN [25] is a generative model, i.e., it learns prototypes and respective distributions au-
tomatically from the training data with no user/problem-specific parameters. We reuse the
publicly available xDNN implementation and experiment with different pre-trained models
described in the subsection on Pretrained Models.4 Table 3 describes each layer of the classifica-
tion head, and the hidden state of L3 is used for feature extraction for similar example searches
over the training dataset.
Table 3
Classification Head parameters for the Example-based method.
Layers Dimension Activation
L1 Feature length * 512 ReLU
L2 512 * 256 ReLU
L3 256 * 128 ReLU
Prediction 128 * Label count Sigmoid
4
https://github.com/Plamen-Eduardo/xDNN---Python