Multimodal detection of hateful memes by applying a vision-language pre-training model.

Detrimental to individuals and society, online hateful messages have recently become a major social issue. Among them, one new type of hateful message, "hateful meme", has emerged and brought difficulties in traditional deep learning-based detection. Because hateful memes were formatted with both text captions and images to express users' intents, they cannot be accurately identified by singularly analyzing embedded text captions or images. In order to effectively detect a hateful meme, the algorithm must possess strong vision and language fusion capability. In this study, we move closer to this goal by feeding a triplet by stacking the visual features, object tags, and text features of memes generated by the object detection model named Visual features in Vision-Language (VinVl) and the optical character recognition (OCR) technology into a Transformer-based Vision-Language Pre-Training Model (VL-PTM) OSCAR+ to perform the cross-modal learning of memes. After fine-tuning and connecting to a random forest (RF) classifier, our model (OSCAR+RF) achieved an average accuracy and AUROC of 0.684 and 0.768, respectively, on the hateful meme detection task in a public test set, which was higher than the other eleven (11) published baselines. In conclusion, this study has demonstrated that VL-PTMs with the addition of anchor points can improve the performance of deep learning-based detection of hateful memes by involving a more substantial alignment between the text caption and visual information.

Introduction

Hateful messages, more commonly known as hate speech, have unfortunately almost become a ubiquitous phenomenon on social media. After reviewing the definitions of hateful messages from authoritative sources, such as YouTube, Facebook, Twitter, etc. A hateful message is defined as a statement that explicitly or implicitly expresses hatred or violence against people with protected characteristics (). This definition distinguishes hateful messages from offensive language by their targets: though offensive language can be directed at either individuals or groups (i.e.,” Get out of this place!”), it does not target them based on their protected characteristics. Hateful messages are detrimental to both individuals and society because they can lead to prejudices against individuals, depreciation of minority’s abilities, alienation of minorities, degradation of individual mental health, rise of suicide, increase in offline hate crimes, and discriminatory practices when allocating public resources [1-6]. Especially due to their massive scale and rapid propagation on the internet, hateful messages have become a critical threat to individuals and the whole society.

With tremendous social pressure to limit the spread of hateful messages, though no legal framework explicitly targeting this subject has yet been developed, most major digital social media platforms have developed their respective moderation policies [7,8]. However, the outdated manual moderation is considered insufficient, as human moderators are slow and expensive. Moreover, content moderators are known to suffer post-traumatic stress disorder (PTSD)-like syndromes after repetitively reviewing violent and exploitative content [9]. As a result, generations of automatic hateful message detection methods have been developed since 2009, from the initial block-word-list approach to the current deep learning-based technology [10-17]. Though earlier methods have been limited in their usefulness as they often can only handle obvious patterns, encouragingly, the present transformer-based deep learning models have reached a much higher accuracy (Acc.) of more than 90% in detecting textual hateful messages due to its powerful capacity in the semantic comprehension [17]. The commercial application of deep learning models effectively reduces the spreading of textual hateful messages.

However, in recent years, a new type of “hateful message”, namely “hateful meme”, has emerged and become rampant due to its expressiveness and subtlety [18]. Hateful memes combine both text captions and background images to express users’ intents. Specifically, hateful memes can be more disguiseful than individual hateful images or hateful texts because many sentences or images that are harmless or even hilarious may become hateful when placed together (). These subtle references are easy for humans to understand yet difficult for machines to detect. So, hateful users thus intentionally publish hateful memes to evade conventional detectors [18]. Traditional automatic text detection completely ignores image features and vice versa. Therefore, the task of detecting multimodal hateful memes is very challenging. In order to develop a detection model to capture the complexity of these memes, the model should be not only able to process every single modality but also capable of fusing the two modalities.

Examples of hateful and non-hateful memes.

In meme A (non-hateful), the image of a nailed tire and the sentence “You have a spare tire today, don’t you?” are both neutral, as well as the combined version; in meme B (hateful), the same caption is inscribed as meme A, but the image part is changed from a nailed tire to the "girl holding flowers", has a reversed implication from neutrality to insulting as hateful. Besides, in meme B (hateful), neither the image part nor the caption expresses hatred alone, yet when combined, this meme expresses hatred. (Copyrighted by author F. Pan).

In the past few years, the advent of transformer-based large-scale Pre-trained Models (PTMs) has shed hope into overcoming the challenge mentioned above. PTMs such as GPT (Generative Pre-trained Transformers) and BERT (Bidirectional Encoder Representation from Transformer) have recently achieved great success in many complex natural language processing (NLP) tasks, demonstrating overall superior performance over conventional deep learning models and becoming a milestone in the wider machine learning community [19]. The rich knowledge stored implicitly in the tremendous amount of model parameters could be leveraged by fine-tuning them for specific downstream tasks. Following the success of PTMs on language tasks, the deep-learning community has proposed various unified Vision-Language PTMs (VL-PTMs), such as ConcatBERT, VilBERT, and Visual BERT, for multimodal tasks [18,20-24]. These models are shown to be able to capture more complex information and can be optimized to directly combine learned vector representations of both the text and image to reach state-of-the-art results in several Vision-Language tasks [18,22,23,25,26]. However, no matter which method was used, the current VL-PTMs presented an Acc. of about 60% to 65% in the hateful meme detection task, which is still far away from human performance (Acc. reported to be 84.7%) [18].

Although current VL-PTMs showed overwhelmingly positive performance in learning text semantics, they do not fare so well when learning image region features, probably due to an issue known as “ambiguity” [27]. To overcome this problem and improve the cross-modal learning in Vision-Language semantic space, additional input of “anchor point” was added to our detection model to bridge the text and vision semantic spaces, which might improve the VL-PTM performance. This idea of “anchor point” is motivated by the observation that modern object detectors can accurately detect the salient objects in an image, which are often mentioned in the paired text data [24]. Utilizing anchor point information has been demonstrated to benefit cross-modal learning in VL-PTM training, as such models have shown an overall improvement over SoTA (state of the art results) in different downstream tasks, such as image retrieval, text retrieval, and image captioning [27]. However, this method has not been used in hateful meme detection yet. In this study, we built up a novel hateful meme detector by applying a transformer-based VL-PTM with a triplet input: image feature, text caption, and object tags identified in the image. Unlike previous VL-PTMs, our model additionally considers object tags of the image background, serving as the anchor points. We speculate that adding the object-tag input can bridge the text caption and related visual image in hateful meme learning and help facilitate visual feature classifications. After training and testing our model, we compare its performance against other published baselines in the same public dataset.

Materials and methods

Dataset

Our model was trained and tested in a public dataset named “Facebook Hateful Meme Dataset” (https://www.kaggle.com/parthplc/facebook-hateful-meme-dataset). In constructing this dataset, researchers at Facebook first reconstruct online memes by placing meme text caption over a new underlying licensed image without the loss of meaning. Then, they hired annotators from a third-party annotation company rather than a crowd-sourcing platform. The annotators were trained for 4 hours to recognize hateful memes according to Facebook guidelines (https://www.facebook.com/communitystandards/hatespeech) (. Then, these annotators reconstructed the memes and made the classification. For every hateful meme, there is always a non-hateful alternative whose caption or image is changed from the original one. This kind of substitution is called “benign confounders”, a technique similar to current strategies of using counterfactual or contrastive examples; after the memes were labeled hateful, “benign confounders” were constructed [18]. Finally, a dataset with a total of 10k memes was set up and categorized into hateful or non-hateful. A dev and test set constituted 5% and 10% of the data, respectively, and the rest served as a training set.

Detection pipeline

In this model, an optical character recognition (OCR) module is applied to extract the text part of the memes, and an object detection (OD) module VinVl (Visual features in Vision-Language) is used to encode the correlated image part of memes. Then, we choose a VL-PTM module named OSCAR+ (Object-SemantiCs Aligned Re-training) to encode the extracted text part, image part, and object tags of memes [24,28]. OSCAR+ was established on the basis of multi-layer Transformers like most other PTMs [29]. However, unlike most existing VL-PTMs only providing a dual input of concatenated image region features and text features, OSCAR+ additionally affords a piece of specific information for object tags to bridge the text caption and related visual image.

We used the Google Colab platform to provide accelerators for inference and training. The hardware accelerators we used are an Nvidia P100 GPU and an 8-core TPU (Tensor 209 Processing Unit) V3. We first fine-tune OSCAR+ on the hateful meme dataset. In this stage, a two-layer Fully-Connected network (FC) is connected to OSCAR+ output (OSCAR+FC). The mini-batch gradient descent is carried out on the training set of 8500 images with a batch size set to 50 and a learning rate of 5.00E-06; the loss function is set to Binary Cross-Entropy Loss with Logits L(x, y) = −(y ln σ(x) + (1 − y) ln (1 − σ(x)). The dev set is used to optimize the model in avoidance of overfitting, and the final model is chosen at the epoch when the maximum area under the receiver operating characteristic (AUROC) is obtained on the dev set [30,31]. After OSCAR+ is fine-tuned, its output is connected to a random forest classifier (RF). Then, the RF is further optimized, consisting of ten decision trees whose maximum depth is set to 10. The trained RF is the final classifier (OSCAR+RF) for recognizing hateful memes. All training was repeated four times. The pipeline of our model construction is illustrated in .

Full pipeline.

FC denotes Fully-Connected network; RF denotes random forest classifier.

Meme preprocessing

Because the meme itself can not be fed directly into OSCAR+, it must be preprocessed into a suitable data format (). In preprocessing, the meme is first input into a VinVl object detector which uses the ResNeXt 152-C4 as the backbone feature extractor [32,33]. In VinVl, the backbone network transforms the input meme into a preliminary feature map. Afterwards, a Region Proposal Network (RPN) outlines regions of interest (ROIs) on the feature map containing predefined categories of objects. The features of ROIs are then pooled into ROI-Pooling vectors of 2048 dimensions. Afterwards, those vectors are passed through two paralleled FCs pathways: 1. the first FC is used to predict the bounding box’s position and size for each ROI-Pooling vector, outputting spatial features; then, each ROI-Pooling vector is concatenated with the corresponding Bounding-Box-Regression vector to pass through the following FC to form an image feature embedding vector with a size of 768; 2. the other paralleled FC is used to predict the category of the object in the corresponding ROIs outputting text embeddings of object tags with an exact size of 768. In general, the VinVl object detector will produce two sequences of vectors after meme input: the text object tags and the corresponding image feature embeddings () [28].

Meme preprocessing.

A. Embeddings of image features and object tags in VinVl. B. Meme preprocessing by VinVl and OCR in parallel results in a triplet input of image feature, object tags, and text caption to OSCAR+.

On the other hand, the meme is input into an OCR module to extract all the caption text that appears in the memes [34]. Then, both text caption extracted by OCR and texted object tags outputted from VinVl were tokenized by OSCAR+, forming tokenized vectors with the same size of 768. At last, the text tokens, the object tag tokens, and the image feature embeddings are then further concatenated in sequence, inserting a special token [SEP] between different sections of OSCAR+ input and appending a special token [CLS] at the start of the whole input sequence. Afterwards, a meme was transferred into an embedded and tokenized triplet for further OSCAR+ encoding () [24].

Encoding process in VL-PTM

The OSCAR+ model consists of 12 same encoder blocks () [24]. The first encoder takes the preprocessed meme embeddings with a matrix of dimension (m + 2n + 3) × 768 as input, which consists of the image features embeddings (m), the object tag embeddings (n), the text caption embeddings (n) and the embedding of special tokens ([CLS], [SEP], and [SEP]). The followed encoder block takes the output embedding sequence produced by the previous encoder as input (). In every encoder block (), the input embedding sequence is passed through 12 self-attention heads in parallel, each outputting a smaller matrix of dimension (m + 2n + 3) × 64. More specifically, in each self-attention head, the input matrix will pass through three paralleled FCs to produce three smaller matrices , and of dimension (m + 2n + 3) × 64. Then, the standard dot product attention operation was carried out as: where d equals 64 () [29]. These intermediate output matrices are again concatenated to form a matrix of the original size, passed through an FC, then added with the original matrix, and normalized by rows. This normalized matrix is given through an FC, added with itself, and normalized by rows. Finally, the encoder block will produce a matrix of dimension (m + 2n + 3) × 768.

Architectures of OSCAR+ and its encoder.

A. OSCAR+ consists of 12 tandem encoders; B. the architecture of an encoder; C. structure of the self-attention head.

Comparisons with baselines

We compared the hateful meme detection Acc. and AUROC between our model and other published eleven (11) baselines trained on the same dataset [18]. The baselines included both unimodal PTMs and multimodal VL-PTMs. The unimodal PTMs pre-trained in text data was BERT (Text BERT) [14]. The unimodal PTMs pre-trained in image data included standard ResNet-152 with average pooling (Image-Grid) and with fine-tuned FC6 layer by using weights of the FC7 layer (Image-Region) [35]. The multimodal VL-PTMs include fused unimodal ResNet-152 and BERT output scores (Late Fusion) [18], concatenate ResNet-152 features with BERT and training an MLP on top (Concat BERT) [18], supervised multimodal bi-transformers using either Image-Grid or Image-Region feature input (MMBT-Grid and MMBT-Region) [36], ViLBERT [21], Visual BERT [37], ViLBERT trained on Conceptual Captions (ViLBERT CC) [38], and Visual BERT trained on COCO dataset (Visual BERT COCO) [39]. We re-evaluated these eleven models in our study on the same Google Colab platform. Batch sizes were chosen to meet memory constraints. As mentioned before, the training of all models was consecutively repeated four times to select four final models. Afterwards, all these final models, including ours and baselines, were evaluated in the test set, and average values of Acc. and AUROC were compared.

Results

The results are summarized in . Our proposed OSCAR+RF model achieved the best performance at the test set among all models with an average Acc. of 0.684 and an average AUROC of 0.768, followed by OSCAR+FC with an average Acc. of 0.677 and an average AUROC of 0.762 with very slight differences. Besides, we observe that the text-only classifier performs slightly better than the vision-only classifier, and multimodal VL-PTMs performed better than the unimodal models. In multimodal detecting models, adding the tag information (OSCAR+RF and +FC) increases both the Acc. and AUROC.

Discussion

In this study, we show that learning cross-modal representations can be improved by introducing object tags detected in memes as anchor points to ease the understanding of semantic alignments between images and text captions. Our results demonstrate that our multimodal VL-PTMs by intaking object tags are better than previous unimodal and multimodal models without anchor points in previous studies.

The detection accuracy of traditional unimodal deep learning methods, such as convolutional neural networks, was limited due to their consideration of only one modality; so, they could not comprehensively understand the subtlety behind the memes [10,16,18,19]. Our study shows that the unimodal models, including Image-Grid, Image-Region, and Text BERT, demonstrate lower Acc. and AUROC than any multimodal models. However, when comparing different unimodal models, the unimodal linguistic model (Text BERT) shows a better performance than visual unimodal models (Image-Grid and Image-Region) in detecting hateful memes, probably because the image region features are less representative of the meme’s intention than word semantics [24]. In our results, Text BERT has an average AUROC of 0.639, higher than Image-Grid and Image-Region with only visual detecting modality (an average AUROC of 0.514 and 0.561, respectively).

On the other hand, multimodal learning of both text and image generally increases detection accuracy over unimodal learning. As mentioned before, many sentences or images that are harmless by themselves may become hateful when combined, or changing the image part or text part in the meme could quickly reverse the hateful intention. Thus, enhancing the cross-modal unification ability of the models can effectively increase the meme classification accuracy [20-24,37]. In addition, injecting more human knowledge into the model might help identify hateful memes more accurately. Encouragingly, the advent of large-scale PTM offered promising directions. Thanks to the immensity of training data and the massive number of model parameters (e.g., for BERT, the pre-training corpus contains 3,300 million words and the Base-version module has 110 million parameters while the Large-version module contains 340 million parameters), transformer-based VL-PTMs demonstrated higher potentials in complex learning because of the rich knowledge infused in the massive amount of model parameters; moreover, some have even surpassed human performance on multiple language understanding benchmarks, such as GLUE [19,35,40-42]. In a previous study, VL-PTMs were shown to be capable of re-learning from additional knowledge datasets, such as the Conceptual Captions dataset learned by ViLBERT CC and COCO dataset learned by Visual BERT COCO, presenting a potential benefit in the hateful meme detection when compared with the same baselines (ViLBERT and Visual BERT) without being trained those datasets [18]. But our study didn’t find apparent discrepancies between these models. We believe pre-training these models with the more comprehensive caption or image entity datasets in future probably will improve the hateful meme detection accuracy because most hateful memes employ a lot of background knowledge which makes the relations between visual and text elements very complex and diverse [25,43,44].

More importantly, our results revealed that by intaking object tags as anchor points, our VL-PTM can achieve better Acc. and AUROC than conventional convolutional neural networks and previous VL-PTMs. This benefit is probably due to the more explicit representation by object tags which provide the model with clues about the corresponding image and caption features. Our model uses an OD module VinVl to encode and output a diverse tag collection of visual objects [28]. Promisingly, more powerful OD modules and web entity recognition Application Programming interfaces (API) have been developed and released over time. Combining these state-to-the-art tools probably facilitates more accurate object tag formation and, as a result, enhances the Vision-Language semantic learning by the VL-PTMs.

At last, we tried to replace the final FC linear classifier at the end of OSCAR+ with an RF classifier. After fine-tuning, the OSCAR+RF model shows a slight improvement in Acc. and AUROC compared with the OSCAR+FC model on the same test set. However, this finding somewhat supports the speculation that the merits of our model are majorly attributed to the addition of object tags input. The memes that invoked the combined visual and textual cues could be classified more correctly after strengthening the cross-modal association. Moreover, we conjecture that utilizing other more specific object tags, such as ethnic groups, nationalities, and religions, can provide further merits, posing a promising future research direction.

Our detection model also has some limitations. First, hateful messages are evolving quickly, so the model cannot keep its detection accuracy if not re-trained in time. For example, modern online communication heavily employs non-standard features, such as emojis and other irregular tokens such as $; and hateful users often try to evade detection by substituting the characters in their messages with symbols very different in terms of machine encodings yet looking or sounding very similar to human beings. One future research direction is to take advantage of these underutilized visual or audio aspects of the text information in order to adapt to more real-life scenarios. Second, the detection accuracy of our models, or similar models, largely depends on the knowledge base that the VL-PTM trained on. As we showed that detecting underlying hateful metaphors requires the system to possess the ability to relate visual and linguistic entities in the image or captions to very detailed real-world knowledge, we expect future VL-PTMs to be supplemented with fine-grained external knowledge bases, such as Dbpedia [45] and wikidata [46] to achieve better performance on this task. Third, a limited capacity for object recognition in our OD module. So far, most contemporary VL-PTMs only take image and text into account, and the training dataset contains only common objects. However, the hateful memes often invoke unusual and specific objects connected to historical or social events not presented in those training sets. Thus, a sufficient dataset related to social or historical events is also in great demand to train a more knowledgeable OD module.

Conclusion

This study has demonstrated that VL-PTMs with the addition of anchor points can improve the detection of hateful memes that involve strongly correlated text and visual information; our proposed models show a better detection performance than previous unimodal and multimodal baselines without anchor point applications.

Definitions of hateful message (or hate speech).

(DOCX)

Click here for additional data file.

14 Jun 2022

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Liejun Wang, Ph.D.

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. We note that S1 Fig and S2 includes an image of a participant in the study.

As per the PLOS ONE policy (http://journals.plos.org/plosone/s/submission-guidelines#loc-human-subjects-research) on papers that include identifying, or potentially identifying, information, the individual(s) or parent(s)/guardian(s) must be informed of the terms of the PLOS open-access (CC-BY) license and provide specific permission for publication of these details under the terms of this license. Please download the Consent Form for Publication in a PLOS Journal (http://journals.plos.org/plosone/s/file?id=8ce6/plos-consent-form-english.pdf). The signed consent form should not be submitted with the manuscript, but should be securely filed in the individual's case notes. Please amend the methods section and ethics statement of the manuscript to explicitly state that the patient/participant has provided consent for publication: “The individual in this manuscript has given written informed consent (as outlined in PLOS consent form) to publish these case details”.

If you are unable to obtain consent from the subject of the photograph, you will need to remove the figure and any other textual identifying information or case descriptions for this individual.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: This study has demonstrated that Vision-Language PTMs with the addition of anchor points can improve the detection of hateful memes that involve strongly correlated textual and visual information. However, I have the following issues:

1. The motivation of the paper is not clear in Section Introduction. What are the challenges involved? What solutions already exist for the problem you want to solve? What are their limitations and drawbacks? It lacks a critical analysis of these works and a positioning of the proposed solution concerning the state-of-the-art works.

2. It is not clear the contribution of this work for the advance of the state-of-the-art in this domain.

3. The experimental results may be discussed in detail fundamentally by providing critical analysis.

4. Add most recent references at appropriate places. Such as:

https://link.springer.com/article/10.1007/s00521-021-06546-x

https://link.springer.com/article/10.1007/s00521-022-07054-2

Reviewer #2: This paper presents a mode for hateful memes detection by integrating vision-language pre-trained models with random forest classifier. Experiments are conducted on Facebook Hateful Meme Dataset to evaluate the proposed model.

Although the main idea is easy to follow, I have several concerns as following:

- The main idea of this work seems a simple combination of two existing works (OSCAR and random forest classifier).

- Several related works for vision-language pre-training are missing:

[A] "Unified Vision-Language Pre-Training for Image Captioning and VQA." AAAI. 2020.

[B] “Scheduled Sampling in Vision-Language Pretraining with Decoupled Encoder-Decoder Network.” AAAI. 2021.

- In Table 1, the performance in Test Acc. of OSCAR+RF is lower than Visual BERT COCO.

- More ablated runs (e.g., Visual BERT COCO+RF) should be included for performance comparison.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

16 Jul 2022

Response to Reviewers

Editors:

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. We note that S1 Fig and S2 includes an image of a participant in the study.

As per the PLOS ONE policy (http://journals.plos.org/plosone/s/submission-guidelines#loc-human-subjects-research) on papers that include identifying, or potentially identifying, information, the individual(s) or parent(s)/guardian(s) must be informed of the terms of the PLOS open-access (CC-BY) license and provide specific permission for publication of these details under the terms of this license. Please download the Consent Form for Publication in a PLOS Journal (http://journals.plos.org/plosone/s/file?id=8ce6/plos-consent-form-english.pdf). The signed consent form should not be submitted with the manuscript, but should be securely filed in the individual's case notes. Please amend the methods section and ethics statement of the manuscript to explicitly state that the patient/participant has provided consent for publication: “The individual in this manuscript has given written informed consent (as outlined in PLOS consent form) to publish these case details”.

If you are unable to obtain consent from the subject of the photograph, you will need to remove the figure and any other textual identifying information or case descriptions for this individual.

R: We appreciate your notifications on the manuscript style and informed consent of the cited figures. In the modification, we revised our manuscript following your comments and added the agreement of the figure publish.

Reviewer #1:

This study has demonstrated that Vision-Language PTMs with the addition of anchor points can improve the detection of hateful memes that involve strongly correlated textual and visual information. However, I have the following issues:

1. The motivation of the paper is not clear in Section Introduction. What are the challenges involved? What solutions already exist for the problem you want to solve? What are their limitations and drawbacks? It lacks a critical analysis of these works and a positioning of the proposed solution concerning the state-of-the-art works.

2. It is not clear the contribution of this work for the advance of the state-of-the-art in this domain.

3. The experimental results may be discussed in detail fundamentally by providing critical analysis.

4. Add most recent references at appropriate places. Such as:

https://link.springer.com/article/10.1007/s00521-021-06546-x

https://link.springer.com/article/10.1007/s00521-022-07054-2

R: We appreciate your efforts to point out these flaws in the manuscript, and we followed your suggestions to revise our draft. First, in the introduction, we explained the importance and dilemma of hateful meme detection nowadays, including limitations and drawbacks of different techniques. Besides, we highlighted our innovation of adding tags as anchor points which might enhance the capability of hateful meme detection.

Second, the idea of the anchor point in VLP is one of the most critical progresses in the past two years. When we started this project, this idea had not been used in the downstream task of hateful meme detection. We believe our findings could inspire fellow researchers to adopt and improve this thought. At the same time, in the discussion section, we also pointed out some future research directions, such as adding more specific tags, including race, skin color, etc., which we believe can improve the efficacy of the detector.

Third, in the last few weeks, we consecutively retrained our model and other eleven baselines four times on the Google Colab platform for a comprehensive comparison. The batch size was set to the largest value given the memory restriction. Then, to reduce the variance of estimating the performance of each model, all of them were trained and tested four times, and the averages of Acc. and AUROC were taken for comparison. We observed that our multimodal model with anchor points performed better than unimodal or other multimodal models without anchor points. These results support our idea mentioned above.

At last, we appreciate your recommending those two related papers. In revision, we cited them to support our method part.

Reviewer #2:

This paper presents a mode for hateful memes detection by integrating vision-language pre-trained models with random forest classifier. Experiments are conducted on Facebook Hateful Meme Dataset to evaluate the proposed model.

Although the main idea is easy to follow, I have several concerns as following:

- The main idea of this work seems a simple combination of two existing works (OSCAR and random forest classifier).

- Several related works for vision-language pre-training are missing:

[A] "Unified Vision-Language Pre-Training for Image Captioning and VQA." AAAI. 2020.

[B] "Scheduled Sampling in Vision-Language Pretraining with Decoupled Encoder-Decoder Network." AAAI. 2021.

- In Table 1, the performance in Test Acc. of OSCAR+RF is lower than Visual BERT COCO.

- More ablated runs (e.g., Visual BERT COCO+RF) should be included for performance comparison.

R: We'd like to express our sincere gratitude for your review and for providing these valuable comments and suggestions. We revised our manuscript thoroughly and added some essential information about our model. First, our model is based on anOSCAR+ backbone, which differs from other vision-language transformers by taking images, anchor point tags, and text caption triplets as inputs. Although our model can be considered as a combination of an optical character recognition (OCR) module, an object detection (OD) module, a fine-tuned OSCAR+ module, and a random forest (RF) classifier module, this idea and our results are inspiring. Still, there is a substantial space for future improvement on the automatic hateful meme detection using deep-learning techniques. The emergence of multimodal transformer shed light on achieving effective detection of hateful memes owing to its capability in cross-modal context semantic learning. From previous studies, multimodal transformers presented superior performances (Acc. about 60-65%) compared to conventional CNN models (Acc. about 50% - 51%) in this task. However, they are still very far away from human performances (Acc. about 85%). Under this milieu, we noticed that cross-modal correlation (so-called "ambiguity" issue) might limit the performance of multimodal transformers. As a solution, anchor points, one of the most critical ideas in multimodal deep learning in the past two years, is adopted in this study to connect image and text semantics. When we started our project, this model type had not been attempted in meme classification yet. We believe the results demonstrated by our study would provide justification for future research on such methods. In future research, we also plan to go further by employing other techniques, such as adding more specific tags, including race, skin tone, etc. These upgraded tags have the potential to improve model performance. In our revision, we explained this issue in the introduction and discussion.

Second, we appreciate you for providing us with those two papers about unified VL transformers with different designs. In the revision, we discussed applying different VL transformers in hateful meme detection.

Third, about the question that the Acc. of OSCAR+RF in the test set is lower than Visual BERT COCO, we admit that the data in Table 1 is not rigorous. In our previous version, we simply cited the results from the published literature instead of running these models on our own. In this revision, we retrained those baseline models and added another two multimodal models (Late fusion and ConcatBERT) on the same Goggle Colab platform with batch sizes set to the largest possible value given memory limitation. In addition, all were consecutively trained and tested four times and the mean values of their Acc. and AUROC were compared, which helped to reduce the variance of estimating model performance. As a result, we can see Transformer-based Vision-Language Pre-Training Models (VL-PTMs) with anchor point tags inputs performed better than conventional VL PTMs without those tags. However, after multiple training, we find the differences of Acc. and AUROC between OSCAR+FC and OSCAR+RF are very small (actually, in previous manuscript, this gap was also small). So, we speculated that the benefit of our model is majorly ascribed to the addition of the anchor point involvement (object tags) but not the last classifier. Following your comments, we also revised our viewpoint and discussion.

Submitted filename: Response_to_Reviewers_2022.7.17.docx

Click here for additional data file.

26 Aug 2022

Multimodal Detection of Hateful Memes by Applying a Vision-Language Pre-Training Model

PONE-D-22-10554R1

Dear Dr. Pan,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Liejun Wang, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

The reviewer has completed reviewing the manuscript and the authors have addressed the reviewer's concerns. Therefore, I recommend publishing the paper.

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: (No Response)

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: (No Response)

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: No

Reviewer #2: (No Response)

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: (No Response)

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Thank you for the response and revisions. You have successfully addressed the issues raised by the reviewers and the manuscript can be accepted for publication.

Reviewer #2: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

**********

2 Sep 2022

PONE-D-22-10554R1

Multimodal Detection of Hateful Memes by Applying a Vision-Language Pre-Training Model

Dear Dr. Pan:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Liejun Wang

Academic Editor

PLOS ONE

Table 1

Comparisons among our models with other published baselines.

Models	Batch size	Detecting modality	Loss	Optimizer	Learning rate	Acc.,	AUROC,	Acc.,	AUROC,
						dev set (n = 500)*	dev set (n = 500) *	test set (n = 1000) *	test set (n = 1000) *
Image-Grid	32	Image	Cross entropy	AdamW	1.00E-05	0.500±0.045(0.436–0.536)	0.516±0.027(0.478–0.543)	0.511±0.023(0.478–0.526)	0.514±0.018(0.498–0.530)
Image-Region	32	Image	Cross entropy	AdamW	5.00E-05	0.513±0.032(0.484–0.548)	0.549±0.030(0.508–0.579)	0.531±0.023(0.502–0.558)	0.561±0.039(0.526–0.617)
Text BERT	64	Text	Cross entropy	AdamW	5.00E-05	0.569±0.020(0.548–0.588)	0.625±0.047(0.579–0.669)	0.586±0.024(0.556–0.612)	0.639±0.006(0.633–0.645)
Late Fusion	32	Image&Text	Cross entropy	AdamW	5.00E-05	0.589±0.031(0.544–0.612)	0.641±0.040(0.613–0.700)	0.619±0.011(0.608–0.630)	0.679±0.018(0.665–0.705)
ConcatBERT	32	Image&Text	Cross entropy	AdamW	1.00E-05	0.576±0.038(0.540–0.616)	0.645±0.012(0.629–0.655)	0.622±0.023(0.588–0.636)	0.682±0.017(0.659–0.696)
MMBT-Grid	32	Image&Text	Cross entropy	AdamW	1.00E-05	0.603±0.042(0.544–0.644)	0.672±0.018(0.654–0.696)	0.631±0.014(0.616–0.650)	0.694±0.006(0.687–0.700)
MMBT-Region	32	Image&Text	Cross entropy	AdamW	5.00E-05	0.605±0.059(0.524–0.652)	0.649±0.067(0.585–0.722)	0.642±0.032(0.608–0.672)	0.690±0.046(0.646–0.735)
ViLBERT	32	Image&Text	Cross entropy	AdamW	1.00E-05	0.633±0.020(0.612–0.656)	0.717±0.035(0.677–0.747)	0.659±0.007(0.652–0.668)	0.732±0.015(0.716–0.753)
Visual BERT	32	Image&Text	Cross entropy	AdamW	5.00E-05	0.638±0.023(0.612–0.668)	0.722±0.010(0.711–0.732)	0.664±0.013(0.656–0.684)	0.748±0.011(0.732–0.757)
ViLBERT CC	32	Image&Text	Cross entropy	AdamW	1.00E-05	0.656±0.009(0.648–0.668)	0.730±0.035(0.691–0.773)	0.664±0.009(0.652–0.674)	0.739±0.016(0.724–0.757)
Visual BERT COCO	32	Image&Text	Cross entropy	AdamW	5.00E-05	0.648±0.032(0.608–0.676)	0.732±0.017(0.711–0.752)	0.664±0.020(0.646–0.692)	0.737±0.025(0.711–0.770)
OSCAR+FC	50	Image&Tag&Text	Cross entropy	AdamW	5.00E-06	0.666±0.038(0.626–0.706)	0.758±0.042(0.703–0.803)	0.677±0.010(0.664–0.689)	0.762±0.016(0.749–0.786)
OSCAR+RF	50	Image&Tag&Text	Cross entropy	AdamW	5.00E-06	0.667±0.034(0.618–0.698)	0.759±0.014(0.745–0.777)	0.684±0.002(0.682–0.686)	0.768±0.021(0.737–0.784)

Footnotes: Acc., accuracy; AUROC, area under the receiver operating characteristic.

*Mean±standard error with the range was calculated from evaluations of four final models.