PaperlayerView exact revision on arXiv

Source-linked AI summary

Pro-Cap: Leveraging a Frozen Vision-Language Model for Hateful Meme Detection

Rui Cao, Ming Shan Hee, Adriel Kuek, Wen-Haw Chong, Roy Ka-Wei Lee, Jing Jiang

arXiv:2308.08088v1cs.CVcs.IRcs.MM

TL;DR

Hateful meme detection must account for both image-text content and their interactions, while generic image captions can omit details relevant to hateful content and fine-tuning larger PVLMs is less feasible. Pro-Cap uses a frozen PVLM to answer hateful-content-related questions and turns the answers into captions. Models using Pro-Cap improve on three benchmarks, and the paper reports a new state of the art on them.

  • Problem

    Hateful meme detection requires understanding both image and text and their interactions, while generic image captions may omit details needed for the task.

    Source: §1 · p. 2

  • Method

    Pro-Cap uses a frozen PVLM in zero-shot VQA, asking hateful-content-related questions and using its answers as image captions for a trainable detector.

    Source: §1 · p. 2

  • Results

    Pro-CapPromptHate improves over PromptHate without additional image tags by about 4, 6, and 3 percentage points on FHM, MAMI, and HarM, respectively, and the paper reports a new state of the art on three benchmarks.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    Probing-based captioning is reported as a more affordable way to obtain image entities or demographic information than adding image tags, while retaining comparable results to PromptHate with those tags.

    Source: §1 · p. 2

    Probe-caption answers can contain minor errors, and the authors note that current language models may perform poorly on complex reasoning.

    Source: §5.5 · p. 8

Abstract

from arXiv · show

Hateful meme detection is a challenging multimodal task that requires comprehension of both vision and language, as well as cross-modal interactions. Recent studies have tried to fine-tune pre-trained vision-language models (PVLMs) for this task. However, with increasing model sizes, it becomes important to leverage powerful PVLMs more efficiently, rather than simply fine-tuning them. Recently, researchers have attempted to convert meme images into textual captions and prompt language models for predictions. This approach has shown good performance but suffers from non-informative image captions. Considering the two factors mentioned above, we propose a probing-based captioning approach to leverage PVLMs in a zero-shot visual question answering (VQA) manner. Specifically, we prompt a frozen PVLM by asking hateful content-related questions and use the answers as image captions (which we call Pro-Cap), so that the captions contain information critical for hateful content detection. The good performance of models with Pro-Cap on three benchmarks validates the effectiveness and generalization of the proposed method.

1 INTRODUCTION

Pro-Cap uses zero-shot questions to make a frozen vision-language model produce hateful-content-relevant captions, addressing the limits of generic captions and the cost of fine-tuning larger models. On three benchmarks, models using Pro-Cap improve over PromptHate without additional image tags.

1 INTRODUCTION

  • Hateful memes target people or communities based on identities such as race, gender, or religion, and their spread may contribute to online discord and potentially hate crimes.Source: §1 · p. 1
  • Hateful meme detection requires modeling image-text interactions, which can be difficult to learn from the limited data available in detection datasets.Fine-tuning larger PVLMs is less feasible because they may have billions of trainable parameters.Source: §1 · pp. 1–2
  • PromptHate converts memes to text using generic image captions, which can omit details such as race and gender that matter for hateful-content detection.Adding entity and demographic tags can improve the model, but generating them may be laborious and costly.Source: §1 · p. 2
  • Pro-Cap queries a frozen PVLM with probing questions about hateful-content-related information and uses its answers as image captions for a trainable detector.The paper calls this process probing-based captioning.Source: §1 · p. 2
  • The method uses a PVLM without adaptation or fine-tuning and generates informative captions instead of relying on costly external image-tagging APIs.The authors describe this as a zero-shot question-answering use of PVLMs for hateful meme detection.Source: §1 · p. 2
  • Pro-CapPromptHate improves over PromptHate without additional image tags by about 4, 6, and 3 percentage points on FHM, MAMI, and HarM, respectively.The paper also reports comparable results to PromptHate with additional image tags and better generalization with Pro-CapBERT.Source: §1 · p. 2

2 RELATED WORK

Prior work addresses hateful meme detection through multimodal classification and through text-based prompting. The text-based approach has achieved state-of-the-art results on two benchmarks, but generic image captions can miss details relevant to hateful content.

2 RELATED WORK

  • One line of work fine-tunes PVLMs on meme-detection data, sometimes using ensembles to improve performance.Source: §2 · p. 2
  • Other approaches combine models such as BERT and CLIP with task-specific architectures and tune them end-to-end.Source: §2 · p. 2
  • PromptHate converts meme information into text and prompts a language model, but its generic image captions can omit factors important to hateful meme detection.The paper addresses this caption limitation with probe-based captioning using hateful-content-centric questions.Source: §2 · p. 2

3 PRELIMINARY

The paper defines hateful meme detection as classifying a meme from its image and accompanying text, and describes zero-shot VQA as a way to generate relevant captions. Its experiments use BLIP-2 as the PVLM.

3 PRELIMINARY

  • Given an image and accompanying text, the detector scores the meme as non-hateful or hateful and chooses the label with the higher score.Source: §3 · p. 3
  • For zero-shot VQA, a PVLM receives an image and a prompt formatted as a question, then generates a sequence of answer tokens.The passage illustrates the process with a question about a person's race in an image.Source: §3 · p. 3
  • The experiments use BLIP-2, which combines frozen image and language models with a lightweight Querying Transformer to bridge the modalities.The paper states that another PVLM capable of zero-shot VQA could replace BLIP-2.Source: §3 · p. 3

4 PROPOSED METHOD

The method generates Pro-Cap by querying a PVLM with zero-shot VQA questions designed to elicit image content and hateful-target details. It combines these answers with meme text for BERT- or PromptHate-based classification.

4.1 Overview

  • Zero-shot VQA produces Pro-Cap from PVLM answers, which are combined with meme text and supplied to either BERT or PromptHate for hateful-meme detection.The paper selects zero-shot VQA because it lacks training data for the intended questions and prior work shows promising performance.Source: §4.1 · p. 3

4.2 Design of VQA Questions

  • Questions cover generic image content, five vulnerable-target categories, and animal presence; person- and animal-existence checks gate whether related answers enter Pro-Cap.The target categories are religion, race, gender, nationality, and disability.Source: §4.2 · p. 3, Table 2 · p. 4

4.3 BERT-based Detection Model

  • BERT encodes concatenated meme text T and Pro-Cap C into representation r, then uses a linear layer to classify the meme.The linear layer uses learnable parameters W and b.Source: §4.3 · pp. 3–4

4.4 PromptHate for Hateful Meme Detection

  • PromptHate concatenates meme text, Pro-Cap, and a prompt template, then scores masked-word label candidates using a language model with positive and negative examples.The method replaces PromptHate's ClipCap captions with Pro-Cap.Source: §4.4 · p. 4

4.5 Model Training and Prediction

  • Training applies binary cross-entropy to the two-class output; prediction selects non-hateful when s0 exceeds s1 and hateful otherwise.The ground-truth vector marks the non-hateful class as [1, 0] and the hateful class as [0, 1].Source: §4.5 · p. 4

5 EXPERIMENT

Across three benchmarks, Pro-Cap models compare favorably with baselines, including when image tags are added. Ablations and case studies examine question content, answer length, and errors.

5.1 Experiment Settings

  • Evaluation covers FHM, MAMI, and HarM; FHM results use dev-seen because test labels are unavailable, and HarM merges partially harmful with very harmful memes.MAMI focuses on memes targeting women, while HarM contains COVID-19-related memes and is used to test generalization from hateful to harmful meme detection.Source: §5.1 · pp. 4–5

5.2 Baselines

  • Baselines include text-only and image-only models, general multimodal models, and systems designed specifically for hateful meme detection.The comparisons include Text-BERT, Image-Region, VisualBERT, ViLBERT, ALBEF, BLIP, MOMENTA, DisMultiHate, and PromptHate.Source: §5.2 · p. 5

5.3 Experiment Results

  • Without augmented tags, Pro-CapBERT outperforms multimodal pre-trained BERT models on all three datasets and is comparable with BLIP and ALBEF, while Pro-CapPromptHate shows advantages over those similarly sized models.Pro-CapBERT surpasses ViLBERT by over 4% on FHM and MMBT-Region by about 3% on HarM; it has fewer parameters than ViLBERT, and is three times smaller than BLIP and ALBEF in trainable parameters.Source: §5.3 · pp. 5–6, Table 4 · p. 6
  • With augmented image information, Pro-CapPromptHate achieves the best performance on all three benchmarks, while Pro-CapBERT remains competitive with larger multimodal models and surpasses them on HarM.Pro-CapPromptHate without tags is comparable to PromptHate with tags on FHM and HarM, and surpasses it on MAMI.Source: Table 5 · p. 6, §5.3 · p. 7

5.4 Ablation Study

  • Target-specific probing questions improve performance over generic image captions on all three datasets, with gains over 2% on FHM and over 3% on HarM but only minor improvement on MAMI.The authors suggest generic captions may already cover gender in MAMI, whereas the other datasets involve greater complexity.Source: §5.4 · p. 7
  • Detection performance is robust across answer lengths, although HarM favors longer answers and MAMI favors shorter ones to a small extent.The ablation varies BLIP-2’s text-decoder length penalty across 1, 2, and 3.Source: §5.4 · p. 8

5.5 Case Study

  • Case studies show that target-specific probes can surface race, country, and religion cues that generic captions miss.Error analysis also reports complex-reasoning failures despite sufficient captions, occasional incorrect VQA answers, and possible model bias toward Muslim-related content.Source: §5.5 · p. 8
  • In one error case, sufficient probe-captions do not prevent prediction failure, which the authors suggest may reflect limited complex reasoning or small hateful-meme training datasets.They also note occasional incorrect zero-shot VQA answers and possible bias in hateful-content predictions.Source: §5.5 · p. 8

6 CONCLUSION

The paper probes a frozen PVLM with hateful-content questions and uses its answers as image captions. It reports strong benchmark performance while identifying question selection and interpretation as open limitations.

6 CONCLUSION

  • Pro-CapPromptHate achieves new state-of-the-art results on three benchmarks.The authors also report that BERT with distilled knowledge from large PVLMs surpasses multimodal pre-trained BERT models of similar scale.Source: §6 · p. 8
  • Zero-shot VQA with a frozen PVLM generates hateful-content-centric image captions without fine-tuning the PVLM.The answers to probing questions are used as Pro-Cap.Source: §6 · p. 8
  • Using answers to all probing questions may be suboptimal because some questions may be irrelevant to the meme target.The authors suggest dynamically selecting questions relevant to meme detection as future work.Source: §6 · p. 8
  • The paper calls for more thorough analysis of how probing questions influence final predictions.A gradient-based interpretation approach is offered as one possible future direction.Source: §6 · p. 8

APPENDIX

The appendix includes a comparison of Pro-CapPromptHate and basic PromptHate on the HarM dataset.

APPENDIX

  • Table 9 compares Pro-CapPromptHate with basic PromptHate on the HarM dataset.The supplied passage identifies the comparison's scope but does not state its results.Source: Table 9 · p. 10

A DETAILS FOR IMPLEMENTATION

The implementation uses PyTorch with CUDA 11.2 and reports model-specific code sources and sequence-length and training settings. The appendix also identifies tables covering model parameters and comparisons without augmented image tags or with a single probing question.

A DETAILS FOR IMPLEMENTATION

  • Implementations use PyTorch with CUDA 11.2 and Tesla V100 GPUs with 32 GB of memory.Hateful-meme-specific models use authors' code; Hugging Face packages are used for BERT, VisualBERT, and BLIP, and author-released code for ViLBERT.Source: A DETAILS FOR IMPLEMENTATION · p. 10
  • Tables 12 and 13 cover comparisons without augmented image tags and performance with a single probing question, respectively.The supplied captions identify the comparisons but do not provide their results.Source: Table 12 · p. 11, Table 13 · p. 11
  • ALBEF and BLIP-2 are implemented using packages from the LAVIS library.The passage identifies the implementation source, not experimental outcomes.Source: A DETAILS FOR IMPLEMENTATION · p. 11
  • Meme text and the generic image caption are limited to a combined length of 65, while each additional question is limited to fewer than 20 tokens.Long concatenations are truncated, shorter ones padded, and all models are trained for 10 epochs.Source: A DETAILS FOR IMPLEMENTATION · p. 11
  • Table 11 summarizes model parameter counts.The supplied passage does not provide the counts.Source: A DETAILS FOR IMPLEMENTATION · p. 11

B FULL ABLATION STUDY RESULTS

The full ablation results are reported in Table 12, while Table 6 presents accuracy-only results.

B FULL ABLATION STUDY RESULTS

  • Table 6 reports accuracy-only ablation results; Table 12 provides the full results, including AUC and accuracy.The passage does not state any ablation outcomes.Source: B FULL ABLATION STUDY RESULTS · p. 11

C VISUALIZATION CASES

The section provides additional Pro-CapPromptHate versus basic PromptHate visualization cases for HarM and MAMI, while omitting examples from two other datasets due to space constraints.

D RESULTS WITH PRO-CAP ABOUT ONE TARGET

Results with answers from individual probing questions show that some single-question models surpass using all questions. The findings also suggest that selecting questions relevant to each meme may avoid irrelevant descriptions.

Related papers

Loading 2308.08088v1…