PaperlayerView exact revision on arXiv

Source-linked AI summary

Visually Dehallucinative Instruction Generation

Sungguk Cha, Jusung Lee, Younghyun Lee, Cheoljong Yang

arXiv:2402.08348v1cs.CV

TL;DR

Repurposing VQA and image-captioning datasets can alleviate data scarcity, but reformatting them yields only one- or two-word answers, limiting their use. CAP2QA uses GPT with image-aligned captions, prompts that constrain QA to caption content, and filtering of inappropriate artifacts. CAP2QA-COCO reduced CHAIRs from 76% to 9% and increased Recall from 21% to 32%, with reported gains in zero-shot VQA and expressiveness.

  • Problem

    Repurposing VQA and image-captioning datasets can alleviate data scarcity, but reformatting them yields only one- or two-word answers, limiting their use.

    Source: §1 · p. 1

  • Method

    CAP2QA uses GPT with image-aligned captions, prompts that constrain QA to caption content, and filtering of inappropriate artifacts.

    Source: §3 · p. 3

  • Results

    CAP2QA-COCO reduced CHAIRs from 76% to 9% and increased Recall from 21% to 32%, with reported gains in zero-shot VQA and expressiveness.

    Source: §4.3 · pp. 3–4, §5 · p. 4

  • Takeaways & Limitations

    Experiments with the same data source support the effectiveness of image-alignness for reducing hallucination and improving zero-shot VQA and expressiveness.

    Source: §5 · p. 4

    Using web-scale datasets remains future work and requires careful text selection to account for caption noise and intended use.

    Source: §5 · p. 4

Abstract

from arXiv · show

In recent years, synthetic visual instructions by generative language model have demonstrated plausible text generation performance on the visual question-answering tasks. However, challenges persist in the hallucination of generative language models, i.e., the generated image-text data contains unintended contents. This paper presents a novel and scalable method for generating visually dehallucinative instructions, dubbed CAP2QA, that constrains the scope to only image contents. Our key contributions lie in introducing image-aligned instructive QA dataset CAP2QA-COCO and its scalable recipe. In our experiments, we compare synthetic visual instruction datasets that share the same source data by visual instruction tuning and conduct general visual recognition tasks. It shows that our proposed method significantly reduces visual hallucination while consistently improving visual recognition ability and expressiveness.

1. INTRODUCTION

Visual instruction tuning depends on multimodal data, but existing QA data often has short answers and GPT-assisted generation can be inaccurate or hallucinative. The paper proposes image-aligned instruction generation and CAP2QA-COCO, reporting reduced hallucination and improved recognition and expressiveness.

1. INTRODUCTION

  • Multimodal datasets are important for VLM alignment, while repurposed QA data often offers only one- or two-word answers and GPT-assisted generation has been criticized as inaccurate and hallucinative.The short answers limit use for descriptive and complex answering tasks.Source: §1 · pp. 1–2
  • CAP2QA is presented as a scalable method that uses image-aligning prompts and filtering to generate image-aligned visual instructions.The paper also introduces CAP2QA-COCO, created from COCO-caption data.Source: §1 · p. 2
  • The authors report reduced visual hallucination and improved zero-shot visual recognition and expressiveness with their method.These are the reported experimental outcomes on object recognition and general recognition tasks.Source: §1 · p. 2
  • CAP2QA-COCO is released as a large-scale multimodal dataset produced with the proposed method and COCO-caption data.The paper identifies the dataset as a separate contribution.Source: §1 · p. 2

2. RELATED WORK

VQA datasets evolved from image-grounded recognition questions to tasks that may require external knowledge, while answer generation shifted from short labels toward sentences. The review identifies a gap in datasets for sentence-level QA grounded in image content.

Evolution of VQA datasets

  • Early VQA focused on image-grounded recognition, whereas OKVQA includes questions requiring external knowledge and reasoning beyond image content.The passage describes the latter as lacking image-alignment.Source: §2 · p. 2

Transition from word-level to sentence-level answers

  • VQA initially favored one-word answers, while LLM-enabled models made sentence-level answers plausible and underscored the need for visual instruction data.InstructBLIP reformatted datasets into instruction-style questions but retained concise answers; LLaVA generated sentence-level answers using image annotations and GPT assistance.Source: §2 · p. 2
  • The review identifies a lack of datasets specifically designed for image-aligned instructive QA, in contrast to LLaVA's image-unaligned instructive QA.Instructive QA is described as using sentence-level answers.Source: §2 · p. 2

3. CAP2QA

CAP2QA uses text-only GPT with image-aligned captions to generate question-answer pairs constrained to caption content. Its prompts prioritize answerability and appropriate QA, and a filtering step removes artifacts such as references to the caption itself.

3. CAP2QA

  • CAP2QA repeats generation and filtering up to a retry limit, stopping when at least one QA remains.The algorithm defines this loop using GPT, a filtering function, and a retry integer.Source: §3 · p. 2
  • CAP2QA supplies image-aligned captions to text-only GPT-3.5-turbo or GPT-4 to generate question-answer pairs.The framework uses prompts with bottom rules, a task description, and conditions.Source: §3 · p. 2
  • Its bottom rules constrain generation to the given context and discourage rhetorical text, while the task description prioritizes questions answerable from the caption.The prompts also tell GPT to avoid excuses such as saying the answer is not specified in the caption.Source: §3 · pp. 2–3
  • Condition prompts discourage unreasonable or answer-implying questions and exclude QA that cannot validly be answered from the caption.The framework then filters inappropriate artifacts, including QA that mention the caption itself.Source: §3 · p. 3

4. EXPERIMENTAL RESULTS

On shared-source visual instruction data, CAP2QA-COCO reduces object hallucination and improves zero-shot visual recognition and caption expressiveness. The evaluation compares tuned InstructBLIP models across hallucination, VQA, and captioning tasks.

4. EXPERIMENTAL RESULTS

  • The comparison fine-tunes the same pretrained InstructBLIP model with LLaVA or CAP2QA-COCO under the same training protocol.The authors also vary language-model scale to examine its relationship with recognition performance and conduct ablations.Source: §4 · p. 3, §4.1 · p. 3

4.2. Evaluation

  • Evaluation covers sentence-level object hallucination with CHAIRs, visual recognition with VQAv2, GQA, and OKVQA, and caption expressiveness with BLEU and CIDEr.CHAIRs is calculated by checking whether object-recognition answers contain nonexistent objects; the VQA tasks use accuracy measures.Source: §4.2 · p. 3

4.3. Results

  • CHAIRs is 9% vs 76%, and object-recognition Recall is 32% vs 21%, for CAP2QA-COCO versus the compared instruction data.The evaluation uses InstructBLIP (ViT-G, OPT-1.3B) on the COCO validation set; Recallw/oH indicates recall excluding hallucinated answers.Source: §4.3 · p. 3, Table 2 · p. 3
  • The worst CAP2QA-COCO model at 1.3B outperforms the best LLaVA model at 6.7B on general visual-recognition benchmarks.The reported zero-shot VQA benchmarks include VQAv2, GQA, and OKVQA.Source: Table 3 · p. 4, §4.3 · p. 4
  • In qualitative examples, red text marks irrelevant or incorrect responses, while blue text shows CAP2QA-COCO responses described as appropriate.The results include comparisons of CAP2QA-COCO and LLaVA, which share MSCOCO as their source data.Source: §4 · p. 3, §4.3 · p. 4
  • CAP2QA-COCO preserves source knowledge more, while LLaVA tends to generate content irrelevant to the source in image captioning.CAP2QA-COCO scores decrease slightly with language-model size, while LLaVA-tuned scores increase.Source: §4.3 · p. 4

5. CONCLUSION

The conclusion reports that CAP2QA advances visually dehallucinative instruction generation, with reduced hallucination and improved zero-shot VQA and expressiveness. It identifies careful text selection as a consideration for future use of web-scale data.

5. CONCLUSION

  • The authors report that image-aligned instruction generation reduces visual hallucination and improves zero-shot VQA and expressiveness.They identify exploring web-scale datasets as future work, with careful selection considering caption noise and intended use.Source: §5 · p. 4

Related papers

Loading 2402.08348v1…