Long-Form Answers to Visual Questions from Blind and Low Vision People

Vision language models can now generate long-form answers to questions about images - long-form visual question answers (LFVQA). We contribute VizWiz-LF, a dataset of long-form answers to visual questions posed by blind and low vision (BLV) users. VizWiz-LF contains 4.2k long-form answers to 600 visual questions, collected from human expert describers and six VQA models. We develop and annotate functional roles of sentences of LFVQA and demonstrate that long-form answers contain information beyond the question answer such as explanations and suggestions. We further conduct automatic and human evaluations with BLV and sighted people to evaluate long-form answers. BLV people perceive both human-written and generated long-form answers to be plausible, but generated answers often hallucinate incorrect visual details, especially for unanswerable visual questions (e.g., blurry or irrelevant images). To reduce hallucinations, we evaluate the ability of VQA models to abstain from answering unanswerable questions across multiple prompting strategies.

Paper

References (82)

Scroll for more · 38 remaining

Similar papers

Reviewer sGq55/10 · confidence 4/52024-04-12

Summary

VizWiz-LF presents a unique dataset of long-form answers from blind and low vision (BLV) users, shedding light on the discourse structure of such responses. It also highlights a discrepancy between model-generated and human-perceived correctness of answers to unanswerable questions. Additionally, the evaluation of VQA models provides insights for improving their performance and mitigating the risk of providing misleading answers.

Rating

5

Confidence

4

Ethics flag

1

Reasons to accept

- VizWiz-LF presents a unique dataset of long-form answers from BLV users, enriching the field of visual question answering. - The evaluation of VQA models on their ability to abstain from answering unanswerable questions provides valuable insights for model development and deployment, aiming to mitigate the risk of providing misleading answers.

Reasons to reject

- poor presentation: This article is written with poor logic. For example, why we should understand how humans evaluate long-form answers, and there is no obvious relation between paragraph 5 and paragraph 6. - The author needs to further clarify the difference between the BLV and traditional VQA dataset with long/short answers. - Is any error in this sentence ‘’Vision language models can now generate long-form answers to questions about images – or, long-form visual question answers (LFVQA).’’ - What is the benefit of understanding the functional structure of long-form answers? - Long-from answers from large models (such as GPT-4V) could introduce hallucinations and are not reliable. - Human evaluation samples are limited. - Human evaluation may not be reliable because people are subjective. Could authors provide the annotation detail (such as instructions for human evaluation)? - More human evaluation samples should be provided .

Reviewer PViS8/10 · confidence 4/52024-04-29

Summary

The paper introduces VizWiz-LF, a new long-form visual question answering (VQA) dataset that focuses on visual questions posed by blind and low-vision (BLV) people. The datasets consists of 600 image-question pairs sampled from the VizWiz dataset. To ensure diversity of questions, the data is sampled from 4 question types, i.e., Identification, Description, Reading, and Others, which is based on the question taxonomy of VizWiz. Among the 600 image-question pairs, 230 are marked as unanswerable. For each image-question pair, 3 long-form answers are constructed -- one by human expert, one by GPT-4V, and the other by Gemini. The resulting answers from all annotations are more than 10x longer than the original VizWiz dataset. Each sentence in the long-form answers is annotated with functional roles (confirmation, answer, answer failure, auxiliary information, auxiliary information failure, explanation, suggestion, and miscelleneous) and information type (image content, image quality, external information). To construct the annotation, 180 long-form answers are annotated by human through a three-way annotations achieving a high fleiss kappa score (0.76 for functional roles and 0.87 for information types). The rest of the data is annotated by GPT-4 through few-shot in-context prompt, which achieves F1-score of 0.74 compared to human F1-score of 0.79). Using the generated data, the paper demonstrates that both human expert and GPT-4V produce diverse functional roles, while Gemini often onlty cover answer roles since the long-answer often only consists of a single sentence. Additionally, Gemini also rarely abstain to answer even when the question is unanswerable. Furthermore, the paper also shows that model-based approaches, i.e., Gemini and GPT-4V, contains a higher percentage of external information which does not appear on the image. Human evaluation is also conducted to measure the human preferences of the VizWizL-LF considering 4 factors including relevance. helpfulness, plausibility, and fluency. The human evaluation is conducted by 20 BLV and 20 sighted people. All evaluator groups favoured GPT-4V and human expert annotations over Gemini. While sighted people give low helpfulness score for Gemini due to incorrect information provided as an answer, BLV group rated high plausability for all annotations. An additional automatic evaluation using ROUGE, METEOR, and BERTScore is also conducted further showing the effectiveness of GPT-4V for generating long-form VQA. The paper further explore on abstention strategy which helps to the model to abstain from answering when the question is unanswerable.

Rating

8

Confidence

4

Ethics flag

1

Reasons to accept

- The paper introduces a simple method to extend the existing VQA dataset into a long-form VQA which is beneficial for blind and low-vision people - The paper introduces a new long-form VQA dataset which is annotated with per sentence functional roles and information type, and showcases its applicability for analyzing the behaviour of synthetic data generated by LLMs. - The paper provides an extensive analysis of the quality of the generated data through both human and automatic evaluations, and provides a clear in-depth analysis of the results - The paper showcases an experiment on abstention to improve the ability of LLMs to abstain from answering unanswerable questions. - The paper provides clear, easy-to-follow, and detailed enough writing (except for the result tables) on each section with complete detailed information in the appendix.

Reasons to reject

- The results tables are difficult to follow with too many numbers and have no clear highlight of which part readers should focus on - Section 6 feels rather detached from the other contents of the paper - It is not unclear how long-form VQA and VizWiz-LF can be beneficial for blind and low-vision people since the human evaluation is only conducted on long-form answers with no comparison on short-answer VQA.

Questions to authors

- Is there any human evaluation conducted to sighted and BLV people? - The column gender and age should be swapped in Table 16 on Appendix A.3.1

Reviewer 1BBk7/10 · confidence 4/52024-05-25

Summary

This paper presents VizWiz-LF, a VQA dataset consisting of long-form answers to visual questions posed by blind and low vision (BLV) users. VizWiz-LF contains 1.8k long-form answers with a total of 5.2k sentences collected from human expert describers and two vision language models (GPT-4V and Gemini). The authors show that long-form answers contain information beyond the answer such as explanations and suggestions, and also conduct a human evaluation of the long-form answers with BLV and sighted people. Overall, the paper presents a useful benchmark, and has insightful findings. [Update: Thank you for the clarification! My concerns have been sufficiently addressed by the author rebuttal, and I'm increasing my score.]

Rating

7

Confidence

4

Ethics flag

1

Reasons to accept

* This paper tackles an important problem of evaluating long-form VQA, which is underlooked in the literature. * The paper is well-presented and clearly written. * The analysis of unanswerability and VQA model abstention, in particular, is novel and important. * The findings that a human expert and GPT-4V’s answers contain different functional roles is interesting, and I thought that investigating the communicative goal that each sentence plays within a long-form answer is a much-needed perspective on VQA.

Reasons to reject

* It isn’t clear to me why answers in the dataset were elicited from GPT-4V and Gemini as well as human experts? Using GPT-4V and Gemini answers as ground-truths could potentially introduce some biases and could lead to quality issues in the dataset. * Related to this concern, is there only one human-elicited long form answer per visual question? My concern is that this will make it difficult to evaluate other models with only one ground-truth for comparison. * 600 questions within the dataset is on the smaller side, although I believe this shortcoming is made up by the detailed, in-depth qualitative analysis and human studies. * Only two closed-source models (Gemini and GPT-4V) were evaluated on the dataset, and the six open-source models were evaluated only on the task of abstention, and not on generating long-form answers.

Questions to authors

* Are the GPT-4V and Gemini answers intended to be used to evaluate other models on this dataset? If so, why not just compare model outputs to human ground-truth responses? * Is there a reason why the six open-source models were not evaluated on the long-form generative VQA task?

Reviewer LFVf6/10 · confidence 3/52024-05-30

Summary

The paper presents a dataset, called VizWiz-LF, which is an extension of the VizWiz dataset with long form answers. VizWiz-LF consists of 1.8k long-form answers with a total of 5.2k sentences collected from human expert describers and two vision language models (GPT-4V and Gemini). The analysis suggested that long form answers contain explanations and suggestions which could be useful for BLV and sighted people. The paper evaluated six VQA models to check the ability of abstaining from answering unanswerable questions.

Rating

6

Confidence

3

Ethics flag

1

Reasons to accept

- The paper is on a very important topic. The paper is well motivated and the organization of the paper is pretty good. - Experiments are quite comprehensive. The analysis and human evaluation are very helpful. It will help future works.

Reasons to reject

This is a straight-forward dataset paper that extends a previously proposed dataset. Some of the annotations are model based, therefore, there may be some noise/inaccuracies, however, the human evaluation shows such chances are quite low. Overall, this is a good work with a very limited novelty.

Questions to authors

The reliability of the dataset largely depends on the accuracy of the task performance of the crowd workers. I may have missed but does the paper provide details about the crowd workers? What are their background, how they are compensated, and how much time they took to finish their tasks?

Reviewer PViS2024-06-04

Thank you for addressing all the questions and comments. Upon considering all the reviews and authors' responses, I have decided to keep my current score.

Reviewer LFVf2024-06-04

Thank you for the rebuttal

Thank you for addressing the comments. After careful consideration, I have decided to retain my score.

Reviewer 1BBk2024-06-04

Reply to Author R

Thank you for the clarification! My concerns have been addressed, and I'm increasing my score.

Reviewer sGq52024-06-04

Thanks for your response

Some of my concerns are addressed, but other concerns still are not addressed (such as hallucinations). Therefore, I tend to improve my score mildly.

Program Chairsdecision2024-07-10

Decision

Accept

© 2026 NYSGPT2525 LLC