
Guidelines:
- Ensure the visual question is specific—it should focus on aspects such as environment, object attributes, facial expressions, or clothing.
- Avoid vague questions like “What is the significant object?"
- Avoid well-known facts such as “What is the color of Superman’s suit?"
- Avoid overly obvious questions like “What is the doctor wearing?" if the answer is trivial.
- Instead, ask detailed and challenging questions like:
“What color is the jacket he is wearing?"
“How is the character’s facial expression?"
“What objects are placed on the table?"
“What is the lighting condition in the background?"
Output:
When the event that causes <EFFECT EVENT> occurs, …
Step 4:
You are given a question that asks about Event B, which leads to Event A, meaning it contains references to two events.
Your task is to rewrite the question so that it reads naturally while ensuring the following modifications:
- Replace any mention of “<REFERENCE>" in the question with its corresponding bridge entity: “<BRIDGEENTITY>".
- Ensure the rewritten question is grammatically correct and sounds natural.
- Maintain the focus on Event B (do not rewrite the question to ask only about Event A).
Input: Question: <question>
Output: Rewrite question:
Figure 14: The prompt for generating image description 2-needle questions (continued).
Image Description 2-Needle Question Options Generation
(Gemini-1.5-pro-002)
You will be given three images along with a question that has a correct answer.
Your task is to generate three additional challenging answer choices.
How to Generate Challenging Options:
- If the correct answer involves attributes like color, pattern, shape, or emotion, create alternative choices with different values within the same category.
- Increase difficulty by using existing objects that appear in the images.
Output format:
<Question>
A. <Option A> B. <Option B> C. <Option C> D. <Option D>
Correct Answer: <Letter>
Input:
Question and its correct answer: <Question>
Output:
Figure 15: The prompt for generating options for image description 2-needle questions.
Question Quality Evaluation Prompts
We present the prompts used for evaluating the question quality (Bridge Entity, Semantic Coherence, and Readability) in Fig. 16
Question Quality Validation Prompts
Factor 1
You are given two causally related sentences and a bridge entity.
You need to finish the following task:
determine whether this bridge entity is a piece of shared information that establishes the causal relationship between the two sentences. You should answer ‘yes’ or ‘no’.
Cause: <Cause>
Effect: <Effect>
Bridge Entity: <Bridge Entity>
.
Factor 2
You are given a bridge entity and its vague reference.
You need to finish the following task:
determine whether this vague reference is more ambiguous than the original bridge entity but still preserves its core meaning. You should answer ‘yes’ or ‘no’.
Bridge Entity: <Bridge Entity>
Vague Reference: <Vague Reference>
.
Factor 3
You are given two causally related sentences and a question.
Please evaluate the semantic coherence of the question w.r.t the sentences —
that is, whether the question matches the facts and events described in both sentences
and does not introduce unrelated or contradictory content.
You can only output a score from 1 to 5, where 5 = excellent semantic coherence.
Cause: <Cause>
Effect: <Effect>
Question: <Question>
.
Factor 4
You will be given a question.
Please rate its readability on a scale from 1 to 5, where 5 means very readable and 1 means very not readable.
Consider the following criteria: Naturalness; Grammar; Clarity.
You can only output one number.
Question: <Question>
Score:
Figure 16: The prompt used for evaluating the quality of generated questions.
Cause-Effect Uniqueness Test Prompts
We present the prompts used to test the uniqueness of the generated questions with respect to multiple candidate cause sentences in Fig. 17
Model Performance Evaluation Prompts
We present the task instruction prompts us