NCA-GENM Question 102
Select 3You are tasked with evaluating the quality of test data in a multimodal model that processes both text and image inputs. During an initial review, you notice inconsistencies between the modalities, such as text descriptions not matching the associated images. Which actions would help ensure data quality and consistency in this multimodal setting?
- A
Standardize text descriptions to ensure they accurately describe the associated images.
- B
Remove all text-image pairs where the modalities do not perfectly align.
- C
Introduce a validation step to cross-check alignment between text and image modalities.
- D
Use synthetic data to replace inconsistent pairs without further validation.
- E
Manually review a subset of the test data to identify patterns of misalignment.
Show answer and explanation
Correct answers: A, C, E
Explanation
Ensuring data quality and consistency in a multimodal setting requires a combination of corrective actions, such as standardizing descriptions, validating alignment, and manual inspection. These steps address potential issues in both modalities and improve the reliability of test results. Simply removing mismatched pairs or relying on unvalidated synthetic data could lead to inaccurate test evaluations or introduce new inconsistencies.
- A. Correct.
Standardizing text descriptions ensures alignment with the images and improves the quality of multimodal inputs, making the data more consistent for testing.
- B. Incorrect.
Removing all mismatched pairs may lead to an unbalanced dataset or insufficient test data, which is not always a practical or effective solution.
- C. Correct.
Introducing a validation step to verify alignment between text and images is a crucial step to maintain test data quality in a multimodal setting.
- D. Incorrect.
Replacing data with synthetic pairs without validation could introduce additional errors and may not address the root causes of misalignment.
- E. Correct.
Manually reviewing a subset of the data helps detect misalignment patterns and inform further corrective actions, ensuring high data quality.