Unstructured data quality assessment and cleaning method fusing multi-modal large model

By combining the FLAVA model and related algorithms, cross-modal feature extraction and adaptive cleaning of multimodal unstructured data were achieved, solving the problems of accuracy and adaptability in multimodal data quality assessment and cleaning, and improving the integrity and reliability of data processing.

CN122364671APending Publication Date: 2026-07-10
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Filing Date
2026-04-15
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing technologies struggle to comprehensively assess and clean multimodal unstructured data, particularly in accurately identifying semantic consistency and content duplication among text, image, audio, and video data, and the cleaning methods lack adaptability.

Method used

A reinforcement learning network employing the FLAVA model, deep canonical correlation analysis, Mahalanobis distance, isolated forest algorithm, and PPO algorithm is used for cross-modal feature extraction, quality assessment, and adaptive cleaning. By constructing quality feature vectors and cleaning strategy sequences, cleaning strategies are dynamically generated and network parameters are updated.

Benefits of technology

It improves the accuracy of multimodal unstructured data quality assessment and the adaptability of cleaning strategies, enhances the integrity and reliability of data processing, and forms a closed-loop optimization process of quality assessment, defect identification, and cleaning decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122364671A_ABST
    Figure CN122364671A_ABST
Patent Text Reader

Abstract

This invention discloses a method for quality assessment and cleaning of unstructured data fused with a multimodal large model, comprising the following steps: S1, acquiring and preprocessing text data, image data, audio data, and video data to construct a multimodal data sequence; S2, extracting cross-modal features and performing fusion encoding to generate a semantic feature sequence; S3, performing quality assessment based on the semantic feature sequence to generate a quality assessment sequence; S4, constructing a quality status sequence and a defect type sequence based on the quality assessment sequence; S5, generating a cleaning strategy sequence based on the quality status sequence and the defect type sequence; S6, performing data cleaning operations according to the cleaning strategy sequence and updating network parameters based on the cleaning results. This invention utilizes models such as FLAVA and possesses advantages such as high accuracy in quality assessment, strong adaptability of the cleaning strategy, and stable multimodal data processing results.
Need to check novelty before this filing date? Find Prior Art