Video retrieval method and model training method, apparatus, device, and medium

By generating semantically identical and different sample text pairs, the training samples are enriched, solving the sparsity problem in video retrieval model training and improving retrieval accuracy and user experience.

CN122173679APending Publication Date: 2026-06-09TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
TENCENT TECHNOLOGY (SHENZHEN) CO LTD
Filing Date
2024-12-06
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

Existing technologies ignore the sparsity dilemma in the annotations of the training set during video retrieval model training, failing to provide key contextual information between potential events in the training set and query sentences, resulting in low retrieval accuracy and poor user experience.

Method used

By acquiring multiple initial samples with labels, a large language model is used to generate semantically identical positive sample texts and semantically different negative sample texts, enriching the training samples and generating sample pairs for model training, thereby improving the accuracy of video clip retrieval.

Benefits of technology

It improves the accuracy of video clip retrieval, enhances the user experience, solves the sparsity problem in model training set annotation, and strengthens the contextual information between potential events and query sentences in the training set.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122173679A_ABST
    Figure CN122173679A_ABST
Patent Text Reader

Abstract

Embodiments of the present application provide a video retrieval method, a model training method, an apparatus, a device and a medium, relating to the technical field of computers, the model training method can comprise: obtaining a plurality of initial samples with labels, for each initial sample, rewriting the sample text in the initial sample through a large language model to generate a positive sample text with the same semantics as the sample text and a negative sample text with different semantics, for each initial sample, taking the initial sample as a sample pair, taking the positive sample text and the sample video as a sample pair, taking the negative sample text corresponding to the initial sample and the sample video as a sample pair, training the initial retrieval model based on the sample pairs corresponding to each initial sample to obtain a trained retrieval model, and further, the accuracy of the retrieval model in performing segment retrieval can be improved.
Need to check novelty before this filing date? Find Prior Art