Word Weighting Accuracy via Pre-trained Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of image search results is compromised due to the low accuracy of word weighting in image search engines, which is attributed to the short text of images, leading to inadequate correlation calculations between query terms and candidate images.
Innovation Solution
A method involving the generation of regression and sorting models through machine learning based on text and visual features extracted from clustered images, using similarity thresholds and filtering techniques to enhance word weighting, thereby improving the accuracy of image search results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If word weighting is performed based on short image text, then the processing speed is fast, but the accuracy of word weighting result is low
Solution Approach 1:
The patent applies preliminary action by pre-training word embedding models offline using large-scale corpus data before actual image search operations. The word2vec or BERT models are trained in advance to capture semantic relationships, so that during image search, the pre-trained models can quickly and accurately weight words without requiring real-time complex processing.
Solution Approach 2:
The patent transitions from traditional short-text weighting methods to a new dimension by incorporating pre-trained word embedding models that capture semantic relationships in high-dimensional vector spaces. This dimensional transformation allows the system to leverage linguistic knowledge from large corpora, improving weighting accuracy beyond what simple short-text statistics can provide.
2Measurement precision
If traditional word weighting techniques are used on short image text, then the system complexity is low, but the accuracy of image search results is low
Solution Approach 1:
The system performs preliminary training of word embedding models offline using extensive text corpora. This pre-training captures semantic relationships and word contexts that can be directly applied to image text weighting, improving search accuracy without adding complexity to the real-time image search pipeline.
Solution Approach 2:
The patent introduces pre-trained word embedding models as intermediaries between the image text and the search ranking system. These models serve as a bridge that transforms raw image text into semantically enriched representations, allowing the search system to leverage linguistic knowledge without directly processing complex language understanding tasks.
3Speed
If word embedding models are trained offline, then real-time processing speed is improved, but training time and computational resources are increased
Solution Approach 1:
The patent implements preliminary action by training word embedding models offline in advance before deployment. This allows the computationally intensive training process to occur separately from real-time operations, so that during actual image search, the system can quickly process queries using the pre-trained models without time-consuming training iterations.
Solution Approach 2:
The patent segments the overall system into offline training phase and online inference phase. The word embedding model training is separated from the real-time image search processing, allowing each to be optimized independently - training can use extensive computational resources when needed, while search operations run efficiently in real-time using the pre-trained models.
Data Source
AI summary
The present invention provides a model generation method, word weighting method, device, apparatus, and computer storage medium. In one aspect, the embodiments of the present invention obtain other images that are identical to a specified image, and use the specified image and said other images as sample images; and then obtain a text clustering based on the texts of the sample images; and then obtain text features based on the text clustering and obtaining visual features based on the sample images; and carry out machine learning based on the text features and the visual features to generate a first regression model and a first sorting model for word weighting on words in the text of an image. The embodiments of the present invention therefore can improve the accuracy of the word weighting result in an image text, thus improving the accuracy of an image search result.