Non-text Content Label Selection via Consensus Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current search systems often include images unrelated to search queries in results due to mischaracterization of textual content associated with images, leading to poor search result quality.
Innovation Solution
The method involves selecting non-text content items by grouping initial labels associated with web pages and identifying n-grams that appear in a threshold number of label groups, using these labels to determine the relevance of the content to a search query and adjust result scores accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If textual content associated with an image is used to determine search relevance, then the search process is simple and fast, but the search result quality deteriorates due to mischaracterization of image content
Solution Approach 1:
The patent introduces visual labels as an intermediary between the image and the search query. Instead of directly using textual content (which may be misleading), the system generates visual labels through image analysis and uses these labels to match against search queries. This intermediary layer ensures that the search relevance is determined based on actual visual content rather than potentially inaccurate text descriptions.
2Measurement precision
If visual labels are generated through clustering and consensus from multiple sources, then the accuracy of image content characterization is improved, but the processing time and computational complexity increase
Solution Approach 1:
The patent performs preliminary actions by generating visual labels from multiple sources (different image analysis methods, multiple web pages) before the actual search query is processed. The clustering and consensus-building are done in advance during index construction, so that when a search query arrives, the system can quickly match against pre-computed, high-quality visual labels without performing complex processing in real-time.
Solution Approach 2:
The patent merges multiple sources of visual labels (from different image analysis techniques, multiple web pages containing the same image) into a unified set of consensus labels. By combining and clustering labels from multiple sources, the system achieves higher accuracy through consensus, where only labels that appear frequently across multiple sources are retained as final visual labels.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting labels for a non-text content item. In one aspect, a method receives a set of initial labels for a non-text content item, wherein the set of initial labels specifies text that has been identified as descriptive of the non-text content item and a web page to which the text corresponds. Initial labels corresponding to sets of matching web pages are grouped into separate initial label groups that correspond to each set of matching web pages. Sets of matching labels are grouped into other separate initial label groups that correspond to the sets of matching labels. One or more words that are included in at least a threshold number of the separate label groups are selected as final labels for the non-text content item.