Query Suggestion via Document-Query Embedding Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Search engines face challenges in identifying relevant query suggestions that may not be directly related to the content of a visited document, as they often rely on frequent query terms and struggle to suggest queries that reflect user intent based on past experiences, especially for infrequent queries or documents with unique content.
Innovation Solution
A method that involves generating feature representations for documents and queries, embedding them in a reduced dimensionality space, and determining a similarity measure to identify potential query suggestions, allowing for the suggestion of queries that may not be directly evident from the document content but are based on past user interactions, using a query suggestion rule trained on past document and query pairs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If search engines rely on frequent query terms to generate suggestions, then the system simplicity is maintained, but the ability to suggest relevant infrequent queries is lost
Solution Approach 1:
The patent transforms the high-dimensional sparse feature space of document and query terms into a low-dimensional dense embedding space. This dimensionality reduction allows the system to capture semantic relationships and user intent patterns that transcend simple term frequency counting, enabling suggestion of infrequent but relevant queries without proportionally increasing system complexity.
Solution Approach 2:
The system changes the representation parameters of queries and documents from discrete term frequencies to continuous embedding vectors. This parameter transformation enables the model to generalize from frequent patterns to infrequent cases by measuring similarity in the embedding space rather than relying on exact term matches, thereby improving adaptability while maintaining computational efficiency.
2Measurement precision
If the system uses large datasets for training query suggestions, then the accuracy of suggestion learning is improved, but the training time and computational resources increase
Solution Approach 1:
The patent employs stochastic gradient descent and mini-batch training methods that process subsets of the large dataset iteratively rather than requiring complete passes through all data. This partial action approach achieves good learning accuracy by focusing on representative samples in each iteration, significantly reducing training time while still capturing patterns from the full dataset over multiple epochs.
Solution Approach 2:
The system performs preliminary dimensionality reduction and feature extraction to create compact embedding representations before the main training process. This preliminary action pre-processes the large dataset into a more manageable form, reducing the computational burden of subsequent training iterations and accelerating the overall learning process while preserving the information needed for accurate query suggestion.
3Productivity
If the system embeds features in reduced dimensionality space, then the computational efficiency is improved, but the information loss may occur
Solution Approach 1:
The patent transforms features from sparse high-dimensional representations to dense low-dimensional embeddings through learned transformation matrices. This parameter change preserves essential semantic information by concentrating it in fewer dimensions, achieving computational efficiency while maintaining the discriminative power needed for accurate query suggestion. The embedding process learns to retain the most informative aspects of the original features.
Solution Approach 2:
The system creates composite embedding representations that combine multiple feature types (document content, query terms, user behavior patterns) into a unified low-dimensional vector. This composite approach synthesizes information from different sources, preserving diverse information aspects in the reduced space and preventing loss of any single feature type while achieving computational efficiency through the consolidated representation.
Data Source
AI summary
Methods and apparatus disclosed include identifying a pair comprising a document visited and a subsequent query, the subsequent query being submitted after visiting the document visited; generating a feature representation for the pair, each of the feature representations including one or more document features representing the content of a document visited, and one or more query features representing the content of a subsequent query; embedding the document features and the query features in a reduced dimensionality space; and determining a similarity measure for the reduced dimensionality space. In some embodiments, the similarity measure is indicative of the likelihood of co-occurrence of the document visited and the subsequent query. Other embodiments include iteratively training a query suggestion rule, applying the query suggestion rule to provide suggested queries in response to a document visited, corresponding systems, apparatus, and computer program products.