Query Completion Suggestion Using Bidirectional N-gram Co-occurrence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional search systems provide query suggestions based solely on co-occurrence of query terms, which may not be useful to users, as they often suggest continuations rather than preceding terms, failing to satisfy user needs for versatile suggestions.
Innovation Solution
A method and server using a machine learning algorithm that trains on past user queries to predict query-completion suggestions by parsing queries into n-grams, selecting candidate n-grams based on pair-based co-occurrence, and generating feature vectors to determine group-based co-occurrence, allowing for both preceding and following suggestions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If query suggestions are provided based solely on co-occurrence of query terms, then the suggestion system is simple to implement, but the usefulness and versatility of suggestions deteriorates
Solution Approach 1:
The patent segments the query completion problem into two distinct tasks: predicting preceding terms and predicting following terms. By dividing the single suggestion generation process into separate preceding-term prediction and following-term prediction modules, the system can independently optimize each direction using appropriate co-occurrence statistics, thereby improving overall versatility while maintaining implementation simplicity through modular design.
2Ease of manufacture
If query suggestions provide only continuations of partial queries, then the suggestion system is easy to implement, but user satisfaction deteriorates due to lack of preceding term suggestions
Solution Approach 1:
The patent applies inversion by not only predicting terms that follow the partial query (traditional approach) but also predicting terms that precede the partial query (inverted approach). By implementing bidirectional prediction - both forward continuations and backward completions - the system satisfies diverse user needs while maintaining reasonable implementation complexity through symmetric processing logic.
3Measurement precision
If the system analyzes group-based co-occurrence patterns, then the accuracy and usefulness of suggestions improves, but the computational complexity increases
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing co-occurrence statistics for query terms and n-grams before the actual query completion task. By preparing these statistical models in advance and caching them, the system reduces real-time computational complexity while maintaining high accuracy in suggestions. The heavy computational work is shifted to the offline training phase, making online prediction efficient.
Data Source
AI summary
A method and server for training a machine learning algorithm (MLA) for determining a query-completion suggestion for a partial query is disclosed. The method comprises receiving and parsing past queries into n-grams. Each one of the n-grams being associated with respective n-gram features, the n-gram features being indicative of a pair-based co-occurrence of n-grams in the past queries. The method also comprises, for a given n-gram of a given past query: selecting at least one candidate n-gram from the n-grams based on the pair-based co-occurrence; generating respective feature vectors for the given n-gram and the at least one candidate n-gram; generating a training set for the given n-gram comprising an input portion and a label portion; and training the MLA based on the training set to determine a predicted group-based co-occurrence of at least one in-use candidate n-gram and at least one in-use n-gram.