Masked Language Model Query Expansion for Search Result Diversity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional query processing systems do not automatically modify user queries to generate alternative queries, resulting in limited variations in search results.
Innovation Solution
A query processing apparatus that uses a masked language model to generate alternative queries by replacing target phrases with mask tokens, and selecting alternative phrases based on phrase embeddings to provide more varied search results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional query processing systems are used, then the system structure remains simple, but the search result variety is limited
Solution Approach 1:
A masked language model is introduced as an intermediary component between the query processing system and search results. The model generates alternative queries by filling mask tokens with nearest neighbor phrases, thereby expanding search result variety without requiring fundamental changes to the core system architecture.
Solution Approach 2:
The system performs preliminary query expansion by generating alternative queries before the actual search execution. The masked language model pre-computes multiple variant queries based on the original query, which are then used to retrieve diverse search results, avoiding the need for complex real-time modifications during search.
2Adaptability or versatility
If automatic query modification is implemented, then alternative queries are generated, but processing time increases
Solution Approach 1:
The system changes the parameter of query phrasing by generating multiple variant versions of the same search query. The masked language model modifies the textual representation of queries by replacing mask tokens with different nearest neighbor phrases, enabling parameter variation without fundamental computational overhead.
Solution Approach 2:
The system creates copies of the original query in the form of alternative queries. The masked language model generates multiple phrased versions of the same search intent by filling mask tokens with different candidate phrases, effectively copying the query structure while varying the content to expand search results.
3Measurement precision
If masked language model is used to generate alternative phrases, then phrase embedding comparison is performed, but computational resources increase
Solution Approach 1:
The system replaces traditional mechanical search methods with a neural network-based embedding comparison mechanism. The masked language model substitutes direct keyword matching with semantic embedding comparisons, where phrases are represented as vector embeddings that capture semantic meaning rather than exact string matches.
Solution Approach 2:
The system changes the representation parameter of phrases from discrete tokens to continuous embeddings. By transforming phrases into embedding vectors that capture semantic semantics, the system enables more accurate phrase matching while the computational cost is managed through efficient embedding comparison algorithms rather than exhaustive search.
Data Source
AI summary
Systems and methods for query processing are described. Embodiments of the present disclosure identify a target phrase in an original query, wherein the target phrase comprises a phrase to be replaced in the original query; replace the target phrase with a mask token to obtain a modified query; generate an alternative query based on the modified query using a masked language model (MLM), wherein the alternative query includes an alternative phrase in place of the target phrase that is consistent with a context of the target phrase; and retrieve a search result based on the alternative query.