Tail Query Reformulation via Head Query Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies for product search in e-commerce struggle with reformulating tail queries to accurately capture purchase intent, especially due to scarce historical behavioral signals and the diverse, frequently changing nature of e-commerce catalogs, leading to diminished performance for infrequent queries.
Innovation Solution
The approach involves mapping tail queries to head queries with similar purchase intent using a machine-learning model that learns from historical head queries and refines embeddings with a persistently noisy oracle, incorporating contextual sampling techniques to avoid biases and leverage user actions as rewards, allowing for online training without human-labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the search engine relies on historical behavioral data for query reformulation, then performance on head queries is good, but performance on tail queries deteriorates due to scarce behavioral signals
Solution Approach 1:
The patent introduces an intermediary system that bridges tail queries and head queries through a two-stage process: first mapping tail queries to semantically similar head queries using a pre-trained language model, then reformulating based on the mapped head query's historical behavioral data. This intermediary mapping mechanism allows tail queries to leverage the rich behavioral signals of head queries without directly relying on their own scarce data.
Solution Approach 2:
The system performs preliminary action by pre-training language models and building query mapping relationships in advance using abundant head query data. This pre-computed mapping structure is then reused during online tail query processing, eliminating the need to collect and process behavioral data for each individual tail query, thus overcoming the data scarcity issue.
2Ease of operation
If query reformulation methods extend or relax the query, then general search performance improves, but e-commerce product search performance deteriorates due to loss of purchase intent
Solution Approach 1:
Instead of extending or relaxing tail queries as conventional methods do, the patent inverts the approach by mapping tail queries to head queries and using the head query's established reformulation patterns. This inversion preserves the original tail query's purchase intent while benefiting from the search engine's optimized processing of head queries.
Solution Approach 2:
The system copies the successful reformulation strategy from head queries and applies it to tail queries through mapping. By copying the query structure and reformulation logic from well-performing head queries, the system maintains purchase intent while achieving better search results for tail queries.
3Measurement precision
If the search engine uses exact match for queries, then relevance is high when exact match exists, but performance deteriorates for tail queries with no exact match in the diverse and frequently changing catalog
Solution Approach 1:
The patent implements a dynamic two-stage reformulation process that adapts to the diversity and changes in the e-commerce catalog. The system dynamically maps tail queries to appropriate head queries based on semantic similarity, and then dynamically applies reformulation strategies based on the mapped query's characteristics and historical performance, rather than using a static exact-match approach.
Data Source
AI summary
Technologies are provided for reformulation of a tail query to a head query with the same purchase intent by mapping the tail query to the head query. In some of the technologies, a reasonable embedding can be learned on historical head queries. The embedding can then be refined by leveraging rewards generated from a persistently noisy oracle that compensates for the lack of historical behavioral signal for tail queries. Further, a contextual sampling technique that uses text-based rewards or oracle-based rewards, or both, can be implemented in order to avoid biases introduced by persistent noise in the oracle. Numerical experiments on large scale e-commerce datasets demonstrate that the provided technologies can outperform several conventional approaches to query reformulation.