Referring Expression Generation via Search Interaction Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conversational systems face challenges in generating high-quality dialog responses due to limited training data, requiring large amounts of annotated data to train predictive models effectively, which can be infeasible and expensive to create, especially when newly deployed.
Innovation Solution
Utilizing historical user interaction data from search engines, such as clickstream and web log data, to train models that map item names to referring expressions, enabling rapid deployment and configuration of conversational systems with high performance and contextual relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large amounts of annotated data are used to train predictive models, then dialog response quality is improved, but data collection cost and time are increased
Solution Approach 1:
The patent uses clickstream data from search engine interactions as a proxy copy of annotated training data. Instead of collecting expensive human-annotated dialog data, the system copies and leverages existing search query and result selection patterns to train the predictive model, achieving similar training objectives at lower cost and time investment
Solution Approach 2:
The patent introduces clickstream data as an intermediary between the available search interaction data and the required training data. This intermediary data type bridges the gap by providing structured information about user preferences and item relationships that can be used to train predictive models without requiring direct annotated dialog data
2Measurement precision
If annotated training data is collected, then model accuracy is improved, but implementation cost is increased
Solution Approach 1:
The patent employs inexpensive, readily available clickstream data instead of expensive annotated training data. The search interaction data is freely obtainable from search engine logs, eliminating the need for costly data annotation processes while still providing sufficient information for training accurate predictive models
Solution Approach 2:
The system copies and repurposes existing search interaction data that would otherwise be unused. By leveraging the structural information already present in search query and result data, the patent achieves model training at minimal implementation cost without sacrificing accuracy
3Reliability
If domain-specific training data is used, then query result relevance is improved, but system deployment time is increased
Solution Approach 1:
The patent performs preliminary training of the predictive model using general search interaction data before domain-specific deployment. The model is pre-trained on broad patterns from search clickstream data, allowing it to be quickly adapted to specific domains without requiring extensive domain-specific annotated data, thus maintaining both relevance and deployment speed
Solution Approach 2:
The patent creates a universal predictive model that can handle multiple domains by training on diverse search interaction data. The model learns general patterns of user preference and item relationships that apply across different domains, enabling rapid deployment to new domains without requiring domain-specific retraining
Data Source
AI summary
A method includes receiving data characterizing historical user interaction with a search engine associated with a plurality of items, the received data including a first search query and a first item name selected by a user in response to receiving a result of a search of the first search query; training, using the received data, a model that maps item names to referring expressions, the training including using the first item name selected by the user as an input to the model and the first search query as a supervisory signal to the model; and providing the model. Related apparatus, systems, techniques, and articles are also described.