Search Intention Tagging via User Behavior Copying
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current intention recognition technologies face challenges in accurately classifying user requirements for search data, especially with short texts lacking context and grammatical structure, leading to difficulties in fine-grained intention recognition, particularly for professional data like medicine-related queries, which often require expert labeling, resulting in high time and cost consumption and limited labeled data.
Innovation Solution
A method that involves obtaining and labeling user search data with intention tags, assigning these tags to similar search data from other users who select the same results within a predetermined time frame, thereby generating a large amount of labeled data for training models without relying on traditional natural language processing techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If expert labeling is used for professional search queries, then intention recognition accuracy is improved, but time consumption and cost increase significantly
Solution Approach 1:
The system uses user search behavior data (clicks, selections, time spent) to automatically generate intention labels for search queries. This self-labeling mechanism eliminates the need for expert manual annotation while maintaining high accuracy, as the labels are derived from actual user interactions with search results.
Solution Approach 2:
The patent copies intention labels from queries with known labels to similar queries based on behavioral patterns. When users exhibit similar search behaviors for different queries, the intention tag from one query is transferred to another, creating labeled data without additional expert labeling effort.
2Measurement precision
If expert labeling is used for professional search queries, then intention recognition accuracy is improved, but cost increases significantly
Solution Approach 1:
The system automatically generates intention labels by analyzing user search behavior data, eliminating the need to pay experts for manual labeling. The labeling process is performed autonomously using existing user interaction data, significantly reducing costs while maintaining accuracy.
Solution Approach 2:
The patent transfers intention labels from labeled queries to unlabeled queries through behavioral pattern matching. This copying approach reuses existing labeled data to create new labeled examples without incurring additional expert labeling costs.
3Productivity
If traditional natural language processing techniques are used, then processing speed is maintained, but data diversity and training quality are limited
Solution Approach 1:
The patent moves from traditional NLP text analysis to a new dimension of user behavior analysis. Instead of processing only text content, the system incorporates click patterns, selection behaviors, and temporal data to create enriched training data with higher diversity and better representativeness of actual user intentions.
Solution Approach 2:
The system copies behavioral patterns across different queries to generate diverse training examples. By replicating user behavior sequences and intention tags across multiple search scenarios, the system creates varied training data without requiring proportional increases in expert labeling.
Data Source
AI summary
A method for data processing is provided. The method includes obtaining first retrieving data associated with a first user and a first retrieving result selected by the first user from at least one retrieving result corresponding to the first retrieving data. The first retrieving data is labelled with an intention tag indicating a retrieving intention of the first user. The method further includes obtaining second retrieving data that is used by a second user to conduct retrieving and selecting the first retrieving result within a predetermined time period. The method further includes assigning the intention tag to the second retrieving data.