Query Intent Classifier Using Self-Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Developers of query intent classifiers face challenges in building accurate classifiers due to the need for adequate labeled training data, as insufficient data can lead to inaccurate classification.
Innovation Solution
A technique is employed to construct a query intent classifier using a data structure that correlates queries to items, where initial labeled queries are used to iteratively label unlabeled queries based on click-through data, expanding the training data set and improving classifier performance through a graph-based learning algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If typical training techniques are used with insufficient labeled training data, then the classifier construction process is simple and fast, but the classification accuracy deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-labeling a small subset of training data before the main training process. This initial labeled data serves as a foundation that guides the self-training process, allowing the classifier to iteratively improve accuracy without requiring large amounts of manually labeled data from the start
Solution Approach 2:
The patent implements feedback through a self-training mechanism where the classifier repeatedly trains on its own predictions. The classifier generates predictions on unlabeled data, selects high-confidence predictions to add to the training set, retrains, and repeats this process. This feedback loop enables continuous improvement of classification accuracy while minimizing the need for extensive manual labeling
Solution Approach 3:
The patent applies self-service by enabling the classifier to train itself using its own predictions. The system automatically identifies which unlabeled examples to label based on prediction confidence, adds them to the training set, and retrains without external intervention. This self-service capability allows the classifier to bootstrap its own improvement process with minimal initial labeled data
2Measurement precision
If more labeled training data is collected manually, then classifier accuracy improves, but the time and resources required for data labeling increase
Solution Approach 1:
The classifier performs self-labeling by generating its own training data through iterative prediction and selection. Instead of requiring manual labeling of extensive datasets, the system automatically expands its training set by selecting high-confidence predictions from unlabeled data, thereby eliminating the time-consuming manual labeling process while still achieving high accuracy
Solution Approach 2:
The patent uses preliminary manual labeling of only a small seed set of data, which then serves as the foundation for automated self-training. This preliminary action minimizes the initial time investment for manual labeling while enabling subsequent automated expansion of the training set through the self-training process
Data Source
AI summary
To construct a classifier, a data structure correlating queries to items identified by the queries is received, where the data structure contains initial labeled queries that have been labeled with respect to predetermined classes, and unlabeled queries that have not been labeled with respect to the predetermined classes. The data structure is used to label at least some of the unlabeled queries with respect to the predetermined classes. Queries in the data structure that have been labeled with respect to the predetermined classes are used as training data to train the classifier.