Search Query Language Identification via Weak-Labeled Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional language identification systems perform poorly with shorter texts such as search queries due to lack of language-identifying signals and the expense and time required for human-labeling sufficient training data.
Innovation Solution
The generation of weak-labeled training data using a seed dictionary and k-Nearest Neighbors (kNN) model to annotate query terms, supplemented with human-annotated evaluation data for model training and tuning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional language identification systems are trained on well edited text with language-identifying signals, then the systems work well for longer texts, but the performance drops significantly for shorter texts like search queries
Solution Approach 1:
The patent changes the training data parameters from well-edited text to search query logs, and adjusts the feature extraction to focus on character n-grams and word patterns specific to short texts. This allows the model to adapt to the different characteristics of short text domains while maintaining language identification accuracy.
Solution Approach 2:
The patent creates a synthetic training dataset by copying and transforming search query log data into a format suitable for language identification training. This synthetic data is generated to mimic the characteristics of short texts while providing sufficient training examples, thereby improving model performance without requiring manual annotation of large volumes of short text data.
2Measurement precision
If human labeling is used to create training data for language identification, then the training data is accurate, but the process is expensive and time-consuming
Solution Approach 1:
The system performs self-labeling by automatically annotating search query terms with language labels based on the trained machine learning model. This self-service approach eliminates the need for manual human labeling of training data, significantly reducing the time and cost required for data preparation while maintaining acceptable accuracy through iterative model improvement.
Solution Approach 2:
The patent implements a feedback loop where the machine learning model is trained on initially weak-labeled data, then used to generate improved labels that are fed back into the training process. This feedback mechanism continuously refines the training data quality without requiring manual intervention, thereby reducing time loss while maintaining measurement precision.
3Quantity of substance
If machine learning models are trained on weak-labeled data generated by computer algorithms, then the training data quantity increases, but the labels are associated with an error rate
Solution Approach 1:
The patent applies partial labeling where not all data points are labeled with high confidence, and uses a combination of weak labels and strong labels in the training process. This partial action approach allows the model to learn from the majority of data points even with some label errors, while the excessive action of generating large volumes of weak-labeled data ensures sufficient training material is available to overcome the noise in the labels.
Data Source
AI summary
Systems and methods are disclosed for search query language identification. One method comprises generating a seed dictionary comprising a plurality of labeled dictionary terms and receiving a plurality of unlabeled sample query terms. The plurality of unlabeled sample query terms are compared to the plurality of labeled dictionary terms at a first time, and a first set of labeled sample query terms are generated by labeling at least a subset of the plurality of unlabeled sample query terms based on the first comparison. Remaining unlabeled sample query terms are then compared with the first set of labeled sample query terms at a second time, and a second set of labeled sample query terms are generated by labeling the remaining unlabeled sample query terms based on the second comparison. The first and second sets of labeled sample query terms are provided to a machine learning model configured for query language prediction.