Tagging Model Training with Query Supplementation and Adversarial Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-trained tagging models struggle to accurately interpret queries due to their substandard performance, leading to irrelevant output results in applications that rely on them, such as serving non-relevant content items in response to user queries.
Innovation Solution
A multi-stage training process for a tagging model, involving pre-training on queries, supplementation with document titles, and adversarial knowledge enhancement, using weak and strong labels, and virtual adversarial training to improve generalization and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a tagging model is trained on queries alone, then the training process is simple and fast, but the model achieves low accuracy in interpreting queries
Solution Approach 1:
The training process is divided into three distinct stages: (1) initial training on queries alone to establish baseline capabilities, (2) supplementary training with document titles to enhance contextual understanding, and (3) adversarial training to improve robustness. This segmentation allows the model to progressively improve accuracy while managing training complexity in controlled increments.
Solution Approach 2:
The model performs preliminary training on queries alone before introducing the more complex document title supplementation. This preliminary action establishes a foundation that enables the model to benefit from subsequent training phases without being overwhelmed by complexity from the start.
2Measurement precision
If document titles are supplemented to queries for training, then the model's contextual understanding improves, but the training data processing time increases
Solution Approach 1:
The training is organized into periodic stages where the model cycles through different training data configurations. The second stage periodically introduces document title supplementation to enhance contextual understanding, while the adversarial training in the third stage periodically challenges the model with perturbed inputs. This periodic structure allows the model to efficiently process training data by alternating between different training regimes rather than continuously processing all data at maximum complexity.
3Adaptability or versatility
If adversarial training is applied to improve robustness, then the model's generalizability increases, but the training complexity and computational resources required increase
Solution Approach 1:
Adversarial training applies preliminary anti-action by intentionally introducing perturbations and challenges to the training data before the model encounters real-world variations. The third stage generates adversarial examples that preemptively stress-test the model, forcing it to learn more robust decision boundaries and improving generalizability before deployment.
4Measurement precision
If multiple training stages are used to improve accuracy, then the model performance increases, but the training time and computational resources increase
Solution Approach 1:
The multi-stage training approach segments the overall training process into three specialized phases, each optimized for specific learning objectives. This segmentation allows efficient resource allocation where each stage builds upon the previous one, improving query interpretation accuracy through progressive specialization rather than attempting to achieve all goals simultaneously.
Data Source
AI summary
A technique is described herein for training a tagging model that is able to successfully interpret queries. The technique trains the tagging model in plural stages. A first stage continues training a pre-trained language model based on a set of queries, to produce a first-stage model. A second stage performs training on the basis of a set of supplemented queries and associated weak labels, to produce a second-stage model. Each supplemented query combines a query with titles of documents that match the query. A third stage performs training on the basis of a set of supplemented queries and associated strong labels, to produce a third-stage model. The third stage also uses adversarial knowledge enhancement that has the effect of making the data presented to the third-stage model more difficult for the third-stage model to interpret. This, in turn, improves the generalization capabilities and robustness of the third-stage model.