Dynamic Taxonomy Evolution for Video Tagging Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual and automated video content tagging systems face inefficiencies due to limitations in annotation taxonomies, including general or granular terms, abstract terms, and cultural barriers, leading to inconsistent tagging and increased human intervention.
Innovation Solution
A performance-based evolution system that uses a machine learning model to classify problematic terms as confusing or flawed, providing comparative samples for clarification and modifying taxonomies by editing terms, their scopes, or substituting them with more intuitive alternatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual tagging is performed by human taggers using a predetermined taxonomy, then tagging accuracy can be maintained through human judgment, but productivity becomes impracticable due to the large number of videos requiring annotation
Solution Approach 1:
The patent introduces an automated tagging system as an intermediary between the video content and the final tags. This automated system uses machine learning models to perform initial tagging, significantly increasing productivity while maintaining acceptable accuracy. Human taggers then focus only on reviewing and correcting automated tags rather than performing all tagging manually, thus resolving the contradiction between maintaining accuracy and increasing throughput.
Solution Approach 2:
The system performs preliminary automated tagging before human review. The automated system pre-processes the video content and generates initial tags, which are then refined by human taggers. This preliminary action by the automated system handles the bulk of tagging work, improving productivity while human reviewers ensure accuracy for edge cases.
2Stability of the object's composition
If a closed set of terms is used in the annotation taxonomy, then tagging consistency is improved, but tagging performance deteriorates due to general or abstract terms that do not capture specific content nuances
Solution Approach 1:
The patent implements a dynamic taxonomy system that evolves over time based on performance feedback. The taxonomy is not static but adapts by adding, removing, or modifying terms based on what works best in practice. This allows the system to maintain consistency through the structured framework while improving specificity as the taxonomy evolves to capture more nuanced content categories.
Solution Approach 2:
The system incorporates feedback loops where tagging performance is continuously monitored and used to refine the taxonomy. Performance metrics from both automated and human tagging are analyzed to identify terms that need refinement. This feedback mechanism allows the taxonomy to evolve, improving specificity while maintaining consistency through systematic updates rather than arbitrary changes.
3Productivity
If automated tagging systems are used to increase productivity, then annotation efficiency is improved, but tagging performance depends on the relevance of the closed term set which limits accuracy
Solution Approach 1:
The patent merges automated tagging systems with human review processes into a hybrid system. The automated system handles high-volume initial tagging to maintain productivity, while human reviewers provide oversight to ensure relevance and accuracy. This combination allows the system to leverage the speed of automation while incorporating human judgment for relevance, resolving the contradiction between efficiency and precision.
Solution Approach 2:
Human reviewers serve as an intermediary layer between the automated tagging system and the final output. The automated system generates candidate tags at high speed, and human reviewers act as a filter to correct inaccuracies and ensure relevance. This intermediary human layer maintains productivity by only reviewing a subset of tags while improving the overall relevance of the final tagged content.
Data Source
AI summary
According to one implementation, a system includes a computing platform having processing hardware, a system memory storing a software code; and a machine learning model based classifier. The processing hardware is configured to execute the software code to receive tagging quality assurance (QA) data including multiple terms applied as tags and corrections to those tags, to identify, using the tagging QA data, a first problematic term, and to classify, using the machine learning model based classifier, the first problematic term as one of confusing or flawed. The processing hardware is further configured to execute the software code to obtain, when the first problematic term is classified as confusing, a comparative sample for clarifying use of the first problematic term, and to obtain, when the first problematic term is classified as flawed, modification data for editing a predetermined annotation taxonomy including the first problematic term.