Distributed Data Annotation System Using Iterative Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current data annotation methods, such as Amazon Mechanical Turk, face inefficiencies in determining accurate ground truth due to reliance on majority votes, which can lead to increased time and cost, and may not correctly identify the ground truth, especially when annotators provide relatively equally applied annotations.

Innovation Solution

A distributed data annotation system that uses iterative maximum a posteriori estimation to determine source data and annotator metadata, selecting annotators based on their characteristics and the difficulty of the data, to estimate ground truth efficiently and accurately.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If majority vote method is used to determine ground truth, then simplicity of implementation is improved, but accuracy of ground truth determination deteriorates

Engineering Contradiction:
Improvesimplicity of implementationVSAvoidaccuracy of ground truth determination
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system changes the parameters used for ground truth determination from simple majority voting to a probabilistic model that incorporates annotator competence scores and data difficulty estimates. This allows the system to weight annotations differently based on annotator performance and task difficulty, improving accuracy while maintaining computational feasibility through iterative maximum a posteriori estimation.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If more annotators are assigned to each piece of source data, then accuracy of ground truth determination is improved, but time and cost increase

Engineering Contradiction:
Improveaccuracy of ground truth determinationVSAvoidtime and cost
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system applies partial action by assigning annotations strategically rather than uniformly. It uses competence-based selection to assign annotations only to the extent necessary for reliable ground truth determination, avoiding redundant annotations. The iterative estimation process determines when sufficient annotations have been collected, preventing excessive annotation work while maintaining accuracy.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The system implements feedback through iterative maximum a posteriori estimation that continuously updates ground truth probabilities based on incoming annotations and annotator performance data. This feedback loop allows the system to adaptively determine when sufficient annotations have been collected, optimizing the balance between accuracy and resource consumption.

Inventive Principle:
Principle #23Feedback

3Ease of operation

If annotator selection is based on random assignment, then ease of task distribution is improved, but quality of annotations deteriorates

Engineering Contradiction:
Improveease of task distributionVSAvoidquality of annotations
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The system performs preliminary action by pre-evaluating annotator competence and data difficulty before task assignment. This allows the system to strategically match annotators with appropriate tasks based on their demonstrated abilities, improving annotation quality from the outset rather than relying on random assignment followed by post-hoc filtering.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentUS9928278B2Systems and methods for distributed data annotation
Publication Date: 2018.03.27 CALIFORNIA INST OF TECH
  • US9928278B2 patent drawing
  • US9928278B2 patent drawing
  • US9928278B2 patent drawing

AI summary

Systems and methods for distributed data annotation in accordance embodiments of the invention are disclosed. In one embodiment of the invention, a distributed data annotation server system includes a storage device configured to store source data, one or more annotators, annotation tasks and a processor, wherein a distributed data annotation application configures the processor to receive source data including one or more pieces of source data, select one or more annotators, create one or more annotation tasks for the selected annotators and source data, request one or more annotations for the source data using the annotation tasks, receive annotations, determine source data metadata for at least one piece of source data using the received annotations, generate annotator metadata for at least one annotator using the received annotations and the source data, and estimate the ground truth for the source data using the source data metadata and the annotator metadata.