A truth inference scheme for crowdsourcing using NLP and swin transformers.

Area of Science:

• Artificial Intelligence
• Data Science
• Natural Language Processing

Background:

• Crowdsourcing offers scalable data collection but faces challenges in ensuring data reliability due to varying contributor expertise and task complexity.
• Truth inference is crucial for deriving accurate answers from noisy, heterogeneous crowdsourced responses.
• Existing methods struggle with the nuances of crowdsourced data, impacting the quality of inferred truths.

Purpose of the Study:

• To propose a novel truth inference model that enhances the accuracy and reliability of crowdsourced data.
• To leverage Natural Language Processing (NLP) and transfer learning with Swin transformers for improved truth inference.
• To dynamically refine contributor reliability and task difficulty estimations for more robust results.

Main Methods:

• Integration of Natural Language Processing (NLP) with transfer learning utilizing Swin transformers.
• Swin transformer's shifted windowing technique to capture local and global contextual features in textual data.
• Fine-tuning embedding representations for the specific nuances of crowdsourced tasks.

Main Results:

• The proposed model demonstrates superior performance compared to state-of-the-art methods in accuracy and scalability.
• Consistent outperformance observed across multiple crowdsourcing datasets, especially under noisy and complex conditions.
• Enhanced robustness in truth inference due to dynamic refinement of contributor reliability and task difficulty.

Conclusions:

• The Swin transformer-based truth inference model significantly improves the quality and reliability of crowdsourced data.
• The approach offers a robust solution for handling noisy and complex crowdsourcing tasks.
• This method represents a significant advancement in deriving accurate information from crowdsourced platforms.