Multimodal information capture based truth inference network in crowdsourcing

Abstract

Truth inference of truth from crowdsourced data presents a formidable challenge that has been widely recognized in the field. Recently, there has been a surge in deep learning and Bayesian methods that rely on task features. However, these methods fail to function effectively in situations where task features are lacking or the relationship between task truth and task features is weak. Traditional data mining methods from crowdsourced triplet data either rely on strong model assumptions with poor data adaptability or use weak assumption models based on worker confusion matrices, neglecting the difficulty differences between tasks. To address this, we propose a novel DS-like model that leverages the strong adaptability of the weak model assumption in the DS model by using a task confusion matrix to describe the impact of task difficulty information. Furthermore, we overcome the data information bottleneck by capturing multimodal information about additional data. Our model exhibits weak coupling characteristics, enabling it to adapt to the features of different data. To tackle the complex issues arising from parameter reduction in our model, we introduce an innovative coordinate ascent algorithm, termed ”twice-EM.” Finally, we substantiate the effectiveness of our proposed approach through a comprehensive series of experiments, highlighting significant improvements in the accurate inference of truth, thereby attesting to the significance of our method.