Reinforcement-Based Denoising for Crowdsourcing Named Entity Recognition in Social Networks

Crowdsourced Named Entity Recognition (CNER) is an effective method in social network data mining, enabling the swift identification and categorization of vast textual data, thereby enhancing the understanding and analysis of user behaviors and communication patterns. However, the diversity of such data often leads to inconsistent annotation quality, compromising recognition accuracy. We propose a novel framework that synergizes reinforcement learning with multi-source domain adaptation to bolster the accuracy of social CNER. By filtering out low-quality instances and considering annotator reliability in the multi-source domain adaptation crowdsourcing method, our framework not only adapts to the varied quality of annotations but also mitigates the impact of annotation noise. Validated on both the social network crowdsourced benchmark dataset FMKKM11 and the general-purpose CoNLL03, our method consistently outperforms state-of-the-art methods in terms of model accuracy and noise robustness.