Named Entity Recognition in Chinese E-commerce Domain Based on Multi-Head Attention

At present, the demand for entity recognition for product query and product recommendation systems of e-commerce domain is growing. Traditional methods that rely on experts to define artificial features and domain knowledge can no longer meet the needs of the domain due to their low recognition accuracy. According to the above background, a Chinese named entity recognition method in the e-commerce domain that incorporates a multi-headed attention mechanism and bidirectional long-short term memory network is proposed. A word vector represents the product title statements of e-commerce platforms. The generated word vector sequences are fed into a bi-directional long-short term memory network that mines their contextual semantic features. A multi-headed attention mechanism is introduced to focus on the entity words in the text to unearth their hidden features. In contrast, the entity recognition results are labeled by calculating the joint probability of sequence labels through conditional random fields. The experimental data show that the method can reach 86.16% accuracy and 86.57% F1 value for entity recognition in the e-commerce domain, which is practical and feasible.