Few-Shot Blockchain Domain Named Entity Recognition with Fused Lexical Features

Deep learning-based blockchain named entity recognition (NER) often necessitates large quantities of labeled samples, but there are few blockchain-labeled text data readily available and the cost of manual labeling is high. The limited availability of labeled data makes it necessary to train systems with limited data. For Chinese NER in particular, transformer models such as BERT have shown to be efficient in encoding character-level texts, but the performance is unstable and leaves room for improvement because of the challenges in adding more features and the lack of word-level features. Chinese text modeling has been rarely studied, and most models in this regard have poor generalization abilities. In this paper, we propose the Fused Lexical Features-Prototypical Networks (LT-PN network), a novel approach for a few Chinese NER, consisting of (1) a flexible word Lattice Transformer encoder with a network transformer that can encode texts at the word level and be flexibly combined with character-level functions, and (2) a gated mixed of experts (MoE) network which overcomes the feature overload problem by training the model to conditionally combine contextual and positional features rather than assigning fixed weights to them.