Chinese predicate recognition method based on POS fusion feature and entity boundary diagnosis

The invention provides a Chinese predicate recognition method based on POS (Point of Sale) fusion feature and entity boundary diagnosis, aiming at the problems that a traditional method is easy to misjudge a Chinese predicate boundary and low in word activity predicate recognition precision, so that error accumulation exists in downstream tasks. According to the method, firstly, RoBERTa and BiLSTM are adopted for word embedding, and text context information is fused; then, a novel POS injection mode is introduced according to a method based on SPAN, a linear attention layer is connected after POS representation, key points are focused on all target SPANs so as to further capture a distribution rule, then SPAN coding sequences and POS attention coding sequences are fused through a gating module, final fusion vector representation is obtained, and semantic information is reserved to the maximum extent; the boundary diagnosis module outputs and transmits BiLSTM codes of SPAN head and tail words to two multi-laye