An Efficient Approach of NER in Social Media using BiLSTM-CRF Model

Early named entity recognition models struggled with the issue of multiple meanings of a word because most text processing concentrated on the representation of individual words and character vectors and paid little attention to the semantic relationships between the preceding and subsequent text in an utterance. Most models use the Transformer model's attention mechanism to deal with the issue of words in texts having various meanings. Due to its completely linked nature, the conventional Transformer model has a substantial computational cost. Therefore, in order to address the issue of the conventional Transformer's low computational efficiency, this research suggests a new model called the CNN-BiLSTM-CRF model. First, the input text is transformed into a character vector using a BERT model to handle word ambiguity. Then, a lightweight Star-Transformer extracts local features from word vectors, while a CNN-BiLSTM model globally extracts features from text context. The extracted feature sequences are combined and then input to a CRF for final result prediction. Experimental results demonstrate significant improvements in precision, recall, and F1 score compared to traditional models, along with a nearly 40% increase in computational efficiency.