NER-NET: A NEW DEEP LEARNING ARCHITECTURE WITH HYBRID OPTIMIZATION FOR NAMED ENTITY RECOGNITION

Identification of textual and visual information in a document plays a pivotal role in comprehending the
information contained in it, thereby providing an effective way to analyze the document. Information
Extraction (IE) is an attractive research area that mainly targets developing techniques for analyzing rich
documents. Named Entity Recognition (NER) is the prime process in IE, and it is a sequence-labeling
process, wherein unstructured data is considered and the data is mapped to pre-defined labels. In this work,
a novel deep learning architecture is developed for performing NER, and it is aimed at identifying the
numerical entities in the input text. Here, a novel deep learning network named NER-Net is proposed which
includes various layers, like tokenization, Inverse Document Frequency (IDF), Convolutional Neural
Network (CNN), Bidirectional Long Short Term Memory (BiLSTM), attention, and NER layers. Further, a
Fire Hawk African Vultures Optimization Algorithm (FHAVOA) is proposed for optimally tuning the layer
dimension of the NER-Net. Moreover, the evaluation of NER-Net-FHAVOA shows that accuracy of 0.936,
Mean Square Error (MSE) of 0.087, Root MSE (RMSE) of 0.294, Positive Predictive Value (PPV) of 0.909,
and Negative Predictive Value (NPV) of 0.877 are attained, thus revealing its superiority.