Deep Information Extractor (DIE): A Multipurpose Information Extractor with Shifted Vectors Pre-processing Methods

Information Extraction (IE) from text is a challenging data mining task. Recently, numerous research studies have been proposed in this domain, but most of them used Pre-Processed Embeddings (PPE) inefficiently. PPE converts Pre-Trained Embedding (PTE) into a different format to achieve a more accurate outcome. Most of the existing PPE methods concatenate the embeddings, leading to high computational costs and low accuracy of the extracted information. This paper introduces a novel deep learning model and PPE methods for IE from any text. The proposed PPE uses information from different pre-trained models via establishing sensible relationships between this information without enhancing the computational cost. Proposed PPE methods exploit Bert and Deberta embeddings. To select the most relevant characters and words from a sentence, the bi-directional encoder–decoder Recurrent Neural Network (RNN) model was used, followed by Multi-head Attention (MHAT) to learn the relationship between words. Therefore, results have significantly improved target information estimation accuracy on the benchmark dataset across domains.