Hierarchical Attention Fuzzy Random Multimodal Network–Based English–Hindi Question Answering System

ABSTRACT Question answering system (QAS) is a natural language processing (NLP) task, which is highly significant for searching for helpful information on massive documents or websites. Recently, there has been a quick development of the multilingual content on the web, and this has posed diverse challenges to conventional QASs. In this research, Hierarchical Attention Fuzzy Random Multimodel Network (HAFRMN) is introduced for English–Hindi QAS. Here, two phases, namely training as well as testing, are conducted. During training, a considered set of questions and passages is fed to the Bidirectional Encoder Representations from Transformers (BERT) model. Then, features such as Term Frequency‐Inverse Document Frequency (TF‐IDF) and n‐gram features are extracted from passage tokens and question tokens. On the other hand, answers are passed to the BERT model to obtain tokens. Thus, the target from tokens is acquired, which is given to HAFRMN along with question tokens, passage tokens and feature vectors from the question as well as passage tokens to accomplish the training process. However, HAFRMN is designed by an incorporation of Hierarchical Attention Network (HAN) with Random Multimodel Deep Learning (RMDL) and fuzzy concept. During testing, passages and questions are subjected to the BERT model. From the obtained passage and question tokens, features are extracted. Finally, passage tokens, question tokens, along with extracted features and outcome from the trained model, are fed to HAFRMN to obtain the exact answer. In addition, HAFRMN achieved a maximal exact match of 0.904, 91.6% of precision, 90.9% of recall, and 90.6% of f ‐measure.