Knowledge-Augmented Conversational AI for Low-Resource Indian Languages

Abstract

The conversational AI is based on the question answering system (QAS). It is a critical component in natural language processing (NLP) for retrieving relevant information from large documents or webpages. With the rapid development of multilingual content on the web, traditional QASs face substantial challenges. This research introduces the Hierarchical Attention Fuzzy Random Multimodal Network (HAFRMN) for English-to-Hindi QAS. During the training phase, a set of questions and passages is fed into the Bidirectional Encoder Representations from Transformers (BERT) model. Features such as Term Frequency-Inverse Document Frequency (TF-IDF) and n-Gram features are extracted from both the passages and the questions. Simultaneously, answers corresponding to the questions and the passages are processed through BERT to generate target tokens. These generated tokens, along with the question and passage tokens and feature vectors from the questions, are used to train the HAFRMN. The HAFRMN combines a Hierarchical Attention Network (HAN) with Random Multimodal Deep Learning (RMDL) and fuzzy logic concepts. During testing, passages and questions are again processed through BERT to extract tokens and features. These tokens, features, and the results from the trained model are fed into the HAFRMN to generate accurate answers. The HAFRMN achieved notable performance metrics with an exact match score of 0.904, precision of 91.6%, recall of 90.9%, and an F-measure of 90.6%.