Evaluation of an Arabic Chatbot Based on Extractive Question-Answering Transfer Learning and Language Transformers

Chatbots are programs with the ability to understand and respond natural language in a way that is both informative engaging. This study explored current trends of using transformers transfer learning techniques on Arabic chatbots. The proposed methods used various semantic embedding models from AraBERT, CAMeLBERT, AraElectra-SQuAD, AraElectra (Generator/Discriminator). Two datasets were for evaluation: one 398 questions, other 1395 questions 365,568 documents sourced Wikipedia. Extensive experimental works conducted, evaluating manually crafted entire set by confidence similarity metrics. Our results demonstrate combining power transformer architecture extractive chatbots can provide more accurate contextually relevant answers Arabic. Specifically, our showed AraElectra-SQuAD model consistently outperformed models. It achieved an average score 0.6422 0.9773 first dataset, 0.6658 0.9660 second dataset. concludes remarkable performance, high confidence, robustness, which highlights its potential practical applications processing tasks suggests be further enhanced tasks, such as specialized chatbots, virtual assistants, information retrieval systems Arabic-speaking users.