Question answering system with text mining and deep networks

Nowadays, the rapid development of technology has led to changes in many aspects of people's lifestyles. Due to the pandemic, face-to-face communication has decreased significantly, particularly in education. Institutions are now utilizing their data to improve their sectors. The use of artificial intelligence technologies adds value to companies. Question answering systems are a sub-branch of artificial intelligence that add value by answering users' questions in natural language. These systems consist of two stages: understanding the questions and finding the correct answers. The analysis of questions utilises natural language processing techniques, while answers are obtained through information extraction methods from relevant data sources. In this study, the design of question answering models using text mining and deep networks is presented. The pre-trained English BERT-base model was fine-tuned with the Stanford Question Answering Dataset (SQuADv1.1) using various hyperparameters and fine-tuning values. The results of the training show high success rates, with an F1 score of 88.13% and an Exact Match (EM) rate of 80.74%, compared to previous studies in the literature. An improvement study was conducted on the Turkish History Question Answering Dataset (THQuADv1.0), which was subsequently updated to THQuADv2.0 by adding questions related to Düzce University units. The pre-trained Turkish BERTurk-base model was then fine-tuned using the THQuADv2.0 dataset and the successful hyperparameters and fine-tuning values obtained from the English model.The training resulted in the creation of the BERTDuQuA (BERT Duzce University Question Answering) model for Turkish question answering, which achieved high performance with an F1 score of 87.10% and an EM of 76.90%. A novel model was devised by incorporating the BiLSTM layer into the BERT model. Training was conducted with the THQuADv2.0 dataset, utilising the fine-tuning technique with the optimal hyperparameters identified in the Turkish model. The outcome of this training was the BERTBiDuQuA (BERT&BiLSTM Duzce University Question Answering) model, which was developed for Turkish question answering. The BERTBiDuQuA model demonstrated high success, achieving an F1 score of 88.84% and an EM of 78.43%.