Abstract—Recent studies on QA (Question Answering) system in English language have been emerged extensively with the composition of Natural Language Processing (NLP) and Information Retrieval (IR) by amplifying miniature sub tasks to accomplish a whole AI-system having capability of answering and reasoning complicated and long questions through understating paragraph. In our proposed study, we present a general heuristic framework, an end-to-end model used for paraphrased question answering using single supporting line which is the initial appearance ever in Bangla language. Corpus dataset was scrapped from Bangla wiki and then questions were generated corresponding context have been used to learn the model. Translated bAbI dataset (1 supporting fact) in Bangla language has been also incorporated with to experiment the proposed model manually. To predict appropriate answer, model is trained with question-answer pair and a supporting line. For comparing our task applying variation of basic Recurrent Neural Network (RNN): Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) different accuracy has been found. For further accomplishment, synthetic and semantic word relevance in high dimension vector space: Bangla word embedding system(word2vec) is added to the system as sentence representation along with Positioning Encoding (PE) and which outperforms both memory network GRU and LSTM precisely. Index Terms—Machine learning, natural language processing, information retrieval, long short time memory, gated recurrent unit. I. INTRODUCTION Intelligent question answering system that offers tasks like asking automated machine any questions and getting appropriate answer from computer automatically has been very important end-user task in recent age which ease human’s life drastically. To enable communication with computer, asking question to computer is indispensable task. Many researches have been done over this particular field of question answering in near years. However, it’s an important task which uses a combination of both NLP [1] and IR that shortens the distance between IR-based search and intelligent assistants that uses information extracting process from data Manuscripts received May 30, 2020; revised July 25, 2020. Md. Mohsin Uddin is with East West University, Dhaka Bangladesh, (e-mail: mmuddin@ewubd.edu). Nazmus Sakib Patwary, Md. Mohaiminul Hasan, and Tanvir Rahman were with East West University, Dhaka, Bangladesh (e-mail: nazmus.ewu@gmail.com, mohaiminul.hasan.ewu@gmail.com, tanvir.rahman.ewu@gmail.com). Mir Tanveer Islam completed his BS in EEE from North South University, Dhaka, Bangladesh (e-mail: mirtanveerislam@Gmail.com). [2]. Different answers can be yielded with a modicum variation in semantically equivalent questions. For example, questions like “who did created Microsoft” and “who did started Microsoft” yield same identity. The question answering model ought to acknowledge the answer from its knowledge base considering both questions semantically equivalent [3]. There has been a lot of research in machine learning that are intended to reasoning and intelligently answering questions. It’s a comprehensive area [4]. Towards answering question and reasoning two grand challenges in intelligence system have been arrived in numerous research that make models which is able to make multiple computational steps for question answering and to make model that adopts and working ability considering long term dependencies in case of sequential data as well as unstructured data [5]. In the circumstance of semantic parsing for answering question in recent time, researchers are highly focused on complicated and long question answering [6]. Abundant number of researches has been performed considering English language related to different question answering task like search-based QA, factoid QA etc. in order to accomplish AI-complete question answering in isolated ways using end-to-end neural network. But in Bangla language no such research and task have been conducted regarding question answering which would contribute AI-complete question answering in Bangla language. In our study we propose a system having consideration of close domain dataset and develop algorithms with variety for understanding language and paraphrased question answering. Different architecture of RNN like LSTM, GRU has been used to build satisfactory model. Question-answer pair from Bangla close domain dataset has been used as training data as well as with one support line and multiple related line. It can answer paraphrased question’s answer too and questions containing ‘who’ (‘কে’), ’where’ (‘কেোথোয়’), ’when’ (‘েখন ’), ’what’ (‘কে’) are answered also whereas [1], [2] considered so straight forward simple dataset that only contain questions containing ‘where’ (‘কেোথোয়’) for their single supporting fact category. II. RELATED WORKS In recent years, various research has been applied on question answering over unstructured paragraph data using end-to-end suitable deep neural network model. It’s a part of NLP as well as use the sub part of information retrieval. In the paper [7], authors brought in a Recurrent neural End-To-End Neural Network for Paraphrased Question Answering Architecture with Single Supporting Line in Bangla Language Md. Mohsin Uddin, Nazmus Sakib Patwary, Md. Mohaiminul Hasan, Tanvir Rahman, and Mir Tanveer Islam International Journal of Future Computer and Communication, Vol. 9, No. 3, September 2020 52 doi: 10.18178/ijfcc.2020.9.3.565