BTSQA: An Architecture for Bangla Textual and Spoken Question Answering

Abstract

Question answering (QA) is a field within natural language processing (NLP) that focuses on developing systems capable of automatically answering questions posed in human language. QA systems aim to understand the meaning and intent behind questions and provide accurate and relevant answers by leveraging large corpora of text data. Short Answer Questioning (SQA) is a specific type of question answering task within natural language processing (NLP) that focuses on generating concise and precise answers to fact-based questions. Unlike traditional QA systems that generate longer, descriptive answers, SQA systems aim to extract short snippets of information directly related to the question. These systems employ techniques such as text com prehension, named entity recognition, and information retrieval to identify the most relevant information and produce brief and accurate responses. SQA finds applications in areas such as search engines, voice assistants, and chatbots, where quick and concise answers are desired. In our thesis, we propose BTSQA, an architecture to perform spoken question answering. We have built the architecture with one general QA model and one ASR model. Then we added a word correction step to improve the performance. Initially the general QA model,T5 transformer model, was used to with F1 score of 73.37%. We used audio dataset on the whisper ASR model with WER of 31.58% and Wev2Vec2 model with WER score 29.64%. When we combined the general QA model with ASR model using word correction the performance F1 score was 53.65%. This models were run on the text dataset we built and they were transferred in audio using GoogleTTS