On solving textual ambiguities and semantic vagueness in MRC based question answering using generative pre-trained transformers.

Abstract

Machine reading comprehension (MRC) is one of the most challenging tasks and active fields in natural language processing (NLP). MRC systems aim to enable a machine to understand a given context in natural language and to answer a series of questions about it. With the advent of bi-directional deep learning algorithms and large-scale datasets, MRC achieved improved results. However, these models are still suffering from two research issues: textual ambiguities and semantic vagueness to comprehend the long passages and generate answers for abstractive MRC systems. To address these issues, this paper proposes a novel Extended Generative Pretrained Transformers-based Question Answering (ExtGPT-QA) model to generate precise and relevant answers to questions about a given context. The proposed architecture comprises two modified forms of encoder and decoder as compared to GPT. The encoder uses a positional encoder to assign a unique representation with each word in the sentence for reference to address the textual ambiguities. Subsequently, the decoder module involves a multi-head attention mechanism along with affine and aggregation layers to mitigate semantic vagueness with MRC systems. Additionally, we applied syntax and semantic feature engineering techniques to enhance the effectiveness of the proposed model. To validate the proposed model's effectiveness, a comprehensive empirical analysis is carried out using three benchmark datasets including SQuAD, Wiki-QA, and News-QA. The results of the proposed ExtGPT-QA outperformed state of art MRC techniques and obtained 93.25% and 90.52% F1-score and exact match, respectively. The results confirm the effectiveness of the ExtGPT-QA model to address textual ambiguities and semantic vagueness issues in MRC systems.