Reciprocating Encoder Portrayal From Reliable Transformer Dependent Bidirectional Long Short-Term Memory for Question and Answering Text Classification

Abstract

Diversity in use of Question and Answering (Q/A) is evolving as a popular application in the area of Natural Language Processing (NLP). The alive unsupervised word embedding approaches are efficient to collect Latent-Semantic data on number of tasks. But certain methods are still unable to tackle issues such as polysemous-unaware with task-unaware phenomena in NLP tasks. GloVe understands word embedding by availing information statistics from word co-occurrence matrices. Nevertheless, word-pairs in the matrices are taken from a pre-established window of local context, which may result in constrained word-pairs and also probably semantic inappropriate word-pairs. SemGloVe employed in this paper, refines semantic co-occurrences from BERT into static GloVe word-embedding with Bidirectional-Long-Short-Term-Memory (BERT- Bi-LSTM) model for text categorization in Q/A. This method utilizes the CR23K and CR1000k datasets for the effective text classification of NLP. The proposed model, with SemGloVe Embedding on BERT combined with Bi-LSTM, produced better results on metrics like accuracy, precision, recall, and F1 Score as 0.92, 0.79, 0.85, and 0.73, respectively, when compared to existing methods of Text2GraphQL, GPT-2, BERT and SPARQL. The BERT model with Bi-LSTM is better in every way for responding to different kinds of questions.

Keywords

• Bidirectional encoder representations from transformer
• natural language processing
• question and answering
• SemGloVe