Improving Text Summarization using Ensembled Approach based on Fuzzy with LSTM

Abstractive text summarization using attentional recurrent neural network (sequence-to-sequence) models have proven to be very effective. In this paper, a novel hybrid approach is presented for generating abstractive text summaries by combining fuzzy logic rules (which selects extractive sentences) with bidirectional long short-term memory (Bi-LSTM) which further produces abstractive summary. Bi-LSTM uses attention mechanism and Adam optimizer for updating network weights. The proposed approach utilizes fuzzy measures and inference to extract textual information from the document to find the most relevant sentences. These relevant sentences are given as input to Bi-LSTM to produce an abstractive summary of the significant sentences. The proposed FLSTM model is evaluated using ROUGE toolkit. The experiment is performed on standard datasets (i.e., DUC and CNN/daily mail). Another salient feature of this work is merging of DUC 2003–2004, DUC 2006–2007 datasets to generate a larger dataset to achieve better results. The FLSTM model is compared with other state-of-the-art models, and the empirical results suggested that the proposed FLSTM model outperforms all other models.