Abstractive Text Summarization via Stacked LSTM

In the past, there have been many models proposed for text summarization via sequence to sequence training (seq2seq), attention mechanism, and transformers. Although these methods achieve an advance regarding the performance, these models fail to create a more complex feature representation of the current input and consequently gain inferior performance for modeling the long staggered sentences and modeling the complex inter-sentence dependencies. In order to address this issue, we utilize a more complex feature representation for summarization via stacked LSTM. In this case, the main reason for stacking LSTM is to allow for greater model complexity. For a simple encoder, we stack layers to create a hierarchical feature representation with attention. We generate the text summaries for any test text in terms of predicting the target sequence. With the proposed method, we achieve a better performance compared to the existing state-of-the-art phrase-based system on the task of text summarization on gigaword dataset. Furthermore, Experimental results on this dataset show that our framework performs well in terms of various ROUGE scores.