Sentence Graph Attention for Content-Aware Summarization

Neural network-based encoder–decoder (ED) models are widely used for abstractive text summarization. While the encoder first reads source document and embeds salient information, decoder starts from such encoding to generate summary word-by-word. However, drawback of ED model is that it treats words sentences equally, without discerning most relevant ones others. Many researchers have investigated this problem provided different solutions. In paper, we define a sentence-level attention mechanism based on well-known PageRank algorithm find sentences, then propagate resulting scores into second word-level layer. We tested proposed CNN/Dailymail dataset, found was able summaries with much higher power than state-of-the-art models, in spite an unavoidable (but slight) decrease terms Rouge scores.