Abstract
We aim to automatically generate natural language narratives about an inputstructured knowledge base (KB). We build our generation framework based on apointer network which can copy facts from the input KB, and add two attentionmechanisms: (i) slot-aware attention to capture the association between a slottype and its corresponding slot value; and (ii) a new table positionself-attention to capture the inter-dependencies among related slots. Forevaluation, besides standard metrics including BLEU, METEOR, and ROUGE, we alsopropose a \textit{KB reconstruction} based metric by extracting a KB from thegeneration output and comparing it with the input KB. We also create a new dataset which includes 106,216 pairs of structured KBs and their correspondingnatural language descriptions for two distinct entity types. Experiments showthat our approach significantly outperforms state-of-the-art methods. Thereconstructed KB achieves 68.8% - 72.6% F-score.