Joint abstractive and extractive method for long financial documentsummarization
Nadhem Zmandar, Abhishek Singh, Mahmoud El‐Haj, Paul Rayson · Lancaster EPrints (Lancaster University) · 2021
In this paper we show the results of our participation in the FNS 2021 shared task. In our work we propose an end to end financial narrative summarization system that first selects salient sentences from the document and then paraphrases extracted sentences. This method generates an overall concise summary that maximises the ROUGE metric with the gold standard summary. The end to end system is developed using a hybrid extractive and abstractive architecture followed by joint training using policy-based reinforcement learning to bridge together the two networks. Empirically, we achieve better scores than the proposed baselines and toplines of FNS 2021 (LexRank, TextRank, MUSE topline and POLY baseline) and we were ranked 2nd in the shared task competition. Keywords: Summarization, Neural networks, Reinforcement learning, sequence to sequence learning; actor-critic methods; policy gradients.