Unsupervised graph-clustering learning framework for financial news summarization

Financial news shows significant influence on the inflection point of stock market. To condense the news texts with exponential growth, Automatic Text Summarization(ATS) becomes urgent. However, ATS tailored for financial news summarization has not been explored. In this paper, we propose a Graph-Clustering framework(FinGC) to extract financial news summarization. Our framework jointly learns the graph embedding and performs clustering in an unsupervised way. By combining graph structure which contains cross-sentence relations into text embedding, FinGC enriches the representation of financial news and highlights important sentences that point to specific events. Clustering is incorporated to group the financial news in terms of diversity and non-redundancy. We evaluated our best FinGC model on Fin-News and standard Multi-News. Experiments demonstrate it achieves state-of-the-art performance on standard datasets by ROUGE scores. Finally, we validated our results with human evaluation and show that our model achieves human performance on financial news.