基于深度学习的金融文书摘要自动生成研究与实现

Research and Implementation of Automatic Generation of Financial Document Abstract Based on Deep Learning

金融文书的自然语言处理是目前金融科技领域的研究热点,相关研究大多数着眼于传统的分词和基于机器学习的语义场景分析这种有监督的学习方法,难以满足当前金融监管行业快速处理金融文本的需求。针对这一问题,本文构建了一个基于多层LSTM的中文金融文书摘要自动生成算法框架,通过Seq2Seq模型,基于注意力机制的强化学习框架,发现最优策略,对文本进行编码、解码,从LSTM编码器中抽取文本序列并输出摘要。实验结果表明,多层LSTM结构相比传统RNN的ROUGE值更高,具有较好的学习能力。

The natural language processing of financial documents is a research hotspot in the field of financial sci?ence and technology at present.Most of the relevant researches focus on the supervised learning method of traditional word segmentation and semantic scenario analysis based on machine learning,which is difficult to meet the needs of the current financial regulatory industry to process financial documents quickly.To solve this problem,this paper constructed an algorithm framework based on multi-layer LSTM(short and long-term memory network)for automatic generation of Chinese financial document abstracts,and used Seq2Seq(sequence to sequence)model.The experimen?tal results show that the multi-layer LSTM structure has better learning ability than the traditional RNN(cyclic neu?ral network)in terms of the value of the rule(semantic recovery).