面向财经应用的文本挖掘若干关键技术研究

面向财经应用的文本挖掘主要包括两个任务：1)通过主题模型等方法构建财经指标体系，2)利用情感分析等方法量化财经指标。然而，利用主题模型、情感分析等方法构建和量化财经指标体系，存在以下难点：1)生成的主题和主题词不能反映财经指标涵义；2)随着时间变化，财经主题和反映财经主题涵义的词语都可能发生变化；3)常用财经指标量化方法存在片面性。针对上述问题：1)提出基于财经领域知识的HDP主题模型，实现从财经文本中抽取能较好地反映财经指标涵义的财经要素词，构建财经指标体系；2)设计基于语义相关度的动态时间窗口划分算法，构建动态主题时间窗口划分的主题演化模型，解决现有研究中设置固定时间窗口主题演化模型的不足；3)提出抽取和表达财经因果关系实体、获取财经领域因果知识的新视角，为提高财经指标体系构建和量化效果提供新途径；4)提出综合利用财经领域因果知识、财经要素词情感值和社会关注度的财经指标量化方案。

Text mining for financial applications mainly includes two tasks: 1) constructing a financial indicator system based on the topic model; 2) quantifying financial indicators based on sentiment analysis. However, there are several difficulties in these tasks: 1) the generated topics and topic words cannot reflect the meaning of financial indicators; 2) both financial topics and topic words may change over time; 3) the commonly used quantifying methods of financial indicators have biases. To solve the above problems, in this project, we first propose a finance domain knowledge based hierarchical Dirichlet processes model FDK-HDP. The model extracts topics and topic words more accurately to better represent financial indicators, based on which the financial indicator system can be established. Second, we design a dynamic time window partition algorithm based on semantic relevance, and design a topic evolution model with variable-sized time window, which solves the deficiency of existing topic evolution models with fixed time window. Third, we propose the new perspective of extracting cause-effect entities from financial text corpora and acquiring causal knowledge in financial domain, which provides a new way to improve the construction and quantification of financial indicator system. Lastly, a novel approach is introduced to quantify the financial indicators system which integrates financial causal knowledge with the sentiment analysis results and social attention.