Research On Financial Events Detection By Incorporating Text And Time-series Data

With the perfection of domestic financial market-related systems, more and more individuals and institutions take part in the stock market. In this process, investors have shown great concerns in the events that may cause fluctuations in the capital market. However, due to the exteme expansion of internet information, it has been more and more diffcult for investors to obtain truly important information from the mass information.Thus, it has increasingly become the research and application focus of related fields on how to get important events or topics from the mass internet information.Traditional methods of event detection methods are mostly based on text processing techniques, few method conducts research which combined with features of financial time-series data. There is a large amount of time-series data such as stock transaction data and stock market index in financial field. This data which is the specific response of capital market to related events i s closely related to financial events. Therefore the thesis conducts research on financial event detection combined with text information and time-series information.First, in view of the situation that lack of corpus in financial field, the thesis designed and constructed a Chinese financial event corpus, and then fine-grainedly annotates the financial text events and raleted information. A financial event corpus within 2500 texts is constructed which will improvethe corpus lacking problem in this field. Then the thesis conducts research on financial event detection method based on text information and proposes a financial event extraction method based on text features and event topic sentences to effectively avoid the interference s of too many atom events and mistakes made by inaccurate analysis of dependency parsing. On this base the thesis also conducts research on financial event detection method combined with text data and time-series data.Through event element normalization, anaphora resolution, time alignment and the generation of event template, the thesis designs and implements a cross-text financial event detection and fusion method. By bringing in sorting factors of hot event and time-series characteristic, the thesis designs and implements financial hot event detection method combined with text features and time-series features. It is found in the experiment that after combining with financial time-series data, hot event detection F-score has improved to 82.45% form the former 77.95% and event type identification F-score has improved to 76.85% form the former 71.93%. It is shown that the method combined with text features and time-series features in the thesis can effectively improve the performance of financial event detection and event type classification.