Chinese Financial Event Extraction Base on Syntactic and Semantic Dependency Parsing

As a sub-task of information extraction, event extraction plays an important role in nature language process applications, such as stock market trend forecast, which can provide strong clues for events users, e.g. investors, managers and government, to analyze the market and make decisions. At present, most of the studies about event extraction pay more attention to the type correctness of triggers and arguments, and not consider the effect and value of event extraction based on application requirements. We call this type of event extraction traditional event extraction. The event types and standards in traditional event extraction are derived from ACE2005 containing 8 categories and 33 sub-categories, KBP2015 and ERE, et al. However, there are some limitations in application of them to event extraction in specific financial domain. For example, there is not the overweight event type in ACE2005, which is a special behavior in the financial domain. In this paper, we focus on the financial news and extract open events without types. In the field of finance and economics, most event users are more concerned with the objects and actions that events affect. Therefore, combined with the application requirement, we propose to extract the financial event ET(Sub, Pred, Obj), where Sub, Pred and Obj represent subject, predicate and object respectively. However, Chinese financial news generally suffers from the event nesting and component default problem, which result in event omission and key element missing of events. To tackle this issue, with the expression habits and characteristics of Chinese linguistics, we build a Chinese event extraction framework based on syntactic and semantic dependency parsing. Then summarize four common default structures and design corresponding completion rules. In particular, at the beginning of this paper, we summarize four prominent phenomena in the extraction of events from the headlines of financial news, and explore the cause of these problems, no in-depth analyzing the relevance of syntactic and semantic structure or lack of it. After that, we employ the syntactic dependency parsing tree and lexical structure, and propose the core verb chains, which make sure that each core verb corresponds to an event solving event leakage problem. Thirdly, we add semantic dependency relation between events on the basis of syntactic dependency tree, which is called Syntactic Semantic Dependency Parsing (SSDP) tree. In order to better separate the detected events and their properties, we adjust and optimize SSDP tree to form the SSDP graph, where the word nodes of the same syntactic structure are at the same level, providing a way for subsequent event extraction. Fourthly, with the division of default structure in linguistic, we summarize four common default structures and propose ten corresponding completion rules to solve the problem of component default. Meanwhile, the whole Chinese event extraction algorithm based SSDP graph is shown at the end of the section. Finally, this paper depicts a detailed experimental situation. The experimental dataset, labeling standard and evaluation index are given. Subsequently, the method in this paper is verified on two datasets, financial news titles and common field news titles. At the end, we conduct comprehensive benchmarks on Chinese financial news titles and CoNLL2009 Chinese Corpus. The experimental results show that the proposed methods are effective. © 2021, Science Press. All right reserved.