Research On Document-level Event Extraction Methods In Chinese Financial Domain

Massive data are available for exploitation in the Internet era.Among them,financial information contains important content that may affect the world landscape and personal wealth,so it is especially important to efficiently assist financial practitioners in extracting financial events from the huge amount of data.The event extraction task in financial domain is developed due to this need,and financial events in the real world are usually described in the form of documents,which makes document-level event extraction methods gradually gain the attention of a lot of researchers.Event extraction tasks in other domains have been achieved at the sentence level,however,the Chinese finance domain at the document level is still challenging because documents in Chinese finance are generally characterized by large volume,large number of words and scattered event elements,where different sentences contain different event elements and multiple events share argument elements.Specifically,the dispersion of events in the document prevents the model from effectively modeling long-distance event semantics,and the sharing of argument entities introduces the problem of overlapping entity extraction.Most of the research works focus on the single-event extraction task at the sentence level,either ignoring the document-level event extraction task or considering only the overall F1 scores and ignoring the multi-event F1 scores.To address these problems,this thesis proposes methods for single-event extraction tasks and multi-event extraction tasks,and is dedicated to improving the performance of single-event extraction tasks and multi-event extraction tasks at the document level.The main contributions are listed as follows:(1)For the document-level single-event extraction task,we propose the Joint Prediction of Event Types and Extraction of Argument Entities(JPEE)method,which can simultaneously consider the relationship between event types and event arguments and extract event instances more effectively.(2)For the task of document-level multi-event extraction,we propose the Flattening Tree-like Events(FTE)method,which flattens the complex tree-like event structure into a sequence structure,and the sequence structure can be easily modeled by the current rich sequential neural network models for modeling.By flattening tree-like events into sequences from the perspective of entities,it also avoids the problem of sharing arguments among multiple events for better encoding of multiple events.In this thesis,extensive experimental and analytical work is conducted on relevant datasets in the Chinese financial domain,and in-depth investigations are carried out to demonstrate the effectiveness of the proposed method.The experimental results show that both the JPEE and FTE methods proposed in this thesis have significant improvements over the baseline model,and also extend the research work on documentlevel event extraction.In addition,we apply the FTE method to an event extraction competition,and the results of the experiments and the rankings show the effectiveness of the FTE method on the document-level event extraction task.