CFERE: Multi-type Chinese financial event relation extraction.

Extracting various types of event relations in financial texts can benefit many downstream applications supporting financial analysis. This paper addresses the multi-type event relation extraction problem in the finance domain focusing on handling several issues in existing studies, including (1) limited event relation types involved, (2) insufficient feasibility when handling non-annotated data, (3) ineffectiveness in recognizing multi-type event relation, and (4) the asynchronous event and event relation extraction process. To tackle these limitations, we carefully define six types of event relations based on the characteristics of financial texts (e.g., abundant numerical words) and further devise an integral framework for Chinese financial event relation extraction. The framework is capable of handling unsupervised event extraction and event relation recognition jointly. Specifically, according to linguistic characteristics, a Core Verb Chain is employed for the event identification. Then, by constructing a Syntactic Semantic Dependency Parsing graph, scattered events are combined into pairs, and event ellipsis elements can be completed to prevent event information loss. Also, to capture more sentence semantics, we formulate an Event Restore module that converts the structured event pairs into event restore sentence pairs and pour the pairs into the BERT model for relation type identification. Finally, to enhance the embeddings for event core elements, an Event Core Embeddings layer is augmented in BERT, and we fine-tune the model on our annotated financial corpus. Extensive experiments are conducted to validate the effectiveness of our method, and the results confirm its superiority over the state-of-the-art baselines. • We define six event relation types based on practical needs in finance. • We propose a framework for multi-type event relation extraction, called CFERE. • Event sentence pairs are used to minimize the loss of input information. • An event core embedding layer is added to enhance the encoding of BERT. • We present a novel corpus and the effectiveness of CFERE on two datasets. [ABSTRACT FROM AUTHOR]