| Event Causality Extraction(ECE)aims to extract causal event pairs with structured information from text,including event types and multiple event elements.In the financial domain,accurate and efficient event and causality extraction is of great importance for downstream tasks such as supporting decision-making and risk warning.Currently,the task has received some attention but still faces the problem of insufficient data.There are two problems in relevant data:1)current data only have event subject descriptions without specific event structures;2)current data lack multi-event pairs which are closer to practical scenarios,i.e.,a text contains multiple causality pairs of events.Meanwhile,there are two main types of related work for event causality extraction.The first type is event causality detection,i.e.,determining whether a causal relationship exists between two given event subject descriptions.This part of the work ignores causality composition and structured event extraction,which is difficult to be helpful in downstream applications.The second category relies on trigger words for event extraction and relationship recognition,but laborious human-annotated trigger words are not sufficient in current data and the annotation of trigger words consumes labor costs.In order to address these issues,we investigate the data and the model levels respectively and classify the model-level objects into "single-instance event type pairs" and "multipleinstance event type pairs" according to the number of instances of "event type pairs" corresponding to "different causal event pairs" in a text to be extracted.Specifically,the main research elements of this paper are as follows:(1)Constructing an event causality extraction dataset for the financial domain.This paper constructs an event causality extraction dataset for the financial domain based on Chinese industry research report text data with manual annotation.To address the problem of the lack of multiple event pairs,we propose a causal event subject pair extraction model to mine and filter the large-scale unlabeled corpus,and obtain a large number of single-instance and multiple-instance raw texts containing causal relationships and target event types.To address the problem of lack of structured events,we follow and optimize the event model of CCKS2021 data,and propose to add "enumerable elements" to normalize the changes of events,which is more convenient for downstream task application,knowledge inference,and computations.Through manual annotation and review of results,we completed the construction of a high-quality event causality extraction dataset,and provide statistical analysis of related task difficulties.(2)Study of event causality extraction for single-instance event type pairs.To solve the problem of single-instance event causality extraction without trigger words,this paper proposes a joint causality event pair extraction model in stages.In the first stage,a cascading structured model is used to identify the event types and the causal relationships between them in the text;in the second stage,a sentence representation with embedded event type information is obtained by using "dual localization" and reading comprehension mechanisms,and the first and last positions of each event element are predicted by a multi-layer binary marker decoder.To alleviate the error propagation problem,the two-stage model is jointly trained by sharing coding layers.Experimental results show that the method proposed in this paper can effectively extract single-instance causal event pairs from trigger-free texts without any manual heuristics.(3)Study of event causality extraction for multi-instance event type pairs.To solve the more complex multi-instance event causality extraction problem,this paper proposes a causal event pair extraction model based on set prediction,which encodes the input text and interacts with a predefined set of embedding matrices to obtain the corresponding set of hidden layer representations of causal event pairs.After decoding each implicit representation,all information of each causal event pair is obtained simultaneously:the event types of the cause and effect events,the type of enumerable elements,and the start and the end positions of non-enumerable elements in the text.The experiments show that this method has significant effectiveness on multi-instance event type pairs data while ensuring the effect of single-instance event type on event causality extraction,and can better perform causality event pairs extraction on trigger-word-free text.In summary,this thesis has made preliminary research results on text event causality extraction in the financial domain from the perspectives of dataset construction,single-instance event type pairs,and multi-instance event type pairs,respectively.We hope that the research in this paper can bring help to natural language processing tasks such as information extraction. |
