The Study Of Document-level Time Argument Extraction Driven By Knowledge

The time argument extraction model mainly refers to two aspects of tasks,the first is the task of extracting the arguments corresponding to the time expression of the given text and the given event label,and the second is to calculate the event according to the text of the marked event and time timing relationship between them.This task plays a decisive role in deeply understanding the development process of the event and judging the current state of the event.However,at this stage,the research on the document-level time argument extraction model is not in-depth in the academic circle.On the one hand,under the premise that deep learning requires a large number of samples to train the model,the correct labeled data corpus is relatively scarce,and the model cannot be trained on a large scale to improve effect;on the other hand,the document-level time argument extraction model is more complex than the sentence unit,and the existing models cannot accurately model the entire document.In response to the above problems,on the basis of the existing scarce data sets,two aspects of research have been carried out: on the one hand,the exploration and research on the word embedding calculation method,through the introduction of external knowledge to improve the model extraction effect,on the other hand,through the time theory Meta-theoretical integration into the model improves the performance of the document-level temporal argument extraction model.The main research work is as follows:(1)Propose a method of using hyperbolic space word embedding,replace European word embedding with hyperbolic word embedding,and train on a model based on the RoBERTa Classifier classifier,so that under the unfavorable conditions of lack of data sets,with a lower The vector dimension of,obtains a semantic representation embedding similar to the high-dimensional Euclidean space word embedding,and the hyperbolic word embedding generated in this way is easier to capture the temporal relationship of events.In addition,the model’s understanding of temporal relationships is enhanced by incorporating external knowledge,and finally the word embedding method is dynamically modified according to the model extraction results.(2)A method of using the time argument theory in linguistics to guide the model extraction effect is proposed.By modifying the word embedding in the pre-training stage,sentence tense,posture and time adverbs are used as additional information to enhance the word embedding expression of event trigger words.The model is also trained with an event-time argument extraction task to construct temporally enhanced textual representations.The event-triggered word embedding in this text expression is used to construct an event temporal relationship extraction matrix,and the temporal relationship is calculated in units of chapters.Both models in this paper calculate the model on the basis of Roberta’s pre-trained word embedding,which has greatly improved the performance of the previous model.