Research On Event Extraction Techniques For Domain Texts

Text event extraction(EE)is a crucial task in information extraction that aims to extract structured event information from unstructured text.The extraction target refers to the occurrence of actions or state changes involving one or more specific roles at a particular time and location.It is an essential component in the field of information extraction.Automatic extraction of structured events from domain-specific texts can save substantial manual reading costs and improve the efficiency of acquiring event information.For instance,EE in the financial and military domains can help decision makers quickly respond to changes in the situation.Structured EE in the legal domain can aid judges in analyzing cases and making judgments.Therefore,studying EE technology for domain-specific texts is of great importance and has received widespread attention.Taking the legal,military and financial domains as representatives,this dissertation conducts an in-depth study of the practical needs,model designs,and training processes for domain-specific EE.At present,most research extracts structured events from open-domain texts according to the event structure and task patterns defined by ACE(Automatic Context Extraction).With the development of deep learning technology,numerous EE methods based on deep learning have emerged and currently yield the best results.However,there are differences between event compositions and textual expressions in specific domains versus the public domain.These differences mean existing methods cannot be readily applied to domain-specific EE tasks.For instance,the traditional event structure cannot capture the fine-grained element information required in the legal,military,and financial domains.Moreover,existing EE models cannot simultaneously solve the problems of extracting domain-specific proper nouns,long-distance dependencies,and reference resolution.Domainspecific EE also grapples with scarce labeling data and high labeling costs.Therefore,studying EE technology suitable for domain-specific texts necessitates comprehensively considering the characteristics of domain-specific texts,the information needs of domain-specific events,the costs of model training,and so on.Given the aforementioned problems,this dissertation proposes three objectives for automatic domain-specific EE technology.First,this dissertation aims to construct a deep learning EE model that can handle different task characteristics across various domains.However,this model cannot address scenarios with little labeled data.Therefore,the second goal is to develop a few-shot learning approach for domain-specific EE by leveraging meta-learning techniques and abstract event knowledge.At the same time,considering domain-specific EE’s requirement for model accuracy,this dissertation’s third objective is to design a novel dynamic active learning strategy premised on incorporating knowledge.This strategy reduces the labeling costs required to achieve the target accuracy by selecting high-quality samples for training.To address the aforementioned challenges and problems and fulfill the three goals,this dissertation conducts the following research:(1)A dynamic hierarchical event structure is proposed in this dissertation.By extending the ACE event structure,arguments can continue to include sub-arguments.This structure retains the ACE event structure’s function and can dynamically adjust the complexity of the event structure according to a domain’s target event type characteristics.On this basis,a dynamic hierarchical EE model based on pedal attention mechanism is proposed in this dissertation.By using the neighboring words of the target word in the dependency parsing tree as the pedal of attention mechanism,this model can effectively construct semantic associations locally,over long distances,and across demonstrative pronouns in complex domain-specific texts.The model also designs a weight inheritance mechanism for hierarchical event types,effectively utilizing domain event types’ potential hierarchical characteristics and improving the model’s extraction efficacy.This method can aptly handle EE in various domain-specific texts by reasonably leveraging the attention mechanism’s characteristics in information interaction and synthesizing the word dependency information provided by dependency syntax.In addition,this dissertation defines the event structure for three domains,and labels the EE datasets.Experiments on datasets across three domains demonstrate this method’s effectiveness.(2)This dissertation proposes a new few-shot approach based on knowledge-enhanced Bayesian meta-learning.Because few-shot EE is limited by the lack of sample diversity in the support set,this dissertation incorporates abstract event definitions into the EE model.Specifically,this method defines hierarchical event knowledge across domains and develops a knowledge encoder to map abstract knowledge into vectors in semantic space.Based on the knowledge vector results,this dissertation designs a Bayesian meta-learning method premised on adaptive knowledge enhancement.This method combines the event information provided by the support set samples and the abstract event knowledge to construct a few-shot domain-specific EE model.Experiments on three domain-specific EE datasets demonstrate this method’s effectiveness.(3)This dissertation proposes a new active learning approach based on dynamic loss prediction that dynamically selects high-value samples to label during model training.This method utilizes a memory unit to dynamically store sample information learned from previously labeled samples and selected unlabeled samples.It assesses sample importance through loss prediction for unlabeled samples.And a batch selection strategy based on internal-external loss sorting is proposed to select unlabeled samples,making this method more suitable for the characteristics of single sample and multitask extraction in domain-specific EE.Simultaneously,the batch selection strategy and memory unit can effectively address information redundancy during sample selection and improve the overall sample quality.Finally,to realize the above functions,this dissertation adopts a new delayed training strategy,realizing end-to-end training of the loss prediction model and EE model.Experiments verified that this method can effectively reduce the demand for labeled data in the training process.