Research On Several Key Technologies In Construction Of Event Ontology

In the computer science,an ontology,which is defined as "a formal,explicit specification of a shared conceptualization",plays an increasingly critical role in artificial intelligence,such as natural language processing,text mining and etc.Traditional ontology models use concepts as knowledge units,reveal the static rules of things in objective world by describing taxonomic relations among things.But they are lack of the ability of describing complex dynamic changing processes among things.Event Ontology takes events as basic knowledge representation units and the lattice structure as its architecture,it accords more with dynamic changing rules among things in real world.Event Ontology can not only overcome technical bottlenecks of traditional ontologies,but also lay a solid foundation for the realization of event-oriented and dynamic natural language processing.The related research on the construction of event ontology provides a novel ideas for the technology development of natural language processing,and has important value on scientific theory and application.Compared with existing ontology models,Event Ontology has remarkable features: The model takes event classes as basic knowledge units,which are more suitable to express dynamic knowledge.The model organizes knowledge with the hierarchical structure of event classes,which makes the knowledge structure more clear,and reduces the complexity of non-taxonomic relations among event classes.Therefore,Event Ontology can better express the dynamic semantic knowledge of natural language texts,and facilitate computers to analyze and understand texts.Event Ontology lays a solid foundation for the realization of an event-based natural language processing system.This dissertation presents a detailed analysis of the key issues in the construction of Event Ontology,and gives specific solutions for these issues.Corresponding experiments fully prove the feasibility of proposed methods.The main contents and novelties of this paper are as following:(1)Event recognition approach based on deep learning.Automatic event recognition is a critical foundation for constructing Event Ontology.In this dissertation,the event recognition problem is transformed into a classification problem based on feature vectors and deep learning as the most popular machine learning approach is employed to fulfill the classification.Then,the sentences in the corpus are segmented as words,and these words are classified according to the labels.These words are analyzed by feature analysis to transformed into feature vectors,and finally classified and recognized by a deep belief network(DBN)through the vectors.In addition,for addressing the existing unsupervised learning way for DBN,this dissertation also proposes two improved DBN(hybrid supervision and dynamic supervision).Among these two,the hybrid supervised DBN,through adding the supervised fine-tuning process for each restricted Boltzmann machine(RBM)layer after its' unsupervised training,so as to optimize the network parameters,and improve the recognition performance.The dynamic supervised DBN,by evaluating the training performance of each RBM layer,decides whether to add the supervised fine-tuning process.All of the above,two networks have improved the event recognition performance and enhance the stability of the system.The recognition approach proposed in this dissertation can achieve better recognition performance than the existing approaches,and can be extended to other event elements recognition,which makes a critical attempt to realize automatic annotation based on deep learning.At the same time,the research can also help the semi-automatic annotation of the corpus,accelerate the construction of the corpus,and provide technical support for the construction of large-scale event-oriented corpus.(2)Formal event analysis: Event Ontology takes the lattice structure as the basic architecture.Existing lattice structure mainly aims at static concepts and their attributes to construct a partial order structure.However the event has its dynamic characteristics and event elements have heterogeneous attributes.This dissertation analyzes and demonstrates the relevant elements of events,according to knowledge characteristics of elements,we constructs formal context description ways for different elements.The action element is described by a binary set,which contains two attributes of degree and direction.The formal description of the participant element is also a binary attribute set,consists of the initiator and receiver of an event.The time element is described by an interval.The environment,assertion and language expression are described in the form of a first-order predicate.Thus a heterogeneous formal event context is formed.In order to automatically generate the event lattice from a formal event context,this dissertation designs a progressive event lattice generation algorithm.The essence is that only the most recently generated event lattice nodes are traversed when a formal event is inserted into the whole event lattice,thus the search space of the algorithm is significantly reduced.At the same time,the location of a new generated event lattice node is determined and also the recently generated event lattice nodes are traversed.Massive experiments reveal that the algorithm can effectively realize automatic generation of heterogeneous event lattices,and consumes less time than existing lattice generation algorithms.(3)Event taxonomic relationship reasoning based on Event Ontology.The logical reasoning ability is a core embodiment of Event Ontology function.The reasoning ability directly affects the applicability and expansibility of Event Ontology.To overcome the limitation of dynamic representation ability of the existing description language(SROIQ),we proposes an event ontology description language(EO-SROIQ),which is applied to construct an event term set and an event assertion set.The two sets together constitute Event Ontology knowledge base EO ALCK-,which is the reasoning basis of Event Ontology.In the event taxonomic relationship reasoning,we employ the instance detection to determine the event class of an event(instance),then obtain the taxonomic relationship between events by relationship between their event classes.In the process of an event instance detection,this dissertation employs Extend-Tableau algorithm to execute class(concept)judgment for object and environment elements in the event,and detects EO ALCK--based consistency of action and assertion elements of the event.In addition,the event taxonomic reasoning algorithm is implemented by Java,and the decidability,completeness and rationality of the algorithm are verified in both theory and practice.