Short Text Classification Of On-Demand Service Network Management Configuration Based On Deep Learning

In network management,professionals record unstructured forms of network management configuration cases which are extremely valuable to utilize.Event extraction can identify and extract event elements in unstructured case texts and organize these elements into structured information tuples to facilitate subsequent configuration analysis and improve the efficiency of network management.There are multiple descriptions of the same network management configuration event.Event coreference resolution aims to find the similarity matching between two event text descriptions,which can assist professionals in discovering the same event from multiple sources of information.There may be multiple interrelated events in the network management event text,and many professional terms are involved.Therefore,event extraction requires a combination of linguistic knowledge,deep learning techniques and text analysis methods.Event coreference resolution not only requires similarity matching between sentences,but also needs to consider the cross features between sentence pairs.For the above two tasks,the main contributions of this thesis include two main parts as follows:1.This thesis proposes a hybrid graph attention network-based event extraction algorithm for network management configurations.To address the problem of sparse features and difficulty in capturing semantic relations in sentence-level short texts,this thesis introduces text co-occurrence relation and syntactic dependency relation into the graph structure,and constructs text co-occurrence graph and syntactic dependency graph.Use two independent graph attention networks to update node features,so that the attention score can be updated dynamically,so as to better aggregate the feature information of event text.This thesis also proposes an adaptive fusion method to learn the importance of different features.The experimental results show that the algorithm improves the F1 score by 2.05%on the basis of the baseline model of the network management configuration event professional data set.2.This thesis proposes a data augmentation based biaffine attention coreference resolution algorithm for configuration processing text in the network management domain.The algorithm selects Roberta model for feature extraction.This thesis enhances the generalization of the model by means of various data augmentations such as:dual data,back translation,FGM for trigger words,EDA for negative examples and negative sample sampling.The text semantic similarity model based on Siamese network,such as Sentence-BERT,ignores the characteristic interaction information between entities.In order to solve this problem,this thesis proposes a fusion model of global and local information,and fusing trigger word consistency,event elements consistency,and event overall consistency for coreference relationship judgment.The biaffine attention mechanism is used for coreference classification to capture the semantic interaction features.Experimental results show that the algorithm improves the accuracy by 2.3%and the F1 score by 1.3%on network management data sets based on the baseline model.