| Semantic role labeling,as a form of shallow semantic analysis,plays a crucial role in enhancing language comprehension.This technique focuses on establishing the predicate-argument structure within a sentence,with predicates serving as the core,and identifying the associated arguments while describing the semantic structure through designated roles.Its primary aim is to facilitate deeper comprehension of natural language for machines.In recent years,driven by advancements in deep learning,semantic role labeling has garnered significant attention within the domain of natural language processing.Its importance has become increasingly evident,demonstrating practical value across various applications such as machine translation,information extraction,and question answering systems.However,Tibetan semantic role labeling faces several challenges.These include an inadequate annotation system,the lack of open-source datasets,the inability to leverage the semantic features of predicate categories,and a dependency on outdated methodologies and technologies.These challenges stem from the relatively late inception of research efforts and insufficient attention directed towards Tibetan natural language processing.Therefore,these limitations impede the advancement and practical application of Tibetan semantic role labeling technology,thereby hindering the overall progress of Tibetan natural language processing.Aiming at the problems and challenges faced by the research on semantic role labeling in Tibetan,this thesis draws on the work on semantic role labeling in English and Chinese,and from the perspective of machine learning,combines the unique grammatical structure and semantic expression of Tibetan,and puts forward four solutions,with a view to providing reference and guidance for further research on semantic role labeling in Tibetan:(1)A specification for semantic role labeling of lightweight Tibetan and a scheme for labeling practice;(2)A predicate semantic feature tagging method for Tibetan semantic role labeling;(3)An end-to-end fragmented Tibetan semantic role labeling method based on graph parsing;and(4)A predicate semantic-aware lightweight Tibetan semantic role labeling method.The main works of this thesis are as follows:Firstly,a lightweight Tibetan semantic role labeling specification and annotation practice scheme is proposed to address the problems of insufficient improvement of the Tibetan semantic role labeling system and the lack of open-source datasets for the time being.In the annotation specification,for various complex language scenarios,firstly,the factors that need to be considered in the development of the specification are summarised,and the selection range of Tibetan predicates is defined;Then,the semantic roles are classified and defined in detail;Finally,the clear annotation priorities of the argument words and an analysis of the difficulties,etc.,are given.In terms of annotation practice,firstly,following the developed specification,the construction of lightweight Tibetan semantic role labeling dataset is completed by carrying out annotation practice;Then,the Kappa coefficient is adopted to quantitatively assess the consistency and accuracy of the annotated data,and it is found that the lowest Kappa coefficient can be maintained at more than 79%,and the average Kappa coefficient can be up to 82.02%,which verifies the specification’s feasibility.Secondly,to address the problems that previous semantic role labeling methods do not adopt the model to automatically give predicates and do not fully utilize the semantic features of predicate categories,By defining Tibetan predicates and their categories of semantic features,a predicate semantic feature tagging method oriented to Tibetan semantic role labeling is proposed on the basis of constructing a Tibetan predicate semantic feature tagging dataset.The method accomplishes the design and training of the predicate semantic feature tagging model for Tibetan by drawing on the respective advantages of the Tibetan pre-trained language model(Ti Uni LM),Self-Attention mechanism(Self-Attention),Bidirectional Long and Short-Term Memory Network(BiLSTM),and Conditional Random Field(CRF)model.Experiments show that the accuracy rate(ACC)of Tibetan predicate semantic feature tagging of this method is as high as 98.69%,which can meet the requirements of practical applications.Then,to address the problems of single method and obsolete technology in Tibetan semantic role labeling research,an end-to-end segment(Span)Tibetan semantic role labeling method based on graph parsing is proposed by formulating a fragment-based Tibetan semantic role labeling system.The method skillfully converts the fragment-based Tibetan semantic role labeling into a word-based graph parsing task,and the process can be divided into two phases: semantic role labeling-to-graph conversion and graph-tosemantic role labeling recovery.In the first stage,the fragment-based Tibetan semantic role labeling are converted into attribute graphs for modeling by designing four graph annotation modes: BES,BE,BIES,and BIE,which in turn generates the graphs corresponding to the semantic role labeling.In the second stage,the generated graphs are recovered into the corresponding semantic role labeling by utilizing graph structure properties.Experiments show that the accuracy of this method can reach 91.85%,which is a significant performance improvement compared to the baseline method.Finally,to address the problem that the current form of Tibetan semantic role labeling is only limited to fragment-based,a lightweight Tibetan semantic role labeling method based on predicate semantic perception is proposed on the basis of introducing lightweight Tibetan semantic role labeling.Firstly,the method utilizes Ti Uni LM to dynamically generate rich semantic embedding vectors;Secondly,by designing a predicate semantic perception algorithm,the model is able to automatically perceive the predicate semantic features;Then,an ordered neuron long and short-term memory network(ON-LSTM)is employed to further model the temporal features and syntactic structural information;Finally,the CRF is employed to predict the optimal labeling sequence.Experiments show that the accuracy of the method can reach 92.78%,which provides a significant performance improvement compared to the baseline method. |
