A Structured Data-to-Text Generation Method Based On Deep Neural Network

Structured data-to-text generation is a natural language processing method of converting structured data to corresponding descriptive text.In recent years,due to the application of end-to-end trained deep neural network,the method of structured datato-text generation has shown great potential in many scenarios,and it is often used in practical application scenarios such as news writing and report generation.However,the existing studies still have the following problems.Firstly,the feature representation is not effective,and there are big defects in the reasoning of data information such as specific values and time in the data,and the structural information between the data cannot be fully used to give reasonable guidance for generation.Secondly,it relies on large-scale datasets strongly.Due to the lack of public sample data for structured data to text generation task,deep model training cannot achieve good generalization performance on small-scale datasets.Finally,the model is applicable to a single domain and has weak migration ability.To address these problems,we study a structured data-to-text generation method based on deep neural network.The main work is as follows:(1)In order to fully capture the feature relationship among many structured data and enrich the representation information of semantic vector,we study an enhanced feature extraction method based on structured data-to-text generation.Firstly,a feature extractor with record encoding layer,content selection layer and content planning layer are constructed to obtain the feature sequences with enhanced feature information.Secondly,the feature sequences are transformed into contextual semantic vectors using an encoder of sequence-to-sequence framework.Finally,sequence generation is used to decode the accurate and coherent description text.The experimental results show that this method can effectively enhance the ability of structured data-to-text generation model to extract information from input source sequences.(2)In order to alleviate the limitation of the generalization ability of deep learning model by the size of datasets,and improving its common-sense understanding ability in different fields,we study a structured data-to-text generation method oriented to train the few-shot samples.Firstly,decoding with dual channel which is introduced the pretraining model,and it can expand the prior knowledge.Secondly,the generation adjustment strategy is used to smoothly switch the dual decoding channels.The experimental results show that this method makes the model have better fitting ability under the few training samples.It can expand the knowledge absorption capacity of the model.In addition,this paper design and implement relevant application based on the proposed method.The effectiveness and practical feasibility of the proposed method are verified through application tests with representative examples.