Research On Data To Text Generation For Social Internet Of Things

Social Internet of Things(SIo T)is an emerging paradigm connecting the Internet of things and social networks,enabling the social functions expanded into human-to-thing or inter-thing.There is an urgent issue of how to realize the social contact for communications between human and things or between things.This requires the natural language generation(NLG),focusing on generating text in human languages from non-linguistic data,which is of vital importance to achieve the interaction without barriers between human and other smart objects in SIo T.However,the existing solutions for data-to-text transformation rely on the specific template and standard neural networks models,which are not flexible when applying to the large-scale structured data with diverse meaning.In the generated text from these models,there are many problems such as inconsistent content and poor narration.This paper introduces two pipeline assisted neural network models,which integrates the traditional pipeline modules and neural generation systems,to conduct data-to-text generation tasks in SIo T.The main research contents and innovation work of this paper are as follows:(1)The basic principle of sequence-to-sequence models are studied in depthIn this paper,the basic concepts of neural networks,recurrent neural networks and their variants are systematically introduced firstly,the general sequence-to-sequence framework and the later development of the framework in the data to text generation task is studied in depth,which lays the theoretical foundation for the subsequent research on data to text generation in the social internet of things.(2)A text generation model based on content selection of gating mechanism is proposedThe current data to text generation models only based on encoder-decoder framework,the user cannot directly control the selection and generation of content,the output text frequently refers to the same record multiple times or mentions erroneous records.Aiming at above problems,a data to text generation model based on content selection of gating mechanism is proposed in this paper,which integrated the content selection module in traditional methods and selectively generated the output content through the ingenious gating mechanism.The experimental results on the Roto Wire dataset show that the model has a great generation performance,which improves the accuracy of content selection,reduces the repetition rate of generation relation and effectively enhances the interpretability and controllability of the output text.The structured data will be transformed into natural language text,so as to realize the social interaction between human and intelligent objects in the field of social internet of things.(3)A text generation model based on surface realization of entity tracking is proposedIn order to solve the problem that model is difficult to find salient content in input data for generating high quality long text in large-scale datasets.A data to text generation model based on surface realization of entity tracking is proposed in this paper,which combines the content planning and surface realization modules in traditional methods.The experimental results on the Roto Wire dataset show that the model has the highest accuracy score of 94.82%in relation generation and improves content ordering scores by 16.03%.The content of the generated text is rich in content and high in fidelity,and the entity tracking method also makes the text more coherent,logical and readable,which is close to the human writing style.