Research And System Construction Of Relation Classification And Text Generation Based On Deep Learning

With the explosive growth of data scale in the Internet in the big data era,we have been able to quickly and easily access information from a variety of tools.However,most of the information is lengthy,redundant and irrelevant,and users spend a lot of time eliminating useless information before they can actually get to their target content.Therefore,it becomes important to use text processing techniques that can automatically extract the core information.This thesis focuses on relation classification and text generation task.Relation classification aims to classify the entity pairs in the sentence into a certain relation,while text generation aims at text analysis,content summarization and generation.This work researches the application of deep learning and attention mechanism in relation classification and text generation,and proposes a relation classification model and text generation model based on the improved attention network.Main contributions of the research include:(1)For relation classification,the latest deep learning-based models still have shortcomings,i.e.,they cannot distinguish the semantic features in different contexts,and cannot make use of the hidden type information of the entities.The relation classification model proposed in this thesis consists of context encoder and entity-aware attention network.Contextual word semantics are learned through the self-attention mechanism.Entity selection is applied to adapt the fact that different entities can determine each other’s importance.Latent types of entities are taken as auxiliary information to make full use of the entities’ hidden features.(2)For text generation,this thesis focuses on the data-to-text generation task,i.e.,generate descriptions for structured data.Current approaches mostly use neural language models to learn alignment between output and input based on the attention mechanism,which is still flawed by the gradual weakening of attention when processing long texts and the inability to utilize the records’ structural information.The text generation model proposed in this thsis consists of field-content selective encoder and descriptive decoder.The text’s structure is fused into its representation and a gated content selector is applied to take advantage of the fact that different records can determine each other’s importance.In the decoding phase,the content selector’s semantic vector output enhances the alignment between output description and records through attention mechanism,and a featured copy mechanism is applied to solve the rare word problem.(3)A newscast system for weather forecasting and sports game is built based on the text generation model.The text-generation model converts the input data on weather and sports game into intuitive text that is made available to users;and the administrator is responsible for maintaining the system.For the users,the system provides quick and easy access to their needed news content,and for the administrator,it helps speed up the workflow.