Key Information Extraction Of Sequence Data Based On Deep Neural Network

Sequence data is the most common data type in the field of natural language processing.A series of natural language processing tasks such as sentiment analysis,named entity recognition,machine translation,and intent recognition are researched based on sequence data.Researchers complete the classification or regression problem by building models to mine the key information in the sequence data.With the rise of deep learning,deep neural networks such as Long Short-Term Memory Networks and Transformer have achieved excellent performance in natural language processing tasks due to their end-to-end learning methods and their good timing information capture performance.With tremendous progress,this paper mainly explores the application of deep neural networks on three tasks,and improves the model from the input layer,hidden layer,and attention mechanism layer according to the characteristics of the task.The specific contents are as follows:1.Research of extraction of sequence key information on input layer.To solve the named entity recognition task whose target is contact information,a data labeling strategy is designed,and a sequence labeling network based on long-short term memory network and conditional random field is established.And the influence of the part-ofspeech and pinyin on the model is explored.Experiments prove that the two features of part-of-speech and pinyin have improved the effect of the model.The reasons for the success or failure of each contact method are analyzed,and the future planning is proposed based on the analysis results.2.Research of extraction of sequence key information on hidden layer.This part aims to explore students' consumption way based on the campus card big data.Based on deep neural networks and self-supervision,a Consume2 Vec model that can deeply mine the timing and correlation of consumer data is proposed,and a consumption anomaly detection model is built based on Consume2 Vec model.Experiments prove that the performance of two specific Consume2 Vec models for the mining of student sequence information is verified.The students are divided into different groups for comparison and analysis from different dimensions,and students' consumption laws and characteristics are found.3.Research on the extraction of sequence key information on the attention mechanism layer.The sentiment classification task under the semi-supervised attention mechanism is explored,and two sentiment analysis models SSA ABSA and ASSA ABSA are proposed.Two known sentiment analysis networks and four public data sets are used to conduct experiments.Experiments show that the human subjective attention mechanism has a promoting effect on the emotion classification,and the training trend of the model attention mechanism is related to the human subjective attention mechanism.