Semantic Understanding Of Chinese Short Sentences Based On Deep Learning

With the development of the Internet and the popularization of digital terminal equipments,Internet text data has been becoming an important source of information.The unstructured nature of text data makes it difficult for people to understand and use valuable information.Natural language processing provides an effective way,that quickly and accurately mines useful information from massive text data,which involves tasks such as text classification,information extraction,and semantic computing.However,natural language processing based on traditional machine learning algorithm has many disadvantages: first,it relies on extremely complex feature engineering to extract text features;then,traditional algorithms have difficult modeling flexible natural language;finally,pipelined model structure may lead to error propagation.The rise of deep learning has brought changes to natural language processing and solved the problems of traditional methods: first,the neural network model automatically learns text features through a large number of parameters;second,the flexible network structure makes it easier to model natural language,bringing new research methods to accomplish tasks such as machine translation;finally,the end-to-end training method avoids error propagation.Based on the task of mining post skills from the massive post recruitment texts on the Internet,the paper designs different deep neural networks to implement the semantic understanding of recruitment requirements.The main contents of this paper are as follows:(1)A multi-granular convolutional neural network(MGCNN)and bidirectional long short term memory network(Bi LSTM)combined network model(MGCNN-Bi LSTM)is proposed for text classification.The model first uses convolution kernels with different “Receptive Field” to extract text features and generate feature maps.Then,feature maps are further modeled by Bi LSTM to encode feature vectors with richer representation capabilities.Experiments show that the average F1-Score of the model is 93.56% in the four-category classification task of recruitment requirement texts,and it is 0.62% and 1% higher than MGCNN and Bi LSTM,respectively.(2)A Siamese Network using MGCNN-Bi LSTM model as presentation layer is proposed to calculate the semantic similarity of texts.In the experiment of post skills deduplication based on semantic similarity,the F1-Score of the Siamese Network using the MGCNN-Bi LSTM model as representation layer is 3.19%,0.04% and 0.76% higher than that of the Siamese Network using MLP model,CNN model and Bi LSTM model as representation layer,respectively.(3)The Pointer attention mechanism is introduced into the encoder-decoder model to build an end-to-end information extraction model.The model could synchronously perform two tasks of combining vocabularies in recruitment requirement texts to generate semantic chunks and predicting the categories of semantic chunks,which can avoid the error propagation caused by traditional pipelined model structure.In the task of generating semantic chunks,the model effectively utilizes the information of the semantic chunks such as length and boundary to improve the accuracy.Experiments show that the model has an average F1-Score of 90.42% in the task of generating semantic blocks and labeling their categories,which is 1.11% higher than traditional encoderdecoder model.(4)The encoder-decoder model based on Pointer attention mechanism is tranferred to other data(target data)with similar text features to the training data(source data)to solve the problem that the model cannot be trained due to the small size of the target data.Experiments show that if the text features such as vocabulary,grammar,or semantics of the target data are similar to the source data,the encoder-decoder model based on Pointer attention mechanism has an average F1-Score of 90.62% for post skills extraction on multiple target data,while the model has an F1-Score of 90.67% on the source data.