Automatic Recognition Of Chinese Compound Sentence Relation Words Based On Neural Network And Sentence Feature Fusion

Relational words are the nominal markers for compound sentence and an important component of the relationship in compound sentence.They greatly affect the semantic of the clause and the recognition of the hierarchical relationship of the compound sentence.The automatic recognition of relational words helps to clarify the grammatical components of sentence and the semantic of compound sentence,and it can improve the accuracy of machine translation.At the same time,compound sentence is bridges to connected texts,and the recognition of relative words is great significance to the study of texts.At present,there are rules-based and statistics-based methods for Chinese compound sentence relational words recognition,relying greatly on the rules summarized manually.This paper discusses the application of deep learning for relationship words recognition.Considering that the recognition methods of compound sentence relational words rely on the sentence features extracted manually,we propose a method for improving relation words automatic recognition application neural networks.This method fuses the features extracted from the compound sentence corpus into the word vector,and input the word vector into the constructed neural network model for training.In order to explore the effects of different combinations of sentence features and different word vectors on the recognition of relational words in the neural network model,the sentence analysis was conducted on the compound sentence in the Corpus of Chinese Compound Sentence(CCCS)using the Harbin Institute of Technology language technology platform(LTP),and four common sentence features were extracted to build a sentence feature library;Then we extract and combine the features from the sentence feature library,fuse the combined sentence features with CBW and BERT word vector matrices respectively,and input them into the neural network model for training;Finally,by analyzing the test results of different training models,we can find the best combination of different word vectors and sentence features.The proposed method is evaluated by constructing a training set and a test set.Without any dependence on the sentence features,the model that only input CBOW and BERT word vector matrix training has over 91%F1 values in the test set.At the same time,the F1 value reaches 92.52%while using sentence feature fusion.The experimental results show that the proposed method based on neural network and feature fusion takes advantage of both the automatic feature extraction in deep learning model and the obvious features summarized manually,improving the recognition efficiency and achieves a higher accuracy rate.