Research And Implementation On Recognition And Semantic Complement Of Chinese "?(de)" Structure

"?(de)",is one of the most frequent functional words in modern Chinese,and its usages are quite flexible.Attached to verbs,nouns,adjectives or phrases,"?(de)" can be used as a structural auxiliary word to form a attributive noun structure of adhesive type,called "?(de)" structure."?(de)" can be used as a modal auxiliary word,usually comes at the end of a sentence,to enhance the mood of the sentence as well."?(de)"structure enriches the hierarchical structures of Chinese sentences,which brings difficulties to text understanding.Researchers should pay attention to "?(de)" structure in the processing of massive Chinese texts.Correctly identifying the left and right boundaries of the "?(de)" structure and finding out the center which is omitted due to constraints of contexts and linguistic economic principles are basic tasks of Natural Language Processing(NLP).Dealing well with these tasks are beneficial for some upper-level applications,such as dependency parsing,Chinese Abstract Meaning Representation(CAMR)parsing,and machine translation.Therefore,based on CAMR corpus and Chinese Penn Tree Bank(CTB)corpus,we use methods based on neural network to carry out experimental researches on recognition and semantic complement of Chinese "?(de)" structure.Specifically,the main works of this thesis are as follows.A neural network method is proposed for automatic boundary recognition of the"?(de)" structure.To incorporate local features of contexts into word representations,densely connected convolutional neural network(DC-CNN)is used to flexibly extract variable n-gram features.And a multi-scale feature attention mechanism is utilized to organize these features in a reasonable way to solve contradictions between granularity and semantics.Densely connected bidirectional long short-term memory(DC-Bi-LSTM)is employed to incorporate global features of contexts into word representations.Based on these representations,the probabilities of boundaries are computed by the softmax function.Moreover,based on joint learning,the tasks of finding out left and right boundaries are completed at the same time.The experimental results verify the effectiveness of the proposed method.An automatic recognition method based on neural network for "?(de)" structure containing the usage of semantic ellipsis is proposed.First,stack Bi-LSTM is used to effectively incorporate contextual syntactic and semantic information into word representations.Second,we extract features of the "?(de)" structure and organize them by using Max-pooling layer or Gated Recurrent Unit(GRU)based multiple attention layers.Third,based on these features,we can recognize the elliptical "?(de)" structure.The experimental results show that the proposed model can achieve accurate results.And as the size of the corpus expands,it is possible to achieve better results.We propose a method based on neural network for automatically complement the omitted center of a elliptical "?(de)" structure.First,motivated by the annotation standard of CAMR that concepts in the Named Entity set can be used to represent the implied semantics of the sentence,we classify omitted centers of elliptical "?(de)"structures into several categories.That's to say,this task is defined as a multi-class task.Second,in our proposed model,DC-Bi-LSTM is employed to incorporate hierarchical and abstract semantic features of contexts into word representations.Moreover,referring to a common linguistic phenomenon of the elliptical "?(de)" structure,we utilize CNN to extract semantic features of n-grams in sentences.Third,based on global and local features mentioned above,we can find out the category of the omitted center by the softmax function.Experimental results show that our model can effectively complement categories for omitted centers of elliptical "?(de)" structures.