| As a key technology in the field of artificial intelligence,natural language pro-cessing can be divided into two basic sub-problems:1)natural language understanding(NLU),and 2)natural language generation(NLG).The primary core of solving NLU and NLG tasks is to obtain the semantic representation of natural language.The natural language text is composed of words,sentences,chapters in sequence,among which the quality of word representations will directly affect the quality of subsequent text repre-sentations.Since sentences are the simplest sequences of semantic units,the methods of semantic representation for sentences are often used to process short chapters.Cur-rently,the semantic representaion of words and sentences based on deep learning has achieved success.However,whether the word embeddings learned form textual corpus based on the distributed semantic hypothesis contain sufficient semantic information still needs further study.For the semantic representation of sentences in NLU,the re-current neural network cannot directly model the latent structure in sentences,so how to efficiently incorporate the latent structure into the sentence modelling is worth our exploring.For the semantic representation of sentences in NLG,how to enhance the effects of input conditions on the latent variable in variational autoencoders(VAEs)is important,and deriving controllable stylish latent variable is also worth further study.This thesis focuses on the semantic representation of natural language based on deep learning.It conducts studies on the semantic representation of words and sen-tences,including:First,the thesis studies the words representations with brain data.By analyzing the correlation between word vectors with the semantic vectors of other modals(i.e.,texutal semantic vectors,visual semantic vectors,and primary visual features)and the brain data,the experimental results confirm that the semantic information contained in the textual semantic word vectors is insufficient.Further,a method of estimating word embeddings with brain data is proposed,in which the brain data is converted to semantic similarity inequalities of word pairs,and these inequalities are used as constraints in the estimating process of texual word embeddings.The experiments show that the proposed method can derive word vectors with better correlation with human behavior data.Second,the thesis studies the sentence representation with latent structure descrip-tion.The commonly used recurrent neural networks are incapable of utilizing the latent structure in sentences,such as syntactic information.And the corresponding imporved tree-structured recursive neural networks have problems of relying on external syntax parser and low efficiency in computation.Therefore,a sequential neural encoder with latent structured description is proposed,which does not rely on any external parser and can learn latent chunking structure embedded in sentences in an end-to-end manner.The experiments show that the proposed model can improve performance on both the natural language inference and sentiment analysis tasks.Third,the thesis studies the sentence representation and conversation response gen-eration based on condition-transforming variational autoencoders.As the effects of the input conditions on latent variable is limited in the conditional variational autoen-coders,a condition-transforming variational autoencoder is proposed for generating di-verse short text conversations,in which the latent variable is sampled by performing a non-linear transformation on the combination of the input conditions and the sam-ples from a condition-independent prior distribution.The proposed methods can derive latent variable samples with better condition-dependency than that of the conditional variational autoencoders,and also improves the performance on generating diverse,in-formative,and topic relevant responses.Finally,the thesis studies the sentence representation and emotional conversation response generation based on emotion-regularized variational autoencoders(VAEs).As the ordinary conditional VAEs cannot generate emotional reponses directly by control-lable sampling of latent variable,and the emotion style and semantic content are natu-rally entangled in the language space.An emotion-regularized conditional variational autoencoder is proposed,in which an entangled latent distribution is adopted and the estimated latent variable is used to recover the emotion and the token sequence of the input response simultaneously.Experiments on short text conversations show that the proposed methods can learn more informative and structured latent space,and output responses with better content and emotion expressiveness. |
