Research On Dialog Generation Method Based On Sentence Similarity And Syntactic Structure

In recent years,as deep learning has been widely used in computer vision,neural machine translation and recommendation systems,and has achieved good results,the field of dialogue systems has also begun to apply deep learning techniques to improve the quality of dialogue systems.Compared with traditional methods,deep learning can obtain more semantic features and dialogue generation rules by learning a large amount of data with a minimum of manual input.At the same time,the rapid development and popularity of the Internet,the accumulated amount of interactive data has provided researchers with rich research materials.The current dialogue system can generally be divided into two types from the way of dialogue generation: the retrieval-based dialogue method and the generation-based dialogue method.At the application level of the actual system,the two methods can be combined organically to achieve complementary effects.Therefore,this paper has made in-depth research on these two methods.In the retrieval-based dialogue method,a traditional approach is to establish an offline index for all candidate sets.When the system receives the input information,a batch of candidate responses is initially recalled by the matching algorithm.The candidate responses are then further filtered as needed,ultimately returning results that meet the requirements.In this paper,by applying the training idea of multi-classification model to the sentence similarity task,the textual features are extracted by the cyclic neural network(RNN)structure and some special Softmax in face verification are applied in the training of the classification model,through the multi-classification model.The coding model of the text is trained to perform the sentence similarity calculation.Compared with the TF-IDF similarity calculation method,the sentence similarity calculation method based on word vector and the method based on deep learning,it is proved that the method proposed in this paper has better effect and stability than the existing method.And it provides more similarity calculation methods for search dialogue generation.In the generation-based dialogue method,in addition to the problem of generating a security response,the existing generation model has a large probability that the generated sentence will be insufficient in fluency.For example,the two verbs are connected together,the lack of structural auxiliary words,or the misplaced problems caused by improper position of structural auxiliary words.In order to solve this problem,this paper proposes two solutions.First,by means of techniques such as part-of-speech tagging and syntactic analysis,based on the SequenceToSequence model,the training process is improved by the method of part-of-speech transfer matrix to improve the problem that the generated sentence structure is not smooth.Second,after getting the generated model,use the BeamSearch method to generate N different responses.Using part-of-speech tagging and syntactic structure analysis techniques,the responses are reordered by calculating the structural confusion,and the reordered results are returned to improve the generation effect.