Research On Multi-round Dialogue Generation Method For Patent Consultation

As one of the nine sectors of science and technology services,science and technology consulting is a key link connecting upstream scientific and technological achievements and downstream industry needs.It is an important support and objective demand for realiz-ing technological innovation to lead industrial upgrading and promote the economy to a mid-to-high-end level.The existing scientific and technological consulting methods are represented by ”multi-round dialogues”.The effectiveness of the results mainly depends on whether experts can accurately screen and match appropriate scientific and techno-logical resources.With the development of the science and technology service industry,the numerous,dynamic heterogeneous,diverse and complex science and technology re-sources make it more and more difficult to quickly and accurately screen them.This is in line with the timeliness of the consultant’s response to questions under the increasingly dynamic and changeable technology demand environment.The contradiction between usability requirements has become increasingly prominent.As a result,an all-weather,instant-available intelligent dialogue solution has become an urgent need and a hot re-search direction for scientific and technological consulting.Among them,patent consulting has an important supporting role in guiding and pro-tecting scientific and technological innovation and promoting the high-quality develop-ment of the physical industry.The intelligent dialogue program for patent consulting has become the primary task of scientific and technological consulting research.Because patent consulting content has a complex network dependency that is multi-dimensional,multi-granularity,and multi-field cross-integration,the existing single simple single-round dialogue,search-style multi-round dialogue and other modes are difficult to solve all the problems of the consultant.Therefore,it has become an urgent need and key task to form an intelligent patent consulting program that can replace the manual methods of experts,independently identify dialogue information and dialogue status,quickly understand the question of the consultant,match the appropriate scientific and technological resources,and generate usable answers.To this end,this article uses the national key research and development program topics ”Support for the development and application of an open and ecological enterprise-level cloud ERP platform”(topic number: 2019YFB1704104)and ”distributed resource giant system and resource synergy theory”(topic number: 2017YFB1400301)For the research background,focus on the subject based on the data-driven intelligent service technology to enhance the goal of integrating science and technology resources into the industry,and explore the sharing and service mode of science and technology resources.The unstructured patent resources in the Shield patent service platform are data-supported and are oriented to the field of patent consulting.In order to solve the problems of mutual autonomy,timeliness,and usability caused by excessive reliance on expert manual con-sultation in the current patent consulting service process,the key research The multi-round dialogue method of patent consulting proposes an end-to-end generative multi-round di-alogue overall technical scheme based on deep reinforcement learning.The program is composed of two parts: patent consulting dialogue material preprocessing and a multi-round dialogue model based on deep reinforcement learning.Researches are carried out on two aspects of patent consulting dialogue material preprocessing and patent consulting methods.The main contents are as follows:(1)In view of the large noise,multiple professional vocabulary,and long sequence of patent consulting corpus,the patent consulting corpus was preprocessed with stop words,noise removal,word segmentation,text vectorization,etc.,to build a patent consulting dialogue for the follow-up The model provides data quality assurance and standardization support for text data.(2)The multi-round dialogue model based on deep reinforcement learning proposed in this paper is composed of three parts: natural language generation module,natural lan-guage understanding module and dialogue management module.Among them,in order to solve the problems of the existing methods that are difficult to deal with the long se-quence dependence in the dialogue prediction and the memory forgetting in the process of encoding the sequence,a natural language generation model based on the dynamic residual network is proposed.According to the current decoding environment,the model dynamically selects an optimal state from a set of historical states to establish a connec-tion with the current state to improve the long-sequence dependency of LSTM.Since the introduction of dynamic residual connections will produce long-distance connection de-pendent words,we propose a new method based on reinforcement learning to simulate the dependence between words and the resulting long-distance dependence Introduced into the training process of the model.(3)In order to support the identification and understanding of user consultation ques-tions,a two-stage natural language understanding model is proposed,including one-stage recognition of user needs and intentions,and two-stage formation of a structured semantic framework to obtain slot sequences and sentences Level of intent.The natural language understanding model uses the IOB(in-out-begin)format to represent the slot label.The input is a dialogue sequence,and the output is a slot sequence and sentence-level intent.The goal of optimization is to maximize the slot under a given word sequence.Condi-tional likelihood of sequence and intent.(4)In order to verify the performance of the multi-round dialogue method based on deep reinforcement learning proposed in this paper,a series of experiments were carried out for verification.Among them,the performance of the overall architecture of the model is experimentally verified on the public dialogue data set,and the experimental results are analyzed;in order to verify that the natural language generation model based on the dy-namic residual network proposed in the natural language generation module in this paper is better than the existing model,Use the public data set of the domain to conduct com-parative experiments,and analyze the results of the experiments to show that the natural language generation model proposed in this paper is superior to the previous methods;in order to verify the performance of the two-stage natural language understanding model proposed in the natural language understanding module,patents are used Contrastive ex-periment analysis was conducted on the consulting dialogue data set to verify its perfor-mance.(5)In order to verify the effectiveness of the multi-round dialogue technical solution proposed in this article in the field of patent consulting,according to the characteristics of the patent consulting corpus involving specific domain knowledge,this article designs an evaluation index for the method of dialogue generation for patent consulting,and uses this indicator in the field of patent consulting.The patent consultation is a practical veri-fication of the multi-round dialogue technical solution for patent consultation proposed in this article on the discourse materials.The multi-round dialogue generation method for patent consulting proposed in this paper provides feasible solutions for the development of technology consulting towards autonomy and intelligence,and provides support for the realization of technology resource data sharing and services.