CN116955613B

CN116955613B - Method for generating product concept based on research report data and large language model

Info

Publication number: CN116955613B
Application number: CN202310689284.9A
Authority: CN
Inventors: 苏淦; 黄凯文; 张骏
Original assignee: Guangzhou Datastory Information Technology Co ltd
Current assignee: Guangzhou Datastory Information Technology Co ltd
Priority date: 2023-06-12
Filing date: 2023-06-12
Publication date: 2024-02-27
Anticipated expiration: 2043-06-12
Also published as: CN116955613A

Abstract

The application provides a method for generating product concepts based on research report data and a large language model, which comprises the following steps: acquiring the topic, the source and the release time of the report data, classifying report contents, constructing an industry knowledge base, and vectorizing the knowledge base; extracting the research report knowledge according to an industry knowledge base, and classifying the product concepts to be generated; performing prompt generation on the product attribute under each product functional characteristic classification according to the functional characteristics of the product to be generated; according to the use habit, skill and knowledge background of a product designer, the structured parameters of the campt suitable for different products are automatically recommended aiming at the functional characteristics of different products; and adjusting the structured parameters in the parameter list according to different user requirements to generate the Prompt content for accurately describing the product.

Description

Method for generating product concept based on research report data and large language model

Technical Field

The invention relates to the technical field of information, in particular to a method for generating product concepts based on research report data and a large language model.

Background

When a product designer performs creative design, accurate product concepts need to be generated according to the functional characteristics of the product. In the conventional product design process, a product designer needs to acquire market and industry information from various different information sources, analyze market and industry trends, and then start to generate product concepts. This process is time consuming and requires extensive market and industry knowledge, skills and experience from the product designer. Modern people have higher and higher requirements for product innovation, and along with explosive growth of internet information, research report data are also more and more abundant. However, in the face of such a vast amount of information, it is often difficult for product designers to quickly and accurately obtain the required knowledge, and to convert it into specific product concepts. Product designers need to propose new product concepts in a short time, while also considering market demands and industry trends. It is critical for the designer to obtain industry knowledge and product information, but obtaining knowledge in different industries takes a lot of time and effort. How to improve the working efficiency and innovation ability of product designers becomes a urgent problem to be solved

Disclosure of Invention

The invention provides a method for generating product concepts based on research report data and a large language model, which mainly comprises the following steps:

the method comprises the steps of obtaining the topic, the source and the release time of the report data, classifying report contents, constructing an industry knowledge base, and vectorizing the knowledge base, and specifically comprises the following steps: based on a word2vec model, converting the report content into vector representation; extracting the research report knowledge according to an industry knowledge base, and classifying the product concepts to be generated; performing prompt generation on the product attribute under each product functional characteristic classification according to the functional characteristics of the product to be generated; according to the use habit, skill and knowledge background of a product designer, the method automatically recommends the structuring parameters of the campt suitable for different products according to the functional characteristics of different products, and specifically comprises the following steps: recommending a simple structured parameter list of designers similar to the past works based on a cosine similarity calculation method; according to different user demands, structural parameters in a parameter list are adjusted to generate the Prompt content for accurately describing the product, and the method specifically comprises the following steps: clustering users based on a collaborative filtering algorithm of the users, and determining the structural parameter values of different user demands; updating an industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base, judging whether the new creative field exceeds the applicable range of the current prompt parameter, updating the industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base, and judging whether the new creative field exceeds the applicable range of the current prompt parameter, wherein the method specifically comprises the following steps: based on a support vector regression model, predicting market potential of a new creative field, constructing a decision tree, and judging whether technical implementation of the new field is realized through technical implementation owned by a current research product; adding new campt structural parameters according to the new creative field and applying the new campt structural parameters to creative design, wherein the new campt structural parameters comprise: based on the decision tree algorithm, the constraint conditions of the structural parameters are determined, and the structural parameters are matched and coordinated with the service attributes.

Further optionally, the acquiring the topic, the source and the release time of the report data, classifying the report content, constructing an industry knowledge base, and vectorizing the knowledge base includes:

inputting keywords by adopting a network search engine to acquire topics of the research report data; acquiring research data sources through the published website information and industry information; confirming the release time of the research report by a research institution website and a bulletin mode; classifying the research reports according to research industries, and marking corresponding classification labels for each research report; a crawler tool is adopted to acquire the research report data, and a data model of an industry knowledge base is established, wherein the data model comprises the topic, the source, the release time and the classification information of the research report data; establishing a data table structure of an industry knowledge base by using database software, and importing research report data; converting the report content into vector representation through a word2vec model; vectorized relevant attributes include a report text, keywords, authors, institutions, and citations; comprising the following steps: based on a word2vec model, converting the report content into vector representation;

the word2vec model-based method for converting the report content into the vector representation specifically comprises the following steps:

And acquiring the report information, cleaning, word segmentation and coding the report text, keywords, authors, institutions and data of the cited documents, and constructing a corresponding data set. Constructing a word2vec model by using a gemim deep learning framework; the codes of the input report text, keywords, authors, institutions and citations are converted into vectors of corresponding dimensions through an embedding layer. The input vector representation is input into a model, and a context-dependent vector representation of each datum is obtained through a multi-layer self-attention mechanism and a forward neural network. And inputting the fused vector representation into a full connection layer to obtain a final vector representation.

Further optionally, the extracting the report knowledge according to the industry knowledge base, and classifying the product concept to be generated includes:

firstly, determining products to be generated and related keywords; inputting keywords related to products to be generated into an industry knowledge base for searching, and screening out research reports related to the products to be generated; for each report, extracting key information by using a natural language processing technology, wherein the key information is a product concept and comprises positioning, functional characteristics and target users of the product; and sorting and classifying the extracted key information to form a product concept classification document.

Further optionally, the generating the product attribute according to the functional characteristics of the product to be generated according to the requirement includes:

according to the product concept classification documents, determining the functional characteristics of the product to be generated; determining corresponding product attributes according to functional characteristics of the product to be generated based on the existing open source knowledge graph; collecting product attribute data in product specifications, comments and news reports, establishing a corpus for different product attributes by utilizing data mining and natural language processing technologies, analyzing key words and phrases in the corpus, and determining corresponding campt words of corresponding product attributes; constructing a vocabulary, and allocating a unique integer number to each word; based on a convolutional neural network model, performing deep learning and model training on the product attribute; inputting product attributes, and automatically identifying and recommending a prompt vocabulary corresponding to the proper product attributes by the trained model; and combining different campt vocabularies according to the requirements of the user, and generating campt content aiming at the functional characteristics of each product by using a natural language generation technology.

Further optionally, the automatically recommending the structured parameters of the campt suitable for different products according to the usage habits, skills and knowledge backgrounds of the product designers and aiming at the functional characteristics of the different products includes:

Firstly, obtaining custom design elements and styles of a designer through past works and design styles of the designer; determining familiar software, design theory and skill through histories and working experience of a designer; recommending design elements and styles suitable for the skill and knowledge background of a designer according to the skill and knowledge background of the designer; determining functional characteristic parameters of the product according to industries, target users and functional characteristics of the product; determining tasks and operation flows which need to be completed by a designer according to the functional characteristics of the product, wherein the tasks and operation flows comprise data input, operation selection and result checking; designing corresponding campt content, including parameter names, value ranges, default values, input modes and verification rules; carrying out structuring treatment on the campt content by using a natural language processing technology to generate a corresponding parameter list; comprising the following steps: recommending a simple structured parameter list of designers similar to the past works based on a cosine similarity calculation method;

the cosine similarity calculation method is used for recommending a template structured parameter list of designers similar to the past works, and specifically comprises the following steps:

obtaining design elements, style software, design theory and skill data of past works of a designer as characteristic data, and carrying out structural processing on the obtained characteristic data; performing similarity calculation on past works of a designer by adopting a cosine similarity calculation method to obtain a similarity matrix; and finding out the designer most similar to the current designer according to the similarity matrix, selecting the past works as recommended objects, and obtaining a sample structured parameter list of the works. A list of campt structured parameters for similar designer works is recommended to the designer.

Further optionally, the adjusting the structured parameters in the parameter list according to different user requirements, and generating the promt content accurately describing the product includes:

obtaining portraits and demands of different user groups through a user investigation and data analysis tool; clustering the users based on a collaborative filtering algorithm of the users according to the historical behavior data of the users, and determining the structural parameter values of different user requirements; judging the structural parameters of the product promt to be adjusted according to the determined structural parameter values, and adjusting the structural parameters of the product promt; generating a promt content for accurately describing the product according to the adjusted parameter list; comprising the following steps: clustering users based on a collaborative filtering algorithm of the users, and determining the structural parameter values of different user demands;

the collaborative filtering algorithm based on the user clusters the users and determines the structural parameter values of different user demands, and the method specifically comprises the following steps:

acquiring historical prompt input data of a user; preprocessing the collected user behavior data, including removing repeated data, processing missing values and normalizing; and calculating the similarity between each two users by using a cosine similarity method according to the user behavior data. Dividing users into K clusters by using a K-means clustering algorithm, wherein each cluster represents a user group, and clustering the users according to the similarity; according to the clustering result, analyzing the user behavior data of each cluster, finding out common points and difference points of the user behavior data, determining the demands of different user groups, determining the structural parameter values of different user demands, and adjusting the structural parameters in the parameter list.

Further optionally, updating the industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base, and determining whether the new creative field exceeds the applicable range of the current campt parameter includes:

updating an industry knowledge base, and acquiring a latest research report and industry research, wherein the latest research report and the industry research comprise newly added research industry and expansion of the existing industry; evaluating and classifying the existing research products, classifying the research products into corresponding creative fields according to the characteristics and application scenes of the research products, and adjusting the classification of the creative fields according to the classification of the research products and the development trend of industry research; for the newly added research industry, the new research industry is classified into the corresponding creative field, and information on industry definition, research range and development trend is acquired; judging whether the new creative field is suitable for the current prompt parameter, if not, adjusting the parameter range; the information of the newly added research industry and the adjusted creative field is arranged and filed; comprising the following steps: predicting market potential of the new creative field based on a support vector regression model; constructing a decision tree, and judging whether the technical implementation in the new field is realized through the technical implementation owned by the current research product;

The method for predicting the market potential of the new creative field based on the support vector regression model specifically comprises the following steps:

user demand description data for new creative field products is collected from various channels and cleaned. And performing word segmentation, stop word removal and word frequency statistics on the requirement description text by adopting a natural language processing technology, and determining keywords. Carrying out vectorization on demand description according to keywords, and establishing a model by adopting a K-means clustering algorithm; and judging the user demand characteristics of each category according to the clustering result and the characteristics of the clustering center. And comparing the characteristics and advantages of the new field with the user demand data to judge whether the new field meets the user demand aimed at by the current research product. And using a support vector regression model in an SVM algorithm, taking the characteristics and advantages of the new field as independent variables, taking market demands and competition conditions as target variables, and predicting the market potential of the new field.

The decision tree is constructed to judge whether the technical implementation in the new field is realized through the technical implementation owned by the current research product, and the method specifically comprises the following steps:

and collecting business attribute contents, including technical implementation of current research products and technical implementation of new fields. And constructing a decision tree, taking the technical implementation of the current research product as a root node, and dividing the tree into two branches which are feasible and infeasible according to whether the technical implementation of the new field is realized through the technical implementation owned by the current research product. The accuracy of the decision tree is assessed, the data set is divided into a training set and a test set using a cross-validation method, the decision tree is trained using the training set, and the accuracy of the decision tree is assessed using the test set. And judging whether the creative implementation in the new field is feasible or not according to the result of the decision tree.

Further optionally, the adding new campt structuring parameters according to the new creative field and applying to the creative design includes:

determining the type of data to be used according to the latest research report and the data source of industry research; setting corresponding limiting conditions for the structural parameters according to limiting conditions of the data source, wherein the limiting conditions comprise a minimum character set, a maximum character set and an allowed character set; then, determining the display format of the structured parameters, acquiring data in a data source, and converting the data into the structured parameters; judging whether the structured parameters are related to other attributes or not, if the structured parameters are linked with another structured parameter or combined with the data set, carrying out corresponding processing; applying the new campt structuring parameters to the creative design; comprising the following steps: determining the limiting conditions of the structural parameters and matching and coordinating with the service attributes based on a decision tree algorithm;

the decision tree algorithm-based method for determining the limiting conditions of the structural parameters and matching and coordinating with the service attributes specifically comprises the following steps:

and acquiring a characteristic attribute and a sample attribute, wherein the characteristic attribute comprises a numerical value type. And determining the category to which each sample belongs through the sample attribute, and determining the attribute of the decision tree, wherein the attribute comprises the attribute type, the value range, the requirement of classification accuracy and the limiting condition. A root node is selected, which corresponds to all data samples. And selecting an attribute, classifying the data samples according to the attribute value, and constructing the child nodes of the attribute until the data samples under all the child nodes belong to the same category or all the attributes are used. And traversing the tree step by step from the root node downwards according to the attribute value on the node until the leaf node is reached, wherein the category represented by the leaf node is the classification result of the new data. And determining the limiting conditions of the structural parameters, matching and coordinating with the service attributes to obtain a decision tree classifier, classifying the new data, and judging the category to which the new data belong.

The technical scheme provided by the embodiment of the invention can have the following beneficial effects:

according to the invention, the topics, sources and release time of the report data are acquired, the report content is classified, and an industry knowledge base is constructed, so that the vectorization and extraction of knowledge in different industries are realized. And extracting the research report knowledge according to the industry knowledge base aiming at the product concepts needing to be generated, and further classifying the product concepts. And generating a prompt according to the functional characteristics of the product to be generated and aiming at the product attribute of each product under the functional characteristic classification. According to the use habit, skill and knowledge background of a product designer, the structured parameters of the campt suitable for different products are automatically recommended according to the functional characteristics of different products, and the campt content accurately describing the products is generated. And meanwhile, according to different user requirements, structural parameters in the parameter list are adjusted, so that more accurate Prompt content is generated. And in the process of updating the industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base. If the new creative field exceeds the applicable range of the current campt parameters, new campt structural parameters are added and applied to creative design. In a word, through the technology, extraction and vectorization of knowledge of different industries are achieved, and the Prompt content for accurately describing the product is generated, so that a product designer is helped to achieve creative design better.

Drawings

FIG. 1 is a flow chart of a method of generating product concepts based on report data and a large language model according to the present invention.

FIG. 2 is a schematic diagram of a method of generating product concepts based on report data and a large language model according to the present invention.

Detailed Description

The technical solutions of the present invention will be clearly and completely described in connection with the embodiments, and it is obvious that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The method for generating product concepts based on the report data and the large language model in the embodiment specifically comprises the following steps:

and step 101, acquiring the topic, the source and the release time of the report data, classifying the report content, constructing an industry knowledge base and vectorizing the knowledge base.

And inputting keywords by adopting a network search engine to acquire the topics of the research data. And acquiring the research data source through the published website information and industry information. And confirming the release time of the research report by a research institution website and a bulletin mode. Classifying the research reports according to research industries, and marking corresponding classification labels for each research report; a crawler tool is adopted to acquire the research report data, and a data model of an industry knowledge base is established, wherein the data model comprises the topic, the source, the release time and the classification information of the research report data; and using database software to establish a data table structure of an industry knowledge base and importing the research report data. The report content is converted into a vector representation by a word2vec model. Vectorized relevant attributes include the report text, keywords, authors, institutions, citations. For example, a keyword "Chinese electronic consumer goods market" is input, a plurality of research reports can be obtained through a search engine, and one of the research reports is selected as an analysis object. The report distribution time is confirmed by the website and announcement mode of the research institution, for example, the distribution time is 2021, 5 months. The research report is classified according to the research objects, and can be classified into consumer goods industry, electronic industry and the like, and corresponding labels such as an electronic consumer goods market, 2021 and Chinese market are marked on the research report. The report data is obtained through a crawler tool, manual screening is carried out, junk information is removed, valuable data is reserved, a report identification library is constructed, and the topic, source, release time and classification information attributes of the report data are stored. The report content is converted into vector representation through a word2vec model, and vectorized relevant attributes such as report text, keywords, authors, institutions and citations are saved.

Based on the word2vec model, the content of the study report is converted into a vector representation.

And acquiring the report information, cleaning, word segmentation and coding the report text, keywords, authors, institutions and data of the cited documents, and constructing a corresponding data set. Constructing a word2vec model by using a gemim deep learning framework; the codes of the input report text, keywords, authors, institutions and citations are converted into vectors of corresponding dimensions through an embedding layer. The input vector representation is input into a model, and a context-dependent vector representation of each datum is obtained through a multi-layer self-attention mechanism and a forward neural network. And inputting the fused vector representation into a full connection layer to obtain a final vector representation. For example, there are 10 keywords, 5 authors, 3 institutions, and 8 citations, and then the encoded vector dimensions are 10, 5, 3, and 8, respectively. For the input text, word embedding is used to convert it into vectors of corresponding dimensions. Finally, the encoded vector is input into the wrd vec model for context-dependent representation learning. Assuming 128-dimensional vectors are used, the hidden layer size is 128 and the number of self-attention heads is 8. A text vector representation is obtained that fuses all data information. And finally, inputting the fused vector representation into a full connection layer to obtain a final vector representation. Assuming a full link layer size of 128 dimensions, the final vector is denoted as [0.2,0.1,..0.3 ], where the dimension of the vector is 128, which characterizes the report content.

And 102, extracting the research report knowledge according to an industry knowledge base, and classifying the product concepts to be generated.

Firstly, determining products to be generated and related keywords; and inputting keywords related to the products to be generated into an industry knowledge base for searching, and screening out research reports related to the products to be generated. For each report, extracting key information by using a natural language processing technology, wherein the key information is a product concept and comprises positioning, functional characteristics and target users of the product; sorting and classifying the extracted key information to form a product concept classification document; for example, it is necessary to generate an intelligent contact lens, first input the keyword "intelligent contact lens" for searching, and find a report. Then extracting key information by using a natural language processing technology; and then sorting the extracted key information to form the following documents: intelligent contact lens market research reports product concepts: product positioning: the intelligent contact lens has the characteristics of intelligence, portability and comfort. Functional characteristics: the health condition of eyes is monitored in real time, eye health care suggestions are provided, and lenses can be automatically adjusted. Target user: people who often wear contact lenses, especially those who have high working pressures and use electronic products for a long time, are required.

Step 103, generating a prompt for the product attribute under each product functional characteristic classification according to the functional characteristics of the product to be generated.

According to the product concept classification documents, determining the functional characteristics of the product to be generated; determining corresponding product attributes according to functional characteristics of the product to be generated based on the existing open source knowledge graph; collecting product attribute data in product specifications, comments and news reports, establishing a corpus for different product attributes by utilizing data mining and natural language processing technologies, analyzing key words and phrases in the corpus, and determining corresponding campt words of corresponding product attributes; a vocabulary is built and each word is assigned a unique integer number. Based on a convolutional neural network model, performing deep learning and model training on the product attribute; inputting product attributes, and automatically identifying and recommending a prompt vocabulary corresponding to the proper product attributes by the trained model; combining different campt vocabularies according to the requirements of a user, and generating campt contents aiming at the functional characteristics of each product by using a natural language generation technology; for example, it is desirable to develop a smart audio product and to determine its functional characteristics and corresponding product attributes. Determining the functional characteristics of the product includes: voice control, music playing, smart home control, weather inquiry. According to the existing knowledge graph, determining the corresponding product attribute comprises: speech recognition, audio decoding, home automation, weather forecast. Relevant product specifications, comments and news reports are collected, a corpus is established by utilizing data mining and natural language processing technologies, and key words and phrases are analyzed. For example, keywords related to voice control include: voice recognition, voice instructions, voice interactions, voice assistants; keywords related to music playback include: audio decoding, sound effect optimization, music library, online music. For example, when a user needs to generate a promt content for the music playing function feature, the following contents are combined according to the promt vocabulary recommended by the model: playing popular songs on the intelligent sound equipment, opening an online music platform and adjusting sound effect optimization.

Step 104, according to the usage habit, skill and knowledge background of the product designer, the structured parameters of the campt suitable for different products are automatically recommended aiming at the functional characteristics of different products.

Firstly, the custom design elements and styles are obtained through past works and design styles of designers. The familiar software, design theory and skill are determined by the histories and working experience of the designer. Based on the skill and knowledge background of the designer, design elements and styles are recommended that fit the skill and knowledge background. And determining functional characteristic parameters of the product according to the industries, target users and functional characteristics of the product. Determining tasks and operation flows which need to be completed by a designer according to the functional characteristics of the product, wherein the tasks and operation flows comprise data input, operation selection and result checking; designing corresponding campt content, including parameter names, value ranges, default values, input modes and verification rules; carrying out structuring treatment on the campt content by using a natural language processing technology to generate a corresponding parameter list; for example, a designer needs to design the home page of an e-commerce web site. Based on their past works and design styles, designers were found to prefer to use plain and plain colors and flattened elements. According to the histories and working experience, familiar software of the user is determined to be Sketch and Photoshop, familiar design theory is planar design and user experience design, and mastered skills are color use and typesetting. Thus, design elements and styles that are better suited to their skill and knowledge background are recommended, favoring planarized icons and succinct and clear fonts. According to the industry of products, namely electronic commerce, the target user is young, and the functional characteristics of the products are browsing goods, purchasing goods, checking orders and the like. Therefore, the task and the operation flow which need to be completed by the designer are the layout, color matching, fonts and the like of the designed website home page, and the tasks comprise data input, operation selection and result checking; the corresponding prompt content is designed to comprise a parameter name, a value range, a default value, an input mode and a verification rule, for example, the page width is 1200px, the font color is black, the button color is red, and the commodity display quantity is 8. The simplet content is structured by using natural language processing technology to generate corresponding parameter lists, such as page width (1200 px), font color (black), button color (red) and commodity display quantity (8).

Based on the cosine similarity calculation method, a simplet structured parameter list of designers similar to the past works is recommended.

Obtaining design elements, style software, design theory and skill data of past works of a designer as characteristic data, and carrying out structural processing on the obtained characteristic data; performing similarity calculation on past works of a designer by adopting a cosine similarity calculation method to obtain a similarity matrix; and finding out the designer most similar to the current designer according to the similarity matrix, selecting the past works as recommended objects, and obtaining a sample structured parameter list of the works. Recommending a simple structured parameter list of similar designer works for a designer; for example, the similarity between designer a and designer C is calculated to be 0.408, and the similarity between designer B and designer C is calculated to be 0. Thus, the designer most similar to the current designer A is designer C, for whom a list of the simple structured parameters of the work that designer C has passed is recommended.

And step 105, adjusting the structured parameters in the parameter list according to different user requirements to generate the Prompt content for accurately describing the product.

Obtaining portraits and demands of different user groups through a user investigation and data analysis tool; and clustering the users based on a collaborative filtering algorithm of the users according to the historical behavior data of the users, and determining the structural parameter values of different user requirements. Judging the structural parameters of the product promt to be adjusted according to the determined structural parameter values, and adjusting the structural parameters of the product promt; generating a promt content for accurately describing the product according to the adjusted parameter list; for example, a product is an online shopping platform, and user groups are found to fall into two categories through user research and data analysis tools: population of young people: mainly purchasing fashionable and personalized clothing and accessories; and the housewives mainly purchase household articles and daily necessities. Based on historical behavior data of users, it is found that young people groups prefer to select low-priced commodities, and housewives groups pay more attention to quality and use experience. Thus, we need to adjust to the needs of different user groups: for the young population, more lower priced merchandise is offered and price and discount information is highlighted in the product template. For housewives, descriptions of commodity quality and use experience are enhanced, and key information of materials, sizes and use modes is highlighted in a product project. Based on the above analysis, the structured parameters of the product template are adjusted, such as adjusting the display location of price and discount information, increasing the length and details of the commodity description.

And clustering the users based on a collaborative filtering algorithm of the users, and determining the structural parameter values of different user requirements.

Acquiring historical prompt input data of a user; preprocessing the collected user behavior data, including removing repeated data, processing missing values and normalizing; and calculating the similarity between each two users by using a cosine similarity method according to the user behavior data. Dividing users into K clusters by using a K-means clustering algorithm, wherein each cluster represents a user group, and clustering the users according to the similarity; according to the clustering result, analyzing user behavior data of each cluster, finding out common points and difference points of the user behavior data, determining requirements of different user groups, determining structural parameter values of different user requirements, and adjusting structural parameters in a parameter list; for example, there are 1000 users who input the prompt data. And acquiring historical prompt input data of each user, and calculating the similarity between each user by using a cosine similarity algorithm. Next, the users are divided into K clusters using a K-means clustering algorithm. Finally, the user behavior data of each cluster is analyzed to find their common points and difference points, e.g., users who find some clusters prefer to purchase high value goods, while users of other clusters pay more attention to price offers. For example, a community of users has been found to prefer to purchase high value goods, and their shopping records generally contain higher priced goods, such as luxury goods, high-end electronics. This community of users is defined as "high-end consumers" and the structural parameters of the platform are adjusted according to their needs, e.g. to increase recommendations of high-value goods, improve the distribution services of the platform.

And step 106, updating an industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base, and judging whether the new creative field exceeds the applicable range of the current prompt parameter.

Updating an industry knowledge base, and acquiring a latest research report and industry research, wherein the latest research report and the industry research comprise newly added research industry and expansion of the existing industry; the existing research products are evaluated and classified, the research products are classified into corresponding creative fields according to the characteristics and application scenes of the research products, and the classification of the creative fields is adjusted according to the classification of the research products and the development trend of industry research. And (3) for the newly added research industry, the new research industry is classified into the corresponding creative field, and information in the aspects of industry definition, research range and development trend is acquired. And judging whether the new creative field is suitable for the current prompt parameter, and if not, adjusting the parameter range. And (5) arranging and archiving the information of the newly added research industry and the adjusted creative field. For example, updating industry knowledge base, recent research reports and industry research show a new creative area, named "smart home", with a research scope including smart home device design, manufacturing, sales, and smart home related services. Among the existing research products, some products can be classified into the field of intelligent home, such as intelligent bulbs, intelligent sockets and intelligent door locks. These products are evaluated and classified into the corresponding creative fields. Meanwhile, according to the classification of research products and the development trend of industry research, the classification of the creative field is adjusted, for example, the intelligent home is subdivided into the sub-fields of intelligent illumination and intelligent security. If the new creative field 'intelligent home' is suitable for the current prompt parameter, the current prompt parameter is added into the parameter range, and the development trend of the field is continuously monitored. For example, the campt parameter is "creative field: intelligent device ", expand it into" creative field: intelligent devices (including smart home, smart wear) ", to include the" smart home "field. And (5) arranging and archiving the information in the field of intelligent home.

Based on the support vector regression model, market potential of the new creative field is predicted.

User demand description data for new creative field products is collected from various channels and cleaned. And performing word segmentation, stop word removal and word frequency statistics on the requirement description text by adopting a natural language processing technology, and determining keywords. Carrying out vectorization on demand description according to keywords, and establishing a model by adopting a K-means clustering algorithm; and judging the user demand characteristics of each category according to the clustering result and the characteristics of the clustering center. And comparing the characteristics and advantages of the new field with the user demand data to judge whether the new field meets the user demand aimed at by the current research product. Using a support vector regression model in an SVM algorithm, taking characteristics and advantages of the new field as independent variables, taking market demands and competition conditions as target variables, and predicting market potential of the new field; for example, a new smart home device is to be brought to the market and it is desired to predict the performance of this product in the future market. A large amount of demand description data is collected from user feedback and market research reports, including "intelligent control home appliances", "automated safety system", "energy saving and environmental protection", "speech recognition". First, these demand descriptions are cleaned and analyzed. And (3) performing word segmentation on the text by using a natural language processing technology, removing stop words, counting word frequencies, and determining keywords. Finally we get 10 keywords: intelligent, control, home appliances, automation, security, system, energy saving, environmental protection, voice, recognition. Next, demand description is vectorized, and a model is built by adopting a K-means clustering algorithm. Setting the clustering number as 3, and classifying the demand descriptions into 3 categories: intelligent home control, safety guarantee, energy conservation and environmental protection. According to the clustering result and the characteristics of the clustering center, the user demand characteristics of each category can be obtained: users of the intelligent home control class pay more attention to convenience and intelligent degree; and then, comparing the characteristics and advantages of the new field with the user demand data to judge whether the new field meets the user demand aimed at by the current research product. The new field is assumed to be intelligent fitness equipment, and keywords such as intelligent control, voice recognition and the like are found to have certain correlation with the intelligent fitness equipment. Finally, a support vector regression model in an SVM algorithm is used, characteristics and advantages of the new field are used as independent variables, market demands and competition conditions are used as target variables, and market potential of the new field is predicted. Assuming that the predicted result is $ 1000, this means that our new field has great market potential and can be considered to bring the product to market.

And constructing a decision tree, and judging whether the technical implementation in the new field is realized through the technical implementation owned by the current research product.

And collecting business attribute contents, including technical implementation of current research products and technical implementation of new fields. And constructing a decision tree, taking the technical implementation of the current research product as a root node, and dividing the tree into two branches which are feasible and infeasible according to whether the technical implementation of the new field is realized through the technical implementation owned by the current research product. The accuracy of the decision tree is assessed, the data set is divided into a training set and a test set using a cross-validation method, the decision tree is trained using the training set, and the accuracy of the decision tree is assessed using the test set. Judging whether creative realization in the new field is feasible or not according to the result of the decision tree; for example, the product of current research is an intelligent home system, and its technical implementation includes functions such as voice recognition, remote control, etc. The technology implementation in the new field is a face recognition technology based on machine learning, and the face recognition technology is used for recognizing family members and automatically adjusting the setting of a home system. Constructing a decision tree: the technology of the intelligent home system is used as a root node, if the face recognition technology in the new field can be realized through the existing voice recognition technology and the like, the technology is judged to be feasible, otherwise, the technology is judged to be not feasible. Evaluating the accuracy of the decision tree: the data set is divided into a training set and a testing set, the decision tree is trained using the training set, and the accuracy of the decision tree is assessed using the testing set. Assume 100 samples in the dataset, 70 of which are used for training and 30 of which are used for testing. Judging whether creative realization of the new field is feasible or not according to the result of the decision tree: according to the result of the decision tree, if the face recognition technology in the new field can be realized by the existing voice recognition technology and the like, the face recognition technology is judged to be feasible, otherwise, the face recognition technology is judged to be not feasible.

Step 107, adding new template structural parameters according to the new creative field, and applying the new template structural parameters to creative design.

The type of data that needs to be used is determined based on the latest research reports and the data sources of the industry research. And setting corresponding limiting conditions for the structural parameters according to the limiting conditions of the data source, including minimum and maximum values and allowed character sets. The display format of the structured parameters is then determined, and the data in the data source is acquired and converted into structured parameters. And judging whether the structural parameters are related to other attributes, and if the structural parameters are linked with another structural parameter or combined with the data set, performing corresponding processing. Applying the new campt structuring parameters to the creative design; for example, a structured parameter is added to a report for selecting a time range for display. A time field is provided in the data source in the format YYYY-MM-DD, with a minimum value of 2010-01-01 and a maximum value of 2022-12-31. The constraints for setting the structuring parameters are: minimum 2010-01-01 and maximum 2022-12-31, only dates in YYYY-MM-DD format are allowed to be entered. Next, a display format of the structured parameters is determined, and a calendar control or text box is selected for entering a date. When the data in the data source is acquired, the time field is converted into a date format and is used as an option for structuring parameters. For example, with all dates in the time field as options for structuring parameters, the user may select any one date. If the structured parameter is associated with other attributes, such as linkage with another structured parameter, e.g., if we need to add a structured parameter in the report for selecting a time range type, such as "week", "month", "quarter", etc., then a corresponding linkage rule is set. For example, if the user selects the "week" option, the timeframe structured parameters need to be automatically filled and set to the start date and end date of the week. Finally, the new campt structured parameters are applied to the creative design and presented in the report. For example, the structured parameters are used to filter the data in the report to display the data within the user selected time frame. If the user selects the time range 2020-01-01 through 2020-12-31, the report will display all data within that time range.

Based on the decision tree algorithm, the constraint conditions of the structural parameters are determined, and the structural parameters are matched and coordinated with the service attributes.

And acquiring a characteristic attribute and a sample attribute, wherein the characteristic attribute comprises a numerical value type. And determining the category to which each sample belongs through the sample attribute, and determining the attribute of the decision tree, wherein the attribute comprises the attribute type, the value range, the requirement of classification accuracy and the limiting condition. A root node is selected, which corresponds to all data samples. And selecting an attribute, classifying the data samples according to the attribute value, and constructing the child nodes of the attribute until the data samples under all the child nodes belong to the same category or all the attributes are used. And traversing the tree step by step from the root node downwards according to the attribute value on the node until the leaf node is reached, wherein the category represented by the leaf node is the classification result of the new data. And determining the limiting conditions of the structural parameters, matching and coordinating with the service attributes to obtain a decision tree classifier, classifying the new data, and judging the category to which the new data belong. For example, there is a data set that contains the following attributes: gender: male, female, age: numerical value, income: numerical value, marital status: married, unmarked, occupation: white collar, blue collar, others, and purchasing power: high, medium, low. First, the attributes are divided into feature attributes and sample attributes. The characteristic attributes are gender, age, income, marital status and occupation, the sample attribute is purchasing power, the numerical attribute is age and income, and the decision tree attribute, namely the requirement and the limiting condition of classification accuracy, is determined according to the purchasing power of the sample attribute. Suppose that classification accuracy is required to reach over 80%. The root node is selected as the purchasing power, which corresponds to all data samples. And selecting an attribute, classifying the data samples according to the attribute value, and constructing the child nodes of the attribute until the data samples under all the child nodes belong to the same category or all the attributes are used. Assuming that age is selected as the first attribute, a sample aged 30 years or less is divided into one child node, and a sample aged 30 years or older is divided into another child node. And calculating the purchasing power classification accuracy of the two child nodes, if the purchasing power classification accuracy reaches more than 80%, continuing to construct the child nodes, otherwise, selecting another attribute for division. Assuming that the marital state is continuously used as a second attribute, dividing the two sub-nodes which are married and not married into new sub-nodes respectively, calculating the purchasing power classification accuracy of each sub-node, if the purchasing power classification accuracy reaches more than 80%, continuously constructing the sub-nodes, otherwise, selecting another attribute for division. And the like, until all the data samples under all the child nodes belong to the same category or all the attributes are used. And traversing the tree step by step from the root node downwards according to the attribute value on the node until the leaf node is reached, wherein the category represented by the leaf node is the classification result of the new data. And determining the limiting conditions of the structural parameters, matching and coordinating with the service attributes to obtain a decision tree classifier, classifying the new data, and judging the category to which the new data belong. For example, a 30 year old, medium income, unmarried, white collar person is predicted, and the purchasing power is medium according to the decision tree classifier.

It will be apparent to those skilled in the art that various modifications and variations can be made to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if such modifications and variations of the embodiments of the present application fall within the scope of the claims and the equivalents thereof, the present application is intended to encompass such modifications and variations.

Claims

1. A method for generating product concepts based on research report data and a large language model, the method comprising:

the method comprises the steps of obtaining the topic, the source and the release time of the report data, classifying report contents, constructing an industry knowledge base, and vectorizing the knowledge base, and specifically comprises the following steps: based on a word2vec model, converting the report content into vector representation; extracting the research report knowledge according to an industry knowledge base, and classifying the product concepts to be generated; performing prompt generation on the product attribute under each product functional characteristic classification according to the functional characteristics of the product to be generated; according to the use habit, skill and knowledge background of a product designer, the method automatically recommends the structuring parameters of the campt suitable for different products according to the functional characteristics of different products, and specifically comprises the following steps: recommending a simple structured parameter list of designers similar to the past works based on a cosine similarity calculation method; according to different user demands, structural parameters in a parameter list are adjusted to generate the Prompt content for accurately describing the product, and the method specifically comprises the following steps: clustering users based on a collaborative filtering algorithm of the users, and determining the structural parameter values of different user demands; updating an industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base, judging whether the new creative field exceeds the applicable range of the current prompt parameter, updating the industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base, and judging whether the new creative field exceeds the applicable range of the current prompt parameter, wherein the method specifically comprises the following steps: based on a support vector regression model, predicting market potential of a new creative field, constructing a decision tree, and judging whether technical implementation of the new field is realized through technical implementation owned by a current research product; adding new campt structural parameters according to the new creative field and applying the new campt structural parameters to creative design, wherein the new campt structural parameters comprise: determining the limiting conditions of the structural parameters and matching and coordinating with the service attributes based on a decision tree algorithm;

The method comprises the steps of generating the product attribute of each product functional characteristic category according to the functional characteristics of the product to be generated according to the need, wherein the method comprises the steps of determining the functional characteristics of the product to be generated according to the product concept category document; determining corresponding product attributes according to functional characteristics of the product to be generated based on the existing open source knowledge graph; collecting product attribute data in product specifications, comments and news reports, establishing a corpus for different product attributes by utilizing data mining and natural language processing technologies, analyzing key words and phrases in the corpus, and determining corresponding campt words of corresponding product attributes; constructing a vocabulary, and allocating a unique integer number to each word; based on a convolutional neural network model, performing deep learning and model training on the product attribute; inputting product attributes, and automatically identifying and recommending a prompt vocabulary corresponding to the proper product attributes by the trained model; and combining different campt vocabularies according to the requirements of the user, and generating campt content aiming at the functional characteristics of each product by using a natural language generation technology.

2. The method of claim 1, wherein the acquiring the topic, source, and release time of the report data classifies report content, constructs an industry knowledge base, and vectorizes the knowledge base, comprising:

acquiring the information of the research report, cleaning, word segmentation and coding the data of the research report text, the keywords, the authors, the institutions and the citations, and constructing a corresponding data set; constructing a word2vec model by using a gemim deep learning framework; the codes of the input report text, keywords, authors, mechanisms and citations are converted into vectors with corresponding dimensions through an embedding layer; inputting the input vector representation into a model, and obtaining a context-dependent vector representation of each datum through a multi-layer self-attention mechanism and a forward neural network; and inputting the fused vector representation into a full connection layer to obtain a final vector representation.

3. The method of claim 1, wherein the extracting the report knowledge from the industry knowledge base to classify the product concepts to be generated comprises:

4. The method of claim 1, wherein automatically recommending the structured parameters of the campt suitable for the different products for the functional characteristics of the different products according to the usage habits, skills and knowledge backgrounds of the product designers comprises:

obtaining design elements, style software, design theory and skill data of past works of a designer as characteristic data, and carrying out structural processing on the obtained characteristic data; performing similarity calculation on past works of a designer by adopting a cosine similarity calculation method to obtain a similarity matrix; finding out the designer most similar to the current designer according to the similarity matrix, selecting the past works as recommended objects, and obtaining a sample structured parameter list of the works; a list of campt structured parameters for similar designer works is recommended to the designer.

5. The method of claim 1, wherein the adjusting the structured parameters in the parameter list to generate the promt content that accurately describes the product according to different user requirements comprises:

acquiring historical prompt input data of a user; preprocessing the collected user behavior data, including removing repeated data, processing missing values and normalizing; according to the user behavior data, calculating the similarity between each user by using a cosine similarity method; dividing users into K clusters by using a K-means clustering algorithm, wherein each cluster represents a user group, and clustering the users according to the similarity; according to the clustering result, analyzing the user behavior data of each cluster, finding out common points and difference points of the user behavior data, determining the demands of different user groups, determining the structural parameter values of different user demands, and adjusting the structural parameters in the parameter list.

6. The method of claim 1, wherein the updating the industry knowledge base, updating and adjusting the creative field of the corresponding research product according to the industry knowledge base, and determining whether the new creative field is beyond a range applicable to the current prompt parameter comprises:

updating an industry knowledge base, and acquiring a latest research report and industry research, wherein the latest research report and the industry research comprise newly added research industry and expansion of the existing industry; evaluating and classifying the existing research products, classifying the research products into corresponding creative fields according to the characteristics and application scenes of the research products, and adjusting the classification of the creative fields according to the classification of the research products and the development trend of industry research; for the newly added research industry, the new research industry is classified into the corresponding creative field, and information on industry definition, research range and development trend is acquired;

Judging whether the new creative field is suitable for the current prompt parameter, if not, adjusting the parameter range; the information of the newly added research industry and the adjusted creative field is arranged and filed; comprising the following steps: predicting market potential of the new creative field based on a support vector regression model; constructing a decision tree, and judging whether the technical implementation in the new field is realized through the technical implementation owned by the current research product;

collecting demand description data of a user on a new creative field product from various channels, and cleaning; performing word segmentation, stop word removal and word frequency statistics on the requirement description text by adopting a natural language processing technology, and determining keywords; carrying out vectorization on demand description according to keywords, and establishing a model by adopting a K-means clustering algorithm; judging the user demand characteristics of each category according to the clustering result and the characteristics of the clustering center; comparing the characteristics and advantages of the new field with the user demand data to judge whether the new field meets the user demand aimed at by the current research product; using a support vector regression model in an SVM algorithm, taking characteristics and advantages of the new field as independent variables, taking market demands and competition conditions as target variables, and predicting market potential of the new field;

collecting business attribute contents, including technical implementation of current research products and technical implementation of new fields; constructing a decision tree, taking the technical implementation of the current research product as a root node, and dividing the tree into two branches which are feasible and infeasible according to whether the technical implementation of the new field is realized through the technical implementation owned by the current research product; evaluating the accuracy of the decision tree, dividing the data set into a training set and a testing set by using a cross-validation method, training the decision tree by using the training set, and evaluating the accuracy of the decision tree by using the testing set; and judging whether the creative implementation in the new field is feasible or not according to the result of the decision tree.

7. The method of claim 1, wherein the adding new campt structuring parameters according to new creative fields and applying to creative designs comprises:

acquiring characteristic attributes and sample attributes, wherein the characteristic attributes comprise gender, age, income, marital status and occupation, and the sample attributes are purchasing power; determining the category of each sample through the sample attribute, determining the attribute of the decision tree, wherein the attribute comprises the attribute type, the value range, the requirement of the classification accuracy and the limiting condition; selecting a root node, wherein the node corresponds to all data samples; selecting a characteristic attribute, classifying the data samples according to the attribute value of the characteristic attribute, and constructing child nodes of the characteristic attribute until the data samples under all child nodes belong to the same category or all characteristic attributes are used completely; traversing the tree step by step downwards from the root node according to the characteristic attribute value on the node until the leaf node is reached, wherein the category represented by the leaf node is the classification result of the new data; and determining the limiting conditions of the structural parameters, matching and coordinating with the service attributes to obtain a decision tree classifier, classifying the new data, and judging the category to which the new data belong.