A method for recommending materials based on Transformer model generation technology

Abstract

The present application relates to artificial intelligence technology, and discloses a method for generating technical recommendation materials based on a Transformer model. It includes: training a Transformer model, learning the associated features ki between basic data and technical recommendation materials in various training samples, and obtaining a feature set K1; learning a mapping feature set between various associated features; selecting various types of technical recommendation materials as type X in turn, and based on the mapping feature set, converting other associated features ki in the feature set K1 into associated features kX to obtain a feature set K2; when applying the Transformer model, using the basic data related to the target user and the corresponding feature set K2 to generate the target type of technical recommendation materials. The present application also discloses a computer device, a control device, and a computer-readable storage medium. The present application aims to reduce the model's dependence on a large number of training samples when training a generation model that supports diversified technical recommendation materials, thereby reducing the cost of model training.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, a control device, a computer device, and a computer-readable storage medium for recommending materials based on Transformer model generation technology. Background Art

[0002] At present, it often takes a lot of time and energy for technical personnel to manually write highly professional technical recommendation materials (such as technical documents, technical project reports, software user manuals, technical briefing documents, etc.), which is particularly difficult for technical personnel with weak writing skills.

[0003] With the development of artificial intelligence technology, more and more people have begun to train and apply artificial intelligence models to automatically generate corresponding technical recommendation materials (the most common ones are the generation models of various project reports) to improve the efficiency of generating technical recommendation materials and reduce the cost of manual writing of technical recommendation materials.

[0004] Generally speaking, if you want the quality of the output results of the artificial intelligence model to be high, then there are certain requirements for the sample size during model training. Since there are many types of technical recommendation materials, if a corresponding amount of samples are collected for each type of technical recommendation material to train the model, the sample size and training volume involved are undoubtedly huge, and the more types of materials that the model can generate, the larger the sample size required, and the higher the cost of model training.

[0005] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0006] The main purpose of this application is to provide a method, control device, computer equipment and computer-readable storage medium for generating technical recommendation materials based on a Transformer model, aiming to reduce the model's dependence on a large number of training samples when training a generation model that supports diverse technical recommendation materials, thereby reducing the cost of model training.

[0007] To achieve the above objectives, the present application provides a method for recommending materials based on Transformer model generation technology, comprising the following steps:

[0008] Train the Transformer model to learn the correlation features ki between the basic data and the technical recommendation materials in various training samples, and obtain the feature set K1 = {k1, k2, ..., kN}; where N is the number of types of technical recommendation materials;

[0009] Based on the Transformer model, the mapping features pj between various associated features ki are learned to obtain the feature set P = {p1, p2, ..., pM}; where M ≧ (N-1);

[0010] Select each type of technical recommendation material as type X in turn, and mark the associated features ki corresponding to the technical recommendation materials of type X as associated features kX; and based on the feature set P, convert the associated features ki in the feature set K1 except the associated features kX into associated features kX, so as to obtain the feature set K2 corresponding to the technical recommendation materials of type X;

[0011] When applying the Transformer model, the target type of technical recommendation materials are generated by utilizing the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials.

[0012] Optionally, after the step of obtaining the feature set K2 corresponding to the technical recommendation material of type X, the step further includes:

[0013] Obtaining other training samples other than the training samples corresponding to the technical recommendation materials of type X;

[0014] The basic data in other training samples are input into the Transformer model for iterative training, so that the Transformer model generates technical recommendation materials of type X based on the feature set K2, and the feature set K2 is updated accordingly according to the iterative training results.

[0015] Optionally, the method for recommending materials based on the Transformer model generation technology further includes:

[0016] In the feature set K2, the first weight corresponding to the original associated feature kX is greater than the second weight corresponding to the converted associated feature kX.

[0017] Optionally, before the step of generating the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the technical recommendation materials of the target type when applying the Transformer model, the step further includes:

[0018] Based on the data crawler, basic data related to the target user is extracted from the imported relevant information of the target user.

[0019] Optionally, after the step of extracting basic data related to the target user from the imported relevant information of the target user based on the data crawler, the method further includes:

[0020] Detect whether the data volume of the basic data extracted by the data crawler meets the preset requirements;

[0021] If not, based on the GPT (Generative Pre-Trained) model, use the relevant information and basic data of the target users to expand and generate more basic data so that the data volume of the basic data meets the preset requirements.

[0022] Optionally, when applying the Transformer model, the step of generating the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials includes:

[0023] When applying the Transformer model, the corresponding technical text content is generated by using the basic data related to the target user and the feature set K2 corresponding to the technical recommendation materials of the target type;

[0024] Generate the target type of technical recommendation materials according to the preset material template corresponding to the technical text content and the target type of technical recommendation materials.

[0025] Optionally, before the step of generating the target type of technology recommendation material according to the technical text content and the preset material template corresponding to the target type of technology recommendation material, the step further includes:

[0026] Based on the GPT model, the technical text content is semantically optimized.

[0027] To achieve the above object, the present application also provides a control device, comprising:

[0028] The first training module is used to train the Transformer model to learn the correlation features ki between the basic data and the technical recommendation materials in various training samples, and obtain the feature set K1 = {k1, k2, ..., kN}; where N is the number of types of technical recommendation materials;

[0029] The second training module is used to learn the mapping features pj between various associated features ki based on the Transformer model, and obtain the feature set P = {p1, p2, ..., pM}; where M ≧ (N-1);

[0030] A feature conversion module is used to select various types of technical recommendation materials as type X in turn, and mark the associated features ki corresponding to the technical recommendation materials of type X as associated features kX; and, based on the feature set P, convert the associated features ki in the feature set K1 except the associated features kX into the associated features kX, so as to obtain the feature set K2 corresponding to the technical recommendation materials of type X;

[0031] The material generation module is used to generate the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials when applying the Transformer model.

[0032] To achieve the above objectives, the present application also provides a computer device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the computer program is executed by the processor, the steps of the method for recommending materials based on the Transformer model generation technology as described above are implemented.

[0033] To achieve the above objectives, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for recommending materials based on the Transformer model generation technology are implemented as described above.

[0034] The method, control device, computer equipment and computer-readable storage medium for generating technology recommendation materials based on the Transformer model provided in the present application use the Transformer model to learn and map the associated features ki between different types of technology recommendation materials and corresponding basic data, and realize the mutual conversion between the associated features ki corresponding to different types of technology recommendation materials, so that the Transformer model only needs to be trained with a small number of training samples to obtain enough associated features ki to generate the corresponding types of technology recommendation materials. This not only enables the Transformer model to support the generation of diversified technology recommendation materials, but also eliminates the need to collect and train a large number of training samples for each type of technology recommendation material during model training, thereby reducing the dependence on a large number of training samples, thereby reducing the cost of model training and improving the efficiency of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the steps of a method for recommending materials based on the Transformer model generation technology in one embodiment of the present application;

[0036] Figure 2 This is a schematic diagram of a control device in an embodiment of the present application;

[0037] Figure 3 A schematic diagram of the internal structure of a computer device according to an embodiment of the present application.

[0038] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0039] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0040] In addition, if the description of "first", "second", etc. is involved in this application, it is only used for descriptive purposes (such as for distinguishing the same or similar features), and cannot be understood as indicating or suggesting their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in this field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0041] Reference Figure 1 In one embodiment, a method for recommending materials based on Transformer model generation technology includes:

[0042] Step S10, training the Transformer model, learning the correlation features ki between the basic data and the technical recommendation materials in various training samples, and obtaining a feature set K1 = {k1, k2, ..., kN}; wherein N is the number of types of technical recommendation materials;

[0043] Step S20: Based on the Transformer model, learn the mapping features pj between various associated features ki to obtain a feature set P = {p1, p2, ..., pM}; where M ≧ (N-1);

[0044] Step S30, sequentially selecting various types of technical recommendation materials as type X, and marking the associated features ki corresponding to the technical recommendation materials of type X as associated features kX; and, based on the feature set P, converting the associated features ki in the feature set K1 except the associated features kX into associated features kX, to obtain a feature set K2 corresponding to the technical recommendation materials of type X;

[0045] Step S40: When applying the Transformer model, the target type of technical recommendation materials are generated by using the basic data related to the target user and the feature set K2 corresponding to the technical recommendation materials of the target type.

[0046] In this embodiment, the execution terminal of the embodiment may be a computer device, or may be other equipment or devices (such as a control device) that controls the computer device.

[0047] As described in step S10, basic data and corresponding technical recommendation materials for training are collected and prepared in advance, and these sample data should cover all types of technical recommendation materials that need to be generated.

[0048] Optionally, assuming that the minimum sample size required for the model to successfully generate a single type of technical recommendation material is Y, and the number of types of technical recommendation materials is N (N≧2), then the minimum sample size of the collected training samples required for each type of technical recommendation material is Y / N.

[0049] It should be noted that in order to ensure that the model can learn a certain amount of relevant features of a single type of technical recommendation material, the Y / N value should not be less than 100; if the Y / N value is less than 100, the sample size required for a single type of technical recommendation material should be supplemented to 100 (even so, the sample size of this model is still greatly reduced compared to the hundreds or thousands of samples required for a single type of technical recommendation material by conventional models).

[0050] Among them, a complete training sample should include a technical recommendation material and the basic data corresponding to the technical recommendation material (the basic data is marked with the information required to generate the corresponding technical recommendation material); training samples containing the same type of technical recommendation materials are classified as training samples of the same category.

[0051] Optionally, the basic data can be enterprise or individual attribute information (such as basic information of the company or individual), product data (such as product parameters, product images, product feature descriptions, software operation interface diagrams, etc.), circuit diagrams, technical specifications and standards, experimental data, test results, industry reports, etc. All technology-related data required to generate technical recommendation materials.

[0052] Input the prepared training samples of various types into the Transformer model built based on artificial intelligence technology for training. During the training process, the model will learn the association features ki (i = {1, 2, ..., N}, N is the number of types of technical recommendation materials) between the basic data and the technical recommendation materials in various training samples. Among them, each type of technical recommendation material corresponds to a type of association feature ki (such as the first type of technical recommendation material, the corresponding type of association feature is k1; the second type of technical recommendation material, the corresponding type of association feature is k2), and the same type of association feature ki has multiple sub-class features. These sub-class features can be semantic information, keywords, syntactic structures, etc. learned by the model from the text, which can characterize the characteristics and content of different types of technical recommendation materials.

[0053] For example: the model learns the semantic relationship between basic data and technical recommendation materials, understands the meaning and logical relationship between them, so as to generate recommendation materials that are coherent and professional; the model can identify and learn keywords that frequently appear in different technical recommendation materials, which can be specific industry terms, product feature descriptions or important indicators of market trends; the model understands the grammatical structure and paragraph organization of different technical recommendation materials to ensure that the generated materials meet the expected technical standards and style in terms of expression; the model will learn the typical content characteristics of different types of technical recommendation materials, such as the structure of industry reports, the format requirements of product manuals or the methodology of market analysis, to ensure that the generated recommendation materials can effectively meet user needs and expectations.

[0054] The associated feature ki is a knowledge representation that the model extracts and understands from the data during the training process. This representation enables the Transformer model to learn the ability to generate a specified type of technical recommendation material based on the basic data. For example, based on the first type of associated feature k1, the model can generate the first type of technical recommendation material; based on the second type of associated feature k2, it can generate the second type of technical recommendation material.

[0055] Optionally, the set of associated features ki corresponding to all types of technical recommendation materials is counted, that is, the feature set K1 = {k1, k2, ..., kN} is obtained.

[0056] It should be noted that the Transformer model is a deep learning model that uses an encoder-decoder structure, in which both the encoder and the decoder contain multiple self-attention layers, allowing the model to perform parallel calculations when processing sequence data, significantly improving computational efficiency and model performance. The core of the Transformer model is its self-attention mechanism. By calculating the correlation scores between different positions in the input sequence, the model can focus on the relevant parts and generate representations, thereby capturing long-distance dependencies in the sentence.

[0057] As described in step S20, in the Transformer model, the mapping features pj between various types of associated features ki are learned through a multi-head attention mechanism and a feedforward neural network. Each pj reflects the interaction or relationship between different types of ki. The Transformer model can effectively capture the complex interactions between global dependencies and features through the self-attention mechanism.

[0058] As the Transformer model is trained, all learned mapping features pj are collected and integrated to form a feature set P = {p1, p2, ..., pM} (j = {1, 2, ..., M}, the number of types M of these features pj is at least N-1).

[0059] The mapping feature pj learned in this way can not only reflect the mapping relationship between different types of associated features ki, but also have a deeper semantic expression. That is, the mapping feature pj is not limited to simple correlation, but can also capture the deeper semantic similarities and differences between the types of technical recommendation materials.

[0060] It should be noted that the feedforward neural network, as part of the Transformer model, is used to process and map the feature representations obtained from the attention mechanism. This network structure has a strong nonlinear modeling capability and can map the input feature space to a higher-dimensional expression space, thereby better capturing and expressing the complex semantic relationship between features. Through such processing, the mapping feature pj can express the relationship between different types of associated features ki more semantically richly and deeply. Taking the selected activation function as ReLU as an example, the generation example of the mapping feature pj is as follows:

[0061] pj = ReLU(w*z+b);

[0062] Among them, w and b are the weights and biases learned by the model, which are used for linear transformation; z is the weighted feature representation between the two categories ki calculated by the model.

[0063] Optionally, during the training process, a semantic similarity metric can be introduced as part of the loss function. This metric can help the model better understand and distinguish the semantic relationship between different types of associated features ki, so that the learned mapping features pj have deeper semantic expression capabilities.

[0064] For example, the cosine similarity is used to calculate the dot product and Euclidean norm between various types of associated features ki, or the Euclidean distance between various types of associated features ki is calculated to extract the corresponding mapping feature pj (this mapping feature pj represents a more generalized and expressive representation of the relationship from one type of ki to another type of ki). This representation can help the model more accurately understand and convert features between different types of materials when generating technical recommendation materials.

[0065] Optionally, since there are multiple sub-class features under each type of associated feature ki, between every two types of associated features ki, only the sub-class features of the same type in the two can be subjected to mapping relationship analysis, while there is no need to perform mapping relationship analysis between sub-class features of different types.

[0066] Optionally, since each type of technical recommendation material has a corresponding type of associated feature ki, and for each type of associated feature ki, if the mapping feature pj between it and other types of associated features ki is learned separately, then N-1 rounds of mapping analysis are required (one round of learning obtains one type of pj), and the total number of analysis rounds is N*(N-1), and the number of types of mapping feature pj obtained is M, which is N*(N-1). In this way, when there are many types of technical recommendation materials, the amount of calculation required by the model is undoubtedly huge.

[0067] Therefore, when the number of types of technical recommendation materials is large (that is, N is greater than or equal to the preset value, such as N≧5), a type of associated feature ki can be selected as the benchmark group, and the mapping feature pj between it and other types of associated features ki can be analyzed. The mapping relationship between other types of associated features ki can be first mapped to the benchmark group, and then further mapped to the specific group to be mapped. In this way, the total analysis rounds are only N-1, and the number of types of mapping feature pj M is N-1.

[0068] For example, the mapping feature between the reference group and the first type of associated feature k1 is p1, and the mapping feature between the reference group and the second type of associated feature k2 is p2, then the mapping relationship between k1 and k2 can be represented by p1 to p2.

[0069] This effectively reduces the computational complexity and amount of operations, while maintaining the ability to understand and express the complex relationships between different types of associated features ki.

[0070] Of course, if the number of types of technical recommendation materials is small (i.e., N is less than the preset value, such as N < 5), then for each type of associated feature ki, the mapping feature pj between it and the associated features ki of other types can be learned separately. In this way, the accuracy of subsequent feature conversion can be improved. Since the number of types is small at this time, the amount of calculation of the mapping feature pj between different types is relatively small, so the mapping relationship between different types of ki can be captured more comprehensively.

[0071] As described in step S30, each type of technical recommendation material is selected as type X in turn, that is, the type corresponding to the currently selected technical recommendation material is marked as X, and the associated feature ki corresponding to the technical recommendation material of type X is marked as associated feature kX. For example, when the first type of technical recommendation material is selected as type X, its corresponding associated feature k1 is also marked as kX and is regarded as the original associated feature kX.

[0072] After determining the technical recommendation material of type X, the Transformer model uses the mapping feature pj in the feature set P to convert the associated features ki in the feature set K1 except the associated feature kX into the associated features kX (that is, the converted associated features kX); and based on the set of the original associated features kX and the converted associated features kX, the feature set K2 can be obtained, and the feature set K2 corresponds to the technical recommendation material currently marked as type X.

[0073] During the training process of the Transformer model, the model learns the complex interactions and dependencies between various associated features ki through the multi-head attention mechanism and feedforward neural network. These relationships are expressed as mapping features pj in the feature space, and these pj not only simply express the correlation between ki, but also capture deeper semantic information and mapping relationships between ki. Therefore, based on the set P of these mapping features pj, the model can effectively capture the global dependencies of each associated feature ki, which means that the model can understand the semantic similarities and differences between different associated features ki, so that it can perform effective transformations in the feature space.

[0074] Taking the mapping feature between k1 and k2 as p1 as an example, if k1 is selected as kX, then converting k2 to kX based on p1 is: kX=p1·k2.

[0075] It should be noted that although p1 is learned by the model through the mapping relationship between k1 and k2, it is a feature vector rather than a simple scalar value. This is because p1 is a mapping feature learned based on the model, which not only contains the information and weights required for the conversion between k1 and k2, but also introduces nonlinear transformation (ReLU) and linear transformation (w*z+b) in the conversion process. Therefore, the dot product of p1·k2 is a new feature representation of the model based on the learned mapping relationship (i.e., the converted kX), rather than simply equal to k1. Instead, it can be considered as a feature representation of the same type as k1, which can be used to generate the same type of technical recommendation materials. The conversion logic between other types of ki can be referred to the description here, so it will not be repeated here.

[0076] After the feature set K2 corresponding to the technical recommendation material currently marked as type X is obtained through conversion, the mark of type X is cancelled; then the next type of technical recommendation material is selected and marked as type X, and step S30 is repeated until all types of technical recommendation materials have obtained their corresponding feature sets K2.

[0077] As described in step S40, after the Transformer model training is completed, the Transformer model is put into the application of generating technical recommendation materials. At this time, the basic data related to the target user can be imported, and the type of technical recommendation materials required to be generated for the target user (i.e., the target type) can be selected, and then step S40 is started.

[0078] Among them, the target user refers to the person or entity who will eventually use the generated technology recommendation materials, which can be an enterprise, engineer, researcher, or any individual or organization that needs relevant technology recommendations.

[0079] Optionally, the Transformer model first encodes the target user's basic data into a format that the model can understand, which usually includes converting text and numerical data into vectors or embedded vectors (collectively referred to as feature vectors) to be compatible with the model's input data. At the same time, the feature set K2 corresponding to the target type of technical recommendation materials is obtained.

[0080] Then, the model generates the target type of technical recommendation materials based on the extracted feature vector and the feature set K2 corresponding to the target type of technical recommendation materials.

[0081] The model generates the logic of technical recommendation materials based on the feature vectors and feature set K2 corresponding to the basic data. The model generates the logic of corresponding technical recommendation materials based on the basic data based on the learned association features ki between the basic data and the technical recommendation materials. However, the feature set K2 is regarded as a set of more association features ki of the same type.

[0082] That is, the Transformer model will integrate or weightedly combine the feature vector corresponding to the target user's encoded data with the features in the feature set K2. Based on the integrated results, the Transformer model applies its learned multi-head attention mechanism and feedforward neural network, as well as the global dependencies learned during the training process, to generate specific technical recommendation materials that meet the requirements of the target type.

[0083] Optionally, the model's application of multi-head attention mechanism includes:

[0084] (1) Self-attention mechanism: The model can perform weighted processing on the input serialized feature fusion results through the self-attention mechanism, focusing on the dependencies between different parts of the sequence, especially performing well in solving long-distance dependency problems;

[0085] (2) Attention weight calculation: For each sequence position, the model calculates its attention weight with other positions so as to focus on relevant information in the subsequent encoding and decoding process.

[0086] Optionally, the model includes multiple stacked encoder layers, each of which contains a multi-head attention mechanism and a feedforward neural network to further encode and abstract the input. The decoder generates content related to the target type based on the previous encoding representation and its own state.

[0087] Optionally, use of model global dependencies includes:

[0088] (1) Feature interaction: The model leverages the global dependencies learned during training to ensure that the generated recommendations are semantically and content-coherent and relevant.

[0089] (2) Semantic information capture: Global dependencies are not limited to the relationship between local features, but also include the understanding and integration of overall semantic information, thereby generating more accurate and user-friendly recommendation content.

[0090] Finally, the content generated by the model is organized into a form suitable for users to receive, such as text description, chart display or other formats for result output.

[0091] In this way, the Transformer model uses the complex relationships and dependencies it has learned, combined with the basic data of the target user and the feature set K2 of the target type, to efficiently generate technical recommendation materials that meet the target type.

[0092] In one embodiment, a Transformer model is used to learn and map the associated features ki between different types of technology recommendation materials and corresponding basic data, so as to realize the mutual conversion between the associated features ki corresponding to different types of technology recommendation materials, so that the Transformer model only needs to be trained with a small number of training samples to obtain enough associated features ki to generate corresponding types of technology recommendation materials. This not only enables the Transformer model to support the generation of diversified technology recommendation materials, but also eliminates the need to collect and train a large number of training samples for each type of technology recommendation material during model training, thereby reducing the dependence on a large number of training samples, thereby reducing the model training cost and improving the efficiency of model training.

[0093] For example, if the minimum sample size required for the model to successfully generate a single type of technical recommendation material is Y, and the number of types of technical recommendation materials is N (N≧2), then the minimum sample size required for each type of technical recommendation material is Y / N (the minimum is not less than 100), and the corresponding total sample size is also at least Y (if the value of Y is not greater than 500, the total sample size can be supplemented to 500), while the total sample size required by the conventional model is N*Y, which is N times that of the Transformer model. It can also be seen from this that, in theory, as the types of technical recommendation materials supported by the model increase, the advantage of reducing the sample size required for training the Transformer model based on this method is more obvious (taking N as 2 and Y as 1000 as an example, the conventional training method needs to use 2000 total samples for model training, while this method only needs to use 1000 total samples; taking N as 3 and Y as 1000 as an example, the conventional training method needs to use 3000 total samples for model training, while this method also only needs to use 1000 total samples).

[0094] It should be noted that, given the different complexities of different types of technical recommendation materials, the Y values ​​required for training different types of technical recommendation materials are not necessarily the same. Here, only for the convenience of understanding the advantages of this method, the case where the Y values ​​required for training different types of technical recommendation materials are consistent is selected for exemplary explanation. However, in any case, since this method implements various types of technical recommendation materials, it can convert the corresponding associated features ki of other types of technical recommendation materials into the associated features ki required by itself, thereby greatly reducing the number of training samples required for the model to train each type of technical recommendation material. Therefore, compared with conventional methods, this method has a higher utilization rate of model training samples, and the minimum sample size required for each type is significantly reduced, thereby reducing the sample collection and training required for training the generation model that supports multiple types of technical recommendation materials, improving the efficiency of model training, and reducing the cost of model training.

[0095] In one embodiment, based on the above embodiment, after the step of obtaining the feature set K2 corresponding to the technical recommendation material of type X, the method further includes:

[0096] Obtaining other training samples other than the training samples corresponding to the technical recommendation materials of type X;

[0097] The basic data in other training samples are input into the Transformer model for iterative training, so that the Transformer model generates technical recommendation materials of type X based on the feature set K2, and the feature set K2 is updated accordingly according to the iterative training results.

[0098] In this embodiment, after obtaining the feature set K2 corresponding to the technical recommendation material of a specific type X, other training samples other than the training samples corresponding to the technical recommendation material of type X are obtained, and these samples include technical recommendation materials of types other than X and their related basic data.

[0099] The basic data from these other types of training samples are input into the trained Transformer model for iterative training. In this process, the model uses the previously learned feature set K2 as prior knowledge and combines it with the input basic data to generate technical recommendation materials of type X. In this iterative training process, new associated features kX are also generated and added to the feature set K2 of type X to update the feature set K2. These newly generated associated features kX reflect the model's latest understanding and expression of the features of type X in the iteration.

[0100] Through this iterative training method, the Transformer model can gradually optimize the generation process of X-type technical recommendation materials, making them more accurate and in line with user needs, and further improving the utilization rate of various training samples; and after performing such operations on the feature set K2 corresponding to each type of technical recommendation material, the quality and accuracy of the model generating various types of technical recommendation materials can be comprehensively improved; it is equivalent to that the model can create new training samples for the generation training of various types of technical recommendation materials on the basis of existing training samples, so that the model can obtain more training data for various types of technical recommendation materials (that is, based on the basic data in other types of training samples, and the feature set K2 of this type (having K2 means having the ability to generate this type of technical recommendation materials), to generate this type of technical recommendation materials), thereby improving the quality of model training.

[0101] Moreover, this embodiment is a method of adding training data for the generation and training of different types of technical recommendation materials for the model. Compared with the method of randomly and automatically generating training samples to increase training data, the basic data used in this embodiment comes from real data in other types of training samples. These real data are more representative and can better cover different situations and application scenarios, which helps the model learn broader and more realistic feature representations.

[0102] In some optional embodiments, the Transformer model may also first output the newly generated technical recommendation materials of type X for confirmation by the engineer responsible for model training (the engineer may check the quality, accuracy and whether the technical recommendation materials meet expectations). If the engineer's confirmation is obtained, the model will then add the new associated features kX generated when the technical recommendation materials of type X are generated to the feature set K2 of type X; otherwise, if the engineer's confirmation is not obtained or the engineer's denial is obtained, the model will discard the new associated features kX generated when the technical recommendation materials of type X are generated, and will no longer add them to the feature set K2 of type X.

[0103] In this way, the update quality of the feature set K2 corresponding to each type of technical recommendation material can be guaranteed, and the feature set K2 can be tuned and updated in the direction approved by engineers, thereby ensuring the generation quality of each type of technical recommendation material by the Transformer model.

[0104] In one embodiment, based on the above embodiment, the method for recommending materials based on the Transformer model generation technology further includes:

[0105] In the feature set K2, the first weight corresponding to the original associated feature kX is greater than the second weight corresponding to the converted associated feature kX.

[0106] In this embodiment, when generating a feature set K2 corresponding to the technical recommendation material of type X, the rule for the model to assign corresponding weights to the associated feature kX in the feature set K2 is: the first weight corresponding to the original associated feature kX is greater than the second weight corresponding to the converted associated feature kX.

[0107] Since the original associated feature kX is obtained by training based on the corresponding samples of the X-type technology recommendation material, its feature representation can more accurately represent the association between the basic data and the X-type technology recommendation material. By assigning a higher weight to the original associated feature kX, the model can focus more on the original associated feature kX in the process of generating the X-type technology recommendation material, which helps to reduce information loss and error, thereby improving the quality of the generated technology recommendation material.

[0108] Optionally, after the step of obtaining the feature set K2 corresponding to the technical recommendation material of type X, the method further includes: determining the second weight according to the similarity of the text layout between the technical recommendation material of type X and other types; wherein the similarity is positively correlated with the second weight.

[0109] Optionally, the Transformer model determines the second weight based on the similarity of the text layout between type X and other types of technical recommendation materials. This similarity can be based on factors such as text structure, topic distribution or semantic association, reflecting the content similarity between different types of recommendation materials.

[0110] Optionally, standardized templates of all types of technical recommendation materials may be obtained, and text layout information corresponding to each standardized template may be extracted, and then similarity between text layouts of various types may be calculated using a similarity metric.

[0111] Alternatively, a manual evaluation method may be used to evaluate the similarity between the layouts of various types of texts (eg, the evaluation may be graded into three levels: high, medium, and low, and each level may correspond to a certain numerical value).

[0112] Optionally, the similarity between the text layouts is used to determine the value of the second weight. And as long as the second weight is less than the first weight, the higher the similarity, the larger the value of the second weight can be (i.e., the similarity between the text layout of type X and other types of technical recommendation materials is positively correlated with the value of the second weight). This means that when other types of technical recommendation materials are more similar to type X technical recommendation materials in text layout, the second weight corresponding to kX converted from the corresponding ki will increase accordingly. This helps to adjust and optimize the model's perception and reflection of the correlation between different types of recommended materials.

[0113] By adjusting the value of the second weight in this way, it can be ensured that when the model generates technology recommendation materials, it can effectively integrate other types of information similar to type X, while maintaining the priority of the original features of technology recommendation materials of type X, thereby further improving the quality of technology recommendation material generation.

[0114] In one embodiment, based on the above embodiment, if the feature set K2 also contains a new associated feature kX generated when the Transformer model is iteratively trained using basic data in other training samples, the third weight corresponding to the newly generated associated feature kX is greater than the second weight. At the same time, the third weight is less than or equal to the first weight (preferably, if the newly generated associated feature kX is added to K2 after confirmation by relevant engineers, the third weight can be equal to the first weight).

[0115] The purpose of this weight allocation strategy is to ensure that the newly generated associated features kX can reflect the priority of their source and generation time when they are added to the feature set K2. Through such a weight allocation rule, the model can better utilize the existing knowledge and the contribution of new data, thereby maintaining and optimizing the overall quality and effect of the feature set K2, and thus improving the quality of the model's subsequent generation of corresponding types of technical recommendation materials.

[0116] In some optional embodiments, the values ​​corresponding to the first weight, the second weight, and the third weight are set to be W1, W2, and W3, and under the premise of satisfying the rule of W1≧W3>W2, the three weight value combinations with different values ​​are generated. Then, for the feature set K2 of type X, the model

[0117] Optionally, the Transformer model generates X-type technical recommendation materials corresponding to each weight value combination based on the basic data in the training samples corresponding to the X-type technical recommendation materials, combined with the X-type feature set K2 of different weight value combinations, and marks them as the first material; and the technical recommendation materials in the training samples to which the basic data used to generate the first material belongs are marked as the second material. Then, the similarity between the first material and the second material is calculated, and the first material with the highest similarity to the second material is selected. The similarity can be calculated by using similarity metrics such as cosine similarity, the inverse of the Euclidean distance, and the Pearson correlation coefficient.

[0118] Optionally, after selecting the first material having the highest similarity to the second material, the weight value combination used when generating the first material is used as the values ​​of the first weight, the second weight and the third weight defined in the feature set K2 of the X type.

[0119] Compared with the engineer responsible for model training setting the first weight, the second weight and the third weight value according to human experience, the optional embodiment adopts the method of automatically generating weight value combinations, which can not only avoid the influence of human subjective preference or experience factors on the accuracy of weight values ​​and ensure that the weight setting is more objective, but also generate multiple weight combinations and explore a wider weight space, which helps to discover potential optimal weight combinations, thereby optimizing the performance of feature set K2 and the overall performance of the model.

[0120] This automated process of selecting the best weight combination can generate and evaluate multiple weight combinations in a shorter time. It is more efficient than manual adjustment, saving human resources while ensuring a faster optimization cycle.

[0121] Moreover, the generation of the optimal weight combination is based on the actual performance of the data and model, rather than relying on experience or preference. This data-driven decision can better reflect the complex relationship between the model and data, and improve the scientificity and reliability of the decision.

[0122] In one embodiment, based on the above embodiment, before the step of generating the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the technical recommendation materials of the target type when applying the Transformer model, the step further includes:

[0123] Based on the data crawler, basic data related to the target user is extracted from the imported relevant information of the target user.

[0124] In this embodiment, since not everyone is good at organizing the basic data required to generate technical recommendation materials, a data crawler can be built in advance. The data crawler can automatically obtain the required information from various data sources to ensure that the model can fully utilize and integrate these data when generating recommendation materials.

[0125] The data crawler can extract necessary information from various data sources such as corporate or personal attribute information, product data, circuit diagrams, technical specifications, etc. from the imported target user-related information based on predefined rules and templates. This information is the necessary input for generating technical recommendation materials.

[0126] At the same time, the data crawler can also crawl the required information from academic websites, corporate information websites, application websites of technology recommendation materials, etc. based on web crawler technology.

[0127] Compared with manually organizing basic data, the automated data capture process can significantly save time and labor costs.

[0128] In one embodiment, based on the above embodiment, after the step of extracting basic data related to the target user from the imported relevant information of the target user based on the data crawler, the step further includes:

[0129] Detect whether the data volume of the basic data extracted by the data crawler meets the preset requirements;

[0130] If not, based on the GPT model, use the relevant information and basic data of the target users to expand and generate more basic data so that the amount of basic data meets the preset requirements.

[0131] In this embodiment, the basic data extracted by the data crawler is quantitatively analyzed to evaluate whether it meets the requirements for generating target type of technical recommendation materials. This may involve evaluation of the number of data entries, the types of information covered, and the quality of the data.

[0132] If it is detected that the amount of basic data is insufficient, a strategy of using the GPT model to expand the data can be adopted. That is, calling the GPT model (such as using the corresponding API interface to call the existing GPT model), combining the relevant information of the target users and the existing basic data, to generate more basic data that meets the preset requirements. These data can include supplementary technical specifications, detailed product descriptions, user feedback, etc., to enrich and expand the existing basic data set.

[0133] By inputting relevant contextual information and possible output format requirements, the GPT model can generate new data related to the input information. These generated data can be further used in the generation process of technical recommendation materials.

[0134] In this way, when the amount of basic data that can be effectively used is insufficient, the powerful generative capabilities of the GPT model can be used to expand and enrich the data set to ensure that the model has sufficient basic data when generating technical recommendation materials.

[0135] In one embodiment, based on the above embodiment, when applying the Transformer model, the step of generating the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the technical recommendation materials of the target type includes:

[0136] When applying the Transformer model, the corresponding technical text content is generated by using the basic data related to the target user and the feature set K2 corresponding to the technical recommendation materials of the target type;

[0137] Generate the target type of technical recommendation materials according to the preset material template corresponding to the technical text content and the target type of technical recommendation materials.

[0138] In this embodiment, materials are recommended in advance for different types of technologies, and corresponding preset material templates are created.

[0139] Clearly define the different types of technical recommendation materials to be generated, such as product recommendation reports, technical solution proposals, industry trend analysis reports, etc. Each type corresponds to different information structures and content requirements. Based on the structure and content requirements analysis, design a preset material template. The template should include clear section divisions and content filling requirements to ensure that all information and chapters required by the material type are covered. For example:

[0140] Product recommendation report template: including product overview, functional features, performance parameters, application scenarios, competitive advantages, etc.

[0141] Technical solution proposal template: including problem definition, solution overview, implementation plan, expected results, etc.

[0142] Industry trend analysis report template: includes market background, trend analysis, competition landscape, future forecasts, etc.

[0143] Integrate the designed preset material template into the generation process based on the Transformer model. When the Transformer model executes step S40, it will first generate the target type of technical text content based on the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation material; then fill the target type of technical text content into the preset material template corresponding to the target type of technical recommendation material. After completing the filling, the generated target type technical recommendation material is finally formatted and adjusted to ensure that it meets the expected layout and style requirements. Finally, the final target type of technical recommendation material is output, which can be a formatted document (such as Word, PDF, etc.) or other deliverable format.

[0144] In this way, the Transformer model can efficiently combine user data and preset material templates to generate technical recommendation materials with clear structure and personalized content. This not only improves the efficiency and quality of generated materials, but also ensures that the materials meet the corresponding writing format requirements.

[0145] In one embodiment, based on the above embodiment, before the step of generating the target type of technical recommendation material according to the technical text content and the preset material template corresponding to the target type of technical recommendation material, the step further includes:

[0146] Based on the GPT model, the technical text content is semantically optimized.

[0147] In this embodiment, after the Transformer model generates the target type of technical text content based on the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials, the GPT model can be called to perform semantic optimization on the technical text content generated by the Transformer model.

[0148] The GPT model can understand and generate natural language text, ensuring that the generated technical content is semantically coherent, well-organized, and in line with language standards in the professional field.

[0149] The technical text content that has been semantically optimized based on the GPT model is filled into the preset material template of the target type to generate technical recommendation materials of the target type.

[0150] By integrating the feature learning and generation capabilities of the Transformer model and the semantic optimization capabilities of the GPT model, the generated technology recommendation materials can reach a high level in language expression and visual presentation, which further improves the generation quality of technology recommendation materials.

[0151] In one embodiment, the GPT model is able to understand natural language text and generate coherent, logically clear content. By applying the GPT model to the generated technical text, the professionalism of the document and the accuracy of the language expression can be improved, ensuring that the document complies with the language standards and user expectations in the technical field. The preset material templates define the structure and content requirements of different types of technical recommendation materials, including the titles of each part, the order of content, etc. The GPT-optimized technical text content is filled into these templates, ensuring that the final generated recommendation material has complete information and a good structure, thereby improving the overall quality and professionalism of the document.

[0152] In this way, even when the number of training samples is small, the Transformer model can still generate high-quality technical recommendation materials through preset templates and GPT semantic optimization during the feature learning and generation process. This approach not only improves generation efficiency, but also ensures the quality and professionalism of the generated content, meeting the expected technical recommendation standards. That is, combining the feature learning ability of the Transformer model, the semantic understanding of the GPT model, and the structured application of preset material templates can effectively optimize the generation process of technical recommendation materials, so that the Transformer model can still obtain high-quality technical recommendation material generation results even when the training data is limited.

[0153] In addition, refer to Figure 2 In an embodiment of the present application, a control device Z10 is further provided, comprising:

[0154] The first training module Z11 is used to train the Transformer model, learn the correlation features ki between the basic data and the technical recommendation materials in various training samples, and obtain the feature set K1 = {k1, k2, ..., kN}; where N is the number of types of technical recommendation materials;

[0155] The second training module Z12 is used to learn the mapping features pj between various associated features ki based on the Transformer model, and obtain the feature set P = {p1, p2, ..., pM}; where M≧(N-1);

[0156] The feature conversion module Z13 is used to select various types of technical recommendation materials as type X in turn, and mark the associated features ki corresponding to the technical recommendation materials of type X as associated features kX; and based on the feature set P, convert the associated features ki in the feature set K1 except the associated features kX into the associated features kX, so as to obtain the feature set K2 corresponding to the technical recommendation materials of type X;

[0157] The material generation module Z14 is used to generate the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials when applying the Transformer model.

[0158] Optionally, the control device Z10 may be a virtual control device (such as a virtual machine) or a physical device (such as a physical device other than a computer device that can execute the corresponding method).

[0159] In addition, a computer device is also provided in an embodiment of the present application. The internal architecture of the computer device can be as follows: Figure 3 As shown, it includes a processor, a memory, a communication interface and an input interface connected by a system bus. Among them, the processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database is used to store data called by the computer program. The communication interface is used to communicate data with an external terminal. The input interface is used to receive a signal input by an external device. When the computer program is executed by the processor, a method for recommending materials based on the Transformer model generation technology as described in the above embodiment is implemented.

[0160] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the present application scheme, and does not constitute a limitation on the computer device to which the present application scheme is applied. For example, in some optional embodiments, the computer device may also include an output interface (not shown in the figure), and the output interface is also connected to the system bus and is used to output corresponding signals to the peripheral device.

[0161] In addition, the present application also proposes a computer-readable storage medium, the computer-readable storage medium includes a computer program, and when the computer program is executed by a processor, the steps of the method for recommending materials based on the Transformer model generation technology as described in the above embodiment are implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0162] In summary, the method, control device, computer equipment and computer-readable storage medium for generating technology recommendation materials based on the Transformer model provided in the embodiments of the present application use the Transformer model to learn and map the associated features ki between different types of technology recommendation materials and the corresponding basic data, and realize the mutual conversion between the associated features ki corresponding to different types of technology recommendation materials, so that the Transformer model only needs to be trained with a small number of training samples to obtain enough associated features ki to generate the corresponding type of technology recommendation materials. In this way, not only can the Transformer model support the generation of diversified technology recommendation materials, but also when training the model, there is no need to collect and train a large number of training samples for each type of technology recommendation material, thereby reducing the dependence on a large number of training samples, thereby reducing the cost of model training and improving the efficiency of model training.

[0163] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0164] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0165] The above description is only a preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for recommending materials based on Transformer model generation technology, characterized in that: include: Train the Transformer model to learn the correlation features ki between the basic data and the technical recommendation materials in various training samples, and obtain the feature set K1={k1, k2, …, kN}; where N is the number of types of technical recommendation materials; Based on the Transformer model, the mapping features pj between various types of associated features ki are learned to obtain a feature set P = {p1, p2, ..., pM}; where M ≧ (N-1); when N is greater than or equal to the preset value, a type of associated feature ki is selected as the benchmark group, and the mapping features pj between the benchmark group and other types of associated features ki are analyzed, and the mapping relationship between other types of associated features ki is first mapped to the benchmark group, and then mapped to the specific group required for mapping; when N is less than the preset value, for each type of associated feature ki, the mapping features pj between the associated features ki of other types are learned separately; Select each type of technical recommendation material as type X in turn, and mark the associated features ki corresponding to the technical recommendation materials of type X as associated features kX; and based on the feature set P, convert the associated features ki in the feature set K1 except the associated features kX into associated features kX, so as to obtain the feature set K2 corresponding to the technical recommendation materials of type X; When applying the Transformer model, the target type of technical recommendation materials are generated by utilizing the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials.

2. The method for recommending materials based on Transformer model generation technology according to claim 1, characterized in that: After the step of obtaining the feature set K2 corresponding to the technical recommendation material of type X, the method further includes: Obtaining other training samples other than the training samples corresponding to the technical recommendation materials of type X; The basic data in other training samples are input into the Transformer model for iterative training, so that the Transformer model generates technical recommendation materials of type X based on the feature set K2, and the feature set K2 is updated accordingly according to the iterative training results.

3. The method for recommending materials based on Transformer model generation technology according to claim 1, characterized in that: The method for recommending materials based on the Transformer model generation technology also includes: In the feature set K2, the first weight corresponding to the original associated feature kX is greater than the second weight corresponding to the converted associated feature kX.

4. The method for recommending materials based on Transformer model generation technology according to claim 1, characterized in that: Before the step of generating the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials when applying the Transformer model, the method further includes: Based on the data crawler, basic data related to the target user is extracted from the imported relevant information of the target user.

5. The method for recommending materials based on Transformer model generation technology according to claim 4, characterized in that: After the step of extracting basic data related to the target user from the imported relevant information of the target user based on the data crawler, the method further includes: Detect whether the data volume of the basic data extracted by the data crawler meets the preset requirements; If not, based on the GPT model, use the relevant information and basic data of the target users to expand and generate more basic data so that the amount of basic data meets the preset requirements.

6. The method for recommending materials based on Transformer model generation technology according to any one of claims 1 to 5, characterized in that: When the Transformer model is applied, the step of generating the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials includes: When applying the Transformer model, the corresponding technical text content is generated by using the basic data related to the target user and the feature set K2 corresponding to the technical recommendation materials of the target type; Generate the target type of technical recommendation materials according to the preset material template corresponding to the technical text content and the target type of technical recommendation materials.

7. The method for recommending materials based on Transformer model generation technology according to claim 6, characterized in that: Before the step of generating the target type of technology recommendation material according to the content of the technical text and the preset material template corresponding to the target type of technology recommendation material, the step further includes: Based on the GPT model, the technical text content is semantically optimized.

8. A control device, characterized in that: include: The first training module is used to train the Transformer model to learn the correlation features ki between the basic data and the technical recommendation materials in various training samples, and obtain the feature set K1={k1, k2, …, kN}; where N is the number of types of technical recommendation materials; The second training module is used to learn the mapping features pj between various types of associated features ki based on the Transformer model, and obtain the feature set P={p1, p2, ..., pM}; wherein, M≧(N-1); when N is greater than or equal to a preset value, a type of associated feature ki is selected as a reference group, and the mapping features pj between the reference group and other types of associated features ki are analyzed, and the mapping relationship between other types of associated features ki is first mapped to the reference group, and then mapped to the specific group to be mapped; when N is less than a preset value, for each type of associated feature ki, the mapping features pj between the associated features ki of other types are learned respectively; A feature conversion module is used to select various types of technical recommendation materials as type X in turn, and mark the associated features ki corresponding to the technical recommendation materials of type X as associated features kX; and, based on the feature set P, convert the associated features ki in the feature set K1 except the associated features kX into the associated features kX, so as to obtain the feature set K2 corresponding to the technical recommendation materials of type X; The material generation module is used to generate the target type of technical recommendation materials by using the basic data related to the target user and the feature set K2 corresponding to the target type of technical recommendation materials when applying the Transformer model.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the method for recommending materials based on Transformer model generation technology as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for recommending materials based on Transformer model generation technology as described in any one of claims 1 to 7.

Images