Large model training and presentation file generation method and device, equipment and medium

Abstract

The invention provides a large model training and presentation generation method, device and equipment and a medium, relates to the field of artificial intelligence, and particularly relates to the technical fields of deep learning, large models, NLP and the like, and the large models comprise an outline generation model and a content generation model. According to the specific implementation scheme, sample document content, a sample presentation outline, sample presentation content and user feedback data which are associated with one another are obtained; processing the content of the sample document by adopting an outline generation model to obtain a predicted presentation outline; processing the sample document content and the sample presentation file outline by adopting a content generation model to obtain predicted presentation file content; and training an outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data, and training a content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

Description

Technical Field

[0001] The present disclosure relates to the field of AI (Artificial Intelligence), specifically to technical fields such as deep learning, big models, and NLP (Natural Language Processing), and especially to methods, devices, equipment, and media for training big models and generating presentations. Background Art

[0002] A professional and diversified document resource supply platform, dedicated to fully meeting users' diversified document needs in the fields of office, education, etc. Users can easily retrieve various types of documents on the diversified document resource supply platform, such as teaching plans, summaries, resumes, etc., to meet the specific needs of individuals or teams. Currently, many users hope that the diversified document resource supply platform can quickly generate a PPT (PowerPoint) or a PPT-like presentation based on the document content.

[0003] Therefore, in order to meet the personalized needs of users, it is very important to quickly generate presentations based on document content. Summary of the invention

[0004] The present disclosure provides a method, apparatus, device and medium for training a large model for generating a presentation and generating a presentation.

[0005] According to a first aspect of the present disclosure, a training method for a large model for generating a presentation is provided, wherein the large model includes an outline generation model and a content generation model, and the method includes:

[0006] Acquire training samples; wherein the training samples include: interrelated sample document content, sample presentation outline, sample presentation content, and user feedback data;

[0007] Processing the sample document content using the outline generation model to obtain a predicted presentation outline;

[0008] Processing the sample document content and the sample presentation outline using the content generation model to obtain predicted presentation content;

[0009] The outline generation model is trained according to the predicted presentation outline, the sample presentation outline and the user feedback data, and the content generation model is trained according to the predicted presentation content, the sample presentation content and the user feedback data.

[0010] According to a second aspect of the present disclosure, a method for generating a presentation is provided, comprising:

[0011] Obtaining target document content associated with the presentation to be generated;

[0012] Calling the outline generation model to process the target document content to obtain a presentation outline;

[0013] Calling a content generation model to process the presentation outline and the target document content to obtain presentation content;

[0014] The presentation content is rendered to obtain the presentation.

[0015] According to a third aspect of the present disclosure, a training device for generating a large model of a presentation is provided, wherein the large model includes an outline generation model and a content generation model, and the device includes:

[0016] A first acquisition module, used to acquire training samples; wherein the training samples include: interrelated sample document content, sample presentation outline, sample presentation content and user feedback data;

[0017] A first processing module, configured to process the sample document content using the outline generation model to obtain a predicted presentation outline;

[0018] A second processing module, configured to process the sample document content and the sample presentation outline using the content generation model to obtain predicted presentation content;

[0019] A first training module, configured to train the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data;

[0020] The second training module is used to train the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

[0021] According to a fourth aspect of the present disclosure, there is provided a device for generating a presentation, comprising:

[0022] An acquisition module, used for acquiring target document content associated with the presentation to be generated;

[0023] A first calling module is used to call an outline generation model to process the target document content to obtain a presentation outline;

[0024] A second calling module is used to call a content generation model to process the presentation outline and the target document content to obtain presentation content;

[0025] A rendering module is used to render the presentation content to obtain the presentation.

[0026] According to a fifth aspect of the present disclosure, there is provided an electronic device, including:

[0027] at least one processor; and

[0028] a memory communicatively connected to the at least one processor; wherein,

[0029] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the training method for a large model for generating a presentation proposed in the first aspect of the present disclosure, or execute the method for generating a presentation proposed in the second aspect of the present disclosure.

[0030] According to the sixth aspect of the present disclosure, a non-transitory computer-readable storage medium of computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the training method for a large model for generating a presentation proposed in the first aspect of the present disclosure, or to execute the method for generating a presentation proposed in the second aspect of the present disclosure.

[0031] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the training method for a large model for generating a presentation proposed in the first aspect of the present disclosure, or, when executed, implements the method for generating a presentation proposed in the second aspect of the present disclosure.

[0032] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0034] Figure 1 A flowchart of a training method for a large model for generating a presentation provided in the first embodiment of the present disclosure;

[0035] Figure 2 A flowchart of a training method for a large model for generating a presentation provided in the second embodiment of the present disclosure;

[0036] Figure 3 A flowchart of a training method for a large model for generating a presentation provided in Embodiment 3 of the present disclosure;

[0037] Figure 4 A flowchart of a training method for a large model for generating a presentation provided in Embodiment 4 of the present disclosure;

[0038] Figure 5 A flowchart of a training method for a large model for generating a presentation provided in Embodiment 5 of the present disclosure;

[0039] Figure 6 A training method for a quality assessment model provided in Embodiment 6 of the present disclosure;

[0040] Figure 7 The configuration interface provided by the embodiment of the present disclosure is shown in FIG. Figure 1 ;

[0041] Figure 8 The configuration interface provided by the embodiment of the present disclosure is shown in FIG. Figure 2 ;

[0042] Fig. 9 The configuration interface provided by the embodiment of the present disclosure is shown in FIG. Figure 3 ;

[0043] Fig.10 A flowchart of a method for generating a presentation provided in Embodiment 7 of the present disclosure;

[0044] Fig.11 A schematic diagram of a display interface on the client side provided by an embodiment of the present disclosure;

[0045] Fig.12 It is a schematic diagram of the implementation principle of the embodiment of the present disclosure;

[0046] Fig.13 A schematic diagram of a data insight method provided by an embodiment of the present disclosure;

[0047] Fig.14 A schematic diagram of the structure of a training device for generating a large model of a presentation provided in Embodiment 8 of the present disclosure;

[0048] Fig.15 A schematic diagram of the structure of a presentation document generation device provided in Embodiment 9 of the present disclosure;

[0049] Fig.16 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0050] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0051] A professional and diversified document resource supply platform, dedicated to fully meeting users' diverse document needs in the fields of office, education, etc. Users can easily retrieve various types of documents on the diversified document resource supply platform, such as teaching plans, summaries, resumes, etc., to meet the specific needs of individuals or teams. However, with the rapid development of information technology and the continuous changes in user needs, traditional document query methods have gradually exposed their limitations and are difficult to accurately match users' personalized needs.

[0052] Currently, many users hope that the diversified document resource supply platform can quickly generate a PPT (PowerPoint) or PPT-like presentation based on specific query conditions or document content. More importantly, they hope to be able to integrate the document content they have collected or written during the presentation generation process to achieve rapid integration and efficient display of information. However, traditional presentation generation methods often fail to meet this demand and provide users with satisfactory solutions.

[0053] In the related art, the presentation is taken as PPT for example, and the following three schemes are mainly used to generate PPT:

[0054] The first one is that the PPT is created manually by the user.

[0055] In the current information-intensive social environment, users often need to manually create PPTs based on documents or retrieved materials to meet specific presentation needs. Although this method can ensure the close relevance and high accuracy of PPT content to user needs and improve the quality of the final generated PPT, this method also has significant limitations, namely, it requires users to invest a lot of time and energy in designing and making PPTs, which undoubtedly increases the time and labor costs of PPT creation.

[0056] Moreover, in the process of creating PPT, users not only need to build a clear and logical outline framework, but also must skillfully integrate the content they create or collect into each chapter. In addition, users need to be proficient in document editing tools such as PPT so that they can flexibly adjust the format to ensure the neatness and beauty of the document. At the same time, users also need to have excellent information screening and organization capabilities to ensure the professionalism and rigor of the PPT content.

[0057] In summary, although the manual creation of PPTs by users can bring a high degree of personalization and accuracy, its high requirements for time and resources make it difficult for many users to bear. In the current context of pursuing both quality and efficiency, how to strike a balance between personalized creation and efficient production has become a major challenge facing users.

[0058] The second method is to search for similar PPTs.

[0059] The advantage of searching for similar PPTs through retrieval is that it can significantly save users' time and energy. However, it should be pointed out that search engines mainly provide retrieval results based on text matching, that is, based on the text similarity between the search text and multiple PPTs, they determine the PPTs similar to the search text from multiple PPTs, which often limits their ability to deeply understand the customized specific content and structure in the PPT.

[0060] In addition, since the writing quality of PPTs on the Internet varies, users may need to invest extra time and energy to screen and modify similar PPTs after searching for them to ensure that they can better meet their own needs. Therefore, although the search method provides users with a convenient way to find, users still need to be cautious when using it and carefully evaluate the accuracy and applicability of the search results.

[0061] The third method is to generate PPT with one click through AI technology.

[0062] Users only need to provide the corresponding documents or texts to quickly generate high-quality PPTs. This technology can not only automatically match the appropriate templates and layouts according to the generated content, but also generate corresponding charts based on the data, thus greatly saving users' time and energy.

[0063] At present, the existing AI technology for generating PPT mainly relies on calling large pre-trained models or generating PPT through supervised fine-tuning (SFT) models. However, these methods still have certain limitations in terms of the quality and personalization of generated PPT content.

[0064] Although AI technologies can quickly generate basic PPT content, these technologies have not yet fully utilized the concept of data flywheel. Among them, data flywheel can significantly improve the quality and personalization of generated content through continuous feedback and improvement of the cycle between data input and model output. Therefore, incorporating the concept of data flywheel into AI-generated PPT technology will become a key direction for improving the quality of PPT generation.

[0065] In summary, although the above three solutions can generate corresponding PPTs based on the search data or documents, each has its own limitations and requirements:

[0066] Regarding the first method of "allowing users to write PPTs by themselves", although a PPT that completely meets the user's own needs can be produced, this method has certain requirements on the user's software operation skills. The user needs to be familiar with the PPT production process and tool usage, which undoubtedly takes a lot of time and effort.

[0067] As for the second method, "providing users with some existing similar PPTs as references", although these similar PPTs may not fully meet the needs of users, they at least provide users with a way to quickly obtain information. However, users still need to modify and adjust the PPT content accordingly based on their own document content to ensure that the PPT content matches their needs.

[0068] As for the third method, "quickly customize a PPT for users based on AI technology", the quality of the PPT generated by this method may not fully meet user needs, and the AI ​​traces are more obvious, the output page format is relatively fixed, and the quality needs to be improved.

[0069] Although these three solutions can generate PPT based on search materials or documents, they all have certain operational and technical requirements, requiring users to have certain professional skills and time investment.

[0070] Therefore, in response to at least one of the above-mentioned problems, the present disclosure proposes a method, device, equipment and medium for training and generating presentations of large models.

[0071] The following describes the method, apparatus, device and medium for training a large model and generating a presentation according to an embodiment of the present disclosure with reference to the accompanying drawings.

[0072] A large model refers to a machine learning model with large-scale parameters and complex computational structures. It is usually built from deep neural networks and has billions or even hundreds of billions of parameters. The purpose of designing a large model is to improve the model's expressiveness and predictive performance, and to be able to handle more complex tasks and data. Large models are widely used in various fields, including natural language processing, computer vision, speech recognition, and recommendation systems.

[0073] Figure 1 A flowchart of a training method for a large model for generating a presentation provided in the first embodiment of the present disclosure.

[0074] The presentation includes but is not limited to: PPT (PowerPoint) and presentations similar to PPT.

[0075] Among them, the big model includes the outline generation model and the content generation model.

[0076] The embodiment of the present disclosure is illustrated by the example that the training method for generating a large model for a presentation is configured in a training device for generating a large model for a presentation. The training device can be applied to any electronic device so that the electronic device can perform the training function of the large model for generating a presentation.

[0077] Among them, the electronic device can be any device with computing capabilities, such as a personal computer, a mobile terminal, a server (or a server, a cloud), etc. The mobile terminal can be, for example, a mobile phone, a tablet computer, a personal digital assistant, a wearable device, and other hardware devices with various operating systems, touch screens and / or display screens.

[0078] like Figure 1 As shown, the training method for generating a large model for a presentation may include the following steps S101 to S104:

[0079] Step S101, obtaining training samples; wherein the training samples include: mutually related sample document contents, sample presentation outlines, sample presentation contents and user feedback data.

[0080] The sample presentation content is the content data in the sample presentation, wherein the sample presentation is generated based on the sample document content and the sample presentation outline, and the sample presentation outline is generated based on the sample document content.

[0081] The user feedback data includes positive feedback (such as user payment, likes, positive evaluation data, etc.) and negative feedback (such as dislikes, negative evaluation data, etc.). The user feedback data is obtained by monitoring the sample presentation.

[0082] In the embodiments of the present disclosure, there is no limitation on the method of obtaining training samples. For example, training samples can be obtained from an existing training set, or training samples can be collected online, such as pulling online log data and extracting training samples from the log data, or the training samples can be provided manually, etc. The embodiments of the present disclosure do not limit this.

[0083] Step S102: Process the sample document content using an outline generation model to obtain a predicted presentation outline.

[0084] There is no restriction on the model structure of the outline generation model. For example, the outline generation model may be a KTO (Kahneman-Tversky Optimization) model, a DPO (Direct Preference Optimization) model, or the like.

[0085] In an embodiment of the present disclosure, an outline generation model may be used to process sample document content to obtain a presentation outline output by the outline generation model, which is referred to as a predicted presentation outline in the present disclosure.

[0086] Step S103: Use the content generation model to process the sample document content and the sample presentation outline to obtain predicted presentation content.

[0087] There is no restriction on the model structure of the content generation model. For example, the content generation model may also be a KTO model, a DPO model, etc.

[0088] In an embodiment of the present disclosure, a content generation model may be used to process sample document content and a sample presentation outline to obtain presentation content output by the content generation model, which is referred to as predicted presentation content in the present disclosure.

[0089] Step S104, training the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data, and training the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

[0090] In the disclosed embodiment, the outline generation model can be trained based on the predicted presentation outline, sample presentation outline and user feedback data, so that the trained outline generation model can efficiently align with the user behavior preferences, that is, learn the presentation outline preferred by the user.

[0091] In the disclosed embodiment, the content generation model can also be trained based on the predicted presentation content, sample presentation content and user feedback data, so that the trained content generation model can efficiently align with the user behavior preferences, that is, learn the presentation content preferred by the user.

[0092] The training method of the big model for generating presentations in the embodiment of the present disclosure, during the training stage of the big model, trains the big model according to user feedback information, so that the trained big model can efficiently align with user behavior preferences, that is, improve the quality of presentations generated by the trained big model, meet the personalized needs of users, and improve the user experience in the presentation generation scenario.

[0093] It should be noted that in the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information are all carried out with the user's consent, and are in compliance with relevant laws and regulations and do not violate public order and good morals.

[0094] In order to clearly illustrate how the outline generation model is used to process the sample document content in any embodiment of the present disclosure to obtain a predicted presentation outline, the present disclosure also proposes a training method for generating a large model for presentations.

[0095] Figure 2 This is a flow chart of a training method for a large model for generating a presentation provided in the second embodiment of the present disclosure, wherein the large model includes an outline generation model and a content generation model.

[0096] like Figure 2 As shown, the training method for generating a large model for a presentation may include the following steps S201 to S207:

[0097] Step S201, obtaining training samples; wherein the training samples include: mutually related sample document contents, sample presentation outlines, sample presentation contents and user feedback data.

[0098] For the explanation of step S201, reference may be made to the relevant description in any embodiment of the present disclosure, and no further details will be given here.

[0099] Step S202, obtaining a first prompt template; wherein the first prompt template includes a first task description and a first output format, the first task description is used to instruct the outline generation model to perform the outline extraction task, and the first output format is used to indicate the output format of the outline generation model.

[0100] In the embodiment of the present disclosure, the first prompt template is pre-set for the outline extraction task, and the first prompt template includes a first task description and a first output format.

[0101] The first task description is used to indicate task information to be executed by the outline generation model, that is, to instruct the outline generation model to execute the outline extraction task.

[0102] Among them, the first output format is used to indicate the output format of the outline generation model. Exemplarily, the first output format can be a markdown (or markdown-it) format. The markdown format is used to indicate that the output structure of the outline generation model needs to be a tree structure, that is, the output presentation outline needs to include a multi-level directory outline.

[0103] It should be noted that the first prompt template may be dynamically updated according to actual application requirements, and the embodiments of the present disclosure do not limit this.

[0104] As an example, the first prompt template may be:

[0105] {

[0106] You are an outline generation expert. **Task description**:\nExtract the outline of [text] and output it in markdown format.\n\n

[0107]

Text

[0108] Word

[0109] Outline

[0110] }

[0111] Step S203: fill in the first prompt template according to the sample document content to obtain first prompt information.

[0112] In the disclosed embodiment, the sample document content may be used to fill in the first prompt template to obtain the first prompt information.

[0113] Still taking the example of the first prompt template as shown in step S202, the sample document content can be filled into the "

Text

[0114] Step S204: calling the outline generation model to process the first prompt information to obtain a predicted presentation outline.

[0115] In an embodiment of the present disclosure, the first prompt information may be input into an outline generation model so that the first prompt information is processed by the outline generation model to obtain a predicted presentation outline output by the outline generation model.

[0116] Step S205: Use the content generation model to process the sample document content and the sample presentation outline to obtain predicted presentation content.

[0117] Step S206, training the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data.

[0118] For explanations of steps S205 to S206, reference may be made to the relevant descriptions in any embodiment of the present disclosure, and will not be repeated here.

[0119] In any embodiment of the present disclosure, the outline generation model can be trained by following steps A to C:

[0120] Step A: According to the difference between the predicted presentation outline and the sample presentation outline, determine the value of the loss function corresponding to the outline generation model (referred to as the first loss value in this disclosure).

[0121] Among them, the first loss value is positively correlated with the above difference, that is, the larger the difference is, the larger the first loss value is, and conversely, the smaller the difference is, the smaller the first loss value is.

[0122] Step B: Based on the first loss value, adjust the model parameters in the outline generation model.

[0123] As an example, based on the first loss value, the model parameters in the outline generation model can be adjusted to minimize the first loss value.

[0124] It should be noted that the above only takes the stopping adjustment condition of the model parameters as an example of minimizing the first loss value. In actual application, other stopping adjustment conditions can also be set. For example, the stopping adjustment conditions can also include: the training time reaches the set time, the number of training times reaches the set number of times, the number of training rounds reaches the set number of rounds, etc.

[0125] Step C: Based on each positive feedback and / or negative feedback in the user feedback data, reinforcement learning is performed on the adjusted outline generation model.

[0126] Among them, the reward function of reinforcement learning can be defined, for example, a like is a positive reward and a dislike is a negative reward.

[0127] In summary, using reinforcement learning methods (such as Policy Gradient (policy gradient, which maximizes the expected cumulative reward by directly optimizing the policy function, which estimates the policy gradient and uses the gradient ascent method to update the policy parameters), Actor-Critic (a reinforcement learning algorithm that combines the policy gradient method and the value function method, which combines the policy function (Actor) and the value function (Critic) to improve learning efficiency and stability by optimizing both at the same time), etc.) to further optimize the outline generation model can make the output of the outline generation model after reinforcement learning or the generated presentation outline more in line with the user's actual preferences, thereby improving the generation quality of the presentation.

[0128] Step S207, training the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

[0129] For the explanation of step S207, reference may be made to the relevant description in any embodiment of the present disclosure, and no further details will be given here.

[0130] In the training method for generating a large model for presentations in the embodiment of the present disclosure, the first prompt information can be used as prior information to indicate the content to be extracted and the output format of the outline generation model, which can improve the prediction accuracy of the outline generation model and enable the format of the presentation outline output by the outline generation model to meet the actual application needs of the user.

[0131] In order to clearly illustrate how the content generation model is used in the above embodiments to process sample document content and sample presentation outline to obtain predicted presentation content, the present disclosure also proposes a training method for generating a large model for presentations.

[0132] Figure 3 This is a flow chart of a training method for a large model for generating a presentation provided in Embodiment 3 of the present disclosure, wherein the large model includes an outline generation model and a content generation model.

[0133] like Figure 3 As shown, the training method for generating a large model for a presentation may include the following steps S301 to S307:

[0134] Step S301, obtaining training samples; wherein the training samples include: mutually related sample document contents, sample presentation outlines, sample presentation contents and user feedback data.

[0135] Step S302: Process the sample document content using an outline generation model to obtain a predicted presentation outline.

[0136] For explanations of steps S301 to S302, reference may be made to the relevant descriptions in any embodiment of the present disclosure and will not be repeated here.

[0137] Step S303, obtaining a second prompt template; wherein the second prompt template includes a second task description and a second output format, the second task description is used to instruct the content generation model to perform the content generation task, and the second output format is used to indicate the output format of the content generation model.

[0138] In the disclosed embodiment, the second prompt template is pre-set for the content generation task, and the second prompt template includes a second task description and a second output format. The second task description is used to indicate the task information to be performed by the content generation model, that is, to instruct the content generation model to perform the content generation task; the second output format is used to indicate the output format of the content generation model. The second output format may be different from the first output format.

[0139] It should be noted that the second prompt template may also be dynamically updated according to actual application requirements, and the embodiments of the present disclosure do not limit this.

[0140] Step S304, filling in the second prompt template according to the sample document content and the sample presentation outline to obtain second prompt information.

[0141] In the disclosed embodiment, the sample document content and the sample presentation outline may be used to fill in the second prompt template to obtain the second prompt information.

[0142] Step S305: calling the content generation model to process the second prompt information to obtain predicted presentation content.

[0143] In an embodiment of the present disclosure, the second prompt information may be input into a content generation model so that the content generation model processes the second prompt information to obtain predicted presentation content output by the content generation model.

[0144] Step S306, training the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data.

[0145] Step S307: training the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

[0146] For explanations of steps S306 to S307, reference may be made to the relevant descriptions in any embodiment of the present disclosure, and will not be repeated here.

[0147] In any embodiment of the present disclosure, the content generation model can be trained by following steps A' to C':

[0148] Step A': Determine the value of the loss function corresponding to the content generation model (referred to as the second loss value in this disclosure) based on the difference between the predicted presentation content and the sample presentation content.

[0149] Among them, the second loss value is positively correlated with the above difference, that is, the larger the difference is, the larger the second loss value is, and conversely, the smaller the difference is, the smaller the second loss value is.

[0150] Step B': Based on the second loss value, adjust the model parameters in the content generation model.

[0151] As an example, based on the second loss value, the model parameters in the content generation model can be adjusted to minimize the second loss value.

[0152] It should be noted that the above only uses the stopping adjustment condition of the model parameters as an example of minimizing the second loss value. In actual application, other stopping adjustment conditions can also be set. For example, the stopping adjustment conditions can also include: the training time reaches the set time, the number of training times reaches the set number of training rounds reaches the set number of rounds, and so on.

[0153] Step C': According to each positive feedback and / or negative feedback in the user feedback data, reinforcement learning is performed on the adjusted content generation model.

[0154] Among them, the reward function of reinforcement learning can be defined, for example, a like is a positive reward and a dislike is a negative reward.

[0155] In summary, using reinforcement learning methods (such as Policy Gradient, Actor-Critic, etc.) to further optimize the content generation model can make the output or generated presentation content of the content generation model after reinforcement learning more in line with user preferences, thereby improving the generation quality of the presentation.

[0156] In the training method for generating a large model for presentations of the disclosed embodiment, the second prompt information can be used as a priori information to indicate the task information and output format to be performed by the content generation model, which can improve the prediction accuracy of the content generation model and enable the format of the presentation content output by the content generation model to meet the actual application needs of the user.

[0157] In order to clearly illustrate how to obtain training samples in any embodiment of the present disclosure, the present disclosure also proposes a training method for generating a large model for presentations.

[0158] Figure 4 This is a flow chart of a training method for a large model for generating a presentation provided in Embodiment 4 of the present disclosure, wherein the large model includes an outline generation model and a content generation model.

[0159] like Figure 4 As shown, the training method for generating a large model for a presentation may include the following steps S401 to S408:

[0160] Step S401, obtaining a sample presentation outline by screening from a plurality of candidate presentation outlines.

[0161] Among them, the number of sample presentation outlines can be at least one.

[0162] In the disclosed embodiment, multiple candidate presentation outlines may be screened to obtain a high-quality sample presentation outline.

[0163] In any embodiment of the present disclosure, a high-quality sample presentation outline can be screened by following steps a to c:

[0164] Step a: Obtain multiple candidate presentation outlines.

[0165] As an example, online log data may be pulled and multiple candidate presentation outlines may be extracted from the log data.

[0166] Step b: scoring the plurality of candidate presentation outlines according to the set first quality verification standard to obtain first quality scores of the plurality of candidate presentation outlines.

[0167] Among them, the first quality verification standard is a pre-set quality verification standard. Exemplarily, the first quality verification standard can be used to verify whether the format of the presentation outline meets the requirements (such as verifying whether the format of the presentation outline is Markdown format, whether there is a first-level title, etc.), verifying whether the symbols before each level of title in the presentation outline meet the requirements (such as there is a "#" sign before the first-level title, two consecutive "#" signs before the second-level title, and three consecutive "#" signs before the third-level title), verifying whether the presentation outline contains sensitive information or prohibited information, etc.

[0168] Exemplarily, if the format of the presentation outline meets the requirements, and the symbols before each level of title meet the requirements, and the presentation outline does not contain sensitive information or prohibited information, the first quality score is relatively high; if the format of the presentation outline does not meet the requirements, and / or the symbols before at least one level of title do not meet the requirements, and / or the presentation outline contains sensitive information or prohibited information, the first quality score is relatively low.

[0169] Step c: determining a sample presentation outline from the plurality of candidate presentation outlines according to the first quality scores of the plurality of candidate presentation outlines.

[0170] Exemplarily, a candidate presentation outline having a first quality score higher than a set score threshold may be used as a sample presentation outline.

[0171] In summary, it is possible to filter or screen the candidate presentation outlines according to their quality, thereby improving the quality of the screened sample presentation outlines, and then train the big model (outline generation model and content generation model) based on the high-quality sample presentation outlines, which can improve the training effect and prediction quality of the big model.

[0172] Step S402, obtaining sample document content according to the sample presentation outline; wherein the sample presentation outline is generated according to the sample document content, and / or the sample presentation outline is generated and updated according to the sample document content.

[0173] Exemplarily, sample document content can be extracted from the collected log data, wherein a sample presentation outline is generated based on the sample document content, and / or the sample presentation outline is generated and updated based on the sample document content. For example, AI technology can be used to generate an initial presentation outline based on the sample document content, and the outline can be provided to the user, who can then update the initial presentation outline based on their own needs to obtain a sample presentation outline.

[0174] It should be noted that the sample presentation outline updated by the user can better represent the user's actual preferences and can improve the training effect of the large model (outline generation model and content generation model).

[0175] Step S403, obtaining a sample presentation, wherein the sample presentation is generated according to the sample document content and the sample presentation outline.

[0176] Exemplarily, a sample presentation may be extracted from the collected log data based on sample document content and / or a sample presentation outline, wherein the sample presentation is generated based on the sample document content and the sample presentation outline.

[0177] Step S404, obtaining sample presentation content from the sample presentation, and obtaining user feedback data monitored for the sample presentation.

[0178] The user feedback data includes positive feedback (such as user payment, likes, positive evaluation data, etc.) and negative feedback (such as dislikes, negative evaluation data, etc.). The user feedback data is obtained by monitoring the sample presentation.

[0179] In the embodiment of the present disclosure, the sample presentation may include multiple pages of presentation content, and the sample presentation content may be extracted from the multiple pages of presentation content in the sample presentation.

[0180] Exemplarily, presentation content with traces of modification can be used as sample presentation content, presentation content added by a user can be used as sample presentation content, presentation content with a relatively high quality score can be used as sample presentation content, and so on.

[0181] Step S405, generating training samples according to the sample document content, the sample presentation outline, the sample presentation content and the user feedback data.

[0182] In the disclosed embodiment, user feedback data monitored on the sample presentation may be obtained, and a training sample may be generated based on the sample document content, the sample presentation outline, the sample presentation content, and the user feedback data.

[0183] Step S406: Process the sample document content using the outline generation model to obtain a predicted presentation outline.

[0184] Step S407: Use the content generation model to process the sample document content and the sample presentation outline to obtain predicted presentation content.

[0185] Step S408, training the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data, and training the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

[0186] For explanations of steps S406 to S408, reference may be made to the relevant descriptions in any embodiment of the present disclosure and will not be repeated here.

[0187] The training method for generating a large model of a presentation in the disclosed embodiment can extract sample document content, sample presentations, and sample presentation content from log data collected online according to a sample presentation outline, thereby effectively constructing training samples based on user feedback data, sample document content, sample presentation outline, and sample presentation content monitored for the sample presentation, thereby improving the effectiveness of the training of the large model (outline generation model and content generation model).

[0188] In order to clearly illustrate any embodiment of the present disclosure, the present disclosure also proposes a training method for generating a large model for presentations.

[0189] Figure 5 This is a flow chart of a training method for a large model for generating a presentation provided in Embodiment 5 of the present disclosure, wherein the large model includes an outline generation model and a content generation model.

[0190] like Figure 5 As shown, the training method for generating a large model for a presentation may include the following steps S501 to S509:

[0191] Step S501, obtaining training samples; wherein the training samples include: mutually related sample document contents, sample presentation outlines, sample presentation contents and user feedback data.

[0192] Step S502: Process the sample document content using an outline generation model to obtain a predicted presentation outline.

[0193] Step S503: Use the content generation model to process the sample document content and the sample presentation outline to obtain predicted presentation content.

[0194] Step S504, training the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data, and training the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

[0195] For explanations of steps S501 to S504, reference may be made to the relevant descriptions in any embodiment of the present disclosure, and will not be repeated here.

[0196] Step S505, obtaining a test sample; wherein the test sample includes: mutually related test document content and a test presentation outline.

[0197] The method of obtaining the test samples is similar to that of the training samples, and will not be described in detail here.

[0198] Step S506, using the trained outline generation model to process the test document content to obtain an output presentation outline.

[0199] Among them, the method of generating the output presentation outline is similar to the method of generating the predicted presentation outline, which will not be described here.

[0200] Step S507: Use the trained content generation model to process the test document content and the test presentation outline to obtain output presentation content.

[0201] Among them, the method of generating the output presentation content is similar to the method of generating the predicted presentation content, which will not be described here.

[0202] Step S508: Use a quality assessment model to perform quality assessment on the output presentation outline and the output presentation content to obtain an assessment result.

[0203] The evaluation results are used to indicate the generation quality of the output presentation outline and the generation quality of the output presentation content.

[0204] As an example, the evaluation results may include a quality score of the output presentation outline (referred to as the second quality score in this disclosure) and a quality score of the output presentation content (referred to as the third quality score in this disclosure), wherein the second quality score is used to indicate the degree of alignment between the output of the trained outline generation model and the user behavior preferences, and the third quality score is used to indicate the degree of alignment between the output of the trained content generation model and the user behavior preferences.

[0205] In the disclosed embodiment, a trained quality assessment model (or referred to as a reward model) may be used to perform quality assessment on the output presentation outline and the output presentation content to obtain an assessment result.

[0206] Step S509: if the evaluation result does not meet the set requirements, the trained outline generation model and / or the trained content generation model are retrained.

[0207] In the embodiment of the present disclosure, when the evaluation result does not meet the set requirements, the trained outline generation model can be retrained, and / or the trained content generation model can be retrained.

[0208] In any embodiment of the present disclosure, when the evaluation result only indicates that the generation quality of the output presentation outline is relatively low, the trained outline generation model can be trained again. For example, other training samples can be used to iteratively optimize the trained outline generation model.

[0209] As an example, when the second quality score of the output presentation outline in the evaluation result is lower than the first score threshold, the trained outline generation model is retrained.

[0210] The first score threshold is a preset score threshold.

[0211] In any embodiment of the present disclosure, when the evaluation result only indicates that the generation quality of the output presentation content is relatively low, the trained content generation model can be trained again. For example, other training samples can be used to iteratively optimize the trained content generation model.

[0212] As another example, when the third quality score of the output presentation content in the evaluation result is lower than the second score threshold, the trained content generation model is retrained.

[0213] The second score threshold is a pre-set score threshold. It should be noted that the embodiment of the present disclosure does not limit the magnitude relationship between the first score threshold and the second score threshold. For example, the first score threshold may be greater than the second score threshold, or the first score threshold may be less than the second score threshold, or the first score threshold may be equal to the second score threshold.

[0214] In any embodiment of the present disclosure, when the evaluation results simultaneously indicate that the generation quality of the output presentation outline is relatively low and the generation quality of the output presentation content is relatively low, the trained outline generation model and the trained content generation model can be retrained.

[0215] As another example, when the second quality score of the output presentation outline in the evaluation results is lower than the first score threshold, and the third quality score of the output presentation content is lower than the second score threshold, the trained outline generation model and the trained content generation model are retrained.

[0216] Therefore, when the prediction accuracy of the trained outline generation model does not meet the requirements, the trained outline generation model can be iteratively optimized to improve the quality of the presentation outline predicted by the outline generation model, and / or, when the prediction accuracy of the trained content generation model does not meet the requirements, the trained content generation model can be iteratively optimized to improve the quality of the presentation content predicted by the content generation model, thereby improving the quality of the final generated presentation and improving the user experience in the presentation generation scenario.

[0217] The training method for the large model for generating presentations of the disclosed embodiments can implement the use of a trained quality assessment model to perform quality assessment on the outputs of a trained outline generation model and a trained content generation model, and if the assessment result is not ideal, the trained outline generation model and / or the trained content generation model can be retrained, which can improve the prediction accuracy of the trained large model (outline generation model and content generation model), thereby improving the quality of the final generated presentation.

[0218] In order to clearly explain how the quality assessment model in the above embodiment is trained, the present disclosure also proposes a training method for the quality assessment model.

[0219] Figure 6 This is a training method for the quality assessment model provided in Embodiment 6 of the present disclosure.

[0220] like Figure 6 As shown, the training method of the quality assessment model may include the following steps S601 to S603:

[0221] Step S601, scoring the sample presentation outline according to the set second quality verification standard to obtain a fourth quality score of the sample presentation outline.

[0222] The second quality verification standard is a preset quality verification standard. The second quality verification standard may be the same as the first quality verification standard, or may be different. The embodiment of the present disclosure does not impose any limitation on this.

[0223] Exemplarily, the second quality verification standard can be used to verify whether the symbols before the titles at each level in the presentation outline meet the requirements (such as there is a "#" sign before the first-level title, two consecutive "#" signs before the second-level title, and three consecutive "#" signs before the third-level title), verify whether the length of the titles at each level meets the requirements (such as the first-level title does not exceed 25 words, the second-level title, the third-level title and the fourth-level title do not exceed 20 words), verify whether the semantics of the titles at each level are related, verify whether the hierarchical structure of the titles at each level meets the requirements, and so on.

[0224] For example, if the symbol before at least one level of title in the presentation outline does not meet the requirements, and / or the length of at least one level of title does not meet the requirements, and / or the semantics of at least two levels of titles are irrelevant, the fourth quality score is relatively low; if the symbols before each level of title in the presentation outline meet the requirements, and the length of each level of title meets the requirements, and the semantics of each level of title are relevant, the fourth quality score is relatively high.

[0225] Step S602: Score the sample presentation content according to the set third quality verification standard to obtain a fifth quality score of the sample presentation content.

[0226] The third quality verification standard is a preset quality verification standard, and the third quality verification standard may be different from the second quality verification standard.

[0227] Exemplarily, the third quality verification standard can be used to verify the text clarity, readability, information accuracy, visual effects (fonts, colors and layout, etc.), quality of multimedia elements, semantic relevance between multimedia elements and sample presentation content, etc. of the sample presentation content.

[0228] For example, the sample presentation content can be scored according to the third quality verification standard to obtain the text clarity, readability index, information accuracy, visual effect index, quality index of multimedia elements, and semantic relevance between multimedia elements and the sample presentation content of the sample presentation content; based on the above indicators, the fifth quality score of the sample presentation content is calculated.

[0229] Step S603: Perform multi-task training on the quality assessment model according to the sample presentation outline and the corresponding fourth quality score, and according to the sample presentation content and the corresponding fifth quality score.

[0230] Among them, the multiple tasks include: outline quality assessment task and content quality assessment task.

[0231] In the disclosed embodiment, the quality assessment model may be multi-task trained based on the sample presentation outline, the fourth quality score of the sample presentation outline, the sample presentation content, and the fifth quality score of the sample presentation content.

[0232] In any embodiment of the present disclosure, the quality assessment model can be trained by following steps a' to c':

[0233] Step a': Use a quality assessment model to perform quality assessment on the sample presentation outline to obtain a quality score of the sample presentation outline, which is recorded as a first assessment score in this disclosure.

[0234] Step b': Use a quality assessment model to perform quality assessment on the sample presentation content to obtain a quality score of the sample presentation content, which is recorded as a second assessment score in this disclosure.

[0235] Step c': performing multi-task training on the quality assessment model according to the difference between the first assessment score and the fourth quality score (hereinafter referred to as the first difference), and according to the difference between the second assessment score and the fifth quality score (hereinafter referred to as the second difference).

[0236] As an example, the value of the loss function corresponding to the quality assessment model can be determined based on the first difference and the second difference, which is recorded as the third loss value in the present disclosure. Thus, the quality assessment model can be multi-task trained based on the third loss value to minimize the third loss value.

[0237] Among them, the third loss value is positively correlated with both the first difference and the second difference.

[0238] It should be noted that the above only uses the termination training condition of the quality assessment model as an example of minimizing the third loss value. In actual application, other termination training conditions can also be set. For example, the termination training conditions can also include: the training time reaches the set time, the number of training times reaches the set number, the number of training rounds reaches the set number of rounds, and so on.

[0239] Therefore, based on the evaluation scores output by the quality assessment model and the differences between the quality scores calculated according to the quality verification standards, multi-task supervised training of the quality assessment model can improve the learning effect and prediction accuracy of the quality assessment model.

[0240] In any embodiment of the present disclosure, before training the outline generation model, the content generation model and the quality assessment model, the outline generation model, the content generation model and the quality assessment model can be manually selected from a plurality of basic models and the hyperparameters of each model can be configured.

[0241] As an example, in response to a configuration operation triggered by relevant personnel, a model to be trained can be selected from multiple basic models; wherein the model to be trained includes an outline generation model, a content generation model, and a quality assessment model, and the hyperparameters of the model to be trained can be configured.

[0242] Among them, hyperparameters include but are not limited to: learning rate (determines the step size of model parameter update in each iteration), iteration round, sequence length (the maximum length of the sequence input to the model), prediction ratio, data split ratio (the ratio of dividing the data set into training set and validation set), etc.

[0243] Among them, the prediction ratio is defined as follows: learning rate warmup (Warmup) refers to gradually increasing the learning rate at the beginning of training until it reaches the set learning rate, and then gradually decreasing it. The prediction ratio refers to the ratio of the warmup steps to the total training steps.

[0244] As an example, you can use Figure 7 Use the configuration interface shown to select the basic model and model version required for training, and use Figures 8 to 9 The configuration interface shown in the figure is used to configure the model's hyperparameters, etc. Figure 8 The useCls in the code refers to whether the features extracted or processed by the model select the position of the cls (Classification) token (character).

[0245] In summary, manually selecting the outline generation model, content generation model, and quality assessment model to be trained from multiple basic models and configuring the hyperparameters of each model can take into account the actual application needs, model training effect, and training efficiency.

[0246] The training method of the quality assessment model of the embodiment of the present disclosure can improve the learning effect and prediction accuracy of the quality assessment model by performing multi-task supervised training on the quality assessment model in advance.

[0247] The above are various embodiments corresponding to the training method of the large model for generating a presentation. The present disclosure also proposes an application method of the large model for generating a presentation (ie, a method for generating a presentation).

[0248] Fig.10 A flowchart of a method for generating a presentation provided in Embodiment 7 of the present disclosure.

[0249] The method for generating a presentation according to the embodiment of the present disclosure may be applied to a client.

[0250] like Fig.10 As shown, the method for generating a presentation may include the following steps S1001 to S1004:

[0251] Step S1001, obtaining target document content associated with the presentation to be generated.

[0252] The presentation includes but is not limited to: PPT (PowerPoint) and presentations similar to PPT.

[0253] In an embodiment of the present disclosure, the client may obtain target document content associated with the presentation to be generated.

[0254] In any embodiment of the present disclosure, the target document content may be manually input by a user at the client side, that is, the client side may obtain the target document content input by the user in response to an input operation triggered by the user.

[0255] Among them, the input method of the target document content includes but is not limited to touch input (such as sliding, clicking, etc.), keyboard input, voice input, etc.

[0256] As an example, taking the presentation as PPT, users can Fig.11 The target document content is input into the input box shown in area 111 in the display interface of the client shown.

[0257] In any embodiment of the present disclosure, a user on the client side may manually upload a document and extract the target document content from the uploaded document. That is, the client may obtain the first document uploaded by the user in response to an upload operation triggered by the user, and perform content analysis on the first document to obtain the target document content.

[0258] There is no restriction on the document format of the first document. For example, the document format includes but is not limited to: PDF (Portable Document Format), WORD, TXT (text document, a text format), PNG (Portable Network Graphics, a bitmap format using a lossless compression algorithm), etc.

[0259] As an example, taking the presentation as PPT, the user can click Fig.11 The upload control shown in the middle area 112 is used to upload the first document, and then the content of the first document can be parsed to obtain the target document content.

[0260] In any embodiment of the present disclosure, the user can also search for the document he needs from the document library by searching, and extract the target document content from the document. That is, the client can query the document library for candidate documents that semantically match the search term in response to the search operation (or search operation) triggered by the user, and determine the second document selected by the user from the candidate documents, and perform content analysis on the second document to obtain the target document content.

[0261] The search term may be manually input by the user, or the search term may be selected by the user from candidate terms. The input method of the search term includes but is not limited to touch input (such as sliding, clicking, etc.), keyboard input, voice input, etc.

[0262] The candidate words include associated words and / or preset words, wherein the associated words, also called automatic associated words or search drop-down words, are determined based on the user's historical search behavior and hot search trends. The historical search behavior includes but is not limited to: historical search keywords, search time, search frequency, etc.; hot search trends can be triggered by various factors such as current hotspots, social events, popular culture, etc.

[0263] Among them, preset words are words or phrases that are pre-set based on factors such as user search behavior, seasonal events, and popular events.

[0264] There is no restriction on the document format of the second document. For example, the document format includes but is not limited to: PDF, WORD, TXT, PNG, etc.

[0265] In summary, it is possible to obtain the target document content associated with the presentation to be generated based on different methods, which can improve the flexibility and applicability of the method.

[0266] Step S1002, calling the outline generation model to process the target document content to obtain a presentation outline.

[0267] Exemplarily, the outline generation model can be deployed on the server side, and the client can call the outline generation model on the server side to process the target document content to obtain the presentation outline. For example, the client can send the target document content to the server side, and the server side uses the outline generation model to process the target document content to obtain the presentation outline, and sends the presentation outline to the client side, and correspondingly, the client can receive the presentation outline.

[0268] Optionally, the client can also display the presentation outline to the user, and the user can determine whether to modify or update the presentation outline according to his or her own needs. If so, the client can perform at least one of the following update operations on the presentation outline when detecting the outline modification operation triggered by the user: updating the position of at least one node in the presentation outline, updating the content of at least one node in the presentation outline, deleting at least one node in the presentation outline, and adding at least one node in the presentation outline.

[0269] Step S1003, calling the content generation model to process the presentation outline and the target document content to obtain the presentation content.

[0270] Exemplarily, the content generation model can be deployed on the server side, and the client can call the content generation model on the server side to process the presentation outline and the target document content to obtain the presentation content. For example, the client can send the presentation outline and the target document content to the server side, and the server side uses the content generation model to process the presentation outline and the target document content to obtain the presentation content, and sends the presentation content to the client side, and correspondingly, the client side can receive the presentation content.

[0271] Optionally, after the user updates the presentation outline, the client may call the content generation model to process the updated presentation outline and the target document content to obtain the presentation content.

[0272] Step S1004, rendering the presentation content to obtain a presentation.

[0273] In the disclosed embodiment, the client may render the presentation content to obtain a presentation.

[0274] As an application scenario, using the presentation as a PPT example, users can Fig.11 The user can input the target document content by clicking the input box shown in area 111 in the display interface of the client shown in the figure, or the user can upload the first document by clicking the upload control shown in area 112, and parse the content of the first document to obtain the target document content. After that, the user can click the "Smart PPT" option shown in area 113 to generate a PPT with one click, that is, call the big model of the server to automatically generate a PPT according to the target document content.

[0275] The method for generating a presentation of the disclosed embodiment uses a trained outline generation model and a content generation model to generate a presentation, which can improve the generation quality and efficiency of the presentation. In addition, the user only needs to provide the target document content to automatically generate the presentation, which can reduce the user's operation steps and lower the technical and usage thresholds.

[0276] In any embodiment of the present disclosure, using a presentation as an example of PPT, a set of PPT generation technology based on a customized data flywheel mode can be developed in combination with advanced big model technology, which can accurately identify the core needs of users, support users to provide various documents and upload document content in multiple formats (such as PDF, WORD, TXT, PNG, etc.), and automatically generate professional and high-quality PPTs based on the document content. In order to ensure the professionalism and accuracy of the PPT generated by the big model, multiple strategies can be adopted, including introducing a posteriori signals and using professional document libraries to improve the accuracy and professionalism of the PPT content.

[0277] The application scenarios of the above technologies include but are not limited to: office environments, educational fields and other scenarios that require efficient information display. They aim to provide users with better-quality and more efficient services, meet the diverse needs of different users in PPT generation, and help users integrate information more efficiently and improve work and learning efficiency.

[0278] AI-generated PPT technology can not only efficiently meet user needs, but also give users great flexibility, allowing users to edit the PPT outline generated by AI technology, thereby creating a unique PPT. How to improve the quality of PPT generated by AI technology through data flywheel technology is a major feature of the present disclosure. For example, the present disclosure provides a PPT generation technology under the data flywheel mode, and its implementation principle is as follows: Fig.12 As shown, it mainly includes the following steps:

[0279] Step 1: Data construction of the data flywheel.

[0280] Step 1.1: PPT outline data: Pull the PPT outline data of online users before and after editing every day, and filter out high-quality PPT outline data that meets user needs through data insights.

[0281] Step 1.2: PPT content data: Pull the PPT content data of online users who modify or insert one page every day, as well as user feedback data including user payment, likes, dislikes, positive comments, and negative comments.

[0282] Step 1.3: Conduct data insights (enhancement, cleaning, and enhancement) on the constructed PPT outline data and PPT content data, and filter out data that meets the requirements.

[0283] Step 1.4: Supplement the prompt (referred to as the first prompt information and the second prompt information in this disclosure) to the data in step 1.3 to construct a training data set required for model training; wherein the training data set includes multiple training samples.

[0284] Step 1.5: Perform data quality assessment on the training samples in the training dataset of step 1.4.

[0285] Step 2: Model training of data flywheel.

[0286] Step 2.1: Training of generation models (including outline generation model and content generation model): Use PPT outline data, PPT content data and user feedback data to train the outline generation model and content generation model respectively (such as KTO (using Kahneman-Tversky optimization method, model training based on user positive feedback and / or negative feedback, and efficient alignment of user behavior preferences) model, DPO (direct preference optimization, which refers to the precise control of large model output by directly optimizing the language model) model).

[0287] Step 2.2: Quality Assessment Model ( Fig.12 Training of the reward model): The quality assessment model is trained based on the data evaluated in step 1.5 to address how to learn and follow human-desired behaviors more effectively.

[0288] Step 3: Model evaluation of the data flywheel.

[0289] Step 3.1: Pull the online log data of the most recent day for cleaning.

[0290] Step 3.2: Use the outline generation model and content generation model trained in step 2.1 to process the log data to obtain the PPT outline and PPT content.

[0291] Step 3.3: Use the quality assessment model trained in step 2.2 to perform quality assessment on the PPT outline and PPT content generated in step 3.2.

[0292] Step 4: The data flywheel model is launched.

[0293] Step 4.1: Put the model in step 2.1 that meets the requirements online, that is, put the model with higher evaluation quality online.

[0294] It should be noted that Fig.12 GSB in this sentence stands for Good vs Samevs Bad (used for comparison between two models, if model A is better than model B, it is good; if model A is the same as model B, it is same; if model A is worse than model B, it is bad).

[0295] As an example, the collected online logs may be shown in Table 1.

[0296] Table 1 Online log

[0297]

[0298] The online log includes document content and PPT outlines. The number of PPT outlines can be multiple. As shown in Table 1, the same online log contains two PPT outlines.

[0299] Among them, the first-level title is on the right side of “#”, the second-level title is on the right side of “##”, and the third-level title is on the right side of “###”.

[0300] As an example, data insights can be Fig.13 As shown, inappropriate data can be cleaned up. For example, training samples with an appropriate number of characters can be selected from multiple training samples. For example, the number of characters in the training samples selected by default is between [1766, 2319], accounting for 24.68% of the samples. The number of characters in the training samples after screening is between [372, 470], accounting for 11.13% of the samples.

[0301] As an example, the construction method of Prompt can be shown in Table 2.

[0302] Table 2 Prompt construction of outline generation model

[0303]

[0304] As an example, the data quality check standard may be shown in Table 3.

[0305] Table 3 Check criteria for PPT outline

[0306]

[0307] As an example, during the training phase of the model, you can use Figure 7 Use the configuration interface shown to select the basic model and model version required for training, and use Figures 8 to 9 Use the configuration interface shown to configure the model's hyperparameters, etc.

[0308] With the above Figures 1 to 6 Corresponding to the training method for generating a large model of a presentation provided in the embodiment, the present disclosure also provides a training device for generating a large model of a presentation. Since the training device for generating a large model of a presentation provided in the embodiment of the present disclosure is similar to the training method for generating a large model of a presentation provided in the embodiment of the present disclosure, Figures 1 to 6The embodiment provides a training method for a large model for generating a presentation, and therefore the implementation method of the training method for a large model for generating a presentation is also applicable to the training device for generating a large model for generating a presentation provided in the embodiment of the present disclosure, and will not be described in detail in the embodiment of the present disclosure.

[0309] Fig.14 This is a schematic diagram of the structure of a training device for generating a large model of a presentation provided in Embodiment 8 of the present disclosure, wherein the large model includes an outline generation model and a content generation model.

[0310] like Fig.14 As shown, the training device 1400 for generating a large model of a presentation may include: a first acquisition module 1410 , a first processing module 1420 , a second processing module 1430 , a first training module 1440 and a second training module 1450 .

[0311] The first acquisition module 1410 is used to acquire training samples; wherein the training samples include: interrelated sample document content, sample presentation outline, sample presentation content, and user feedback data;

[0312] A first processing module 1420 is used to process the sample document content using the outline generation model to obtain a predicted presentation outline;

[0313] A second processing module 1430 is used to process the sample document content and the sample presentation outline using a content generation model to obtain predicted presentation content;

[0314] A first training module 1440, for training an outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data;

[0315] The second training module 1450 is used to train the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

[0316] In a possible implementation of the embodiment of the present disclosure, the first processing module 1420 is used to: obtain a first prompt template; wherein the first prompt template includes a first task description and a first output format, the first task description is used to instruct the outline generation model to perform an outline extraction task, and the first output format is used to indicate the output format of the outline generation model; fill the first prompt template according to the sample document content to obtain first prompt information; call the outline generation model to process the first prompt information to obtain a predicted presentation outline.

[0317] In a possible implementation of the embodiment of the present disclosure, the first training module 1440 is used to: determine a first loss value based on the difference between the predicted presentation outline and the sample presentation outline; adjust the model parameters in the outline generation model based on the first loss value; and perform reinforcement learning on the adjusted outline generation model based on each positive feedback and / or negative feedback in the user feedback data.

[0318] In a possible implementation of the embodiment of the present disclosure, the second processing module 1430 is used to: obtain a second prompt template; wherein the second prompt template includes a second task description and a second output format, the second task description is used to instruct the content generation model to perform the content generation task, and the second output format is used to indicate the output format of the content generation model; fill in the second prompt template according to the sample document content and the sample presentation outline to obtain second prompt information; call the content generation model to process the second prompt information to obtain predicted presentation content.

[0319] In a possible implementation of the embodiments of the present disclosure, the second training module 1450 is used to: determine a second loss value based on the difference between the predicted presentation content and the sample presentation content; adjust the model parameters in the content generation model based on the second loss value; and perform reinforcement learning on the adjusted content generation model based on each positive feedback and / or negative feedback in the user feedback data.

[0320] In a possible implementation of the embodiment of the present disclosure, the first acquisition module 1410 is used to: filter a sample presentation outline from multiple candidate presentation outlines; obtain sample document content based on the sample presentation outline; wherein the sample presentation outline is generated based on the sample document content, and / or the sample presentation outline is generated and updated based on the sample document content; obtain a sample presentation, and obtain sample presentation content from the sample presentation; wherein the sample presentation is generated based on the sample document content and the sample presentation outline; obtain user feedback data monitored on the sample presentation, and generate a training sample based on the sample document content, the sample presentation outline, the sample presentation content and the user feedback data.

[0321] In a possible implementation of the embodiment of the present disclosure, the first acquisition module 1410 is used to: acquire multiple candidate presentation outlines; score the multiple candidate presentation outlines according to a set first quality verification standard to obtain first quality scores of the multiple candidate presentation outlines; and determine a sample presentation outline from the multiple candidate presentation outlines according to the first quality scores of the multiple candidate presentation outlines.

[0322] In a possible implementation of the embodiment of the present disclosure, the training device 1400 for generating a large model of a presentation further includes:

[0323] The second acquisition module is used to acquire a test sample; wherein the test sample includes: interrelated test document content and test presentation outline;

[0324] A third processing module is used to process the test document content using the trained outline generation model to obtain an output presentation outline;

[0325] A fourth processing module, for processing the test document content and the test presentation outline using the trained content generation model to obtain output presentation content;

[0326] An evaluation module, used for using a quality evaluation model to perform quality evaluation on the output presentation outline and the output presentation content to obtain an evaluation result;

[0327] The third training module is used to retrain the trained outline generation model and / or the trained content generation model when the evaluation result does not meet the set requirements.

[0328] In a possible implementation of the embodiment of the present disclosure, the evaluation result includes a second quality score of the output presentation outline and a third quality score of the output presentation content, the second quality score is used to indicate the degree of alignment between the output of the trained outline generation model and the user behavior preference, and the third quality score is used to indicate the degree of alignment between the output of the trained content generation model and the user behavior preference;

[0329] The third training module is used to: retrain the trained outline generation model when the second quality score is lower than the first score threshold; and / or retrain the trained content generation model when the third quality score is lower than the second score threshold.

[0330] In a possible implementation of the embodiment of the present disclosure, the quality assessment model is trained using the following modules:

[0331] A first scoring module is used to score the sample presentation outline according to a set second quality verification standard to obtain a fourth quality score of the sample presentation outline;

[0332] A second scoring module is used to score the sample presentation content according to a set third quality verification standard to obtain a fifth quality score of the sample presentation content;

[0333] The fourth training module is used to perform multi-task training on the quality assessment model according to the sample presentation outline and the corresponding fourth quality score, and according to the sample presentation content and the corresponding fifth quality score.

[0334] In a possible implementation of the embodiments of the present disclosure, the fourth training module is used to: use the quality assessment model to perform quality assessment on the sample presentation outline to obtain a first assessment score; use the quality assessment model to perform quality assessment on the sample presentation content to obtain a second assessment score; and perform multi-task training on the quality assessment model based on the difference between the first assessment score and the fourth quality score, and based on the difference between the second assessment score and the fifth quality score.

[0335] In a possible implementation of the embodiment of the present disclosure, the training device 1400 for generating a large model of a presentation further includes:

[0336] A selection module, configured to select a model to be trained from a plurality of basic models in response to a configuration operation; wherein the model to be trained includes an outline generation model, a content generation model, and a quality assessment model;

[0337] Configuration module, used to configure the hyperparameters of the model to be trained.

[0338] The training device for generating a large model of a presentation in the embodiment of the present disclosure, during the training stage of the large model, trains the large model according to user feedback information, so that the trained large model can be efficiently aligned with user behavior preferences, that is, the quality of the presentation generated by the trained large model is improved, the personalized needs of the user are met, and the user experience in the presentation generation scenario is improved.

[0339] With the above Fig.10 Corresponding to the method for generating a presentation provided in the embodiment, the present disclosure further provides a device for generating a presentation. Fig.10 The method for generating a presentation provided in the embodiment corresponds to the method, so the implementation of the method for generating a presentation is also applicable to the device for generating a presentation provided in the embodiment of the present disclosure, and will not be described in detail in the embodiment of the present disclosure.

[0340] Fig.15 This is a structural diagram of a presentation generation device provided in Embodiment 9 of the present disclosure.

[0341] like Fig.15 As shown, the presentation generating device 1500 may include: an acquisition module 1510 , a first calling module 1520 , a second calling module 1530 and a rendering module 1540 .

[0342] The acquisition module 1510 is used to acquire the target document content associated with the presentation to be generated;

[0343] The first calling module 1520 is used to call the outline generation model to process the target document content to obtain a presentation outline;

[0344] The second calling module 1530 is used to call the content generation model to process the presentation outline and the target document content to obtain the presentation content;

[0345] The rendering module 1540 is used to render the presentation content to obtain a presentation.

[0346] In a possible implementation of the embodiment of the present disclosure, the acquisition module 1510 is configured to perform any of the following:

[0347] In response to the input operation, obtaining the input target document content;

[0348] In response to the upload operation, the uploaded first document is obtained, and the content of the first document is parsed to obtain the content of the target document;

[0349] In response to the search operation, candidate documents that semantically match the search term are searched, a second document is determined from the candidate documents, and content of the second document is parsed to obtain the target document content.

[0350] The presentation generation device of the embodiment of the present disclosure uses a trained outline generation model and a content generation model to generate a presentation, which can improve the generation quality and efficiency of the presentation. In addition, the user only needs to provide the target document content to automatically generate the presentation, which can reduce the user's operation steps and lower the technical and usage thresholds.

[0351] In order to implement the above embodiments, the present disclosure also provides an electronic device, which may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the training method for generating a large model for a presentation or the method for generating a presentation proposed in any of the above embodiments of the present disclosure.

[0352] In order to implement the above embodiments, the present disclosure also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the training method for generating a large model for a presentation or the method for generating a presentation proposed in any of the above embodiments of the present disclosure.

[0353] In order to implement the above embodiments, the present disclosure also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the training method for generating a large model for generating a presentation or the method for generating a presentation proposed in any of the above embodiments of the present disclosure.

[0354] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0355] Fig.16 A schematic block diagram of an example electronic device that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0356] like Fig.16 As shown, the electronic device 1600 includes a computing unit 1601, which can perform various appropriate actions and processes according to a computer program stored in a ROM (Read-Only Memory) 1602 or a computer program loaded from a storage unit 1607 to a RAM (Random Access Memory) 1603. In RAM 1603, various programs and data required for the operation of the device 1600 can also be stored. The computing unit 1601, ROM 1602, and RAM 1603 are connected to each other via a bus 1604. An I / O (Input / Output) interface 1605 is also connected to the bus 1604.

[0357] A number of components in the device 1600 are connected to the I / O interface 1605, including: an input unit 1606, such as a keyboard, a mouse, etc.; an output unit 1607, such as various types of displays, speakers, etc.; a storage unit 1608, such as a disk, an optical disk, etc.; and a communication unit 1609, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1609 allows the device 1600 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0358] The computing unit 1601 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1601 include, but are not limited to, a CPU (Central Processing Unit), a GPU (Graphic Processing Units), various dedicated AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, a DSP (Digital Signal Processor), and any appropriate processor, controller, microcontroller, etc. The computing unit 1601 performs the various methods and processes described above, such as the training method for generating a large model of a presentation or the method for generating a presentation. For example, in some embodiments, the training method for generating a large model of a presentation or the method for generating a presentation may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1608. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 1600 via the ROM 1602 and / or the communication unit 1609. When the computer program is loaded into RAM 1603 and executed by computing unit 1601, one or more steps of the training method for generating a large model of a presentation or the method for generating a presentation described above may be performed. Alternatively, in other embodiments, computing unit 1601 may be configured to perform the training method for generating a large model of a presentation or the method for generating a presentation described above in any other appropriate manner (e.g., by means of firmware).

[0359] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application Specific Standard Products), SOCs (System On Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that may be executed and / or interpreted on a programmable system including at least one programmable processor that may be a dedicated or general-purpose programmable processor that may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0360] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0361] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only-Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0362] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0363] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: LAN (Local Area Network), WAN (Wide Area Network), the Internet, and blockchain networks.

[0364] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services (Virtual Private Server). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0365] It should be noted that artificial intelligence is a discipline that studies how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.), and includes both hardware-level and software-level technologies. Artificial intelligence hardware technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, and big data processing; artificial intelligence software technologies mainly include computer vision technology, speech recognition technology, natural language processing technology, as well as machine learning / deep learning, big data processing technology, knowledge graph technology, and other major directions.

[0366] According to the technical solution of the embodiment of the present disclosure, during the training stage of the big model, the big model is trained according to user feedback information, so that the trained big model can be efficiently aligned with user behavior preferences, that is, the quality of the presentation generated by the trained big model is improved, the personalized needs of the user are met, and the user experience in the presentation generation scenario is improved.

[0367] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0368] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for training a large model for generating presentations, wherein: The large model includes an outline generation model and a content generation model, and the method includes: Acquire training samples; wherein the training samples include: interrelated sample document content, sample presentation outline, sample presentation content, and user feedback data; Processing the sample document content using the outline generation model to obtain a predicted presentation outline; Using the content generation model to process the sample document content and the sample presentation outline to obtain predicted presentation content; The outline generation model is trained according to the predicted presentation outline, the sample presentation outline and the user feedback data, and the content generation model is trained according to the predicted presentation content, the sample presentation content and the user feedback data.

2. The method according to claim 1, wherein: The step of using the outline generation model to process the sample document content to obtain a predicted presentation outline includes: Acquire a first prompt template; wherein the first prompt template includes a first task description and a first output format, the first task description is used to instruct the outline generation model to perform an outline extraction task, and the first output format is used to indicate an output format of the outline generation model; Filling the first prompt template according to the sample document content to obtain first prompt information; The outline generation model is called to process the first prompt information to obtain the predicted presentation outline.

3. The method according to claim 1 or 2, wherein: The step of training the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data comprises: determining a first loss value based on a difference between the predicted presentation outline and the sample presentation outline; Based on the first loss value, adjusting model parameters in the outline generation model; According to each positive feedback and / or negative feedback in the user feedback data, reinforcement learning is performed on the adjusted outline generation model.

4. The method according to claim 1, wherein: The using the content generation model to process the sample document content and the sample presentation outline to obtain predicted presentation content includes: Acquire a second prompt template; wherein the second prompt template includes a second task description and a second output format, the second task description is used to instruct the content generation model to perform the content generation task, and the second output format is used to indicate the output format of the content generation model; Filling the second prompt template according to the sample document content and the sample presentation outline to obtain second prompt information; The content generation model is called to process the second prompt information to obtain the predicted presentation content.

5. The method according to claim 1 or 4, wherein: The training of the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data comprises: determining a second loss value based on a difference between the predicted presentation content and the sample presentation content; Based on the second loss value, adjusting a model parameter in the content generation model; According to each positive feedback and / or negative feedback in the user feedback data, reinforcement learning is performed on the adjusted content generation model.

6. The method according to claim 1, wherein: The obtaining of training samples comprises: Screening and obtaining the sample presentation outline from a plurality of candidate presentation outlines; According to the sample presentation outline, the sample document content is obtained; wherein the sample presentation outline is generated according to the sample document content, and / or the sample presentation outline is generated and updated according to the sample document content; Obtaining a sample presentation, and obtaining the sample presentation content from the sample presentation; wherein the sample presentation is generated according to the sample document content and the sample presentation outline; The user feedback data monitored for the sample presentation is obtained, and the training sample is generated according to the sample document content, the sample presentation outline, the sample presentation content and the user feedback data.

7. The method according to claim 6, wherein: The step of screening the sample presentation outline from a plurality of candidate presentation outlines comprises: Obtain multiple candidate presentation outlines; Scoring the plurality of candidate presentation outlines according to a set first quality verification standard to obtain first quality scores of the plurality of candidate presentation outlines; The sample presentation outline is determined from the plurality of candidate presentation outlines according to the first quality scores of the plurality of candidate presentation outlines.

8. The method according to claim 1, wherein: The method further comprises: Acquire a test sample; wherein the test sample includes: interrelated test document content and test presentation outline; Processing the test document content using the trained outline generation model to obtain an output presentation outline; Processing the test document content and the test presentation outline using a trained content generation model to obtain output presentation content; Using a quality assessment model to perform quality assessment on the output presentation outline and the output presentation content to obtain an assessment result; When the evaluation result does not meet the set requirements, the trained outline generation model and / or the trained content generation model are retrained.

9. The method according to claim 8, wherein: The evaluation result includes a second quality score of the output presentation outline and a third quality score of the output presentation content, wherein the second quality score is used to indicate the degree of alignment between the output of the trained outline generation model and the user behavior preference, and the third quality score is used to indicate the degree of alignment between the output of the trained content generation model and the user behavior preference; When the evaluation result does not meet the set requirements, retraining the trained outline generation model and / or the trained content generation model comprises: When the second quality score is lower than the first score threshold, retraining the trained outline generation model; and / or, In a case where the third quality score is lower than a second score threshold, the trained content generation model is retrained.

10. The method according to claim 8, wherein: The quality assessment model is trained by the following steps: Scoring the sample presentation outline according to the set second quality verification standard to obtain a fourth quality score of the sample presentation outline; Scoring the sample presentation content according to the set third quality verification standard to obtain a fifth quality score of the sample presentation content; The quality assessment model is multi-task trained according to the sample presentation outline and the corresponding fourth quality score, and according to the sample presentation content and the corresponding fifth quality score.

11. The method according to claim 10, wherein: The multi-task training of the quality assessment model according to the sample presentation outline and the corresponding fourth quality score, and according to the sample presentation content and the corresponding fifth quality score, comprises: Using the quality assessment model to perform quality assessment on the sample presentation outline to obtain a first assessment score; Using the quality assessment model to perform quality assessment on the sample presentation content to obtain a second assessment score; The quality assessment model is multi-task trained according to a difference between the first assessment score and the fourth quality score, and according to a difference between the second assessment score and the fifth quality score.

12. The method according to claim 1, wherein: The method further comprises: In response to the configuration operation, a model to be trained is selected from a plurality of basic models; wherein the model to be trained includes the outline generation model, the content generation model and the quality assessment model; Configure the hyperparameters of the model to be trained.

13. A method for generating a presentation, comprising: Obtaining target document content associated with the presentation to be generated; Calling the outline generation model to process the target document content to obtain a presentation outline; Calling a content generation model to process the presentation outline and the target document content to obtain presentation content; The presentation content is rendered to obtain the presentation.

14. The method according to claim 13, wherein: The step of obtaining the target document content associated with the presentation to be generated includes any of the following: In response to an input operation, acquiring the input target document content; In response to the upload operation, the uploaded first document is acquired, and the content of the first document is parsed to obtain the content of the target document; In response to the search operation, candidate documents semantically matching the search term are searched, a second document is determined from the candidate documents, and content of the second document is parsed to obtain the target document content.

15. A training device for generating a large model of a presentation, wherein: The large model includes an outline generation model and a content generation model, and the device includes: A first acquisition module, used to acquire training samples; wherein the training samples include: interrelated sample document content, sample presentation outline, sample presentation content and user feedback data; A first processing module, configured to process the sample document content using the outline generation model to obtain a predicted presentation outline; A second processing module, configured to process the sample document content and the sample presentation outline using the content generation model to obtain predicted presentation content; A first training module, configured to train the outline generation model according to the predicted presentation outline, the sample presentation outline and the user feedback data; The second training module is used to train the content generation model according to the predicted presentation content, the sample presentation content and the user feedback data.

16. The device according to claim 15, wherein: The first processing module is used to: Acquire a first prompt template; wherein the first prompt template includes a first task description and a first output format, the first task description is used to instruct the outline generation model to perform an outline extraction task, and the first output format is used to indicate an output format of the outline generation model; Filling the first prompt template according to the sample document content to obtain first prompt information; The outline generation model is called to process the first prompt information to obtain the predicted presentation outline.

17. The device according to claim 15 or 16, wherein: The first training module is used to: determining a first loss value based on a difference between the predicted presentation outline and the sample presentation outline; Based on the first loss value, adjusting model parameters in the outline generation model; According to each positive feedback and / or negative feedback in the user feedback data, reinforcement learning is performed on the adjusted outline generation model.

18. The device according to claim 15, wherein: The second processing module is used to: Acquire a second prompt template; wherein the second prompt template includes a second task description and a second output format, the second task description is used to instruct the content generation model to perform the content generation task, and the second output format is used to indicate the output format of the content generation model; Filling the second prompt template according to the sample document content and the sample presentation outline to obtain second prompt information; The content generation model is called to process the second prompt information to obtain the predicted presentation content.

19. The device according to claim 15 or 18, wherein: The second training module is used to: determining a second loss value based on a difference between the predicted presentation content and the sample presentation content; Based on the second loss value, adjusting a model parameter in the content generation model; According to each positive feedback and / or negative feedback in the user feedback data, reinforcement learning is performed on the adjusted content generation model.

20. The device according to claim 15, wherein: The first acquisition module is used to: Screening and obtaining the sample presentation outline from a plurality of candidate presentation outlines; According to the sample presentation outline, the sample document content is obtained; wherein the sample presentation outline is generated according to the sample document content, and / or the sample presentation outline is generated and updated according to the sample document content; Obtaining a sample presentation, and obtaining the sample presentation content from the sample presentation; wherein the sample presentation is generated according to the sample document content and the sample presentation outline; The user feedback data monitored for the sample presentation is obtained, and the training sample is generated according to the sample document content, the sample presentation outline, the sample presentation content and the user feedback data.

21. The device according to claim 20, wherein: The first acquisition module is used to: Obtain multiple candidate presentation outlines; Scoring the plurality of candidate presentation outlines according to a set first quality verification standard to obtain first quality scores of the plurality of candidate presentation outlines; The sample presentation outline is determined from the plurality of candidate presentation outlines according to the first quality scores of the plurality of candidate presentation outlines.

22. The device according to claim 15, wherein: The device also includes: The second acquisition module is used to acquire a test sample; wherein the test sample includes: interrelated test document content and test presentation outline; A third processing module is used to process the test document content using the trained outline generation model to obtain an output presentation outline; A fourth processing module, configured to process the test document content and the test presentation outline using the trained content generation model to obtain output presentation content; An evaluation module, used for using a quality evaluation model to perform quality evaluation on the output presentation outline and the output presentation content to obtain an evaluation result; The third training module is used to retrain the trained outline generation model and / or the trained content generation model when the evaluation result does not meet the set requirements.

23. The device according to claim 22, wherein: The evaluation result includes a second quality score of the output presentation outline and a third quality score of the output presentation content, wherein the second quality score is used to indicate the degree of alignment between the output of the trained outline generation model and the user behavior preference, and the third quality score is used to indicate the degree of alignment between the output of the trained content generation model and the user behavior preference; The third training module is used to: When the second quality score is lower than the first score threshold, retraining the trained outline generation model; and / or, In a case where the third quality score is lower than a second score threshold, the trained content generation model is retrained.

24. The device according to claim 22, wherein: The quality assessment model is trained using the following modules: A first scoring module, used for scoring the sample presentation outline according to a set second quality verification standard to obtain a fourth quality score of the sample presentation outline; A second scoring module is used to score the sample presentation content according to a set third quality verification standard to obtain a fifth quality score of the sample presentation content; The fourth training module is used to perform multi-task training on the quality assessment model according to the sample presentation outline and the corresponding fourth quality score, and according to the sample presentation content and the corresponding fifth quality score.

25. The device according to claim 24, wherein: The fourth training module is used to: Using the quality assessment model to perform quality assessment on the sample presentation outline to obtain a first assessment score; Using the quality assessment model to perform quality assessment on the sample presentation content to obtain a second assessment score; The quality assessment model is multi-task trained according to a difference between the first assessment score and the fourth quality score, and according to a difference between the second assessment score and the fifth quality score.

26. The device according to claim 15, wherein: The device also includes: A selection module, configured to select a model to be trained from a plurality of basic models in response to a configuration operation; wherein the model to be trained includes the outline generation model, the content generation model and the quality assessment model; A configuration module is used to configure the hyperparameters of the model to be trained.

27. A device for generating a presentation, comprising: An acquisition module, used for acquiring target document content associated with the presentation to be generated; A first calling module is used to call an outline generation model to process the target document content to obtain a presentation outline; A second calling module is used to call a content generation model to process the presentation outline and the target document content to obtain presentation content; A rendering module is used to render the presentation content to obtain the presentation.

28. The device according to claim 27, wherein The acquisition module is used to perform any of the following: In response to an input operation, acquiring the input target document content; In response to the upload operation, the uploaded first document is acquired, and the content of the first document is parsed to obtain the content of the target document; In response to the search operation, candidate documents semantically matching the search term are searched, a second document is determined from the candidate documents, and content of the second document is parsed to obtain the target document content.

29. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the training method for a large model for generating a presentation described in any one of claims 1-12, or the method for generating a presentation described in any one of claims 13-14.

30. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the training method for a large model for generating a presentation according to any one of claims 1-12, or to execute the method for generating a presentation according to any one of claims 13-14.

31. A computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the training method for a large model for generating a presentation according to any one of claims 1-12, or, when executed, implements the method for generating a presentation according to any one of claims 13-14.

Images