Content generation method and device and electronic equipment

Abstract

The invention provides a content generation method. The content generation method comprises the following steps: acquiring input data; in response to inputting input data to a pre-trained content generation model, first generated content is obtained, the first generated content is composed of a plurality of first content units, and each first content unit has a reliable probability value representing the confidence of the corresponding content unit. And according to the reliable probability value of each first content unit, displaying the plurality of first content units differently.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a content generation method, device and electronic device. Background Art

[0002] Large models usually refer to deep learning models with a large number of parameters and high computing resource requirements, such as BERT, T5, etc. These models have achieved significant performance improvements in natural language processing, computer vision and other fields, but users generally do not understand the confidence of the generated content.

[0003] Currently, the format of content generated by large models mainly adopts conventional forms such as titles and subtitles, which makes it difficult for users to determine the reliability of the generated content. Summary of the invention

[0004] One aspect of the present disclosure provides a content generation method, comprising: obtaining input data. In response to inputting the input data into a pre-trained content generation model, obtaining first generated content, the first generated content being composed of a plurality of first content units, each of the first content units having a reliable probability value representing the confidence of the corresponding content unit. Based on the reliable probability value of each first content unit, the plurality of first content units are displayed differently.

[0005] Optionally, the content generation model includes a generation sub-model and a confidence sub-model, the generation sub-model is used to generate the first generated content, and the confidence sub-model is used to generate a reliable probability value of each first content unit in the first generated content.

[0006] Optionally, the content generation model further includes a feature extraction module, and in response to inputting input data into the pre-trained content generation model, obtaining the first generated content includes at least one of the following: in response to inputting text data into the feature extraction module, obtaining text extraction features, inputting the text extraction features into the generation sub-model, obtaining text recognition results, inputting the text recognition results and the text extraction features into the confidence sub-model, and obtaining reliable probability values ​​of each first content unit corresponding to the text recognition results. In response to inputting image data into the feature extraction module, obtaining image extraction features, inputting the image extraction features into the generation sub-model, obtaining image recognition results, inputting the image recognition results and the image extraction features into the confidence sub-model, and obtaining reliable probability values ​​of each first content unit corresponding to the image recognition results. In response to inputting voice data into the feature extraction module, obtaining voice extraction features, inputting the voice extraction features into the generation sub-model, obtaining voice recognition results, inputting the voice recognition results and the voice extraction features into the confidence sub-model, and obtaining reliable probability values ​​of each first content unit corresponding to the voice recognition results.

[0007] Optionally, each first content unit in the first generated content is displayed in a different display form, and the display form includes at least one of color, grayscale, font, size and specific mark.

[0008] Optionally, the first content unit represents at least one of a sentence and a paragraph, and the reliable probability value is an average of reliable probability values ​​of at least one of characters and words contained in the first content unit.

[0009] Optionally, the first content unit represents at least one of a character and a word, and the content generation method further comprises: determining a plurality of weight ranges. For first content units in the same weight range, the first content units are displayed in the same display form. For first content units in different weight ranges, the first content units are displayed in different display forms.

[0010] Optionally, the content generation model includes a generation sub-model and a confidence sub-model, and the content generation model is trained in the following manner: first training data and second training data are obtained, the first training data and the second training data are corresponding sample pairs, and the second training data is the true label of the first training data. The first training data is input into the initial generation sub-model to obtain first inference output data. The first training data and the second training data are input into the initial confidence sub-model to obtain the first weight of each second content unit in the second training data. According to the difference information and the first weight between the second training data and the first inference output data, the parameters of the initial generation sub-model and the initial confidence sub-model are adjusted to obtain the generation sub-model and the confidence sub-model.

[0011] Optionally, the content generation model includes a generation sub-model and a confidence sub-model, and the training of the content generation model includes: obtaining third training data, the third training data includes training samples and control samples, and the training samples are samples obtained by partially masking the control samples. The training samples are input into the initial generation sub-model to obtain second reasoning output data. The training samples and the control samples are input into the initial confidence sub-model to obtain the second weight of each third content unit in the control sample. According to the difference information and the second weight between the control sample and the second reasoning output data, the parameters of the initial generation sub-model and the initial confidence sub-model are adjusted to obtain the generation sub-model and the confidence sub-model.

[0012] Another aspect of the present disclosure provides a content generation device, including: an acquisition module for acquiring input data; a generation module for obtaining first generated content in response to inputting the input data into a pre-trained content generation model, wherein the first generated content is composed of a plurality of first content units, each of which has a reliable probability value representing the confidence of the corresponding content unit; and a display module for displaying the plurality of first content units differently according to the reliable probability value of each first content unit.

[0013] Another aspect of the present disclosure provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute any of the above-mentioned content generation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The above and other objects, features and advantages of the present disclosure will become more apparent through the following description of the embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0015] Figure 1 The application scenario diagram of the content generation method according to the embodiment of the present disclosure is schematically shown;

[0016] Figure 2 A flowchart of a content generation method according to an embodiment of the present disclosure is schematically shown;

[0017] Figure 3A A flowchart of a method for generating first generated content according to an embodiment of the present disclosure is schematically shown; Figure 3B The following schematically shows an architecture diagram of a content generation model according to an embodiment of the present disclosure;

[0018] Figure 4 A flowchart of a content generation method according to another embodiment of the present disclosure is schematically shown;

[0019] Figure 5A A flowchart of a method for training content generation model according to an embodiment of the present disclosure is schematically shown; Figure 5B The following schematically shows an architecture diagram of a content generation model trained according to an embodiment of the present disclosure;

[0020] Figure 6 A flowchart of a method for training a content generation model according to another embodiment of the present disclosure is schematically shown;

[0021] Figure 7 The structure block diagram of the content generation device according to the embodiment of the present disclosure is schematically shown;

[0022] Figure 8 A block diagram of an electronic device suitable for implementing a content generating method according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0023] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0024] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0025] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0026] Some block diagrams and / or flow charts are shown in the accompanying drawings. It should be understood that some blocks or combinations thereof in the block diagrams and / or flow charts may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that these instructions, when executed by the processor, may create a device for implementing the functions / operations described in these block diagrams and / or flow charts.

[0027] Therefore, the technology of the present disclosure can be implemented in the form of hardware and / or software (including firmware, microcode, etc.). In addition, the technology of the present disclosure can take the form of a computer program product on a computer-readable medium storing instructions, which can be used by an instruction execution system or in combination with an instruction execution system. In the context of the present disclosure, a computer-readable medium can be any medium that can contain, store, transmit, propagate, or transmit instructions. For example, a computer-readable medium can include, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, device, or propagation medium. Specific examples of computer-readable media include: magnetic storage devices, such as magnetic tape or hard disk (HDD); optical storage devices, such as compact disk (CD-ROM); memory, such as random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0028] The embodiments of the present disclosure provide a content generation method for improving the display effect of large model training and generated content. The content generation method includes: obtaining input data. In response to inputting the input data into a pre-trained content generation model, a first generated content is obtained, and the first generated content is composed of multiple first content units, and each first content unit has a reliable probability value that characterizes the confidence of the corresponding content unit. According to the reliable probability value of each first content unit, the multiple first content units are displayed differently. By assigning a reliable probability value to each content unit of the generated content, the weight of the reliability of the content unit can be characterized when training the large model, so that the model can adaptively learn reliable content with high weight, thereby improving the training effect of the large model, and then improving the accuracy of the content generated by the large model. At the same time, the display effect of the generated content is also improved by displaying content units with different reliable probabilities, which is convenient for users to identify the reliability of the generated content.

[0029] Figure 1 The application scenario diagram of the content generation method according to the embodiment of the present disclosure is schematically shown.

[0030] like Figure 1 As shown, the application scenario 100 according to this embodiment may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0031] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as security applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only for example).

[0032] The terminal devices 101 , 102 , and 103 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers.

[0033] The server 105 may be a server that provides various services, such as a background management server (only an example) that provides support for websites browsed or applications logged in by users using the terminal devices 101, 102, and 103. The background management server may analyze and process the received data such as user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal device.

[0034] It should be noted that the content generation method provided in the embodiment of the present disclosure can generally be executed by the server 105 or by the terminal device. Accordingly, the content generation device provided in the embodiment of the present disclosure can generally be set in the server 105. The content generation method provided in the embodiment of the present disclosure can also be executed by a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105. Accordingly, the content generation device provided in the embodiment of the present disclosure can also be set in a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105.

[0035] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to the implementation requirements.

[0036] The following will be based on Figure 1 The scene described by Figure 2~Figure 6 The content generation method of the disclosed embodiment is described in detail.

[0037] Figure 2 The flowchart of the content generation method according to the embodiment of the present disclosure is schematically shown.

[0038] According to the embodiments of the present disclosure, Figure 2 As shown, the content generation method of this embodiment includes, for example, operations S210 to S230.

[0039] In operation S210, input data is acquired.

[0040] For example, receiving user input or data generated within the system as the basis for generating content. This data can be any form of information such as text, images, audio, etc., depending on the specific application scenario.

[0041] In operation S220, in response to inputting input data into a pre-trained content generation model, first generated content is obtained, the first generated content being composed of a plurality of first content units, each of the first content units having a reliable probability value representing a confidence level of the corresponding content unit.

[0042] The acquired input data is passed to a pre-trained content generation model. This model may be based on deep learning technology, such as recurrent neural network (RNN), long short-term memory network (LSTM), Transformer, etc., and can generate corresponding content based on the input data. After receiving the input data, the model performs internal processing and generates the first generated content.

[0043] The first generated content is a series of information generated by the model based on the input data, which is organized into multiple first content units. Each first content unit represents an independent or related part of the generated content, such as a sentence, paragraph, image fragment, etc.

[0044] Each first content unit is accompanied by a reliable probability value, which is used to characterize the confidence level of the corresponding content unit. That is, the model has an internal assessment of the accuracy and reliability of the content unit it generates.

[0045] In operation S230, a plurality of first content units are displayed differently according to the reliability probability value of each first content unit.

[0046] The system will decide how to display these content units based on the reliability probability value of each first content unit. For example, content units with high reliability probability values ​​can be displayed first or highlighted in a highlighted or bold manner; while content units with low reliability probability values ​​may be displayed in a darker, smaller font, or with additional warning information to remind users that the credibility of these contents may be low.

[0047] The content generation model can be a large model and can be applied in multiple application scenarios.

[0048] For example, when automatically generating news summaries, the big model may generate multiple summary paragraphs or sentences based on the news content. Due to the complexity and diversity of news content, the accuracy of different summary paragraphs may vary. Therefore, the confidence of each summary paragraph can be displayed differently based on the big model, such as highlighting high-confidence summaries or adding warning labels to low-confidence summaries.

[0049] For example, when automatically generating the abstract, introduction, or conclusion of an academic paper, a large model may generate multiple related paragraphs or sentences. Since academic papers have extremely high requirements for accuracy and rigor, the reliability of different content units needs to be clearly distinguished. Content units with high confidence can be used directly in the paper, while content units with low confidence need further verification or modification.

[0050] For example, when using a large model for creative writing or story generation, the model may generate multiple plot segments or character descriptions. Due to the diversity and subjectivity of creative content, the appeal or rationality of different content units may vary. At this time, the content units can be displayed differently based on the confidence of the large model, such as giving priority to high-confidence plot segments, or appropriately modifying or discarding low-confidence content.

[0051] For example, when automatically generating legal documents or contracts, a large model may generate multiple clauses or paragraphs. Since legal documents have extremely high requirements for accuracy and rigor, the reliability of each content unit needs to be strictly evaluated. High-confidence content units can be used directly in legal documents, while low-confidence content units need to be further reviewed and modified by legal experts.

[0052] For example, when automatically generating product descriptions or advertising copy, a large model may generate multiple descriptive sentences or phrases. In order to attract consumers' attention and improve conversion rates, the attractiveness of different content units needs to be evaluated and displayed differently. For example, high-confidence and attractive content units can be displayed in a prominent position on the advertisement or product page.

[0053] For example, in the medical field, large models may be used to assist in diagnosis or generate treatment recommendations. Due to the sensitivity and importance of medical information, strict reliability assessments need to be performed on different content units generated by large models. High-confidence diagnosis results or treatment recommendations can be directly provided to doctors or patients for reference, while low-confidence content needs to be further verified or avoided.

[0054] For example, in the financial field, large models may be used for tasks such as market analysis, risk assessment, or stock prediction. Due to the complexity and uncertainty of financial markets, the reliability of different content units (such as forecast results, investment advice, etc.) generated by large models may vary. Therefore, these content units need to be displayed separately so that investors or decision makers can make more informed decisions.

[0055] The input data is received through a pre-trained content generation model, and a first generated content containing multiple content units is generated. Each content unit is assigned a reliable probability value representing its confidence. Subsequently, the system displays different content units differently based on these reliable probability values ​​to improve the user experience and the perceived credibility of the content. This method is particularly suitable for application scenarios that require automatic content generation and evaluation of its reliability, such as news summary generation, text creation assistance, etc.

[0056] According to an embodiment of the present disclosure, the content generation model includes a generation sub-model and a confidence sub-model, the generation sub-model is used to generate first generated content, and the confidence sub-model is used to generate a reliable probability value of each first content unit in the first generated content.

[0057] In this embodiment, the content generation model is further refined to include two main parts: a generation sub-model and a confidence sub-model. These two sub-models work together to provide a more sophisticated and reliable content generation experience.

[0058] The generation sub-model is responsible for receiving input data and generating first generated content according to its internal algorithm and training data. The first generated content is composed of multiple first content units, which can be sentences, paragraphs, image fragments or other forms of information, depending on the application scenario.

[0059] While the generation sub-model generates the first generated content, the confidence sub-model works in parallel to evaluate and generate a reliable probability value for each first content unit. This probability value reflects the confidence of the generation sub-model in the content unit, that is, to what extent the content unit is considered accurate or reasonable.

[0060] In some embodiments, the input data is fed into a generation sub-model, which generates first generated content based on the input data and an algorithm.

[0061] At the same time, the confidence sub-model receives the output of the generation sub-model (ie, the first generated content), and independently evaluates each first content unit to generate a corresponding reliable probability value.

[0062] These two processes can be synchronous or asynchronous, depending on how the model is implemented.

[0063] After obtaining the reliability probability values ​​of the first generated content and each content unit, the system will display the content units differently according to these probability values. For example, a high-confidence content unit can be presented to the user in a highlighted, bold, or priority display manner; while a low-confidence content unit may be displayed in a dimmed, reduced font, or with additional warning information to remind the user of its reliability.

[0064] In some application scenarios, the system may also allow users to interact with displayed content units, such as choosing whether to view low-confidence content units, rating content units, or providing feedback, etc. These interaction and feedback data can be used to further train and optimize the generation sub-model and the confidence sub-model to improve their performance and reliability.

[0065] The content generation method is further refined by introducing the generation sub-model and the confidence sub-model. The generation sub-model is responsible for generating content, while the confidence sub-model is responsible for evaluating the reliability of each unit of the generated content. This division of labor and cooperation makes the content generation process more refined and reliable.

[0066] Figure 3A A flow chart of a method for generating first generated content according to an embodiment of the present disclosure is schematically shown. Figure 3B The following schematically shows an architecture diagram of a content generation model according to an embodiment of the present disclosure.

[0067] According to an embodiment of the present disclosure, the content generation model also includes a feature extraction module, such as Figure 3A As shown, for example, by performing at least one of operations S321 to S323, in response to inputting input data into a pre-trained content generation model, first generated content is obtained.

[0068] In this embodiment, the content generation model is further expanded by adding a feature extraction module to support the processing of multiple types of input data, including text, image, and speech. This design enables the model to be more flexibly applied to different scenarios and needs.

[0069] like Figure 3B As shown in Figure 1, the feature extraction module is responsible for receiving different types of input data (text, image, speech) and extracting features that are useful for generating content. For text data, features may include vocabulary, syntactic structure, etc.; for image data, features may include color, texture, shape, etc.; for speech data, features may include phonemes, intonation, speaking speed, etc.

[0070] The generation sub-model receives the extracted features from the feature extraction module and generates first generated content based on the features. The type of the generated content depends on the type of input data, such as text recognition results, image recognition results, or speech recognition results.

[0071] After the generation sub-model generates the first generated content, the confidence sub-model receives the generation result and the corresponding extracted features, evaluates and generates reliable probability values ​​for each first content unit, and these probability values ​​reflect the confidence of the generation sub-model in the content unit.

[0072] In operation S321, in response to inputting text data into the feature extraction module, text extraction features are obtained, the text extraction features are input into the generation sub-model to obtain text recognition results, the text recognition results and the text extraction features are input into the confidence sub-model to obtain reliable probability values ​​of each first content unit corresponding to the text recognition results.

[0073] When text data is input, the feature extraction module extracts text features, and then these features are sent to the generation sub-model to generate text recognition results. Then, the text recognition results and the text extraction features are sent to the confidence sub-model together to obtain the reliable probability values ​​of each first content unit corresponding to the text recognition results.

[0074] For example, in an application scenario of an intelligent writing assistant, the user is a writer who is writing a novel. He uses an intelligent writing assistant to assist in writing. He enters a descriptive text, such as "On a stormy night, a lonely traveler walked into a deserted castle."

[0075] The feature extraction module extracts key words, syntactic structures and other features from the text. The generation sub-model generates text recognition results based on these features, such as automatically expanding the storyline, describing the environment atmosphere, etc. The confidence sub-model evaluates the reliability of the generated content and provides a reliable probability value for each content unit.

[0076] The user receives an expanded text that includes a detailed description of the castle's interior and the traveler's psychological activities. The system also marks which parts are highly confident and which may need further modification.

[0077] In operation S322, in response to inputting image data into the feature extraction module, image extraction features are obtained, the image extraction features are input into the generation sub-model to obtain image recognition results, the image recognition results and the image extraction features are input into the confidence sub-model to obtain reliable probability values ​​of each first content unit corresponding to the image recognition results.

[0078] When image data is input, the process is similar, but the feature extraction module extracts image features, the generation sub-model generates image recognition results, and the confidence sub-model evaluates the reliability of each content unit in these image recognition results.

[0079] For example, in a medical image analysis application scenario, doctors use medical image analysis systems to process patients' X-rays or CT scan images during diagnosis.

[0080] The feature extraction module extracts features such as color, texture, shape, etc. from the image. The generation sub-model identifies abnormal areas in the image based on these features, such as tumors, fractures, etc. The confidence sub-model evaluates the reliability of the recognition results and provides a reliable probability value for each abnormal area.

[0081] The doctor received a detailed diagnostic report, which marked the abnormal areas in the image and their reliable probability values, helping the doctor to judge the condition more accurately.

[0082] In operation S323, in response to inputting voice data into the feature extraction module, voice extraction features are obtained, the voice extraction features are input into the generation sub-model to obtain voice recognition results, the voice recognition results and the voice extraction features are input into the confidence sub-model to obtain reliable probability values ​​of each first content unit corresponding to the voice recognition results.

[0083] When voice data is input, the feature extraction module extracts voice features, the generation sub-model generates voice recognition results, and the confidence sub-model evaluates the reliability of each content unit in these voice recognition results.

[0084] For example, in an application scenario of an intelligent customer service system, a user dials a customer service number to inquire about a problem, and the intelligent customer service system interacts with the user through voice recognition technology.

[0085] The feature extraction module extracts features such as phonemes, intonation, and speech speed from the speech. The generation sub-model recognizes the user's voice instructions or question content based on these features. The confidence sub-model evaluates the reliability of the recognition results and provides a reliable probability value for each recognition result.

[0086] The intelligent customer service system gives corresponding answers or solutions based on the recognition results and reliable probability values. At the same time, for recognition results with low confidence, the user may be asked to restate or transferred to manual customer service for processing.

[0087] The system will display each content unit in the first generated content differently based on the reliable probability value generated by the confidence sub-model. High-confidence content units are displayed in a more prominent manner, while low-confidence content units are displayed in a less prominent manner or with additional warning information.

[0088] Users can interact with the generated content, such as viewing, editing, or providing feedback. These feedback data can be used to further optimize the performance of the feature extraction module, the generation sub-model, and the confidence sub-model.

[0089] By introducing the feature extraction module, the functionality of the content generation model is expanded to enable it to handle multiple types of input data (text, images, speech). The feature extraction module provides rich input features for the generation sub-model, which helps to generate more accurate and richer first generated content. This design makes the content generation method more flexible and powerful, and suitable for a wider range of application scenarios.

[0090] It should be noted that the design of the content generation model is flexible and can be further extended to process other types of data. For example, it can be modified to process video data (by extracting video frames as image data and combining it with time series analysis), gesture data (capturing gestures through sensors and extracting features), and even brain wave data (in neuroscience research, extracting features of brain wave signals to analyze and generate content). These extensions require customization and optimization of feature extraction modules and generation sub-models for specific data types.

[0091] According to an embodiment of the present disclosure, each first content unit in the first generated content is displayed in a different display form, and the display form includes, for example, at least one of color, grayscale, font, size, and specific mark.

[0092] In this embodiment, the output display mode of the content generation method is further enriched, especially for each first content unit in the first generated content, different display forms are used to distinguish and display. This design helps users to more intuitively understand the reliability, importance or other specific attributes of the generated content.

[0093] The content generation model first generates first generated content according to the input data. These contents are composed of multiple first content units, which can be sentences, paragraphs, image segments, voice segments, etc.

[0094] Next, the confidence sub-model performs confidence evaluation on each first content unit in the first generated content to generate a corresponding reliable probability value.

[0095] When displaying the first generated content, the system uses different display forms to display it according to the reliability probability value of each first content unit or other preset conditions (such as importance, urgency, etc.). These display forms include but are not limited to:

[0096] Color, high confidence content units may be represented by bright shades such as green or blue, while low confidence content units may be represented by warning shades such as yellow or red.

[0097] Grayscale,For some application scenarios, it may be preferable to use grayscale to distinguish the reliability of content units, such as high-confidence content units are brighter and low-confidence content units are darker.

[0098] Fonts, highlight important or low-confidence content units by changing the font style (such as bold, italics, underline).

[0099] Size adjusts the display size of content units to make high-confidence content more prominent and low-confidence content smaller.

[0100] Specific marking, adding specific icons, symbols or annotations next to content units to visually indicate their reliability or other attributes.

[0101] Users can view detailed confidence information for a content unit, modify display settings, or provide feedback by clicking, hovering, or otherwise interacting.

[0102] For example, when generating news summaries, the system can display high-confidence news points in bold and green fonts, while low-confidence information is displayed in gray and smaller fonts with a "pending verification" mark next to it.

[0103] For example, when generating a preliminary draft of an academic paper, the system can highlight reliable cited data in blue, while indicating assumptions that require further verification with a yellow background and italics.

[0104] For example, when generating a medical diagnosis report, the system can highlight confirmed diseases in red bold font and mark suspected symptoms with a yellow background and specific icons to help doctors quickly identify and make judgments.

[0105] By introducing differentiated display forms, the user experience and practicality of the content generation method are further improved. Through a variety of methods such as color, grayscale, font, size and specific marks, the system can intuitively display the reliability, importance or other specific attributes of each first content unit in the first generated content, helping users to understand and evaluate the generated content more quickly.

[0106] According to an embodiment of the present disclosure, the first content unit represents at least one of a sentence and a paragraph, and the reliable probability value is an average of reliable probability values ​​of at least one of characters and words contained in the first content unit.

[0107] In this embodiment, the calculation method of the reliable probability value is further refined, especially when processing the first content unit including sentences and paragraphs.

[0108] The content generation model generates first generated content according to input data. The first generated content consists of multiple first content units. The units can be a single sentence, a paragraph consisting of multiple sentences, or other forms of text fragments.

[0109] For each word and phrase in the first content unit, the system first performs a separate confidence evaluation to generate their respective reliable probability values, which reflect the accuracy of the model's recognition or understanding of these words and phrases.

[0110] The system then calculates the overall reliability probability value of the content unit based on the reliability probability values ​​of all the characters and words in the content unit. This is usually achieved by taking the average or weighted average of these probability values. This step ensures that the reliability probability value of the content unit accurately reflects the confidence level of the words within it.

[0111] When displaying the first generated content, the system uses differentiated display methods (such as color, grayscale, font, size, specific mark, etc.) according to the reliability probability value of each first content unit. This display method helps users quickly identify the reliability of the content and make corresponding judgments or decisions.

[0112] For example, when writing an academic paper, the system can generate preliminary paragraph content. For each paragraph (i.e., the first content unit), the system will calculate the overall reliability probability value of the paragraph based on the reliability probability values ​​of each sentence and each word in the paragraph. Then, the system may use different colors or grayscales to distinguish paragraphs of different reliability, and attach specific notes or icons next to paragraphs with low reliability, prompting users to further verify or modify them.

[0113] For another example, when generating news reports, the system can automatically identify and extract key information to form sentences and paragraphs. By calculating the reliability probability value of each sentence and paragraph, the system can distinguish which information is highly reliable (such as official news) and which information needs further confirmation (such as unverified rumors). When displaying news reports, the system can use differentiated display methods to highlight these content units of different reliability.

[0114] By introducing word-level confidence assessment and content unit-level confidence aggregation methods, the accuracy and reliability of the content generation model are further improved. This design not only helps improve user experience, but also helps enhance user trust and satisfaction with generated content.

[0115] Figure 4 The flowchart of a content generation method according to another embodiment of the present disclosure is schematically shown.

[0116] According to an embodiment of the present disclosure, the first content unit represents at least one of a character and a word, such as Figure 4 As shown, the content generating method, for example, further includes operations S410 to S430.

[0117] In operation S410, a plurality of weight ranges are determined.

[0118] In this embodiment, the concept of weight range is further introduced to achieve more detailed and flexible display control of the first content unit (representational character, word or combination thereof).

[0119] The content generation model generates output content including a plurality of first content units according to input data. These first content units may be single characters, words, or a combination thereof (such as phrases, sentences, etc.).

[0120] For each first content unit, the system calculates its reliable probability value, which reflects the accuracy of the model's recognition or understanding of the content unit.

[0121] The system predefines or dynamically determines multiple weight ranges based on the distribution of reliable probability values. These ranges can be fixed (such as high, medium, and low) or dynamically divided based on data distribution.

[0122] In operation S420, the first content units within the same weight range are displayed in the same display form.

[0123] Each first content unit is assigned to a corresponding weight range according to its reliable probability value.

[0124] The first content units within the same weight range are displayed in the same display form, which may include using the same color, font, size, background or other visual elements.

[0125] In operation S430, first content units within different weight ranges are displayed in different display forms.

[0126] Different display forms are used to distinguish the first content units in different weight ranges. For example, a highly reliable content unit may be highlighted in green, while a low-reliability content unit may be displayed in red italics.

[0127] Users can view detailed reliable probability value information, modify display settings, or provide feedback by interacting with the generated content. This feedback data will be used to optimize the content generation model and display strategy.

[0128] For example, when providing learning materials for students, the system can generate documents containing key concepts and explanatory texts. For each concept or explanatory text (as the first content unit), the system assigns it to different weight ranges according to its reliability probability value. When displaying the document, high-reliability content units (such as authoritatively verified concepts) are displayed in a clear and easy-to-read format, while low-reliability content units (such as unconfirmed opinions) are displayed in a more cautious and eye-catching way to remind students to verify it themselves.

[0129] For example, when assisting authors in writing articles or reports, intelligent editing tools can automatically check each word, sentence or paragraph in the text (as the first content unit), and classify and display them differently according to their reliable probability values. This helps authors quickly identify and deal with possible problems or uncertainties, thereby improving the quality and accuracy of the text.

[0130] By introducing the concept of weight range, a more detailed and flexible differentiated display of the first content unit is achieved. This approach not only helps to improve the user experience and enable users to understand the reliability of the content more intuitively, but also helps to improve the practicality and credibility of the content generation model.

[0131] Figure 5A The flowchart of the method for training content generation model according to an embodiment of the present disclosure is schematically shown. Figure 5B The following schematically shows an architecture diagram of a content generation model trained according to an embodiment of the present disclosure.

[0132] According to an embodiment of the present disclosure, the content generation model includes a generation sub-model and a confidence sub-model, such as Figure 5A As shown, for example, the content generation model is trained by operations S510 to S540.

[0133] In operation S510 , first training data and second training data are acquired, where the first training data and the second training data are corresponding sample pairs, and the second training data is a true label of the first training data.

[0134] In this embodiment, the training method of the content generation model is further deepened, especially how to further improve the performance and robustness of the model by introducing additional supervisory signals and complex data structures.

[0135] like Figure 5B As shown, the first training data is the raw data used to input the generation sub-model to generate content during the model training process. These data are usually unlabeled or only contain preliminary processed data, which are used to trigger the generation sub-model to generate corresponding outputs.

[0136] In order to enhance the generalization ability of the model, the first training data should cover multiple fields, styles and language types. These data are usually raw data that have not been deeply processed, such as unlabeled text, images or speech.

[0137] In order to train a model with superior performance, a large amount of first training data is usually required.

[0138] For example, in a text generation task, the first training data may be a series of unlabeled articles, book paragraphs, web page texts, etc. In an image generation task, the first training data may be a series of unlabeled pictures or image datasets.

[0139] The second training data is the “true label” or “standard answer” in the sample pair corresponding to the first training data, which is used to evaluate the output quality of the generation sub-model and is used as one of the inputs of the confidence sub-model to train its evaluation ability.

[0140] The second training data must be accurate to ensure that the performance of the generated sub-model can be correctly evaluated. The second training data needs to correspond to the first training data one by one to form a sample pair. These data usually contain detailed annotations or explanations of the first training data, such as classification labels for text, segmentation masks for images, etc.

[0141] For example, in a text generation task, the second training data may be an accurate translation, summary, or rewrite of each sentence or paragraph in the first training data. In an image classification task, the second training data may be the classification label of each image (such as "cat", "dog", "car", etc.).

[0142] It should be noted that the above examples are hypothetical enumerations based on the characteristics of general training data. In actual applications, the specific contents of the first training data and the second training data may vary depending on the specific tasks and data sets.

[0143] In addition to the first training data and the second training data (true labels), other training data, such as from different domains, styles, or languages, can be introduced to increase the generalization ability of the model.

[0144] To train more complex models, training data can be organized into complex structures such as trees, graphs, or sequence-to-sequence (seq2seq) to capture richer contextual information and dependencies.

[0145] In operation S520, first training data is input into the initial generation sub-model to obtain first inference output data.

[0146] First, we build the initial generation sub-model and confidence sub-model. The generation sub-model is responsible for generating content, while the confidence sub-model evaluates the reliability of the generated content.

[0147] The training process is divided into two stages: pre-training and fine-tuning. In the pre-training stage, the two sub-models are trained independently on their respective training sets to obtain basic generation and evaluation capabilities. In the fine-tuning stage, the two sub-models are trained jointly to optimize the synergy between them.

[0148] For example, the pre-trained dataset is input into the initial generative sub-model, and the generative sub-model is trained through self-supervised learning (such as language model tasks, masked language model tasks, etc.) so that it can generate coherent and meaningful text or content.

[0149] In operation S530, the first training data and the second training data are input into the initial confidence sub-model to obtain a first weight of each second content unit in the second training data.

[0150] The first training data and the second training data are sequentially or simultaneously input into the initial confidence sub-model. Here, each second content unit (such as a sentence, an image block, etc.) in the second training data will be processed as an independent input.

[0151] The confidence sub-model evaluates each second content unit based on the input data content using the internal learning mechanism (such as neural network layer, weight parameters, etc.) and outputs the corresponding first weight. This weight represents the model's evaluation result of the quality or credibility of the content unit.

[0152] The calculation of the first weight may be based on a variety of factors, such as the grammatical correctness of the content unit, semantic coherence, relevance to the topic, naturalness of generation, etc. These factors will be comprehensively considered through the algorithms and parameters within the confidence sub-model.

[0153] In operation S540, parameters of the initial generation sub-model and the initial confidence sub-model are adjusted according to the difference information between the second training data and the first inference output data and the first weight to obtain a generation sub-model and a confidence sub-model.

[0154] The first inference output data is compared with the second training data (usually the true or expected output) to calculate the difference between the two. This difference can be measured in many ways, such as loss function value (such as mean squared error, cross entropy loss, etc.), similarity score (such as cosine similarity) or other evaluation indicators.

[0155] The first weight is the evaluation result of the confidence sub-model on each content unit (such as a text paragraph, an image area, etc.) in the second training data, representing the quality or credibility of these content units.

[0156] In the parameter adjustment process, the first weight can be used as a weighting factor to weight the difference information. In this way, the model can pay more attention to those content units with higher confidence (i.e., larger weight), and thus give them more importance when adjusting parameters.

[0157] By adjusting the parameters of the generation sub-model and the confidence sub-model, the difference between the first inference output data and the second training data is reduced, and the accuracy of the confidence sub-model in evaluating the generated content is improved.

[0158] For example, an optimization algorithm such as gradient descent can be used to back-propagate the error to each layer of the model and adjust the parameters layer by layer based on the difference information and the weighted result of the first weight.

[0159] Since the generation sub-model and the confidence sub-model are usually interrelated, a joint optimization approach can be used to adjust the parameters of the two models at the same time. This helps the model find a better balance between generating content and evaluating content.

[0160] In addition to the model parameters, some hyperparameters (such as learning rate, batch size, number of iterations, etc.) can also be adjusted to further optimize the training process.

[0161] After multiple iterations and optimizations, when the difference between the first inference output data and the second training data reaches a preset threshold or meets other convergence conditions, the training process stops. The generated sub-model and confidence sub-model obtained at this time are the final training results.

[0162] During the training process, feedback from human experts can be introduced as additional supervisory signals. This feedback can be a direct evaluation of the generated content or a correction to the confidence sub-model evaluation results.

[0163] Advanced optimization algorithms such as Adam, RMSprop, etc. can be used to speed up the training process and improve convergence performance.

[0164] You can also fine-tune the model's hyperparameters through methods such as grid search, random search, or Bayesian optimization to find the optimal parameter configuration.

[0165] During the training process, you can also use the validation set to regularly evaluate the model to monitor the performance changes of the model. And according to the evaluation results, you can iteratively optimize the model, including adjusting the model structure, loss function, optimization algorithm, etc.

[0166] By introducing multi-source training data, complex data structures, joint training strategies, additional supervisory signals, optimization algorithms, and hyperparameter adjustments, the training effect and performance of the content generation model are improved. This training method not only improves the generation and evaluation capabilities of the model, but also enhances the generalization and robustness of the model. In practical applications, these training strategies and techniques can be flexibly selected and adjusted according to the characteristics of specific tasks and data sets to obtain the best model performance.

[0167] Figure 6 A flowchart of a method for training a content generation model according to another embodiment of the present disclosure is schematically shown.

[0168] According to an embodiment of the present disclosure, the content generation model includes a generation sub-model and a confidence sub-model, such as Figure 6 As shown, for example, the content generation model is trained through operations S610 to S640.

[0169] In this embodiment, the training method of the content generation model is further explored, especially based on a strategy combining mask training and weighted evaluation.

[0170] In operation S610, third training data is acquired, where the third training data includes a training sample and a control sample, where the training sample is a sample obtained by partially masking the control sample.

[0171] The third training data includes training samples and control samples. The training samples are obtained by partially masking the control samples, that is, randomly masking a part of the content units (such as words, sentences or image areas) in the control samples to simulate the situation of missing or to be generated.

[0172] In operation S620, the training sample is input into the initial generation sub-model to obtain second inference output data.

[0173] The processed training samples are input into the initial generation sub-model, and the generation sub-model attempts to restore or generate the masked content units according to the remaining unmasked content units to obtain the second inference output data.

[0174] In operation S630, the training sample and the control sample are input into the initial confidence sub-model to obtain the second weight of each third content unit in the control sample.

[0175] At the same time, the training sample (including the masked part) and the complete control sample are input into the initial confidence sub-model. The confidence sub-model evaluates each third content unit (which can be a word, sentence, image block, etc.) in the control sample and generates a second weight, which represents the credibility or importance of each content unit.

[0176] In operation S640, parameters of the initial generation sub-model and the initial confidence sub-model are adjusted according to the difference information between the control sample and the second reasoning output data and the second weight, so as to obtain the generation sub-model and the confidence sub-model.

[0177] Calculate the difference between the control sample and the second inference output data. This is usually measured by some loss function, such as cross entropy loss, mean square error, etc.

[0178] The difference information is weighted using the second weight generated by the confidence sub-model. For content units with higher confidence, their difference information will receive a greater weight during training, prompting the model to pay more attention to the generation quality of these key parts.

[0179] Based on the weighted difference information, the initial generation sub-model and the initial confidence sub-model are jointly optimized. This usually involves adjusting the parameters of the model using a back-propagation algorithm to minimize the weighted loss function.

[0180] During the training process, you may also need to adjust some hyperparameters, such as learning rate, batch size, number of iterations, etc., to improve the training effect and convergence speed.

[0181] During the training process, the performance of the model is evaluated regularly to ensure that the generation sub-model can generate high-quality content and that the confidence sub-model can accurately assess the quality of the generated content.

[0182] When the model performance reaches the preset threshold or meets other convergence conditions, the training is stopped. The generated sub-model and confidence sub-model obtained at this time are the trained models.

[0183] By introducing the strategies of mask training and weighted evaluation, the training effect of the content generation model is effectively improved. Specifically, by generating training samples through partial mask control samples, the generation sub-model learns to restore or generate missing content during the training process; at the same time, the confidence sub-model is used to perform weighted evaluation on the control samples, so that the model pays more attention to the generation quality of key parts during the training process. This training method not only improves the accuracy of generated content, but also enhances the model's ability to evaluate the quality of generated content.

[0184] Based on the above method, the present disclosure also provides a content generation device. Figure 7 The content generating apparatus is described in detail.

[0185] Figure 7 The structural block diagram of the content generating device according to the embodiment of the present disclosure is schematically shown.

[0186] like Figure 7 As shown, the content generation device 1000 of this embodiment includes, for example, an acquisition module 710, a generation module 720, and a display module 730. The content generation device 1000 can execute the above reference Figure 2~Figure 6 The method described is used to adjust the display effect of generated content.

[0187] Specifically, the acquisition module 710 is used to acquire input data. In one embodiment, the acquisition module 710 can be used to perform the operation S210 described above, which will not be described in detail here.

[0188] The generation module 720 is used to obtain first generated content in response to inputting input data into a pre-trained content generation model, wherein the first generated content is composed of a plurality of first content units, each of which has a reliable probability value representing the confidence of the corresponding content unit. In one embodiment, the generation module 720 can be used to perform the operation S220 described above, which will not be described in detail here.

[0189] The display module 730 is used to display the multiple first content units according to the reliability probability value of each first content unit. In one embodiment, the display module 730 can be used to perform the operation S230 described above, which will not be described in detail here.

[0190] It is understood that the acquisition module 710, the generation module 720 and the display module 730 can be combined in one module, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present disclosure, at least one of the acquisition module 710, the generation module 720 and the display module 730 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented in hardware or firmware in any other reasonable way of integrating or packaging the circuit, or in a proper combination of software, hardware and firmware. Alternatively, at least one of the acquisition module 710, the generation module 720 and the display module 730 can be at least partially implemented as a computer program module, and when the program is run by a computer, the function of the corresponding module can be executed.

[0191] Figure 8 A block diagram of an electronic device suitable for implementing a content generating method according to an embodiment of the present disclosure is schematically shown.

[0192] like Figure 8 As shown, the electronic device 800 according to an embodiment of the present disclosure includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage part 808 to a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a dedicated microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include an onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0193] In RAM 803, various programs and data required for the operation of electronic device 800 are stored. Processor 801, ROM 802 and RAM 803 are connected to each other via bus 804. Processor 801 performs various operations of the method flow according to the embodiment of the present disclosure by executing the program in ROM 802 and / or RAM 803. It should be noted that the program can also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 can also perform various operations of the method flow according to the embodiment of the present disclosure by executing the program stored in the one or more memories.

[0194] According to an embodiment of the present disclosure, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to the bus 804. The electronic device 800 may further include one or more of the following components connected to the I / O interface 805: an input portion 806 including a keyboard, a mouse, etc.; an output portion 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage portion 808 including a hard disk, etc.; and a communication portion 809 including a network interface card such as a LAN card, a modem, etc. The communication portion 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 810 as needed, so that a computer program read therefrom is installed into the storage portion 808 as needed.

[0195] The present disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or may exist independently without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the content generation method according to the embodiment of the present disclosure is implemented.

[0196] According to an embodiment of the present disclosure, the computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, it may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, an apparatus or a device. For example, according to an embodiment of the present disclosure, the computer-readable storage medium may include the ROM 802 and / or RAM 803 described above and / or one or more memories other than ROM 802 and RAM 803.

[0197] The embodiment of the present disclosure also includes a computer program product, which includes a computer program, and the computer program contains program code for executing the method shown in the flowchart. When the computer program product is run in a computer system, the program code is used to enable the computer system to implement the content generation method provided by the embodiment of the present disclosure.

[0198] The above functions defined in the system / device of the embodiment of the present disclosure are executed when the computer program is executed by the processor 801. According to the embodiment of the present disclosure, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0199] In one embodiment, the computer program may rely on tangible storage media such as optical storage devices, magnetic storage devices, etc. In another embodiment, the computer program may also be transmitted and distributed in the form of signals on a network medium, and downloaded and installed through the communication part 809, and / or installed from a removable medium 811. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0200] In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 809, and / or installed from the removable medium 811. When the computer program is executed by the processor 801, the above functions defined in the system of the embodiment of the present disclosure are performed. According to the embodiment of the present disclosure, the system, device, means, module, unit, etc. described above can be implemented by a computer program module.

[0201] According to an embodiment of the present disclosure, the program code for executing the computer program provided by the embodiment of the present disclosure can be written in any combination of one or more programming languages. Specifically, these computing programs can be implemented using high-level process and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, Java, C++, python, "C" language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on the remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect through the Internet).

[0202] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0203] It will be appreciated by those skilled in the art that the features described in the various embodiments and / or claims of the present disclosure may be combined and / or combined in a variety of ways, even if such combinations and / or combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments and / or claims of the present disclosure may be combined and / or combined in a variety of ways without departing from the spirit and teachings of the present disclosure. All of these combinations and / or combinations fall within the scope of the present disclosure.

[0204] The embodiments of the present disclosure are described above. However, these embodiments are only for the purpose of illustration and are not intended to limit the scope of the present disclosure. Although the embodiments are described above separately, this does not mean that the measures in the various embodiments cannot be used in combination to advantage. The scope of the present disclosure is defined by the attached claims and their equivalents. Without departing from the scope of the present disclosure, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present disclosure.

Claims

1. A content generation method, comprising: Get input data; In response to inputting the input data into a pre-trained content generation model, first generated content is obtained, wherein the first generated content is composed of a plurality of first content units, each of the first content units having a reliable probability value representing a confidence level of the corresponding content unit; The plurality of first content units are displayed differently according to the reliability probability value of each of the first content units.

2. The method according to claim 1, wherein: The content generation model includes a generation sub-model and a confidence sub-model, the generation sub-model is used to generate the first generated content, and the confidence sub-model is used to generate a reliable probability value of each of the first content units in the first generated content.

3. The method according to claim 2, wherein: The content generation model further includes a feature extraction module, and the step of obtaining the first generated content in response to inputting the input data into the pre-trained content generation model includes at least one of the following: In response to inputting text data into the feature extraction module, obtaining text extraction features, inputting the text extraction features into the generation sub-model to obtain text recognition results, inputting the text recognition results and the text extraction features into the confidence sub-model to obtain reliable probability values ​​of each first content unit corresponding to the text recognition results; In response to inputting image data into the feature extraction module, obtaining image extraction features, inputting the image extraction features into the generation sub-model to obtain image recognition results, inputting the image recognition results and the image extraction features into the confidence sub-model to obtain reliable probability values ​​of each first content unit corresponding to the image recognition results; In response to inputting speech data into the feature extraction module, speech extraction features are obtained, the speech extraction features are input into the generation sub-model to obtain speech recognition results, the speech recognition results and the speech extraction features are input into the confidence sub-model to obtain reliable probability values ​​of each first content unit corresponding to the speech recognition results.

4. The method according to claim 1, wherein: Each of the first content units in the first generated content is displayed in a different display form, and the display form includes at least one of color, grayscale, font, size and specific mark.

5. The method according to claim 4, wherein: The first content unit represents at least one of a sentence and a paragraph, and the reliable probability value is an average of reliable probability values ​​of at least one of characters and words included in the first content unit.

6. The method according to claim 4, wherein: The first content unit represents at least one of a character and a word, and the method further comprises: Determine multiple weight ranges; Displaying the first content units within the same weight range in the same display form; The first content units within different weight ranges are displayed in different display forms.

7. The method according to claim 1, wherein: The content generation model includes a generation sub-model and a confidence sub-model, and the content generation model is trained in the following manner: Acquire first training data and second training data, where the first training data and the second training data are corresponding sample pairs, and the second training data is a true label of the first training data; Inputting the first training data into the initial generation sub-model to obtain first inference output data; Inputting the first training data and the second training data into an initial confidence sub-model to obtain a first weight of each second content unit in the second training data; According to the difference information between the second training data and the first inference output data and the first weight, parameters of the initial generation sub-model and the initial confidence sub-model are adjusted to obtain the generation sub-model and the confidence sub-model.

8. The method according to claim 1, wherein: The content generation model includes a generation sub-model and a confidence sub-model, and the training of the content generation model includes: Acquire third training data, the third training data comprising a training sample and a control sample, the training sample being a sample obtained by partially masking the control sample; Inputting the training sample into the initial generation sub-model to obtain second inference output data; Inputting the training sample and the control sample into an initial confidence sub-model to obtain a second weight of each third content unit in the control sample; According to the difference information between the control sample and the second inference output data and the second weight, the parameters of the initial generation sub-model and the initial confidence sub-model are adjusted to obtain the generation sub-model and the confidence sub-model.

9. A content generating device, comprising: An acquisition module, used to obtain input data; A generation module, configured to obtain first generated content in response to inputting the input data into a pre-trained content generation model, wherein the first generated content is composed of a plurality of first content units, each of the first content units having a reliable probability value representing a confidence level of the corresponding content unit; The display module is configured to display the plurality of first content units differently according to the reliability probability value of each of the first content units.

10. An electronic device, comprising: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors execute the content generating method according to any one of claims 1 to 8.

Images