Evaluation model construction method, model generation information evaluation method, and apparatus

Abstract

The embodiments of the present disclosure relate to an evaluation model construction method, a model generation information evaluation method, and an apparatus, said construction method comprising: acquiring a plurality of multimedia samples generated by a target generation model in respect of a prompt information sample, and acquiring evaluation result labels respectively corresponding to the plurality of multimedia samples in various preset evaluation dimensions; by means of an information processing network, separately fusing the prompt information sample with target multimedia samples, so as to obtain multiple pieces of sample fusion information; on the basis of the multiple pieces of sample fusion information and by means of a plurality of evaluation networks to be trained, generating network evaluation results respectively corresponding to the plurality of multimedia samples in the various evaluation dimensions; on the basis of the evaluation result labels and the network evaluation results respectively corresponding to the plurality of multimedia samples in the various evaluation dimensions, training the plurality of evaluation networks; and, on the basis of the information processing network and the plurality of trained evaluation networks, constructing an evaluation model. Thus, the embodiments of the present disclosure effectively ensure the reliability of evaluating model generation information.

Description

Evaluation model construction method, model generation information evaluation method and device

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on August 30, 2023, with application number 202311110444.6 and invention name “Method for constructing an evaluation model, method and device for evaluating model generation information”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present disclosure relates to the field of computer technology, and in particular to a method for constructing an evaluation model, and a method and device for evaluating model-generated information. Background Art

[0003] With the prevalence of generative models, they are being used in a variety of fields, such as gaming and academia, to quickly and easily generate the required multimedia information. For example, image generation models can be used to generate realistic virtual images, and text generation models can be used to generate text containing the required content. The reliability of the multimedia information generated by the generative model will directly affect the user experience. Therefore, it is necessary to evaluate the multimedia information generated by the generative model so that the generative model can be optimized and adjusted based on the evaluation results, enabling the generative model to generate more reliable multimedia information to meet user needs. However, related technologies have poor evaluation methods for multimedia information generated by generative models.

[0004] Summary of the Invention

[0005] In order to solve the above technical problems or at least partially solve the above technical problems, the present disclosure provides a method for constructing an evaluation model, and a method and device for evaluating model generation information.

[0006] In a first aspect, an embodiment of the present disclosure provides a method for constructing an evaluation model, comprising: obtaining a plurality of multimedia samples generated by a target generation model for a prompt information sample, and obtaining evaluation result labels corresponding to the plurality of multimedia samples under a plurality of preset evaluation dimensions; wherein the plurality of multimedia samples are all presented in the form of targets; fusing the prompt information sample with the target multimedia samples in the plurality of multimedia samples corresponding to the prompt information sample under each of the evaluation dimensions through a preset information processing network to obtain a plurality of sample fusion information; based on the plurality of sample fusion information, generating network evaluation results corresponding to the plurality of multimedia samples under the plurality of evaluation dimensions through a plurality of evaluation networks to be trained; wherein the plurality of evaluation networks correspond one-to-one to the plurality of evaluation dimensions; training the plurality of evaluation networks based on the evaluation result labels and network evaluation results corresponding to the plurality of multimedia samples under the plurality of evaluation dimensions; and constructing an evaluation model based on the information processing network and the trained plurality of evaluation networks.

[0007] In a second aspect, an embodiment of the present disclosure provides an evaluation method for model-generated information, comprising: obtaining target multimedia information generated by a target generation model for target prompt information; based on the target prompt information, using a pre-constructed evaluation model to generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions; wherein the evaluation model is constructed using the evaluation model construction method described in the first aspect.

[0008] In a third aspect, an embodiment of the present disclosure further provides a device for constructing an evaluation model, comprising: a sample acquisition module for acquiring multiple multimedia samples generated by a target generation model for a prompt information sample, and acquiring evaluation result labels corresponding to the multiple multimedia samples under a preset plurality of evaluation dimensions; wherein the multiple multimedia samples are all presented in the form of targets; a fusion module for fusing the prompt information sample with the target multimedia samples in the multiple multimedia samples corresponding to the prompt information sample under each of the evaluation dimensions through a preset information processing network to obtain a plurality of sample fusion information; a network evaluation module for generating network evaluation results corresponding to the multiple multimedia samples under the multiple evaluation dimensions through a plurality of evaluation networks to be trained based on the multiple sample fusion information; wherein the multiple evaluation networks correspond one-to-one to the multiple evaluation dimensions; a network training module for training the multiple evaluation networks based on the evaluation result labels and network evaluation results corresponding to the multiple multimedia samples under the multiple evaluation dimensions; and a model construction module for constructing an evaluation model based on the information processing network and the trained multiple evaluation networks.

[0009] In a fourth aspect, an embodiment of the present disclosure further provides an evaluation device for model output information, comprising: a generation information acquisition module for acquiring target multimedia information generated by a target generation model for target prompt information; a multi-dimensional evaluation module for generating, based on the target prompt information, network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions using a pre-constructed evaluation model; wherein the evaluation model is constructed using the evaluation model construction method described in the first aspect.

[0010] In a fifth aspect, an embodiment of the present disclosure further provides an electronic device, comprising: a processor; a memory for storing executable instructions of the processor; the processor for reading the executable instructions from the memory and executing the instructions to implement a method for constructing an evaluation model or a method for evaluating model output information as provided in an embodiment of the present disclosure.

[0011] In a sixth aspect, an embodiment of the present disclosure further provides a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is used to execute the method for constructing an evaluation model or the method for evaluating model output information as provided in an embodiment of the present disclosure.

[0012] In a seventh aspect, an embodiment of the present disclosure further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method for constructing an evaluation model or the method for evaluating model output information provided by an embodiment of the present disclosure.

[0013] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0015] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0016] FIG1 is a flow chart of a method for constructing an evaluation model provided by an embodiment of the present disclosure;

[0017] FIG2 is a schematic diagram of the structure of an evaluation model provided by an embodiment of the present disclosure;

[0018] FIG3 is a schematic diagram of the structure of an evaluation model provided by an embodiment of the present disclosure;

[0019] FIG4 is a schematic diagram of constructing an evaluation model provided by an embodiment of the present disclosure;

[0020] FIG5 is a schematic diagram of constructing an evaluation model provided by an embodiment of the present disclosure;

[0021] FIG6 is a flow chart of a method for evaluating model generation information provided by an embodiment of the present disclosure;

[0022] FIG7 is a schematic diagram of the structure of an evaluation model construction device provided by an embodiment of the present disclosure;

[0023] FIG8 is a schematic diagram of the structure of a device for evaluating model output information provided by an embodiment of the present disclosure;

[0024] FIG9 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0025] In order to more clearly understand the above-mentioned objectives, features and advantages of the present disclosure, the scheme of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features therein can be combined with each other in the absence of conflict.

[0026] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, rather than all of the embodiments.

[0027] Figure 1 is a flow chart of a method for constructing an evaluation model provided by an embodiment of the present disclosure. The method can be executed by a device for constructing an evaluation model, wherein the device can be implemented using software and / or hardware and can generally be integrated into an electronic device. As shown in Figure 1, the method mainly includes the following steps S102 to S110:

[0028] Step S102: Obtain multiple multimedia samples generated by the target generation model for the prompt information sample, and obtain evaluation result labels corresponding to the multiple multimedia samples under multiple preset evaluation dimensions. The number of the multiple multimedia samples can be two or more, and the number of the multiple evaluation dimensions can be two or more, and can be flexibly set based on needs.

[0029] In the embodiment of the present disclosure, the multimedia information generated by the target generation model for the prompt information sample is referred to as a multimedia sample. In the embodiment of the present disclosure, there is no restriction on the form of multimedia information that can be generated by the target generation model. Any generation model can be used as a target generation model. For a certain target generation model, the multiple multimedia samples generated for the prompt information sample are all presented in a target form. For example, the target form includes a text form, an image form, an audio form or a video form. That is, the multiple multimedia samples can all be text, or all be images, or all be audio, etc. In other words, the target generation model can be a text generation model, an image generation model, an audio generation model or a video generation model, and the target generation model can be flexibly set according to specific needs.

[0030] In actual applications, the prompt information sample can be input into the target generation model so that the target generation model generates a multimedia sample corresponding to the prompt information sample. The embodiment of the present disclosure does not limit the form and content of the prompt information sample. The prompt information sample is used to indicate the relevant prompt information of the multimedia information to be generated by the target generation model. The more detailed and specific the information described by the prompt information sample is, the more the multimedia information generated by the target generation model can meet the requirements. The method for the target generation model to generate corresponding multimedia information for the prompt information can refer to the relevant technology and will not be repeated here. In the embodiment of the present disclosure, the target generation model can be controlled to generate multiple different multimedia samples by adjusting the parameters of the target generation model or inputting noise.

[0031] In the embodiment of the present disclosure, in order to be able to more reliably and comprehensively evaluate the multimedia information of the target generation model, multiple evaluation dimensions can be set, that is, to measure the quality of the multimedia information from multiple angles, or to evaluate the multiple attributes of the multimedia information separately. The evaluation dimensions corresponding to different forms of multimedia information may be different. For example, if the multimedia information is in text form, the corresponding multiple evaluation dimensions may be fluency dimension, logic dimension, content matching dimension, etc.; if the multimedia information is in image form, the corresponding multiple evaluation dimensions may be clarity dimension, aesthetic dimension, image-text matching dimension, exposure dimension, texture dimension, etc.; if the multimedia information is in audio form, the corresponding multiple evaluation dimensions may be sound quality dimension, content matching dimension, fluency dimension, etc.; if the multimedia information is in video form, the corresponding multiple evaluation dimensions may be picture dimension, content matching dimension, fluency dimension, etc. It should be noted that the above are all exemplary descriptions. In actual applications, the corresponding multiple evaluation dimensions can be flexibly determined according to the form of multimedia information that can be generated by the target generation model, and there is no limitation here.

[0032] The disclosed embodiments can obtain evaluation result labels corresponding to multiple multimedia samples under multiple preset evaluation dimensions. The evaluation result labels are used to represent relatively accurate results corresponding to the multiple multimedia samples under the multiple preset evaluation dimensions, such as results that are more consistent with human perception. For example, the evaluation result labels can be determined based on the evaluation of multiple multimedia samples by users or professional evaluators under the multiple preset evaluation dimensions. In this case, the evaluation result labels are obtained based on the manual evaluation results. The evaluation result labels are used to reflect human feedback on the multiple multimedia samples. In addition, the evaluation result labels can also be obtained based on a trained neural network model, which is not limited here. In actual applications, for each evaluation dimension, the evaluation result labels corresponding to the multiple multimedia samples in the evaluation dimension can be represented by, for example, label evaluation scores or label evaluation levels, or by ranking the label evaluations of the multiple multimedia samples from best to worst. It is also possible to select only the multimedia samples with the best performance and the multimedia samples with the worst performance under the evaluation dimension. The above are only examples. The expression form of the evaluation result labels can be set according to specific needs and is not limited here. The embodiment of the present disclosure obtains evaluation result labels corresponding to multiple multimedia samples under each evaluation dimension. By presenting multiple multimedia samples to evaluators such as annotators or users, it is convenient for evaluators to evaluate different multimedia samples by comparison, and the obtained evaluation result labels are also more objective. In addition, by obtaining the evaluation result labels corresponding to multiple multimedia samples under multiple evaluation dimensions, the embodiment of the present disclosure can also more comprehensively and reliably present human perception feedback on multiple multimedia samples from different angles, and the obtained evaluation result labels are also more precise and accurate.

[0033] In step S104, the preset information processing network fuses the prompt information sample with the target multimedia sample from the multiple multimedia samples corresponding to the prompt information sample under each evaluation dimension, thereby obtaining multiple sample fusion information. Specifically, the number of sample fusion information is the same as the number of multimedia samples, and each sample fusion information corresponds one-to-one with the target multimedia sample. Each sample fusion information incorporates the characteristics of the prompt information sample and the characteristics of the target multimedia sample corresponding to the sample fusion information.

[0034] The embodiments of the present disclosure do not limit the method of determining the target multimedia samples among the multiple multimedia samples corresponding to the prompt information sample under each evaluation dimension. For example, a multimedia sample that meets the preset conditions can be used as the target multimedia sample. For example, for each evaluation dimension, the positive sample and / or negative sample corresponding to the prompt information sample in the multiple multimedia samples is used as the target multimedia sample, wherein the positive and negative samples can be determined based on the evaluation result label. Under each evaluation dimension, the number of positive samples and the number of negative samples corresponding to the prompt information sample can also be flexibly set, and are not limited here. Taking the evaluation dimension as the clarity dimension and the multimedia sample as an example, the top N clearest images and the top M blurry images among the multiple images generated by the target generation model for the prompt information sample can be used as the target multimedia samples respectively.

[0035] Exemplarily, the information processing network can be directly implemented using a pre-trained feature processing network, which can extract the first feature of the prompt information sample and the second feature of each target multimedia sample respectively, and then fuse the first feature with each second feature respectively to obtain multiple sample fusion information. The sample fusion information can be presented in the form of a feature vector.

[0036] Step S106: Based on the fusion information of multiple samples, network evaluation results corresponding to multiple multimedia samples under multiple evaluation dimensions are generated through multiple evaluation networks to be trained; wherein different evaluation networks correspond to different evaluation dimensions, multiple evaluation networks correspond one-to-one to multiple evaluation dimensions, and the number of evaluation networks is the same as the number of evaluation dimensions.

[0037] For example, for each sample fusion information, the sample fusion information is input into multiple evaluation networks, so that each evaluation network generates a network evaluation result corresponding to the sample fusion information. The network evaluation result is considered to be the network evaluation result of the evaluation network for the multimedia sample corresponding to the sample fusion information. Similarly, the network evaluation result can be represented by a network evaluation score or a network evaluation grade, or by ranking the network evaluations of multiple multimedia samples from best to worst. It is also possible to select only the multimedia samples with the best performance and the multimedia samples with the worst performance under the evaluation dimension corresponding to the evaluation network. The above is only an example, and the specific expression form of the network evaluation result can be set according to needs and is not limited here.

[0038] To facilitate application, in some specific implementation examples, each evaluation network can generate an evaluation result of the corresponding multimedia sample under the evaluation dimension corresponding to the evaluation network for each sample fusion information. The evaluation result can be represented by a score or grade, that is, each evaluation network can clearly and unambiguously output the network evaluation result of each multimedia sample, thereby demonstrating the performance of each multimedia sample under the evaluation dimension corresponding to the evaluation network.

[0039] In practical applications, for each evaluation network, each sample fusion information in multiple sample fusion information can be input into the evaluation network separately. The evaluation network can extract the features corresponding to the prompt sample information and the features corresponding to the multimedia sample from the sample fusion information. With the assistance of the features corresponding to the prompt sample information, the evaluation network can better evaluate the quality of the multimedia sample under the evaluation dimension corresponding to the evaluation network, and present the corresponding network evaluation results in preset forms such as scores.

[0040] Step S108 : training multiple evaluation networks based on the evaluation result labels and network evaluation results corresponding to the multiple multimedia samples under the multiple evaluation dimensions.

[0041] For example, for each evaluation dimension, the evaluation network corresponding to that evaluation dimension is trained based on the evaluation result labels and network evaluation results corresponding to multiple multimedia samples under that evaluation dimension. That is, different evaluation networks are trained separately. It is understood that the aforementioned information processing network can be a pre-trained network, that is, a network that can effectively and reliably obtain fused information samples corresponding to the prompt information sample and each multimedia sample. On this basis, the parameters of the aforementioned information processing network can be fixed, and the parameters of each evaluation network can be directly adjusted based on the network evaluation results corresponding to the multiple multimedia samples output by each evaluation network and the pre-acquired evaluation result labels, thereby optimizing each evaluation network to make the network evaluation results output by each evaluation network more accurate. For example, when the evaluation result labels are based on human evaluation results, the network evaluation results obtained in this manner can be closer to human feedback, that is, help users perceive the network evaluation results as more accurate and more consistent with their perceptions. The training processes of different evaluation networks can be independent of each other. This approach not only makes training more efficient, but also better ensures that each evaluation network can accurately output the evaluation results for the corresponding evaluation dimension without being affected by the performance of multimedia samples under other evaluation dimensions.

[0042] Step S110: constructing an evaluation model based on the information processing network and the trained multiple evaluation networks.

[0043] In some embodiments, the information processing network and multiple trained evaluation networks together constitute an evaluation model. This evaluation model is used to evaluate the received multimedia information generated by the target generation model based on multiple evaluation dimensions and output network evaluation results corresponding to each of the multiple evaluation dimensions. In other words, using this evaluation model, model-generated information can be evaluated from multiple perspectives simultaneously. This approach is more convenient and allows for a comprehensive, multi-dimensional evaluation of model-generated information at one time.

[0044] In other embodiments, the information processing network and each trained evaluation network are separately constructed into evaluation models, resulting in multiple evaluation models. For example, if there are N evaluation networks in total, N evaluation models can be constructed. Each evaluation model is used to evaluate the multimedia information generated by the target generation model based on the evaluation dimension corresponding to the evaluation network in the evaluation model, and output the network evaluation result corresponding to the evaluation dimension. This approach is more flexible, and users can select the evaluation model corresponding to the desired evaluation dimension to evaluate the model-generated information according to their needs.

[0045] In summary, the method for constructing the above-mentioned evaluation model provided by the embodiment of the present disclosure can train the evaluation network according to the evaluation result labels and network evaluation results corresponding to multiple multimedia samples generated by the target generation model for the prompt information sample under multiple evaluation dimensions. Therefore, it can effectively ensure that the multiple evaluation networks after training can generate relatively accurate evaluation results, and the obtained network evaluation results are also multi-dimensional, that is, the final evaluation model can provide relatively accurate evaluation results for the multimedia information generated by the target generation model under multiple evaluation dimensions, which fully guarantees the reliability of the evaluation of the model-generated information.

[0046] To ensure the reliability of the fused information, the disclosed embodiments further provide an implementation example of an information processing network, wherein the information processing network includes a feature encoding layer and a feature fusion layer; the disclosed embodiments do not limit the specific structures of the feature encoding layer and the feature fusion layer. On this basis, the prompt information sample is fused with each multimedia sample through the preset information processing network, including the following steps a and b:

[0047] Step a: obtaining, through a feature coding layer, first features of the prompt information sample and second features of a target multimedia sample among a plurality of multimedia samples corresponding to the prompt information sample under various evaluation dimensions.

[0048] Step b: The first feature is fused with each second feature separately through a feature fusion layer. For example, for each second feature, the feature fusion layer can perform a dot product operation on the first feature and the second feature, thereby aggregating and interacting the prompt information with the multimedia information to obtain the desired sample fusion information.

[0049] Taking into account that the presentation forms of the prompt information and the multimedia information may be the same or different, in order to facilitate feature extraction, based on the different presentation forms of the prompt information sample and the multimedia sample, for example, the presentation form of the prompt information sample is in text form, while the presentation form of the multimedia information is in image form, the feature coding layer may include a first encoder and a second encoder; the first encoder and the second encoder are used to process information in different presentation forms. The first encoder is used to obtain the first feature, and the second encoder is used to obtain the second feature. In other words, the first encoder is used to process information in the first presentation form, such as for feature extraction of text; the second encoder is used to process information in the second presentation form, such as for processing images. The above is only an example. Specifically, corresponding encoders can be set based on the form of the prompt information and the form of the multimedia information, so that the prompt information and multimedia information can be encoded more reliably and effective features can be extracted.

[0050] The embodiments of the present disclosure do not limit the structure of the evaluation network. The structures of multiple evaluation networks can be the same or different. For example, multiple evaluation networks can all be implemented using a multilayer perceptron (MLP).

[0051] Based on the foregoing, the present disclosure further illustrates the training method of the evaluation network as follows.

[0052] In order to train the evaluation network efficiently and reliably, in some specific implementation examples, the evaluation result labels corresponding to multiple multimedia samples under each evaluation dimension are used to indicate the first sample and the second sample corresponding to the multiple multimedia samples under the evaluation dimension; wherein, the first sample corresponding to each evaluation dimension is the target multimedia sample represented as the best among the multiple multimedia samples by the evaluation result label information under the evaluation dimension; the second sample corresponding to each evaluation dimension is the target multimedia sample represented as the worst among the multiple multimedia samples by the evaluation result label information under the evaluation dimension. That is, based on the evaluation result labels, the best and worst samples corresponding to multiple multimedia samples under each evaluation dimension can be obtained. The above method can effectively ensure the accuracy of the evaluation result labels. Specifically, the embodiment of the present disclosure fully takes into account that there may be similar samples in the multiple multimedia samples generated by the target generation model. It is usually difficult for humans to accurately distinguish which of these similar samples is better, so it is not convenient to give evaluation results for each multimedia sample separately. In other words, even if evaluation results are given separately for each multimedia sample, it may not be objective and accurate. However, humans can usually intuitively select the best and worst performing samples under a certain evaluation dimension from multiple multimedia samples through comparison, etc. For example, under the clarity dimension, humans can easily select the clearest image and the most blurred image from 10 images. However, it is difficult to clearly distinguish and compare the clarity of multiple images with average clarity one by one and give accurate evaluation results. Therefore, the best and worst samples selected by manual evaluation in the embodiment of the present disclosure are relatively reliable and more efficient and quick. In summary, the evaluation result labels corresponding to the multiple multimedia samples obtained in each evaluation dimension in the embodiment of the present disclosure are used to indicate the best samples and the worst samples corresponding to the multiple multimedia samples under the evaluation dimension, thereby fully ensuring the accuracy of the obtained evaluation result labels.

[0053] In order to enable the evaluation network to output corresponding evaluation results for each multimedia sample in a targeted manner, in some specific implementation examples, the network evaluation results corresponding to multiple multimedia samples under each evaluation dimension are used to indicate the network evaluation score corresponding to each target multimedia sample under the evaluation dimension. Different from the above-mentioned evaluation result labels, the embodiment of the present disclosure enables the evaluation network to output corresponding network evaluation scores for each target multimedia sample under each evaluation dimension, thereby accurately measuring the performance of each target multimedia sample under each evaluation dimension through the scores, so that in subsequent applications, the evaluation network provided by the embodiment of the present disclosure can also output corresponding scores for individual multimedia information (such as an image) generated by the target generation model.

[0054] In order to more conveniently train the evaluation network with the help of evaluation result labels, the method provided in the embodiment of the present disclosure further includes: for each evaluation dimension, setting the label evaluation score of the first sample corresponding to the evaluation dimension to a preset maximum value, and setting the label evaluation score of the second sample corresponding to the evaluation dimension to a preset minimum value; wherein the network evaluation score is not higher than the preset maximum value, and the network evaluation score is not lower than the preset minimum value. For example, the preset maximum value is 1 and the preset minimum value is 0, so the network evaluation score output by any evaluation network for any multimedia information to be evaluated is within the interval [0,1]. It can be understood that in the label evaluation, only the first sample (optimal sample) and the second sample (worst sample) corresponding to multiple multimedia samples under each evaluation dimension are obtained by comparison, wherein the first sample and the second sample are both target multimedia samples. The evaluation network can output a corresponding network evaluation score for each target multimedia sample under each evaluation dimension. In order to train the evaluation network based on the evaluation result labels and the network evaluation results, the first sample and the second sample can be assigned the highest score and the lowest score respectively, so that the loss can be calculated later with the network evaluation score.

[0055] On the basis of the foregoing, for each evaluation dimension, based on the evaluation result labels and network evaluation results corresponding to multiple multimedia samples under the evaluation dimension, a specific implementation method for training the evaluation network corresponding to the evaluation dimension can be: based on the difference between the label evaluation score of the first sample corresponding to the evaluation dimension and the network evaluation score of the first sample corresponding to the evaluation dimension, and based on the difference between the label evaluation score of the second sample corresponding to the evaluation dimension and the network evaluation score of the second sample corresponding to the evaluation dimension, the evaluation network corresponding to the evaluation dimension is trained. Exemplarily, the above differences can be determined based on cross-entropy loss, and the evaluation network is trained with the goal of narrowing the above differences, thereby enabling the evaluation network to generate network evaluation results consistent with the label evaluation.

[0056] The above technical solution provided by the embodiment of the present disclosure can preset multiple evaluation dimensions and construct an evaluation model using the preset information processing network and the trained evaluation networks corresponding to the multiple evaluation dimensions. In the above solution, the evaluation network can be trained based on the evaluation result labels and network evaluation results corresponding to the multiple multimedia samples generated by the target generation model for the prompt information sample under multiple evaluation dimensions. Therefore, it can be effectively guaranteed that the trained multiple evaluation networks can generate relatively accurate evaluation results, and the obtained network evaluation results are also multi-dimensional, that is, the final evaluation model can provide relatively accurate evaluation results for the multimedia information generated by the target generation model under multiple evaluation dimensions, effectively ensuring the reliability of the evaluation of the model-generated information.

[0057] In order to facilitate the understanding of the evaluation model finally obtained by the embodiment of the present disclosure, the embodiment of the present disclosure first provides a structural diagram of an evaluation model as shown in Figure 2. The evaluation model illustrated in Figure 2 directly includes an information processing network and multiple evaluation networks (evaluation network 1 to evaluation network N). The multiple evaluation networks can be regarded as different branch networks of the evaluation model, and the different evaluation networks do not affect each other. In addition, please refer to the structural diagram of an evaluation model as shown in Figure 3, which illustrates N evaluation models, each of which includes an information processing network and an evaluation network. The evaluation networks and corresponding evaluation dimensions corresponding to the N evaluation models are different, but the information processing networks in the N evaluation models can be the same. In actual applications, the above method can be flexibly selected to construct the evaluation model according to needs.

[0058] Based on Figure 2, the embodiment of the present disclosure also illustrates a structural diagram of an evaluation model as shown in Figure 4, which further illustrates that the information processing network includes a feature coding layer and a feature fusion layer. The feature coding layer and the feature fusion layer can refer to the aforementioned related content and will not be described in detail here. The information processing network in Figure 3 can also further include a feature coding layer and a feature fusion layer, which can be specifically implemented with reference to the information processing network in Figure 4, and will not be further illustrated here.

[0059] In some specific implementation examples, taking the prompt information sample as a text and the multimedia information as an image as an example, based on Figure 4, a schematic diagram of the construction of an evaluation model as shown in Figure 5 is further given. In this evaluation model, multiple evaluation dimensions are clarity dimension, aesthetic dimension and image-text matching dimension as examples, and the evaluation network is evaluated in the form of scoring. In Figure 5, for ease of understanding, the data layer for providing data, the feature processing layer for performing information processing based on data (corresponding to the aforementioned information processing network) and the scoring layer for scoring and evaluating the data (corresponding to the aforementioned evaluation network) are vividly illustrated. Among them, the feature processing layer and the scoring layer can be used as the main network layer of the evaluation model. In Figure 5, each evaluation network is further taken as an example of an MLP network. The clarity MLP network, the aesthetic MLP network and the image-text matching MLP network are illustrated in the scoring layer (corresponding to the aforementioned evaluation network), and the network evaluation results such as clarity score, aesthetic score, image-text matching score, etc. can be output respectively. The following is a detailed explanation of Figure 5:

[0060] Figure 5 illustrates the data layer used as input for the evaluation model to be constructed: clarity training samples, aesthetic training samples, and image-text matching training samples. This allows for the training of clarity MLP networks, aesthetic MLP networks, and image-text matching MLP networks, respectively, that can output accurate scores. Taking the clarity training samples as an example, they include information prompt samples and multiple images (e.g., four) output by the target generation model based on these information prompt samples (e.g., prompt text). The clarity training samples are annotated with evaluation result labels, specifically, the image with the best clarity and the image with the worst clarity among the four images. The clarity training sample is input to the feature processing layer, and the text encoder (corresponding to the aforementioned first encoder) encodes the prompt text in the clarity training sample, and the image encoder (corresponding to the aforementioned second encoder) encodes the target image in the clarity training sample (such as the clearest image and the blurriest image corresponding to the prompt text). The encoded image features and the encoded text features are fused separately through the multimodal interaction layer (corresponding to the aforementioned feature fusion layer) to obtain multiple fusion features. Each fusion feature is used to represent the multimodal fusion information of an image and the prompt text; the multiple fusion features are then output to the clarity MLP network for the clarity MLP network to score the clarity of the target image in the clarity training sample based on the fusion features, and adjust the parameters of the clarity MLP network based on the clarity scoring results and the evaluation result labels. The remaining dimensions are similar and will not be repeated here.

[0061] It should be noted that FIG5 is only an example, and in actual applications, the evaluation dimensions can be flexibly changed or expanded according to the form of multimedia information and actual scenarios, and are not limited here.

[0062] In summary, the evaluation model provided by the embodiments of the present disclosure can provide relatively accurate evaluation results for the multimedia information generated by the target generation model under multiple evaluation dimensions, that is, provide fine-grained evaluation results. The embodiments of the present disclosure do not limit the application scenarios of the evaluation model. For example, the evaluation model can be used to evaluate the multimedia information output by the target generation model, thereby optimizing the target generation model based on the evaluation results; for example, the multimedia information output by multiple generation models can be compared, thereby comparing the performance of different generation models under multiple evaluation dimensions or understanding the advantages of different generation models.

[0063] Based on the above, the embodiments of the present disclosure further provide a method for evaluating model generation information. For example, see FIG6 , which is a flow chart of a method for evaluating model generation information provided by the embodiments of the present disclosure. The method can be performed by a device for evaluating model generation information, wherein the device can be implemented using software and / or hardware and can generally be integrated into an electronic device. As shown in FIG6 , the method mainly includes the following steps S602 to S604:

[0064] Step S602: Obtain target multimedia information generated by the target generation model for the target prompt information. The number of target multimedia information can be only one, and there is no limit on the number of target multimedia information. For example, if the target generation model is an image generation model, the target multimedia information can be only one image.

[0065] Step S604: Based on the target prompt information, a pre-built evaluation model is used to generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions; wherein the evaluation model is constructed using the aforementioned evaluation model construction method.

[0066] In the above manner, the evaluation model can provide relatively accurate evaluation results for the target multimedia information generated by the target generation model under multiple evaluation dimensions, thereby fully ensuring the reliability of the evaluation of the model-generated information.

[0067] For example, the above step S604 may be performed with reference to the following steps 1 and 2:

[0068] Step 1: The target prompt information and target multimedia information are fused together using the pre-built information processing network in the evaluation model to obtain target fused information. The structure of the information processing network can be found in the aforementioned related content. The specific processing method for the target prompt information and target multimedia information is consistent with the processing method for the prompt information samples and multimedia samples described above, and will not be repeated here.

[0069] Step 2: Based on the target fusion information, multiple evaluation networks in the evaluation model are used to generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions. The specific implementation method can be referred to the above related content and will not be repeated here.

[0070] In practical applications, the method provided by the embodiment of the present disclosure also includes: obtaining the weight corresponding to each evaluation dimension; generating a comprehensive evaluation result corresponding to the target multimedia information based on the weight corresponding to each evaluation dimension and the network evaluation result corresponding to the target multimedia information under each evaluation dimension.

[0071] That is, the embodiments of the present disclosure can further integrate the network evaluation results under multiple evaluation dimensions to obtain a comprehensive evaluation result, such as integrating the network evaluation scores corresponding to multiple evaluation dimensions to obtain a comprehensive score corresponding to the target multimedia information. This comprehensive score can better present the comprehensive performance of the target multimedia information under multiple dimensions.

[0072] In summary, the above method provided by the embodiment of the present disclosure can provide evaluation results close to human feedback for the multimedia information generated by the target generation model under multiple evaluation dimensions, and can further comprehensively evaluate the model generation information based on the network evaluation results of multiple dimensions, fully ensuring the reliability of the evaluation of the model generation information.

[0073] Corresponding to the aforementioned method for constructing an evaluation model, an embodiment of the present disclosure provides a device for constructing an evaluation model. FIG7 is a schematic structural diagram of a device for constructing an evaluation model provided by an embodiment of the present disclosure. The device can be implemented by software and / or hardware and can generally be integrated into an electronic device. As shown in FIG7 , the device for constructing an evaluation model includes:

[0074] The sample acquisition module 702 is configured to acquire multiple multimedia samples generated by the target generation model for the prompt information sample, and to acquire evaluation result labels corresponding to the multiple multimedia samples under multiple preset evaluation dimensions; wherein the multiple multimedia samples are all presented in the form of targets;

[0075] A fusion module 704 is configured to fuse the prompt information sample with a target multimedia sample in a plurality of multimedia samples corresponding to the prompt information sample under each evaluation dimension through a preset information processing network to obtain fusion information of the plurality of samples;

[0076] The network evaluation module 706 is used to generate network evaluation results corresponding to multiple multimedia samples under multiple evaluation dimensions based on the fusion information of multiple samples through multiple evaluation networks to be trained; wherein different evaluation networks correspond to different evaluation dimensions;

[0077] A network training module 708 is configured to train multiple evaluation networks based on evaluation result labels and network evaluation results corresponding to multiple multimedia samples under multiple evaluation dimensions;

[0078] The model construction module 710 is used to construct an evaluation model based on the feature processing network and the trained multiple evaluation networks.

[0079] In the above-mentioned device, the evaluation network can be trained according to the evaluation result labels and network evaluation results corresponding to multiple multimedia samples generated by the target generation model for the prompt information sample under multiple evaluation dimensions. Therefore, it can be effectively guaranteed that the multiple evaluation networks after training can generate relatively accurate evaluation results, and the obtained network evaluation results are also multi-dimensional, that is, the final evaluation model can provide relatively accurate evaluation results for the multimedia information generated by the target generation model under multiple evaluation dimensions, which fully guarantees the reliability of the evaluation of the model-generated information.

[0080] In some implementations, the network training module 708 is specifically configured to: for each evaluation dimension, train the evaluation network corresponding to the evaluation dimension based on the evaluation result labels corresponding to the multiple multimedia samples under the evaluation dimension and the network evaluation results.

[0081] In some embodiments, the evaluation result labels corresponding to the multiple multimedia samples under each evaluation dimension are used to indicate the first sample and the second sample corresponding to the multiple multimedia samples under the evaluation dimension; wherein, the first sample corresponding to each evaluation dimension is the target multimedia sample represented as the best among the multiple multimedia samples by the evaluation result label information under the evaluation dimension; and the second sample corresponding to each evaluation dimension is the multimedia sample represented as the worst among the multiple multimedia samples by the evaluation result label information under the evaluation dimension.

[0082] In some implementations, the network evaluation results corresponding to the multiple multimedia samples under each evaluation dimension are used to indicate the network evaluation score corresponding to each target multimedia sample under the evaluation dimension.

[0083] In some embodiments, the device also includes a setting module for setting, for each of the evaluation dimensions, the label evaluation score of the first sample corresponding to the evaluation dimension to a preset maximum value, and setting the label evaluation score of the second sample corresponding to the evaluation dimension to a preset minimum value; wherein, the network evaluation score is not higher than the preset maximum value, and the network evaluation score is not lower than the preset minimum value; on this basis, the network training module 708 is specifically used to: train the evaluation network corresponding to the evaluation dimension based on the difference between the label evaluation score of the first sample corresponding to the evaluation dimension and the network evaluation score of the first sample corresponding to the evaluation dimension, and based on the difference between the label evaluation score of the second sample corresponding to the evaluation dimension and the network evaluation score of the second sample corresponding to the evaluation dimension.

[0084] In some embodiments, the information processing network includes a feature coding layer and a feature fusion layer; the fusion module 704 is specifically used to: obtain the first feature of the prompt information sample and the second feature of the target multimedia information in the multiple multimedia samples corresponding to the prompt information sample under each of the evaluation dimensions through the feature coding layer; and fuse the first feature with each of the second features through the feature fusion layer.

[0085] In some embodiments, the presentation form of the prompt information sample is different from the presentation form of the multimedia information, and the feature encoding layer includes a first encoder and a second encoder; wherein the first encoder is used to obtain the first feature, and the second encoder is used to obtain the second feature.

[0086] In some implementations, the target form includes text form, image form, audio form, or video form.

[0087] The evaluation model construction device provided in the embodiments of the present disclosure can execute the evaluation model construction method provided in any embodiment of the present disclosure, and has the corresponding functional modules and beneficial effects of the execution method.

[0088] Corresponding to the aforementioned model output information evaluation method, an embodiment of the present disclosure provides a model output information evaluation device. FIG8 is a schematic structural diagram of a model output information evaluation device provided by an embodiment of the present disclosure. The device can be implemented by software and / or hardware and can generally be integrated into an electronic device. As shown in FIG8 , the model output information evaluation device includes:

[0089] A generated information acquisition module 802 is used to acquire target multimedia information generated by the target generation model based on the target prompt information;

[0090] The multi-dimensional evaluation module 804 is used to generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions based on the target prompt information using a pre-built evaluation model; wherein the evaluation model is constructed using the evaluation model construction method.

[0091] In the above device, the evaluation model can provide evaluation results close to human feedback for the target multimedia information generated by the target generation model under multiple evaluation dimensions, fully ensuring the reliability of the evaluation of the model-generated information.

[0092] In some embodiments, the multi-dimensional evaluation module 804 is specifically used to: fuse the target prompt information with the target multimedia information through the information processing network in the pre-constructed evaluation model to obtain target fusion information; based on the target fusion information, generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions through multiple evaluation networks in the evaluation model.

[0093] In some embodiments, the device also includes a comprehensive evaluation module for obtaining the weight corresponding to each of the evaluation dimensions; based on the weight corresponding to each of the evaluation dimensions and the network evaluation results corresponding to the target multimedia information under each of the evaluation dimensions, a comprehensive evaluation result corresponding to the target multimedia information is generated.

[0094] The evaluation device for model output information provided in the embodiments of the present disclosure can execute the evaluation method for model output information provided in any embodiment of the present disclosure, and has functional modules and beneficial effects corresponding to the execution method.

[0095] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device embodiment can refer to the corresponding process in the method embodiment, and will not be repeated here.

[0096] An embodiment of the present disclosure provides an electronic device, which includes: a storage device storing a computer program; and a processing device configured to execute the computer program in the storage device to implement the steps of any one of the methods in the present disclosure.

[0097] Reference is now made to FIG9 , which illustrates a schematic diagram of the structure of an electronic device 900 suitable for implementing embodiments of the present disclosure. Terminal devices in embodiments of the present disclosure may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. The electronic device illustrated in FIG9 is merely an example and should not limit the functionality or scope of use of embodiments of the present disclosure.

[0098] As shown in Figure 9, the electronic device 900 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage device 908 into a random access memory (RAM) 903. Various programs and data required for the operation of the electronic device 900 are also stored in the RAM 903. The processing device 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0099] Typically, the following devices may be connected to the I / O interface 905: an input device 906 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 907 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 908 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 909. The communication device 909 may allow the electronic device 900 to communicate with other devices wirelessly or by wire to exchange data. Although FIG9 shows the electronic device 900 with various devices, it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0100] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication device 909, or installed from the storage device 908, or installed from the ROM 902. When the computer program is executed by the processing device 901, the above-mentioned functions defined in the method of the embodiment of the present disclosure are performed.

[0101] In addition to the above-mentioned methods and devices, the embodiments of the present disclosure may also be a computer program product, which includes computer program instructions, which, when executed by a processor, cause the processor to perform the image processing method provided by the embodiments of the present disclosure. The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present disclosure, the programming languages ​​including object-oriented programming languages ​​such as Java, C++, etc., and also conventional procedural programming languages ​​such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0102] In addition, the embodiments of the present disclosure may also be a computer-readable storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, the processor is caused to execute the method provided by the embodiments of the present disclosure.

[0103] The computer-readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can, for example, include but is not limited to a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0104] The embodiments of the present disclosure further provide a computer program product, including a computer program / instruction, which implements the method in the embodiments of the present disclosure when executed by a processor.

[0105] It is understandable that before using the technical solutions disclosed in the various embodiments of this disclosure, the type, scope of use, usage scenarios, etc. of the personal information involved in this disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0106] For example, in response to a user's active request, a prompt message is sent to the user to clearly inform the user that the operation requested will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the electronic device, application, server, storage medium, or other software or hardware that performs the operations of the disclosed technical solution based on the prompt message.

[0107] As an optional but non-limiting implementation, in response to receiving a user's active request, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. Furthermore, the pop-up window may also contain a selection control for the user to select "agree" or "disagree" to provide personal information to the electronic device.

[0108] It is understandable that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0109] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0110] The foregoing description is intended only to provide specific embodiments of the present disclosure, intended to enable those skilled in the art to understand and implement the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure is not intended to be limited to the embodiments described herein, but rather to be construed in the broadest manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing an evaluation model, comprising: Acquire multiple multimedia samples generated by the target generation model for the prompt information sample, and acquire evaluation result labels corresponding to the multiple multimedia samples under multiple preset evaluation dimensions; wherein the multiple multimedia samples are all presented in the form of targets; The prompt information sample is respectively fused with the target multimedia sample in the multiple multimedia samples corresponding to the prompt information sample under each of the evaluation dimensions through a preset information processing network to obtain multiple sample fusion information; Based on the plurality of sample fusion information, generating network evaluation results corresponding to the plurality of multimedia samples under the plurality of evaluation dimensions respectively through the plurality of evaluation networks to be trained; wherein the plurality of evaluation networks correspond one to one to the plurality of evaluation dimensions; Based on the evaluation result labels and network evaluation results respectively corresponding to the multiple multimedia samples under the multiple evaluation dimensions, the multiple evaluation networks are trained; An evaluation model is constructed based on the information processing network and the multiple evaluation networks after training.

2. The method according to claim 1, wherein the training of the multiple evaluation networks based on the evaluation result labels and network evaluation results corresponding to the multiple multimedia samples under the multiple evaluation dimensions comprises: For each evaluation dimension, based on the evaluation result labels corresponding to the multiple multimedia samples under the evaluation dimension and the network evaluation results, the evaluation network corresponding to the evaluation dimension is trained.

3. The method according to claim 2, wherein the evaluation result labels corresponding to the plurality of multimedia samples under each evaluation dimension are used to indicate the first sample and the second sample corresponding to the plurality of multimedia samples under the evaluation dimension; in, The first sample corresponding to each evaluation dimension is a target multimedia sample whose evaluation result label information under the evaluation dimension is represented as the best among the multiple multimedia samples; the second sample corresponding to each evaluation dimension is a target multimedia sample whose evaluation result label information under the evaluation dimension is represented as the worst among the multiple multimedia samples.

4. The method according to claim 3, wherein the network evaluation results corresponding to the multiple multimedia samples under each evaluation dimension are used to indicate the network evaluation score corresponding to each of the target multimedia samples under the evaluation dimension.

5. The method according to claim 4, wherein the method further comprises: For each of the evaluation dimensions, the label evaluation score of the first sample corresponding to the evaluation dimension is set to a preset maximum value, and the label evaluation score of the second sample corresponding to the evaluation dimension is set to a preset minimum value; wherein the network evaluation score is not higher than the preset maximum value, and the network evaluation score is not lower than the preset minimum value; The training of the evaluation network corresponding to the evaluation dimension based on the evaluation result labels corresponding to the plurality of multimedia samples under the evaluation dimension and the network evaluation results includes: Based on the difference between the label evaluation score of the first sample corresponding to the evaluation dimension and the network evaluation score of the first sample corresponding to the evaluation dimension, and based on the difference between the label evaluation score of the second sample corresponding to the evaluation dimension and the network evaluation score of the second sample corresponding to the evaluation dimension, the evaluation network corresponding to the evaluation dimension is trained.

6. The method according to claim 1, wherein the information processing network comprises a feature encoding layer and a feature fusion layer; the fusing the prompt information sample with the target multimedia sample in the plurality of multimedia samples corresponding to the prompt information sample under each of the evaluation dimensions through the preset information processing network comprises: Obtaining, through the feature coding layer, respectively the first feature of the prompt information sample and the second feature of the target multimedia sample among the multiple multimedia samples corresponding to the prompt information sample under each of the evaluation dimensions; The first feature is fused with each of the second features respectively through the feature fusion layer.

7. The method according to claim 6, wherein the presentation form of the prompt information sample is different from the presentation form of the multimedia sample, and the feature coding layer includes a first encoder and a second encoder; wherein, The first encoder is used to obtain the first feature, and the second encoder is used to obtain the second feature. The method according to claim 1 , wherein the target form comprises a text form, an image form, an audio form or a video form.

9. A method for evaluating model-generated information, comprising: Obtain target multimedia information generated by the target generation model according to the target prompt information; Based on the target prompt information, a pre-constructed evaluation model is used to generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions; wherein the evaluation model is constructed using the evaluation model construction method described in any one of claims 1 to 8.

10. The method according to claim 9, wherein the step of using a pre-built evaluation model to generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions based on the target prompt information comprises: The target prompt information is fused with the target multimedia information through an information processing network in a pre-constructed evaluation model to obtain target fusion information; Based on the target fusion information, network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions are generated through multiple evaluation networks in the evaluation model.

11. The method according to claim 9, wherein the method further comprises: Obtaining the weight corresponding to each evaluation dimension; Based on the weight corresponding to each of the evaluation dimensions and the network evaluation result corresponding to the target multimedia information under each of the evaluation dimensions, a comprehensive evaluation result corresponding to the target multimedia information is generated.

12. A device for constructing an evaluation model, comprising: A sample acquisition module, used to acquire multiple multimedia samples generated by the target generation model for the prompt information sample, and to acquire evaluation result labels corresponding to the multiple multimedia samples under multiple preset evaluation dimensions; wherein the multiple multimedia samples are all presented in the form of targets; a fusion module, configured to fuse the prompt information sample with a target multimedia sample in the multiple multimedia samples corresponding to the prompt information sample under each of the evaluation dimensions through a preset information processing network to obtain multiple sample fusion information; A network evaluation module, configured to generate network evaluation results corresponding to the plurality of multimedia samples under the plurality of evaluation dimensions respectively through the plurality of evaluation networks to be trained based on the plurality of sample fusion information; wherein the plurality of evaluation networks correspond to the plurality of evaluation dimensions one by one; A network training module, configured to train the multiple evaluation networks based on the evaluation result labels and network evaluation results respectively corresponding to the multiple multimedia samples under the multiple evaluation dimensions; The model building module is used to build an evaluation model based on the information processing network and the multiple evaluation networks after training.

13. A device for evaluating model output information, comprising: A generated information acquisition module is used to acquire target multimedia information generated by a target generation model for target prompt information; A multi-dimensional evaluation module is used to generate network evaluation results corresponding to the target multimedia information under multiple preset evaluation dimensions based on the target prompt information using a pre-constructed evaluation model; wherein the evaluation model is constructed using the evaluation model construction method described in any one of claims 1 to 8.

14. An electronic device, comprising: a storage device having a computer program stored thereon; A processing device is used to execute the computer program in the storage device to implement the method for constructing an evaluation model as described in any one of claims 1 to 8 or the method for evaluating model output information as described in any one of claims 9 to 11.

15. A computer-readable storage medium, wherein the storage medium stores a computer program, wherein the computer program is used to execute the method for constructing an evaluation model as described in any one of claims 1 to 8 or the method for evaluating model output information as described in any one of claims 9 to 11.

Images