Method and device for model evaluation, equipment and storage medium

Abstract

The embodiment of the invention provides a method and device for model evaluation, equipment and a storage medium. The method comprises the steps that multiple inputs in an input set are provided for a first generative model to obtain a first output set output by the first generative model, and the first output set comprises multiple outputs corresponding to the multiple inputs; obtaining respective first annotation information for a plurality of outputs in the first output set, wherein the first annotation information indicates an annotated quality level in a plurality of quality levels divided by each output under each quality evaluation dimension in a plurality of quality evaluation dimensions; and determining a first overall quality score of the first generative model at least based on the respective first annotation information of the plurality of outputs in the first output set and the respective quality scores corresponding to the plurality of quality grades divided under the plurality of quality evaluation dimensions.

Description

Technical Field

[0001] Example embodiments of the present disclosure generally relate to the field of computer technology, and more particularly to methods, devices, apparatuses, and computer-readable storage media for model evaluation. Background Art

[0002] As machine learning and deep learning technologies have been widely used in many fields, a generative model refers to a model input based on a given modality that can generate new model outputs. Generative models are mainly used in natural language processing, machine translation, speech synthesis, image generation and other fields. In the future, they can be applied to more fields, such as medical care, finance, education, etc. For the output of the generative model, it is necessary to be able to accurately evaluate the performance of the model. Summary of the invention

[0003] In a first aspect of the present disclosure, a method for model evaluation is provided. The method includes: providing multiple inputs in an input set to a first generative model respectively to obtain a first output set output by the first generative model, the first output set including multiple outputs corresponding to the multiple inputs respectively; obtaining first annotation information for each of the multiple outputs in the first output set, the first annotation information indicating the quality level of each output in the multiple quality levels divided under each quality assessment dimension in multiple quality assessment dimensions; and determining a first overall quality score of the first generative model based at least on the first annotation information of each of the multiple outputs in the first output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions.

[0004] In a second aspect of the present disclosure, a device for model evaluation is provided. The device includes: a first input providing module, configured to provide multiple inputs in an input set to a first generative model respectively, so as to obtain a first output set output by the first generative model, wherein the first output set includes multiple outputs corresponding to the multiple inputs respectively; a first annotation obtaining module, configured to obtain first annotation information for each of the multiple outputs in the first output set, wherein the first annotation information indicates the quality level of each output that is annotated in multiple quality levels divided under each quality assessment dimension in multiple quality assessment dimensions; and a first overall quality determination module, configured to determine a first overall quality score of the first generative model based at least on the first annotation information of each of the multiple outputs in the first output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions.

[0005] In a third aspect of the present disclosure, an electronic device is provided. The device includes at least one processing unit; and at least one memory, the at least one memory is coupled to the at least one processing unit and stores instructions for execution by the at least one processing unit. When the instructions are executed by the at least one processing unit, the device executes the method of the first aspect.

[0006] In a fourth aspect of the present disclosure, a computer-readable storage medium is provided, wherein a computer program is stored on the medium, and when the computer program is executed by a processor, the method of the first aspect is implemented.

[0007] It should be understood that the content described in this section is not intended to limit the key features or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:

[0009] Figure 1 A schematic diagram showing an example environment in which embodiments of the present disclosure can be implemented;

[0010] Figure 2 A block diagram of a model evaluation system according to some embodiments of the present disclosure is shown;

[0011] Figure 3 A comparison graph of overall quality scores of models according to some embodiments of the present disclosure is shown;

[0012] Figure 4 A flowchart showing a process for model evaluation according to some embodiments of the present disclosure is shown;

[0013] Figure 5 A block diagram showing an apparatus for model evaluation according to some embodiments of the present disclosure; and

[0014] Figure 6 A block diagram of an electronic device is shown in which one or more embodiments of the present disclosure may be implemented. DETAILED DESCRIPTION

[0015] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0016] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The term "some embodiments" should be understood as "at least some embodiments". Other explicit and implicit definitions may also be included below.

[0017] It is understandable that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and relevant provisions.

[0018] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present 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.

[0019] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information, so that the user can independently choose whether to provide personal information to software or hardware such as electronic devices, applications, servers or storage media that execute operations of the technical solution of the present disclosure based on the prompt message.

[0020] As an optional but non-limiting implementation, in response to receiving an active request from the user, the prompt information is sent to the user in a manner such as a pop-up window, in which the prompt information can be presented in text form. In addition, the pop-up window can also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0021] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0022] As used herein, the term "model" can learn the association between the corresponding input and output from the training data, so that after the training is completed, the corresponding output can be generated for a given input. The generation of the model can be based on machine learning technology. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs by using multi-layer processing units. A neural network model is an example of a model based on deep learning. In this article, "model" may also be referred to as "machine learning model", "learning model", "machine learning network" or "learning network", and these terms are used interchangeably in this article.

[0023] A "neural network" is a machine learning network based on deep learning. A neural network is capable of processing inputs and providing corresponding outputs, and typically includes an input layer and an output layer and one or more hidden layers between the input layer and the output layer. Neural networks used in deep learning applications typically include many hidden layers, thereby increasing the depth of the network. The layers of a neural network are connected in sequence so that the output of the previous layer is provided as input to the next layer, where the input layer receives the input of the neural network and the output of the output layer serves as the final output of the neural network. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each of which processes input from the previous layer.

[0024] Generally, machine learning can be roughly divided into three stages, namely the training stage, the testing stage, and the application stage (also called the inference stage). In the training stage, a given model can be trained using a large amount of training data, and the parameter values ​​are continuously updated iteratively until the model can obtain consistent inferences that meet the expected goals from the training data. Through training, the model can be considered to be able to learn the association between input and output (also called input-to-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing stage, the test input is applied to the trained model to test whether the model can provide the correct output, thereby determining the performance of the model. The testing stage can sometimes be integrated into the training stage. In the application stage, the trained model can be used to process the actual model input based on the parameter values ​​obtained through training to determine the corresponding model output.

[0025] Generative models refer to content generation technologies implemented by models. Generative models can automatically generate meaningful and coherent content based on input content, including natural language text or images, audio, video, etc. Driven by large-scale pre-trained models, the generation quality of generative models has reached a level that makes people feel "written like a human" in many application fields. Although generative models have made significant progress in generating content, how to objectively and effectively evaluate the quality of these generated contents remains a key challenge. In many current applications, quality assessment methods for generative models are crucial to ensuring the accuracy and usability of generated content.

[0026] Current evaluation methods are generally divided into automatic evaluation and manual evaluation. Automatic evaluation methods mainly rely on various metrics to evaluate the quality of generated content. BLEU (Bilingual Evaluation Understudy) is an evaluation metric widely used in machine translation tasks, which mainly measures the n-gram overlap between generated content and reference text. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a metric used to evaluate automatic summarization tasks, which mainly calculates the n-gram overlap between generated content and reference summaries. METEOR (Metric for Evaluation of Translation with Explicit ORdering) is also an evaluation metric used for machine translation. It not only calculates the word and phrase overlap between generated content and reference text, but also considers stems and synonyms. Perplexity is an indicator used to measure the predictive ability of language generation models. The lower its value, the more the text generated by the model conforms to grammatical rules and actual text. Although these indicators can evaluate the quality of generated text to a certain extent, they mainly focus on the overlap of text content and do not fully consider other aspects such as the logic and information quality of the generated content. In addition, such indicators cannot be extended to non-text modal content. In terms of manual evaluation, evaluators are required to evaluate the quality of generated content based on predefined evaluation criteria.

[0027] Since it is aimed at generative models, automatic evaluation methods cannot accurately evaluate whether a model is usable. Taking text as an example, the grammatical errors, text style, text logic and text bias in the text output by the model are beyond the scope of automatic evaluation methods. However, relying solely on manual evaluation may lead to subjective bias, resulting in inaccurate model evaluation, and the evaluation results of different models by different people may not be comparable.

[0028] In an embodiment of the present disclosure, an improved model evaluation scheme is proposed. In this scheme, the output generated by the generative model to be evaluated for each input in the reference input set is determined based on a reference input set. On multiple predetermined quality assessment dimensions, each output generated by the model is labeled, and the quality level to which each output is divided on each quality assessment dimension is labeled. Each quality assessment dimension can be divided into multiple quality levels at a certain granularity. Each quality level has a corresponding quality score. In this way, the overall quality score of the generative model can be determined based on the labeling information of multiple outputs and the quality score corresponding to each quality level. By using the same quality assessment dimension, reference input set and labeling standard, the quality of the model-generated content can be accurately and objectively measured, and the quality of different models can be compared.

[0029] Some example embodiments of the present disclosure will be described below with reference to the accompanying drawings.

[0030] Figure 1 A schematic diagram of an example environment 100 in which embodiments of the present disclosure can be implemented is shown. In the environment 100, a model evaluation system 110 is configured to perform model performance evaluation on one or more generative models 120-1, 120-2, ... 120-N, the quality of model generated content. For ease of discussion, these generative models are collectively or individually referred to as generative models 120. Each generative model 120 can process an input 122 and provide a corresponding output 124.

[0031] Depending on the model configuration, the input 122 of the generative model 120 may include content in a textual modality, and the output 124 may also be content in a textual modality. For example, the generative model 120 may be a natural language processing model or a translation model, which can generate an output text sequence based on an input text sequence. In some example embodiments below, for the purpose of discussion, a text generation model is used as an example for illustration.

[0032] In some embodiments, the input 122 of the generative model 120 may also include other modalities, such as image modality, video modality, audio modality, etc., and the output 124 may also include other models other than text modality, such as image modality, video modality, audio modality. For example, the generative model 120 can be an image generation model that can generate new images based on input text and / or images. The generative model 120 can also be a speech synthesis model that can generate speech based on input text. The generative model 120 can be any appropriate unimodal or multimodal content generation model.

[0033] In the environment 100, the model evaluation system 110 can be any type of terminal device or server device. Examples of terminal devices include mobile terminals, fixed terminals or portable terminals, including mobile phones, desktop computers, laptop computers, notebook computers, netbook computers, tablet computers, media computers, multimedia tablets, personal communication systems (PCS) devices, personal navigation devices, personal digital assistants (PDAs), audio / video players, digital cameras / camcorders, positioning devices, television receivers, radio broadcast receivers, e-book devices, gaming devices, or any combination of the foregoing, including accessories and peripherals of these devices or any combination thereof. In some embodiments, the model evaluation system 110 can also support any type of interface for users (such as "wearable" circuits, etc.). Examples of server devices include, but are not limited to, mainframes, edge computing nodes, computing devices in cloud environments, and the like.

[0034] It should be understood that the structure and function of the various elements in the environment 100 are described for exemplary purposes only and do not imply any limitation on the scope of the present disclosure.

[0035] Figure 2 A schematic block diagram of a model evaluation system 110 according to some embodiments of the present disclosure is shown. As shown, the input set used as the reference input includes a plurality of inputs 212-1, 212-2, ... 212-M (collectively or individually referred to as inputs 212). The modality of the input 212 depends on the requirements of the generative model 120 to be evaluated. The input set can be collected from any appropriate data source for performing model evaluation.

[0036] If the quality of the output generated by the generative model 120-1 is to be evaluated, multiple inputs 212 in the input set can be provided to the generative model 120-1 respectively to obtain an output set output by the generative model 120-1, and the output set includes multiple outputs 222-1, 222-2, ... 222-M corresponding to the multiple inputs 212-1, 212-2, ... 212-M respectively.

[0037] In some embodiments, if multiple generative models 120 are to be evaluated, or the model performance of multiple generative models 120 is to be compared, multiple inputs 212 in the same input set can also be provided to other generative models, such as generative models 120-2, ... 120-N, to obtain an output set output by the generative model 120-2, which includes multiple outputs 224-1, 224-2, ... 224-M (collectively or individually referred to as outputs 224) corresponding to the multiple inputs 212-1, 212-2, ... 212-M, and an output set output by the generative model 120-N, which includes multiple outputs 226-1, 226-2, ... 226-M (collectively or individually referred to as outputs 226) corresponding to the multiple inputs 212-1, 212-2, ... 212-M. Depending on the configuration of the generative model, each input in the input set may include data of at least one of the following modalities: text modality, image modality, video modality, audio modality. Each output in the output set may include data in at least one of the following modalities: text modality, image modality, video modality, audio modality.

[0038] Further, the output set of each generative model 120 is provided to the annotation subsystem 230 to annotate each output according to a plurality of predetermined quality assessment dimensions. Before describing the specific annotation, the predetermined quality assessment dimensions are briefly described.

[0039] In some embodiments, the quality assessment dimensions for the generative model may include the following dimensions.

[0040] The accuracy dimension is used to evaluate the accuracy of the answer text output by the model for the input question text. The answer text should be free of grammatical errors and factual errors. If the correct answer is provided as a reference, the generated content can be roughly estimated to be aligned with the provided answer, thereby judging the accuracy of the answer.

[0041] Clarity / Explicitness: The response text should be easy to understand and as concise as possible. Clarity can be demonstrated by being descriptive and using appropriate vocabulary. The argument of the response sentence should be clear and uncomplicated, in other words, each sentence should have one argument, not multiple arguments at once.

[0042] The completeness dimension is used to assess whether the response text should cover the question comprehensively; that is, it should include all the parts that should be included in the answer. The response should provide enough details for the user to clearly understand what he should do based on the response.

[0043] The specificity dimension was used to assess whether the response text should be focused and evidence-based (supported by facts rather than generalities), and the response text should not be vague.

[0044] The politeness dimension is used to assess whether the reply text should be polite, kind, and positive to the user.

[0045] The safety dimension is used to evaluate whether the reply text should be safe for users, should not use harmful words or generate sensitive content, and should not require personal user information in the answer.

[0046] The caring dimension is used to evaluate whether the reply text should pay attention to the different perspectives, backgrounds, emotions, mentalities, and knowledge bases of users and give different reply contents.

[0047] In some embodiments, the above-mentioned quality assessment dimensions can be used not only to measure the text content output by the generative model, but also to measure the content of other modalities. In some embodiments, the selection of multiple quality assessment dimensions for evaluating the model can be based on at least one of the following: the modality included in the output of the generative model, or the modality included in the input of the generative model. For example, if the output of the generative model 120 includes an image, then the quality assessment dimensions may include the above-mentioned multiple dimensions, and may also include image clarity, which is used to evaluate the clarity of the output image. In addition, if the input of the generative model 120 includes text and the output includes an image, that is, the generative model 120 is a text-to-image generative model, then the quality assessment dimensions may include the above-mentioned multiple dimensions, image clarity, and the degree of match between text and image. The quality assessment dimensions can be selected according to the configuration and application needs of the specific model.

[0048] After the quality assessment dimension is selected, in an embodiment of the present disclosure, each quality assessment dimension can also be divided into multiple quality levels, each quality level corresponding to the quality range of the output of the model under the dimension. For example, the quality assessment dimension can be divided into three dimensions: positive, neutral, and negative, corresponding to high, medium, and low quality levels, respectively. The positive indicates that the output of the model meets the requirements of the quality assessment dimension, the neutral indicates that the output of the model basically meets the requirements of the quality assessment dimension, and the negative indicates that the output of the model does not meet the requirements of the quality assessment dimension. Of course, finer or coarser quality levels can also be divided as needed. The granularity of the quality levels divided by different quality assessment dimensions can be the same, or can be different.

[0049] After the quality assessment dimensions are defined and the quality levels of each quality assessment dimension are divided, the annotation subsystem 230 obtains the annotation information for each of the multiple outputs in the output set of each generative model 120. The annotation information indicates the quality level of each output of the generative model 120 that is annotated in the multiple quality levels divided under each of the multiple quality assessment dimensions. For example, under each quality assessment dimension, each output can be annotated as positive, neutral, or negative.

[0050] In some embodiments, the output of the generative model 120 can be manually annotated. Before the annotation work begins, the definition of each quality assessment dimension, the division and definition of the quality level can be provided to the annotator so that the annotator can implement and understand the quality standards. In addition, the input of the generative model 120 and its corresponding output (for example, for text modality, the input can be a question and the output can be an answer) can be presented to the annotator at the same time. In this way, before making a judgment on the output under a certain dimension, it can be ensured that the annotator clearly understands the corresponding input. In some embodiments, for each quality assessment dimension, the annotator can simply select one of the multiple quality levels under the dimension, for example, the annotator determines whether the reply is a positive, negative or neutral reply. For example, if the model input is the question “what is the difference between ipv4 vs. ipv6 routing?” and a generative model 120 generates the answer “The main difference between IPv4 and IPv6 is that IPv4 is 32-bit IP address whereas IPv6 is 128-bit IP address.... IPv4 supports only 32-bit addresses, whereas IPv6 supports 128-bit addresses. IPv4 supports binary digits, whereas IPv.” then the answer will be labeled negatively for completeness because the answer generation is incomplete.

[0051] In some embodiments, in addition to manual annotation, annotation can also be performed by automatic tools. For example, a dialogue model (also belonging to a class of generative models) whose model performance is tested to meet the requirements can be selected as a benchmark generative model. Such a generative model can complete the corresponding operation based on the request input by the user. Therefore, the output of other generative models is annotated using the benchmark generative model. When annotating each output, the definition of each quality assessment dimension, the division and definition of the quality level, together with the input and corresponding output of the model, can be provided to the benchmark generative model, and the generative model is required to annotate the output. Since the multiple generative models 120 to be evaluated are evaluated based on the same benchmark model, the quality assessment results finally obtained can also reasonably reflect the relative quality of multiple generative models 120.

[0052] In some embodiments, during the labeling process, if positive, negative or neutral is selected for a certain dimension, the rationality of the selection can be noted, that is, why the dimension is negative and other dimensions are positive. This information can help the subsequent model evaluation system to fine-tune the score.

[0053] According to the above labeling process, labeling information of multiple quality levels (positive, negative, neutral) of each output corresponding to each input in each quality assessment dimension can be obtained. The labeling information of the output set of the generative model 120 is further provided to the score calculation subsystem 240 in the model evaluation system 110. For each generative model 120, the score calculation subsystem 120 determines the overall quality score of the generative model 120 based on at least the labeling information of each of the multiple outputs in the output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions.

[0054] Each quality level can be assigned a corresponding quality score. For example, the positive score can be set to 5 points, the neutral score to 2 points, and the negative score to 0 points. In this way, by counting the annotation results of each output in multiple quality assessment dimensions, the quality score of the output can be determined. Further, by counting the quality scores of multiple outputs of each generative model 120, the overall quality score of the generative model 120 can be determined. Specifically, the quality scores corresponding to the quality levels annotated for each output under each dimension can be summed to obtain the quality score of the output. In some embodiments, the overall quality score of the generative model can be determined by aggregating the quality scores of multiple outputs (for example, taking the average). For examples of determining the quality scores of outputs and the overall quality scores, see Table 1 below. In the example of Table 1, a positive score of 5 points, a neutral score of 2 points, and a negative score of 0 points are used as an example for illustration.

[0055] Table 1

[0056]

[0057] In some embodiments, when the generative model 120 is measured according to multiple quality assessment dimensions, the weights corresponding to each of the multiple quality assessment dimensions can also be determined based on the priorities of the multiple quality assessment dimensions. In this way, the model evaluation system 110 also determines the overall quality score of the generative model 120 based on the weights corresponding to each of the multiple quality assessment dimensions. For example, different priorities can be set for different quality assessment dimensions. For example, if the generative model is very concerned about the dimension of accuracy, then this dimension is set to p0 priority, while completeness is set to p1 priority, politeness, security, etc. are set to p2 priority, and different priorities represent different weights. In different application scenarios, the degree of emphasis on different quality assessment dimensions may be different, and weights can be set to highlight the impact of high-priority quality assessment dimensions on the quality of model output.

[0058] In the case of considering the weights of multiple quality assessment dimensions, when calculating each generated quality score, the quality score of the output is determined based on the first annotation information of the output, the quality scores corresponding to the multiple quality levels divided under the multiple quality assessment dimensions, and the weights. The determination of the quality score of the output can be expressed as follows:

[0059]

[0060] Where P i Represents the weight corresponding to the quality assessment dimension, Dimension j It represents the quality score of the output under the jth quality assessment dimension (corresponding to the marked quality level). The final quality score of the output can be determined by multiplying the weight and the score of the dimension and adding the results under multiple instruction evaluation dimensions.

[0061] In some embodiments, the quality score of each output may be normalized to a predetermined score interval (e.g., a percentage interval). However, the overall quality score may be determined by aggregating the quality scores of multiple outputs, and the overall quality score also falls within the predetermined score interval (e.g., a percentage interval). Of course, after aggregating the quality scores of multiple outputs, the aggregated quality scores may be normalized to a predetermined score interval to obtain the final overall quality score of each generative model 120.

[0062] In some embodiments, the model performance of the generative model 120 may also be evaluated by dimension. Specifically, for a given quality assessment dimension, the quality score of the generative model 230 in the dimension is determined based on at least the annotation information of the multiple outputs of the generative model 120 in the given quality assessment dimension and the quality scores corresponding to the multiple quality levels divided in the given quality assessment dimension. The quality score in each dimension may be normalized to a predetermined score interval.

[0063] In some embodiments, after calculating the overall quality scores of the respective generative models 120 , the model performances of the generative models 120 can be determined by comparing the overall quality scores, thereby judging the overall model quality. Figure 3 A comparison chart 300 of the overall quality scores of models according to some embodiments of the present disclosure is shown. By calculating the overall quality score, the quality trend of multiple models can be easily determined. Since the overall quality score is calculated based on the same input set and the same quality evaluation criteria, such a model evaluation result is more accurate and objective.

[0064] In some embodiments, in addition to comparing the overall quality scores of multiple generative models 120, for a single generative model 120, whether the model meets expectations can also be measured by relying on the overall quality score. For example, a quality score threshold can be set to determine whether a generative model 120 can be applied.

[0065] In some embodiments, in addition to aggregating multiple quality assessment dimensions to measure the overall quality score, the performance of multiple generative models 120 on different quality assessment dimensions can be compared by dimension, or the performance of a single generative model 120 on a certain quality assessment dimension can be measured to see whether it meets expectations.

[0066] In some embodiments, the model evaluation system 110 may also include an evaluation fine-tuning function. The model evaluation system 110 may receive feedback on the quality score of at least one of the multiple outputs in the output set. The feedback may indicate whether the quality score of the corresponding output can accurately reflect the quality of the output. For example, it is possible to return to the specific input and output for verification to determine whether the output with a high quality score is indeed of high quality, and whether the output with a low quality score is indeed of low quality. This process requires human judgment, which can help optimize the entire quality evaluation system so that the quality evaluation results are increasingly fitted to the actual generation quality of the model. Specifically, if the feedback indicates that at least one quality score fails to accurately reflect the quality of at least one output, one or more aspects involved in the quality evaluation system may be fine-tuned. In some embodiments, the weights corresponding to each of the multiple quality evaluation dimensions may be updated to make the calculated quality score more accurate. In some embodiments, alternatively or additionally, the multiple quality levels divided under each quality evaluation dimension may be updated. For example, for one or more quality evaluation dimensions, fine-grained quality levels may be divided. In some embodiments, alternatively or additionally, the quality scores corresponding to the multiple quality levels divided under each quality evaluation dimension may be updated. For example, the original setting of the positive score is 5 points, the neutral score is 2 points, and the negative score is 0 points. It can be updated later to a positive score of 5 points, a neutral score of 2.5 points, and a negative score of 0 points.

[0067] In some embodiments, for the same input, the generative model 120 may provide different outputs. In some embodiments, by providing a certain input (referred to as: first input) to the generative model 120 multiple times, multiple first outputs corresponding to the first input can be obtained. However, the annotation information of each of the multiple first outputs corresponding to the first input can be obtained by the annotation subsystem 230, and the annotation information indicates the quality level of each output that is annotated in the multiple quality levels divided under each quality assessment dimension in the multiple quality assessment dimensions. The quality scores of each of the multiple first outputs can be determined by the score calculation subsystem 240 based on the score determination method discussed above. In this way, for the same input, the quality scores of the multiple inputs given by the generative model 120 can be judged.

[0068] Furthermore, the corresponding generative model 120 can be updated based on the ranking of the quality scores of the multiple first outputs. The goal of the update is to enable the generative model 120 to preferentially output inputs with higher quality scores in subsequent operation. In this way, the generative model 120 can be guided to generate and provide higher quality outputs.

[0069] In some embodiments, during the reinforcement learning phase of the generative model 120 , the ranking of the quality scores of the plurality of first outputs may be fed back to the generative model 120 , so that the generative model 120 learns to output results with higher quality scores.

[0070] Figure 4 FIG. 4 is a flowchart of a process 400 for model evaluation according to some embodiments of the present disclosure. The process 400 may be implemented in Figure 1 The model evaluation system 110 is provided.

[0071] In box 410, the model evaluation system 110 provides multiple inputs in the input set to the first generative model respectively to obtain a first output set output by the first generative model, where the first output set includes multiple outputs corresponding to the multiple inputs respectively.

[0072] In block 420 , the model evaluation system 110 obtains first labeling information for each of the plurality of outputs in the first output set, the first labeling information indicating a labeled quality level of each output in a plurality of quality levels divided under each quality evaluation dimension in a plurality of quality evaluation dimensions.

[0073] In box 430, the model evaluation system 110 determines a first overall quality score of the first generative model based at least on the first annotation information of each of the multiple outputs in the first output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality evaluation dimensions.

[0074] In some embodiments, process 400 also includes: for a given quality assessment dimension among the multiple quality assessment dimensions, determining the quality score of the first generative model under the given quality assessment dimension based at least on first annotation information of multiple outputs in the first output set under the given quality assessment dimension and quality scores corresponding to each of multiple quality levels divided under the given quality assessment dimension.

[0075] In some embodiments, process 400 also includes: providing multiple inputs from the input set to a second generative model respectively to obtain a second output set output by the second generative model, the second output set including multiple outputs corresponding to the multiple inputs respectively; obtaining second labeling information for each of the multiple outputs in the second output set, the labeling information indicating the quality level labeled for each output in the multiple quality levels divided under each quality assessment dimension in the multiple quality assessment dimensions; determining a second overall quality score of the second generative model based at least on the second labeling information for each of the multiple outputs in the second output set and the quality scores corresponding to the multiple quality levels divided under the multiple quality assessment dimensions; and comparing the model performance of the first generative model and the second generative model by at least comparing the first overall quality score and the second overall quality score.

[0076] In some embodiments, determining the overall quality score of the first generative model includes: determining the weights corresponding to each of the multiple quality assessment dimensions based on the priorities of the multiple quality assessment dimensions; and determining the first overall quality score based on the weights corresponding to each of the multiple quality assessment dimensions.

[0077] In some embodiments, determining the first overall quality score based on the weights corresponding to each of the multiple quality assessment dimensions includes: for each of the multiple outputs in the first output set, determining the quality score of the output based on the first annotation information of the output, the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions, and the weights; and determining the first overall quality score by aggregating the quality scores of the multiple outputs.

[0078] In some embodiments, process 400 also includes: receiving feedback on the quality score of at least one output among the multiple outputs in the first output set; if the feedback indicates that the at least one quality score fails to accurately reflect the quality of the at least one output, determining an update to at least one of the following: the weights corresponding to each of the multiple quality assessment dimensions, the multiple quality levels divided under each quality assessment dimension, or the quality scores corresponding to the multiple quality levels divided under each quality assessment dimension.

[0079] In some embodiments, process 400 also includes: obtaining multiple first outputs corresponding to the first input by providing the first input to the first generative model multiple times; obtaining third labeling information of each of the multiple first outputs corresponding to the first input, the third labeling information indicating the quality level labeled for each output in the multiple quality levels divided under each quality assessment dimension in the multiple quality assessment dimensions; determining the quality scores of each of the multiple first outputs based at least on the third labeling information of each of the multiple first outputs and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions; and updating the first generative model based on the ranking of the quality scores of each of the multiple first outputs.

[0080] In some embodiments, each input in the input set includes data of at least one of the following modalities: text modality, image modality, video modality, audio modality. In some embodiments, each output in the first output set includes data of at least one of the following modalities: text modality, image modality, video modality, audio modality.

[0081] In some embodiments, the selection of the plurality of quality assessment dimensions is based on at least one of: a modality included in an output of the first generative model, or a modality included in an input of the first generative model.

[0082] Figure 5 A block diagram of an apparatus 500 for model evaluation according to some embodiments of the present disclosure is shown. The apparatus 500 may be implemented as or included in the model evaluation system 110. Each module / component in the apparatus 500 may be implemented by hardware, software, firmware, or any combination thereof.

[0083] As shown in the figure, the device 500 includes a first input providing module 510, which is configured to provide multiple inputs in the input set to the first generative model respectively to obtain a first output set output by the first generative model, and the first output set includes multiple outputs corresponding to the multiple inputs respectively. The device 500 includes a first annotation obtaining module 520, which is configured to obtain first annotation information for each of the multiple outputs in the first output set, and the first annotation information indicates the quality level of each output labeled in the multiple quality levels divided under each quality assessment dimension in multiple quality assessment dimensions. The device 500 also includes a first overall quality determination module 530, which is configured to determine the first overall quality score of the first generative model based on at least the first annotation information of each of the multiple outputs in the first output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions.

[0084] In some embodiments, the device 500 also includes: a dimension score determination module, which is configured to determine the quality score of the first generative model under the given quality assessment dimension for a given quality assessment dimension among multiple quality assessment dimensions, based at least on the first annotation information of multiple outputs in the first output set under the given quality assessment dimension and the quality scores corresponding to each of the multiple quality levels divided under the given quality assessment dimension.

[0085] In some embodiments, the device 500 also includes: a second input providing module, configured to provide multiple inputs from the input set to the second generative model respectively to obtain a second output set output by the second generative model, the second output set including multiple outputs corresponding to the multiple inputs respectively; a second annotation obtaining module, configured to obtain second annotation information for each of the multiple outputs in the second output set, the annotation information indicating the quality level of each output annotated in the multiple quality levels divided under each quality assessment dimension in the multiple quality assessment dimensions; a second overall quality determination module, configured to determine the second overall quality score of the second generative model based at least on the second annotation information of each of the multiple outputs in the second output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions; and a quality comparison module, configured to compare the model performance of the first generative model and the second generative model by at least comparing the first overall quality score and the second overall quality score.

[0086] In some embodiments, the first overall quality determination module 530 is further configured to determine weights corresponding to each of the multiple quality assessment dimensions based on their priorities; and to determine a first overall quality score based on the weights corresponding to each of the multiple quality assessment dimensions.

[0087] In some embodiments, the first overall quality determination module 530 is also configured to determine, for each of the multiple outputs in the first output set, a quality score of the output based on the first annotation information of the output, the quality scores corresponding to each of the multiple quality levels divided under multiple quality assessment dimensions, and the weights; and to determine the first overall quality score by aggregating the quality scores of the multiple outputs.

[0088] In some embodiments, the apparatus 500 further includes: a feedback receiving module configured to receive feedback on a quality score of at least one of the multiple outputs in the first output set; and an updating module configured to determine an update to at least one of the following items if the feedback indicates that the at least one quality score fails to accurately reflect the quality of the at least one output: weights corresponding to each of the multiple quality assessment dimensions, multiple quality levels divided under each quality assessment dimension, or quality scores corresponding to the multiple quality levels divided under each quality assessment dimension.

[0089] In some embodiments, the device 500 also includes: a repeated providing module, configured to obtain multiple first outputs corresponding to the first input by providing the first input to the first generative model multiple times; a third annotation obtaining module, configured to obtain third annotation information of each of the multiple first outputs corresponding to the first input, the third annotation information indicating the quality level of each output labeled in the multiple quality levels divided under each quality assessment dimension in the multiple quality assessment dimensions; a multi-output score determination module, configured to determine the quality scores of each of the multiple first outputs based at least on the third annotation information of each of the multiple first outputs and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions; and a model updating module, configured to update the first generative model based on the ranking of the quality scores of each of the multiple first outputs.

[0090] In some embodiments, each input in the input set includes data in at least one of the following modalities: text modality, image modality, video modality, audio modality. In some embodiments, each output in the first output set includes data in at least one of the following modalities: text modality, image modality, video modality, audio modality.

[0091] In some embodiments, the selection of the plurality of quality assessment dimensions is based on at least one of: a modality included in the output of the first generative model, or a modality included in the input of the first generative model.

[0092] Figure 6 1 shows a block diagram of an electronic device 600 in which one or more embodiments of the present disclosure may be implemented. It should be understood that Figure 6 The electronic device 600 shown is merely exemplary and should not constitute any limitation on the functionality and scope of the embodiments described herein. Figure 6 The electronic device 600 shown can be used to implement Figure 1 content management system 120 and / or content distribution platform 120, Figure 2 The model evaluation system 110, the model supply system 220 and / or the project delivery system 230. Figure 6 The electronic device 600 shown can be used to implement Figure 5 Device 500.

[0093] like Figure 6As shown, the electronic device 600 is in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to, one or more processors or processing units 610, a memory 620, a storage device 630, one or more communication units 640, one or more input devices 650, and one or more output devices 660. The processing unit 610 may be an actual or virtual processor and is capable of performing various processes according to a program stored in the memory 620. In a multi-processor system, multiple processing units execute computer executable instructions in parallel to improve the parallel processing capabilities of the electronic device 600.

[0094] The electronic device 600 typically includes a plurality of computer storage media. Such media can be any available media accessible to the electronic device 600, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 620 can be a volatile memory (e.g., registers, caches, random access memory (RAM)), a non-volatile memory (e.g., a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), flash memory) or some combination thereof. The storage device 630 can be a removable or non-removable medium, and can include a machine-readable medium, such as a flash drive, a disk, or any other medium, which can be used to store information and / or data and can be accessed within the electronic device 600.

[0095] The electronic device 600 may further include additional removable / non-removable, volatile / non-volatile storage media. Figure 6 As shown in , a disk drive for reading or writing from a removable, non-volatile disk (e.g., a "floppy disk") and an optical drive for reading or writing from a removable, non-volatile optical disk may be provided. In these cases, each drive may be connected to the bus (not shown) by one or more data media interfaces. The memory 620 may include a computer program product 625 having one or more program modules that are configured to perform various methods or actions of various embodiments of the present disclosure.

[0096] The communication unit 640 implements communication with other electronic devices through a communication medium. Additionally, the functions of the components of the electronic device 600 can be implemented with a single computing cluster or multiple computing machines that can communicate through a communication connection. Therefore, the electronic device 600 can operate in a networked environment using a logical connection with one or more other servers, a network personal computer (PC), or another network node.

[0097] The input device 650 may be one or more input devices, such as a mouse, a keyboard, a tracking ball, etc. The output device 660 may be one or more output devices, such as a display, a speaker, a printer, etc. The electronic device 600 may also communicate with one or more external devices (not shown) through the communication unit 640 as needed, such as a storage device, a display device, etc., communicate with one or more devices that allow a user to interact with the electronic device 600, or communicate with any device that allows the electronic device 600 to communicate with one or more other electronic devices (e.g., a network card, a modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).

[0098] According to an exemplary implementation of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is also provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, and the computer-executable instructions are executed by a processor to implement the method described above.

[0099] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices, equipment, and computer program products implemented according to the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0100] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0101] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, so that the instructions executed on the computer, other programmable data processing apparatus, or other device implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0102] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple implementations of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of a module, program segment or instruction includes one or more executable instructions for realizing the logical function of the specification. In some implementations as replacements, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be realized by a special hardware-based system that performs the function or action of the specification, or can be realized by a combination of special hardware and computer instructions.

[0103] The above descriptions of various implementations of the present disclosure are exemplary, non-exhaustive, and not limited to the disclosed implementations. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described implementations. The selection of terms used herein is intended to best explain the principles of the implementations, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the various implementations disclosed herein.

Claims

1. A method for model evaluation, comprising: Providing a plurality of inputs in the input set to a first generative model respectively, to obtain a first output set output by the first generative model, wherein the first output set includes a plurality of outputs corresponding to the plurality of inputs respectively; Obtaining first labeling information for each of the plurality of outputs in the first output set, the first labeling information indicating a quality level of each output labeled in a plurality of quality levels divided under each quality assessment dimension in a plurality of quality assessment dimensions; as well as A first overall quality score of the first generative model is determined based at least on the first annotation information of each of the multiple outputs in the first output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions.

2. The method according to claim 1, further comprising: For a given quality assessment dimension among the plurality of quality assessment dimensions, The quality score of the first generative model under the given quality assessment dimension is determined based at least on first annotation information of multiple outputs in the first output set under the given quality assessment dimension and quality scores corresponding to each of multiple quality levels divided under the given quality assessment dimension.

3. The method according to claim 1, further comprising: Provide a plurality of inputs from the input set to a second generative model respectively, so as to obtain a second output set output by the second generative model, wherein the second output set includes a plurality of outputs corresponding to the plurality of inputs respectively; Obtaining second labeling information for each of the plurality of outputs in the second output set, the labeling information indicating a labeled quality level of each output in a plurality of quality levels divided under each quality assessment dimension in a plurality of quality assessment dimensions; Determine a second overall quality score of the second generative model based at least on the second annotation information of each of the multiple outputs in the second output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions; as well as Model performance of the first generative model and the second generative model are compared by at least comparing the first overall quality score and the second overall quality score.

4. The method of claim 1 , wherein determining an overall quality score of the first generative model comprises: Determining weights corresponding to the multiple quality assessment dimensions based on the priorities of the multiple quality assessment dimensions; as well as The first overall quality score is also determined based on weights corresponding to each of the plurality of quality assessment dimensions.

5. The method according to claim 4, wherein determining the first overall quality score based on the weights corresponding to the plurality of quality assessment dimensions comprises: For each of the multiple outputs in the first output set, determine a quality score of the output based on the first annotation information of the output, the quality scores and weights corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions; as well as The first overall quality score is determined by aggregating the quality scores of the plurality of outputs.

6. The method according to claim 5, further comprising: receiving feedback on a quality score of at least one of the plurality of outputs in the first set of outputs; If the feedback indicates that the at least one quality score does not accurately reflect the quality of the at least one output, determining an update to at least one of: The weights corresponding to the multiple quality assessment dimensions respectively, Multiple quality levels for each quality assessment dimension, or The quality scores corresponding to the multiple quality levels divided under each quality assessment dimension.

7. The method according to claim 1, further comprising: Obtaining a plurality of first outputs corresponding to the first input by providing the first input to the first generative model multiple times; Obtaining third labeling information of each of the plurality of first outputs corresponding to the first input, the third labeling information indicating a quality level labeled for each output in a plurality of quality levels divided under each quality assessment dimension in the plurality of quality assessment dimensions; Determine a quality score of each of the plurality of first outputs based at least on the third annotation information of each of the plurality of first outputs and a quality score corresponding to each of the plurality of quality levels divided under the plurality of quality assessment dimensions; as well as The first generative model is updated based on a ranking of the quality scores of each of the plurality of first outputs.

8. The method of claim 1, wherein each input in the input set comprises data in at least one of the following modalities: a text modality, an image modality, a video modality, an audio modality; and Each output in the first output set includes data in at least one of the following modalities: text modality, image modality, video modality, and audio modality.

9. The method of claim 8, wherein the selection of the plurality of quality assessment dimensions is based on at least one of: a modality included in the output of the first generative model, or a modality included in the input of the first generative model.

10. A device for model evaluation, comprising: A first input providing module is configured to provide a plurality of inputs in the input set to the first generative model respectively, so as to obtain a first output set output by the first generative model, wherein the first output set includes a plurality of outputs respectively corresponding to the plurality of inputs; A first labeling obtaining module is configured to obtain first labeling information for each of the plurality of outputs in the first output set, wherein the first labeling information indicates a quality level of each output labeled in a plurality of quality levels divided under each quality assessment dimension in a plurality of quality assessment dimensions; as well as The first overall quality determination module is configured to determine a first overall quality score of the first generative model based at least on the first annotation information of each of the multiple outputs in the first output set and the quality scores corresponding to each of the multiple quality levels divided under the multiple quality assessment dimensions.

11. An electronic device, comprising: at least one processing unit; as well as At least one memory, the at least one memory being coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions causing the device to perform the method according to any one of claims 1 to 9 when executed by the at least one processing unit.

12. A computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the method according to any one of claims 1 to 9 when executed by a processor.

Images