Model evaluation method and device, electronic equipment and computer readable storage medium

Abstract

The invention discloses a model evaluation method and device, electronic equipment and a computer readable storage medium. The method comprises the steps of obtaining a to-be-evaluated answer text and a corresponding target question text; wherein the to-be-evaluated answer text is generated by the target model based on the target question text; determining question attributes of the target question text; selecting at least one referee model matched with the question attribute from a referee model pool as a target referee model; and evaluating the answer text to be evaluated by using each target judgment model to obtain a first target evaluation result. In this way, the model evaluation accuracy can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of model evaluation, in particular to a model evaluation method and device, electronic equipment and computer readable storage medium. BACKGROUND

[0002] At present, with the vigorous development of artificial intelligence and natural language processing technology, various text generation systems based on target models have been widely applied. These target models can automatically generate corresponding answer texts according to input question texts, and play an important role in intelligent customer service, question and answer systems, content creation assistance and other fields.

[0003] At present, most existing evaluation methods use a single evaluation model or fixed evaluation standards, and when evaluating the answer texts generated by the target model, the evaluation accuracy is difficult to guarantee. SUMMARY

[0004] The technical problem solved by the present application is to provide a model evaluation method, device, electronic equipment and computer readable storage medium, which can improve the model evaluation accuracy.

[0005] To solve the above technical problems, the first aspect of the present application provides a model evaluation method, which comprises: obtaining a to-be-evaluated answer text and a corresponding target question text; wherein the to-be-evaluated answer text is generated by a target model based on the target question text; determining the question attribute of the target question text; selecting at least one adjudication model matched with the question attribute from a pool of adjudication models as a target adjudication model; and using each target adjudication model to evaluate the to-be-evaluated answer text to obtain a first target evaluation result.

[0006] Among them, the target adjudication model is at least two; using each target adjudication model to evaluate the to-be-evaluated answer text to obtain a first target evaluation result, comprising: assigning a first evaluation weight to each target adjudication model; and using each target adjudication model to evaluate the to-be-evaluated answer text to obtain a second target evaluation result corresponding to each target adjudication model; based on the first evaluation weight corresponding to each target adjudication model, weighting and fusing each second target evaluation result to obtain the first target evaluation result.

[0007] The problem attribute is a mixed attribute of subjectivity and objectivity; before selecting at least one adjudication model matching the problem attribute from the adjudication model pool as a target adjudication model, the model evaluation method further comprises: decomposing the answer text to be evaluated to obtain a first sub-answer text corresponding to the objective attribute and a second sub-answer text corresponding to the subjective attribute; selecting at least one adjudication model matching the problem attribute from the adjudication model pool as a target adjudication model, comprising: selecting at least one adjudication model matching the objective attribute from the adjudication model pool as a first target model; and selecting at least one adjudication model matching the subjective attribute from the adjudication model pool as a second target model; weighting and fusing each second target evaluation result based on the first evaluation weight corresponding to each target adjudication model to obtain a first target evaluation result, comprising: weighting and fusing each first result based on the first evaluation weight corresponding to each first target model to obtain a second result; wherein each first result is obtained by evaluating the first sub-answer text by each first target model; and weighting and fusing each third result based on the first evaluation weight corresponding to each second target model to obtain a fourth result; wherein each third result is obtained by evaluating the second sub-answer text by each second target model; obtaining the first target evaluation result based on the second result and the fourth result.

[0008] The first target evaluation result is obtained based on the second result and the fourth result, comprising: determining second evaluation weights corresponding to the objective attribute and the subjective attribute respectively; weighting and fusing the second result and the fourth result based on the second evaluation weights corresponding to the objective attribute and the subjective attribute to obtain the first target evaluation result.

[0009] The first evaluation weight is allocated to each target adjudication model, comprising: obtaining a model historical performance and a parameter configuration vector of each target adjudication model; determining the first evaluation weight of each target adjudication model based on the model historical performance and the parameter configuration vector of each target adjudication model.

[0010] The problem attribute is represented by a problem attribute vector, and selecting at least one adjudication model matching the problem attribute from the adjudication model pool as a target adjudication model is performed by a first evaluation executor; before selecting at least one adjudication model matching the problem attribute from the adjudication model pool as a target adjudication model, the model evaluation method further comprises: selecting a second evaluation executor meeting the evaluation requirement as the first evaluation executor based on the problem attribute vector from a plurality of second evaluation executors; and generating a parameter configuration vector based on the problem attribute vector; wherein the parameter configuration vector is used to guide the first evaluation executor to select the target adjudication model; and routing the parameter configuration vector to the first evaluation executor.

[0011] The second evaluation executor meeting the evaluation requirements is selected as the first evaluation executor from the second evaluation executors based on the problem attribute vector, including: obtaining a system performance vector representing historical performance of the system; splicing the system performance vector and the problem attribute vector to obtain a composite state vector; predicting a total long-term reward corresponding to each second evaluation executor based on the composite state vector by using an executor decision model; and taking the second evaluation executor corresponding to the maximum total long-term reward as the first evaluation executor.

[0012] The problem attribute is subjective or subjective and objective mixed; after the first target evaluation result is obtained by using each target judge model to evaluate the answer text to be evaluated, the model evaluation method further includes: reviewing the first target evaluation result in at least one review dimension to obtain a target review result; and adjusting the first target evaluation result based on the target review result to obtain a third target evaluation result.

[0013] The at least one review dimension includes at least one of the following: compliance, user preference.

[0014] The first target evaluation result is reviewed in at least one review dimension to obtain a target review result by using a review model; at least one judge model matched with the problem attribute is selected from a judge model pool as a target judge model by the first evaluation executor; the first evaluation executor is selected from the second evaluation executors by using an executor decision model; and the model evaluation method further includes at least one of the following steps: fine-tuning the review model by using a sample review result and a real review result; the sample review result is obtained by reviewing a sample evaluation result in at least one review dimension by the review model; the sample evaluation result is obtained by evaluating the model evaluation model based on a sample answer text and a corresponding sample question text; and fine-tuning the executor decision model by using the sample review result.

[0015] To solve the above technical problems, the second aspect of the present application provides a model evaluation device, which comprises an acquisition module, a determination module, a selection module and an evaluation module; the acquisition module is used to acquire an answer text to be evaluated and a corresponding target question text; the answer text to be evaluated is generated by a target model based on the target question text; the determination module is used to determine the problem attribute of the target question text; the selection module is used to select at least one judge model matched with the problem attribute from a judge model pool as a target judge model; and the evaluation module is used to evaluate the answer text to be evaluated by using each target judge model to obtain a first target evaluation result.

[0016] To solve the above technical problems, the third aspect of the present application provides an electronic device, which comprises a memory and a processor, the memory is used to store program instructions, and the processor is used to execute the program instructions to realize the model evaluation method described above.

[0017] To solve the above technical problems, the fourth aspect of the present application provides a computer readable storage medium, which is used to store program instructions, and the program instructions can be executed to realize the model evaluation method described above.

[0018] The above technical solution uses the target judgment model matched with the problem attribute of the target problem text to evaluate the to-be-evaluated answer text corresponding to the target problem text, and the target judgment model can more accurately capture the key features of the to-be-evaluated answer text, thereby improving the evaluation accuracy of the to-be-evaluated answer text.

[0019] Further, in the case that the judgment models matched with the problem attribute of the target problem text selected from the judgment model pool are at least two, the to-be-evaluated answer text corresponding to the target problem text is evaluated by using at least two target judgment models matched with the problem attribute of the target problem text, and different target judgment models evaluate the to-be-evaluated answer text from different evaluation angles, so that the to-be-evaluated answer text can be more comprehensively evaluated. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flowchart of an embodiment of the model evaluation method provided by the present application; Figure 2 is a schematic diagram of the framework structure of an embodiment of the model evaluation system provided by the present application; Figure 3 is Figure 1 is a flowchart of an embodiment of the step S14 shown in the figure; Figure 4 is a schematic diagram of the structure of an embodiment of the model evaluation device provided by the present application; Figure 5 is a schematic diagram of the structure of an embodiment of the electronic device provided by the present application; Figure 6 is a schematic diagram of the structure of an embodiment of the computer readable storage medium provided by the present application. DETAILED DESCRIPTION

[0021] The schemes of the embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0022] In the following description, specific details such as specific system structures, interfaces, techniques, etc. are presented in order to thoroughly understand the present application, but are not intended to limit the present application.

[0023] The term "and / or", used herein only describes an associated relationship, which means that there can be three relationships, for example, A and / or B can represent three cases: A exists alone, A and B exist together, and B exists alone. In addition, the character " / " herein generally represents an "or" relationship between the associated objects. In addition, "multiple" herein means two or more than two. In addition, the term "at least one" herein means any one of multiple or any combination of at least two of multiple, for example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0024] Please refer to Figure 1 , Figure 1 is a flowchart of an embodiment of the model evaluation method provided by the present application. It should be noted that the embodiment is not limited to the order of the flowchart shown in Figure 1 . As shown in Figure 1 , the embodiment includes: Step S11: obtaining an answer text to be evaluated and a corresponding target question text.

[0025] In the present embodiment, the answer text to be evaluated and the corresponding target question text are obtained; wherein the answer text to be evaluated is generated by the target model based on the target question text.

[0026] In an embodiment, the target model is a large language model, then the model evaluation method provided by the present application is to evaluate the answer text generated by the large language model based on the question text. Of course, in other embodiments, the target model can also be a small model with less parameters, which is not limited here.

[0027] In an embodiment, the answer text to be evaluated and the corresponding target question text can be obtained from local storage or cloud storage. Of course, in other embodiments, the target question text can also be obtained in real time and the answer text to be evaluated can be generated based on the target question text using the target model, which is not limited here.

[0028] Step S12: determining the problem attribute of the target question text.

[0029] In the present embodiment, the problem attribute of the target question text is determined. Wherein, the problem attribute of the target question text can be subjective attribute (subjectivity), objective attribute (objectivity), and mixed attribute (subject-object-mixedness). That is, the target question text is quantitatively analyzed, from the binary classification of "subjectivity / objectivity" of the target question text to the multi-classification of "subject-object-mixedness", to realize the accurate quantitative analysis of the target question text.

[0030] In an embodiment, the question property of the target question text is represented by a question property vector, which is obtained by using a question property analysis model to analyze the target question text. The analysis process of the question property analysis model on the target question text is specifically represented as follows:

[0031] wherein, represents the question property vector; represents the objectivity weight component; represents the subjectivity weight component; represents the mixed subjectivity weight component; and T represents the target question text. represents the question property analysis model.

[0032] It should be noted that, The higher the value is, the more the target question text tends to examine the ability of the target model to master and output objective facts, deterministic knowledge, and logical rules. The higher the value is, the more the target question text tends to examine the ability of the target model to create, express emotions, write in a stylized manner, and make value judgments without standard answers. The higher the value is, the more the target question text requires the target model to have both objective knowledge and subjective ability. It is an index for measuring the complexity and comprehensiveness of the target question text.

[0033] In a specific embodiment, the question property analysis model is a specific CNN-Attention cascaded network architecture. The question property analysis model first extracts local semantic features through a convolutional neural network, then introduces a self-attention mechanism to capture global contextual relationships, and finally outputs a question property vector through a weight prediction layer. The question property analysis model is specifically represented as follows:

[0034] wherein, represents the question property vector; represents the objectivity weight component; represents the subjectivity weight component; represents the mixed subjectivity weight component; and T represents the target question text. represents the hidden representation vector of objectivity; represents the hidden representation vector of subjectivity; represents the hidden representation vector of mixed subjectivity; and MLP represents a multi-layer perceptron.

[0035] It should be noted that the hidden representation vector of objectivity , the hidden representation vector of subjectivity Hidden representation vectors of subjective and objective characteristics It is extracted from text features through a self-attention mechanism (attention pooling mechanism). These text features are obtained by extracting features from the target question text using a convolutional neural network. Furthermore, the weight components of the question attribute vector satisfy... Furthermore, all weighted components are non-negative.

[0036] In one implementation, such as Figure 2 As shown, Figure 2 This is a schematic diagram of the framework structure of an embodiment of the model evaluation system provided in this application. The model evaluation system includes a module one, in which the problem attributes of the target problem text are determined by using the problem attribute parsing model. Figure 2 (1-Task Attribute Analysis).

[0037] Step S13: Select at least one referee model from the referee model pool that matches the problem attributes as the target referee model.

[0038] In this implementation, at least one judge model matching the question attributes is selected from the judge model pool as the target judge model. Different question attributes have different requirements for the answer. For example, objective questions require accurate answers, subjective questions emphasize creativity and expression, and questions with mixed subjective and objective attributes require consideration of both. Therefore, by selecting at least one judge model matching the question attributes as the target judge model, the subsequent evaluation of the corresponding answer text using the target judge model matching the question attributes of the target question text can more accurately capture the key features of the answer text, thereby improving the evaluation accuracy of the answer text. It should be noted that since the answer text is generated by the target model based on the target question text, this can also be understood as improving the evaluation accuracy of the target model.

[0039] There is no limit to the number of referee models selected from the referee model pool that match the problem attributes; this can be set according to actual usage needs. For example, it can be 1, 2, 5, or all referee models in the pool that match the problem attributes.

[0040] Different judge models in the judge model pool matched with the question attribute have different evaluation angles for the answer. For example, one judge model matched with the question attribute pays more attention to the emotional expression of the answer, and another judge model matched with the question attribute pays more attention to the creativity of the answer. Therefore, by selecting at least two judge models matched with the question attribute from the judge model pool as target judge models, subsequent evaluation of the to-be-evaluated answer text corresponding to the target question text is performed by using at least two target judge models matched with the question attribute of the target question text. Different target judge models evaluate the to-be-evaluated answer text from different evaluation angles, so that the to-be-evaluated answer text can be more comprehensively evaluated. It should be noted that since the to-be-evaluated answer text is generated by the target model based on the target question text, it can also be understood that the target model evaluation is realized from multiple angles and more comprehensively.

[0041] In an embodiment, the question attribute is represented by a question attribute vector. Before selecting at least one judge model matched with the question attribute from the judge model pool as a target judge model, a parameter configuration vector is generated based on the question attribute vector, where the parameter configuration vector is used to guide the selection of the target judge model. At this time, at least one judge model matched with the question attribute is selected from the judge model pool as the target judge model, specifically: based on the parameter configuration vector, at least one judge model matched with the question attribute is selected from the judge model pool as the target judge model. The parameter configuration vector is specifically represented as:

[0042] wherein, represents the parameter configuration vector; K represents a learnable weight matrix; b represents a learnable bias vector; represents a nonlinear activation function; represents element-wise multiplication; represents a question attribute mask, which is a binary vector (composed of 0 and 1) related to the question attribute. For example, for a pure objective question, the dimension related to subjective evaluation in the control signal can be forced to be 0, so as to shield unnecessary control and achieve more accurate control.

[0043] It should be noted that the parameter configuration vector is an auxiliary continuous vector used to finely and task-adaptively configure the hyperparameters (such as the evaluation weight of the judge model) inside the evaluation executor.

[0044] In a specific embodiment, as Figure 2 shown, the model evaluation system includes module one, and module one further includes a feedforward controller (DRL feedforward control-decision maker) that generates a parameter configuration vector based on a question attribute vector.

[0045] In an embodiment, as shown in FIG. 1, the model evaluation system comprises module two, in which at least one adjudication model matching the problem attribute is selected from the adjudication model pool as the target adjudication model by the evaluation executor based on the parameter configuration vector. Figure 2 Figure 2 In an embodiment, the “3-multi-adjudication collaborative evaluation based on dynamic weight” in the above embodiment.

[0046] In an embodiment, the problem attribute is represented by a problem attribute vector; the selection of at least one adjudication model matching the problem attribute from the adjudication model pool as the target adjudication model is performed by a first evaluation executor; before the selection of at least one adjudication model matching the problem attribute from the adjudication model pool as the target adjudication model, a second evaluation executor meeting the evaluation requirement is selected from a plurality of second evaluation executors as the first evaluation executor based on the problem attribute vector; and a parameter configuration vector is generated based on the attribute vector; wherein the parameter configuration vector is used to guide the first evaluation executor to select the target adjudication model; and the parameter configuration vector is routed to the first evaluation executor. That is, based on the problem attribute vector, the optimal evaluation executor is determined, and the parameter configuration information is routed to the optimal evaluation executor.

[0047] In an embodiment, the selection of the second evaluation executor meeting the evaluation requirement from the plurality of second evaluation executors as the first evaluation executor based on the problem attribute vector is specifically: obtaining a system performance vector representing the historical performance of the system; concatenating the system performance vector and the problem attribute vector to obtain a composite state vector; predicting the total long-term reward corresponding to each second evaluation executor based on the composite state vector by using an executor decision model; and selecting the second evaluation executor corresponding to the maximum total long-term reward as the first evaluation executor. Wherein, the composite state vector is specifically represented as:

[0048] Wherein, represents the composite state vector; represents the problem attribute vector; represents vector concatenation or vector splicing; represents the system performance vector representing the historical performance of the system, which is integrated according to the global feedback signal of module four.

[0049] Specifically, the executor decision model outputs a vector based on the composite state vector , wherein each value in the vector represents the predicted total long-term reward or expected long-term return of selecting each second evaluation executor a t in the state.

[0050] ​In an embodiment, the executor decision model adopts a Double DQN architecture to ensure the stability and unbiasedness of the value function learning.

[0051] In an embodiment, the training learning process of the executor decision model is guided by a structured reward function r t The key feature of the guidance is that when module four identifies and corrects systematic biases, a non-sparse, high-amplitude negative reward signal is returned to the agent, establishing a direct attribution link from system performance degradation to front-end decision strategy adjustment, forcing the executor decision model to learn strategies that can avoid such systematic risks.

[0052] Through interaction with the evaluation environment, experience tuples (s t , a t , r t , s t+1 ) are collected, and the Bellman optimality equation of Q-learning is used to update the parameters of the executor decision model. For Double DQN, the calculation of target Q value Y t is as follows:

[0053] wherein, represents the target Q value; represents the reward function; represents the discount factor; represents the next composite state vector; represents the composite state vector all possible second evaluation executors; represents the parameters of the online network; represents the Q value function of the online network; represents the parameters of the target network; represents the Q value function of the target network; represents the maximization parameter, finding the second evaluation executor that maximizes the function value (i.e. ).

[0054] In an embodiment, as shown in Figure 2 , the model evaluation system includes module one, in which the executor decision model selects a second evaluation executor that meets the evaluation requirements from a plurality of second evaluation executors as a first evaluation executor based on a problem attribute vector, and routes a parameter configuration vector to the first evaluation executor (“2-DRL feedforward control-decision” in Figure 2 ).

[0055] Step S14: evaluating the answer text to be evaluated by using each target judging model to obtain a first target evaluation result.

[0056] In this embodiment, the answer text to be evaluated is evaluated by using each target judging model to obtain a first target evaluation result. Each target judging model is a judging model selected from the judging model pool and matched with the question attribute. Therefore, the answer text to be evaluated is evaluated by using the target judging model matched with the question attribute of the target question text, so that the target judging model can more accurately capture the key features of the answer text to be evaluated, thereby improving the evaluation accuracy of the answer text to be evaluated. It should be noted that, since the answer text to be evaluated is generated by the target model based on the target question text, it can also be understood as improving the evaluation accuracy of the target model.

[0057] In addition, in the case where the judging model matched with the question attribute selected from the judging model pool is at least two, the answer text to be evaluated corresponding to the target question text is evaluated by using at least two target judging models matched with the question attribute of the target question text. Different target judging models evaluate the answer text to be evaluated from different evaluation angles, so that the answer text to be evaluated can be more comprehensively evaluated. It should be noted that, since the answer text to be evaluated is generated by the target model based on the target question text, it can also be understood as achieving multi-angle and more comprehensive target model evaluation.

[0058] In an embodiment, the question attribute is subjective or subjective-objective mixed; after the answer text to be evaluated is evaluated by using each target judging model to obtain a first target evaluation result, the first target evaluation result is reviewed in at least one review dimension to obtain a target review result; based on the target review result, the first target evaluation result is adjusted to obtain a third target evaluation result. The review dimension includes but is not limited to user preference, compliance.

[0059] The answer corresponding to the subjective question or the answer corresponding to the subjective-objective mixed question often does not have a clear standard answer, because the answer corresponding to the subjective question or the answer corresponding to the subjective-objective mixed question may involve high-dimensional expression and user preference; therefore, it is usually difficult to completely reflect the importance of human perspective only by the evaluation result of the judging model. By reviewing the first target evaluation result from the perspective of user preference, and adjusting the first target evaluation result based on the review result, the second target evaluation result generated is combined with the evaluation result from the perspective of user preference.

[0060] By simulating human judgment criteria, the emotional inclination, creative expression, and user-friendliness in the answer to be evaluated are captured, ensuring that the second target evaluation result is combined with user preferences. For example, a high-quality answer to be evaluated not only depends on logical correctness, but also needs to show content appeal or interactive adaptability. By further analyzing the first target evaluation result and adjusting the first target evaluation result based on the analysis result, the generated second target evaluation result is combined with the evaluation result from the user preference perspective.

[0061] To generate an evaluation result combined with user preference, refined human preference check items are designed for different task types (such as reasoning, task solving, fact verification, translation, etc.). In the actual evaluation process, these check items are automatically loaded into the review evaluation according to the task type as part of the evaluation criteria. In this way, various task scenarios can be flexibly adapted, and the multi-dimensional requirements of human preferences can be injected into the evaluation process to enhance the applicability of subjective task evaluation.

[0062] Considering that the answers to subjective questions or the answers to mixed subjective and objective questions may involve safety and ethical issues, the first target evaluation result is reviewed from the compliance perspective, and the first target evaluation result is adjusted based on the review result, so that the generated second target evaluation result is combined with the answer compliance perspective evaluation result.

[0063] In a specific embodiment, the first target evaluation result is reviewed by a review model in at least one review dimension to obtain a target review result. The target review result is a structured error signal, and the structured error signal is specifically as follows:

[0064] Wherein, represents the structured error signal; represents the preference judgment result, for example, “Answer A is better than Answer B” or “The first target evaluation result fails to accurately reflect the advantage of Answer A in creativity”, and this judgment is generated by the review model based on its internal user preference criteria; represents the confidence score of the judgment, which is the degree of confidence of the review model in the preference judgment result generated by itself; The error type label selected from the preset error category set is a qualitative error of the review model for the first target evaluation result, for example, it judges that the error type is # fact error # or # lack of creativity #, and this label comes from a preset set and is a “classification” of the review model for the question; represents the error attribution text for judgment, which is the review model's reason for giving the preference judgment result and error type labels The structured error signal provides a rich, interpretable, actionable data foundation for the system's closed-loop calibration.

[0065] In one embodiment, as shown in Figure 2 FIG. 6, the model evaluation system includes Module 3, in which the first target evaluation result is reviewed by a review model in at least one review dimension to obtain a target review result, which is a structured error signal (e.g., "5-Review model correction based on human preference" and "6-Generation of structured error signal" in Figure 2

[0066] In one embodiment, the review of the first target evaluation result in at least one review dimension to obtain a target review result is performed by a review model, and the review model is fine-tuned using sample review results and true review results. The sample review results are obtained by reviewing sample evaluation results in at least one review dimension by the review model, and the sample evaluation results are obtained by the model evaluation model based on sample answer texts and corresponding sample question texts. That is, the sample evaluation results and the true evaluation results constitute a training set for the review model, and the model parameters of the review model are updated using the training set to fine-tune the review model.

[0067] Specifically, the loss of the review model is determined based on the difference between the sample review results and the true review results using a loss function, and the model parameters of the review model are updated by minimizing the loss of the review model. The updating process of the model parameters of the review model is specifically formalized as:

[0068] wherein, represents the model parameters of the review model; E represents the structured error signal; represents the loss function, which measures the difference between the sample review results and the true review results; represents the partial derivative of the loss function with respect to the model parameters , which points to the direction in the model parameter space that makes the loss decrease the fastest; represents the learning rate, which controls the step size of the model parameter update (e.g., = 0.001), to avoid too large or too small update steps.

[0069] ​​To reduce the computational resource consumption, in an embodiment, the current model parameters of the review model are frozen, and a low-rank adapter (LoRA) is added to the network layers of the review model, and only the low-rank adapter is fine-tuned. Wherein, the formula is as follows:

[0070] Wherein, W0 represents the current model parameters of the review model, specifically the weight matrix of the current review model; represents the low-rank adapter; B and A represent the low-rank adaptive weight matrix of ΔW.

[0071] The current network parameters or original parameters of the review model are kept unchanged, and the low-rank adapter is added to the review model. That is, without directly modifying the original parameters of the review model, the low-rank adapter is introduced to adjust the review model, on the one hand, the review ability of the review model is retained; on the other hand, the model performance of the review model is improved. In addition, the current network parameters or original parameters of the review model are unchanged and do not participate in parameter adjustment, and only the low-rank adapter is adjusted subsequently, which reduces the training parameter amount and avoids the high computational cost of full parameter adjustment.

[0072] In an embodiment, multiple branch adapters can be trained for different error types, and dynamically called through a gating routing mechanism, which reduces the memory overhead and improves the pertinence of correction. An adversarial robust term is introduced during training to ensure that the output of the review model is stable and consistent under synonymous or perturbation input.

[0073] In an embodiment, as shown in Figure 2 , the model evaluation system includes module four, in which the review model is fine-tuned based on the structured error signal output by the review model (7a-knowledge editing and correction fine-tune review model in Figure 2 ).

[0074] In an embodiment, as shown in Figure 2 , the model evaluation system includes module four, in which the systematic bias integrator performs systematic bias integral adjustment (7b-systematic bias integral in Figure 2 ), and the systematic bias integrator maintains the systematic bias vector , each component of the systematic bias vector corresponds to a type of error. The systematic bias vector is iteratively updated according to the received structured error signal:

[0075] Wherein, represents the systematic bias vector at time t; Represents the systematic deviation vector at time t+1; Indicates the forgetting factor; This represents a function that converts the error type label L in the current structured error signal into a one-hot vector. It should be noted that this iterative update process calculates the exponential moving average of various error types to identify persistent systematic biases.

[0076] In one specific implementation, the actuator decision model is fine-tuned using the sample review results. Specifically, Figure 3 As shown, the model evaluation system includes module four. In module four, when the systematic deviation vector... any component b i Exceeding the preset threshold τ i At that time, a global feedback signal G is generated. Figure 3 (8-Generate a high-amplitude negative reward for the global feedback signal). The global feedback signal G is transmitted in reverse to Module 1. If the decision core of Module 1 is a reinforcement actuator decision model, then the global feedback signal G can serve as a huge negative reward value r. neg This is used to update the Q-value function of the executor decision model, thereby forcing it to adjust the diversion strategy from the source of the decision.

[0077] Please see Figure 1 , Figure 3 yes Figure 3 The diagram shows a flowchart of one embodiment of step S14. It should be noted that if substantially the same result is achieved, this embodiment does not necessarily follow the same pattern. Figure 2 The illustrated process sequence is limited. For example... Figure 2 As shown, there are at least two target referee models, and this embodiment includes: Step S31: Assign the first evaluation weight to each target referee model.

[0078] In this implementation, a first evaluation weight is assigned to each target referee model.

[0079] In one embodiment, assigning a first evaluation weight to each target referee model specifically involves: obtaining the historical performance and parameter configuration vector of each target referee model; and determining the first evaluation weight of each target referee model based on its historical performance and parameter configuration vector. The process of determining the first evaluation weight of each target referee model based on its historical performance and parameter configuration vector is formally expressed as follows:

[0080] in, This represents the first evaluation weight of the j-th target referee model; Denotes the j-th target referee model; a parameter configuration vector; a historical performance matrix representing the historical performance of the jth target adjudication model.

[0081] The historical performance of the target adjudication model directly reflects its performance in past answer evaluation. Using the historical performance as the basis for weight distribution can ensure that the target adjudication model with better performance plays a greater role in the evaluation of the answer text to be evaluated, thereby improving the evaluation accuracy of the answer text to be evaluated.

[0082] In a specific embodiment, as shown in Figure 2 The model evaluation system includes module two, in which the evaluation executor determines the first evaluation weight of each target adjudication model based on the historical performance of each target adjudication model and the parameter configuration vector. Figure 2 “3-Adjudication Coordination Evaluation Based on Dynamic Weighting”).

[0083] Step S32: Using each target adjudication model to evaluate the answer text to be evaluated to obtain the second target evaluation result corresponding to each target adjudication model.

[0084] In this embodiment, each target adjudication model is used to evaluate the answer text to be evaluated to obtain the second target evaluation result corresponding to each target adjudication model. Different target adjudication models have different evaluation angles for the answer text to be evaluated. Therefore, using each target adjudication model to evaluate the answer text to be evaluated to obtain the second target evaluation result corresponding to each target adjudication model can evaluate the answer text to be evaluated from different evaluation angles to obtain evaluation results from different evaluation angles.

[0085] Step S33: Based on the first evaluation weight corresponding to each target adjudication model, the second target evaluation result is weighted and fused to obtain the first target evaluation result.

[0086] In this embodiment, the second target evaluation results are weighted and fused based on the first evaluation weights corresponding to the target judgment models to obtain the first target evaluation result. The second target evaluation results are evaluation results obtained by evaluating the answer text to be evaluated from different evaluation angles. Therefore, the second target evaluation results are weighted and fused to integrate the evaluation results of different evaluation angles of the answer text to be evaluated, which can comprehensively evaluate the answer text to be evaluated from multiple evaluation angles and avoid one-sidedness of a single evaluation angle, so that a more comprehensive first target evaluation result is obtained, and more comprehensive evaluation of the answer text to be evaluated is realized. In addition, the second target evaluation results are weighted and fused based on the first evaluation weights corresponding to the target judgment models, which can ensure that the more optimal target judgment model plays a greater role in the final evaluation result, thereby improving the evaluation accuracy of the answer text to be evaluated. Therefore, the first target evaluation result obtained by evaluating the answer text to be evaluated is accurate and comprehensive, and more comprehensive and accurate evaluation of the answer text to be evaluated is improved.

[0087] In an embodiment, as shown in FIG. 1, Figure 4 the model evaluation system includes module two, in which the answer text to be evaluated is evaluated by using the target judgment models to obtain the second target evaluation results corresponding to the target judgment models, and the second target evaluation results are weighted and fused based on the first evaluation weights corresponding to the target judgment models to obtain the first target evaluation result (as shown in Figure 4 “3-Multi-Judge Cooperative Evaluation Based on Dynamic Weight”).

[0088] In an embodiment, the problem attribute is a mixed attribute of subjectivity and objectivity. Before selecting at least one judgment model matching the problem attribute from the judgment model pool as a target judgment model, the answer text to be evaluated is decomposed to obtain a first sub-answer text corresponding to an objective attribute and a second sub-answer text corresponding to a subjective attribute. At this time, at least one judgment model matching the problem attribute is selected from the judgment model pool as a target judgment model, specifically: at least one judgment model matching the objective attribute is selected from the judgment model pool as a first target model; and at least one judgment model matching the subjective attribute is selected from the judgment model pool as a second target model.

[0089] The requirements for answers of objective attribute questions and subjective attribute questions are different, for example, objective attribute questions require accurate answers, subjective attribute questions value the creativity and expression of answers more, and mixed attribute questions require both. Therefore, the mixed attribute question corresponding to the to-be-evaluated answer text is divided into a first sub-answer text corresponding to the objective attribute and a second sub-answer text corresponding to the subjective attribute, and at least one adjudication model matched with each sub-answer text is selected from the adjudication model pool, so that the subsequent evaluation of each sub-answer text by the adjudication model matched with each sub-answer text can more accurately capture the key features of the sub-answer text, thereby improving the evaluation accuracy of the sub-answer text and further improving the evaluation accuracy of the to-be-evaluated answer text.

[0090] Further, the first target evaluation result is obtained by weighted fusion of each second target evaluation result based on the first evaluation weight corresponding to each target adjudication model, specifically: the second result is obtained by weighted fusion of each first result based on the first evaluation weight corresponding to each first target model; wherein each first result is obtained by evaluation of the first sub-answer text by each first target model; and the fourth result is obtained by weighted fusion of each third result based on the first evaluation weight corresponding to each second target model; wherein each third result is obtained by evaluation of the second sub-answer text by each second target model; and the first target evaluation result is obtained based on the second result and the fourth result.

[0091] Each first result is an evaluation result obtained by evaluating the first sub-answer text from different evaluation angles; therefore, the weighted fusion of each first result is the fusion of evaluation results of different evaluation angles of the first sub-answer text, which can integrate evaluation results of multiple evaluation angles and avoid one-sidedness of a single evaluation angle, thereby obtaining a more comprehensive second result and achieving more comprehensive evaluation of the first sub-answer text. In addition, the weighted fusion of each first result based on the first evaluation weight corresponding to each first target model can ensure that better first target models play a greater role in the final evaluation result, thereby improving the evaluation accuracy of the first sub-answer text. Therefore, the second result obtained by evaluating the first sub-answer text is accurate and comprehensive, and the more comprehensive and accurate evaluation of the first sub-answer text is improved.

[0092] The third result is a result of evaluating the second sub-answer text from different evaluation angles. Therefore, the weighted fusion of the third result is a result of fusing the evaluation results of different evaluation angles of the second sub-answer text, which can comprehensively evaluate the evaluation results of multiple evaluation angles and avoid one-sidedness of a single evaluation angle, so as to obtain a more comprehensive fourth result and realize more comprehensive evaluation of the second sub-answer text. In addition, the weighted fusion of the third result based on the first evaluation weight corresponding to each second target model can ensure that a more optimal second target model plays a greater role in the final evaluation result, thereby improving the evaluation accuracy of the second sub-answer text. Therefore, the fourth result obtained by evaluating the second sub-answer text is accurate and comprehensive, and the more comprehensive and accurate evaluation of the second sub-answer text is improved.

[0093] Since the second result obtained by evaluating the first sub-answer text is accurate and comprehensive, and the fourth result obtained by evaluating the second sub-answer text is accurate and comprehensive, the first target evaluation result obtained based on the second result and the fourth result is accurate and comprehensive, that is, the evaluation of the to-be-evaluated answer text corresponding to the target problem text of the subjective and objective mixed nature is accurate and comprehensive.

[0094] In a specific embodiment, the first target evaluation result is obtained based on the second result and the fourth result, specifically: determining the second evaluation weight corresponding to the objective attribute and the subjective attribute respectively; based on the second evaluation weight corresponding to the objective attribute and the subjective attribute, the second result and the fourth result are weighted and fused to obtain the first target evaluation result. Wherein, the specific formula is as follows:

[0095] Wherein, represents the first target evaluation result; represents the second result; represents the fourth result; represents the parameter configuration vector; represents a neural network with attention gate mechanism, which can adaptively adjust the second evaluation weight corresponding to the objective attribute and the subjective attribute according to the parameter configuration vector.

[0096] Please refer to Figure 5 , Figure 5is a structural schematic diagram of an embodiment of the model evaluation device provided in the present application. The model evaluation device 40 comprises an acquisition module 41, a determination module 42, a selection module 43 and an evaluation module 44; the acquisition module 41 is configured to acquire a to-be-evaluated answer text and a corresponding target question text; wherein the to-be-evaluated answer text is generated by a target model based on the target question text; the determination module 42 is configured to determine a question attribute of the target question text; the selection module 43 is configured to select at least one adjudication model matching the question attribute from a pool of adjudication models as a target adjudication model; and the evaluation module 44 is configured to evaluate the to-be-evaluated answer text by using each target adjudication model to obtain a first target evaluation result.

[0097] wherein the target adjudication model is at least two; the evaluation module 44 is configured to evaluate the to-be-evaluated answer text by using each target adjudication model to obtain a first target evaluation result, comprising: assigning a first evaluation weight to each target adjudication model; and evaluating the to-be-evaluated answer text by using each target adjudication model to obtain a second target evaluation result corresponding to each target adjudication model; and performing weighted fusion on each second target evaluation result based on the first evaluation weight corresponding to each target adjudication model to obtain the first target evaluation result.

[0098] wherein the question attribute is a mixed nature of subjectivity and objectivity; before the selection module 43 selects at least one adjudication model matching the question attribute from the pool of adjudication models as a target adjudication model, the selection module 43 is configured to decompose the to-be-evaluated answer text to obtain a first sub-answer text corresponding to an objective attribute and a second sub-answer text corresponding to a subjective attribute; the selection module 43 is configured to select at least one adjudication model matching the objective attribute from the pool of adjudication models as a first target model, and select at least one adjudication model matching the subjective attribute from the pool of adjudication models as a second target model; and the evaluation module 44 is configured to perform weighted fusion on each second target evaluation result based on the first evaluation weight corresponding to each target adjudication model to obtain the first target evaluation result, comprising: performing weighted fusion on each first result based on the first evaluation weight corresponding to each first target model to obtain a second result; wherein each first result is obtained by evaluating the first sub-answer text by using each first target model; and performing weighted fusion on each third result based on the first evaluation weight corresponding to each second target model to obtain a fourth result; wherein each third result is obtained by evaluating the second sub-answer text by using each second target model; and obtaining the first target evaluation result based on the second result and the fourth result.

[0099] The evaluation module 44 is configured to obtain a first target evaluation result based on the second result and the fourth result, including: determining second evaluation weights corresponding to the objective attribute and the subjective attribute respectively; and performing weighted fusion on the second result and the fourth result based on the second evaluation weights corresponding to the objective attribute and the subjective attribute, to obtain the first target evaluation result.

[0100] The evaluation module 44 is configured to assign a first evaluation weight to each target adjudication model, including: obtaining a model historical performance and a parameter configuration vector of each target adjudication model; and determining the first evaluation weight of each target adjudication model based on the model historical performance and the parameter configuration vector of each target adjudication model.

[0101] The problem attribute is represented by a problem attribute vector, and at least one adjudication model matched with the problem attribute is selected from the adjudication model pool as a target adjudication model by the first evaluation executor; the model evaluation device 40 further includes a routing module 45, which is configured to, before selecting at least one adjudication model matched with the problem attribute from the adjudication model pool as a target adjudication model, include: selecting a second evaluation executor meeting the evaluation requirement as the first evaluation executor from a plurality of second evaluation executors based on the problem attribute vector; and generating a parameter configuration vector based on the problem attribute vector; wherein the parameter configuration vector is used to guide the first evaluation executor to select the target adjudication model; and routing the parameter configuration vector to the first evaluation executor.

[0102] The routing module 45 is configured to select a second evaluation executor meeting the evaluation requirement as the first evaluation executor from a plurality of second evaluation executors based on the problem attribute vector, including: obtaining a system performance vector representing a system historical performance; splicing the system performance vector and the problem attribute vector to obtain a composite state vector; predicting a total long-term reward corresponding to each second evaluation executor based on the composite state vector by using an executor decision model; and selecting a second evaluation executor corresponding to the maximum total long-term reward as the first evaluation executor.

[0103] The problem attribute is subjective or subjective and objective mixed; the model evaluation device 40 further includes a review module 46, which is configured to, after obtaining the first target evaluation result by using each target adjudication model to evaluate the answer text to be evaluated, include: performing review on the first target evaluation result in at least one review dimension to obtain a target review result; and adjusting the first target evaluation result based on the target review result to obtain a third target evaluation result.

[0104] The at least one review dimension includes at least one of the following: compliance, user preference.

[0105] The review of the first target evaluation result in at least one review dimension is performed by using a review model, at least one adjudication model matched with the problem attribute is selected from a pool of adjudication models as a target adjudication model by a first evaluation executor, the first evaluation executor is selected from a plurality of second evaluation executors by using an executor decision model, and the model evaluation apparatus further comprises a training module configured to perform at least one of the following steps: fine-tuning the review model by using a sample review result and a true review result, wherein the sample review result is obtained by reviewing a sample evaluation result in at least one review dimension by using the review model, and the sample evaluation result is obtained by evaluating a sample answer text and a corresponding sample question text based on a model evaluation model; and fine-tuning the executor decision model by using the sample review result.

[0106] Please refer to Figure 6 , Figure 6 is a structural schematic diagram of an embodiment of an electronic device provided in the application. The electronic device 50 comprises a memory 51 and a processor 52 coupled with each other. The processor 52 is configured to execute program instructions stored in the memory 51 to implement the steps of any of the model evaluation method embodiments described above. In a specific implementation scenario, the electronic device 50 can include but is not limited to a microcomputer, a server, and in addition, the electronic device 50 can also include a notebook computer, a tablet computer and other mobile devices, which are not limited herein.

[0107] Specifically, the processor 52 is configured to control itself and the memory 51 to implement the steps of any of the model evaluation method embodiments described above. The processor 52 can also be referred to as a CPU (Central Processing Unit). The processor 52 can be an integrated circuit chip having a processing capability of signals. The processor 52 can also be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. In addition, the processor 52 can be jointly implemented by integrated circuit chips.

[0108] Please refer to ​ , ​is a structural schematic diagram of an embodiment of the computer readable storage medium provided in the application. The computer readable storage medium 60 of the embodiment of the application stores program instructions 61, which, when executed, implement the method provided by any embodiment of the model evaluation method of the application and any non-conflicting combination. Among them, the program instructions 61 can form a program file and be stored in the above computer readable storage medium 60 in the form of a software product, so that a computer device (which can be a personal computer, a server, or a network device, etc.) executes all or part of the steps of the method of each embodiment of the application. The aforementioned computer readable storage medium 60 includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes, or a computer, a server, a mobile phone, a tablet, and other terminal devices.

[0109] If the technical solution of the application involves personal information, the product applying the technical solution of the application has been explicitly informed of the personal information processing rules before processing the personal information, and has obtained the personal independent consent. If the technical solution of the application involves sensitive personal information, the product applying the technical solution of the application has obtained the personal independent consent before processing the sensitive personal information, and at the same time meets the requirement of "explicit consent". For example, at the personal information collection device such as camera, a clear and prominent mark is set to inform that the personal information collection range has been entered, and the personal information will be collected. If the individual voluntarily enters the collection range, it is regarded as agreeing to collect the personal information; or on the device for processing personal information, through the pop-up information or by uploading the personal information by the individual, the personal authorization is obtained under the condition that the personal information processing rules are informed by obvious mark / information. The personal information processing rules can include personal information processor, personal information processing purpose, processing method, and personal information type, etc.

[0110] The above is only an embodiment of the application, and does not limit the patent scope of the application. Any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the application.

Claims

1. A model evaluation method, characterized in that, The method includes: Obtain the text of the answer to be evaluated and the corresponding target question text; wherein, the text of the answer to be evaluated is generated by the target model based on the target question text; Determine the problem attributes of the target problem text; From the pool of judge models, select at least one judge model that matches the attribute of the problem as the target judge model; The target judge model is used to evaluate the answer text to be evaluated, and the first target evaluation result is obtained.

2. The method according to claim 1, characterized in that, The target judge model is at least two; the step of using each of the target judge models to evaluate the answer text to be evaluated, and obtaining a first target evaluation result, includes: Assign a first evaluation weight to each of the target referee models; Furthermore, the target judge model is used to evaluate the answer text to be evaluated, and the second target evaluation result corresponding to each target judge model is obtained; Based on the first evaluation weights corresponding to each of the target judge models, the evaluation results of each of the second targets are weighted and fused to obtain the evaluation result of the first target.

3. The method according to claim 2, characterized in that, The problem attribute is a mixture of subjective and objective; before selecting at least one judge model matching the problem attribute from the judge model pool as the target judge model, the method further includes: The text of the answer to be evaluated is decomposed to obtain a first sub-answer text corresponding to the objective attribute and a second sub-answer text corresponding to the subjective attribute; The step of selecting at least one judge model from the judge model pool that matches the problem attribute as the target judge model includes: From the pool of judge models, at least one judge model that matches the objective attribute is selected as the first target model; And, from the pool of judge models, at least one judge model that matches the subjective attribute is selected as the second target model; The weighted fusion of the evaluation results of each second target based on the first evaluation weight corresponding to each target referee model to obtain the evaluation result of the first target includes: Based on the first evaluation weights corresponding to each first target model, the first results are weighted and fused to obtain the second result; wherein, each first result is obtained by evaluating the first sub-answer text by each first target model. Furthermore, based on the first evaluation weights corresponding to each of the second target models, each of the third results is weighted and fused to obtain a fourth result; wherein, each of the third results is obtained by evaluating the second sub-answer text by each of the second target models; Based on the second result and the fourth result, the evaluation result of the first target is obtained.

4. The method according to claim 3, characterized in that, The process of obtaining the first target evaluation result based on the second result and the fourth result includes: Determine the second evaluation weights corresponding to the objective attribute and the subjective attribute, respectively; Based on the second evaluation weights corresponding to the objective attributes and the subjective attributes, the second result and the fourth result are weighted and fused to obtain the first target evaluation result.

5. The method according to claim 2, characterized in that, Assigning first evaluation weights to each of the target referee models includes: Obtain the historical performance and parameter configuration vectors of each target referee model; Based on the historical performance and parameter configuration vector of each target referee model, the first evaluation weight of each target referee model is determined.

6. The method according to claim 1, characterized in that, The problem attribute is represented by a problem attribute vector; the step of selecting at least one judge model matching the problem attribute from the judge model pool as the target judge model is executed by the first evaluation executor; before selecting at least one judge model matching the problem attribute from the judge model pool as the target judge model, the method further includes: Based on the problem attribute vector, a second evaluation executor that meets the evaluation requirements is selected from several second evaluation executors and used as the first evaluation executor; Furthermore, a parameter configuration vector is generated based on the problem attribute vector; wherein the parameter configuration vector is used to guide the first evaluation executor in selecting the target referee model; The parameter configuration vector is routed to the first evaluation executor.

7. The method according to claim 6, characterized in that, The step of selecting a second evaluation executor that meets the evaluation requirements from a plurality of second evaluation executors based on the problem attribute vector, and using it as the first evaluation executor, includes: Obtain the system performance vector that characterizes the historical performance of the system; The system performance vector and the problem attribute vector are concatenated to obtain a composite state vector. The total long-term reward corresponding to each of the second evaluation executors is predicted using the executor decision model based on the composite state vector. The second evaluation executor corresponding to the maximum total long-term reward is used as the first evaluation executor.

8. The method according to claim 1, characterized in that, The question attribute is subjective or a mixture of subjective and objective; after evaluating the answer text to be evaluated using each of the target judge models to obtain the first target evaluation result, the method further includes: The evaluation results of the first target are reviewed in at least one dimension to obtain the target review results; Based on the review results of the target, the evaluation results of the first target are adjusted to obtain the evaluation results of the third target.

9. The method according to claim 8, characterized in that, The at least one review dimension includes at least one of the following: compliance, user preference.

10. The method according to claim 8, characterized in that, The review of the first target evaluation result across at least one review dimension to obtain the target review result is performed using a review model. The selection of at least one referee model matching the problem attribute from the referee model pool, as the target referee model, is performed by a first evaluation executor. The first evaluation executor is selected from several second evaluation executors using an executor decision model. The method includes at least one of the following steps: The review model is fine-tuned using sample review results and real review results; wherein, the sample review results are obtained by the review model reviewing the sample evaluation results in at least one review dimension, and the sample evaluation results are obtained by the model evaluation model based on the sample answer text and the corresponding sample question text; The actuator decision model is fine-tuned using the results of the sample review.

11. A model evaluation device, characterized in that, The device includes: The acquisition module is used to acquire the text of the answer to be evaluated and the corresponding target question text; wherein, the text of the answer to be evaluated is generated by the target model based on the target question text; The determination module is used to determine the problem attributes of the target problem text; The selection module is used to select at least one referee model from the referee model pool that matches the problem attribute as the target referee model. The evaluation module is used to evaluate the answer text to be evaluated using the target judge models to obtain the first target evaluation result.

12. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory being used to store program instructions, and the processor being used to execute the program instructions to implement the model evaluation method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program instructions that can be executed to implement the model evaluation method as described in any one of claims 1-10.

Images