Correction methods, devices, storage media and electronic equipment

Abstract

This application provides a grading method, apparatus, storage medium, and electronic device, relating to the field of artificial intelligence. The method includes: acquiring a question stem text, and corresponding answer text and response text; based on the question stem text, and corresponding answer text and response text, using a grading model to generate grading text corresponding to the response text, the grading model is trained based on a response text recognition model, which is used to recognize the response text corresponding to the question stem text. The grading model in this application is an end-to-end model, realizing one-stop, full-process grading of response results. Furthermore, since the grading model is based on the response text recognition model, it can improve the recognition accuracy of response results and further improve the accuracy of grading.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a method, apparatus, storage medium, and electronic device for editing. Background Technology

[0002] Currently, there are generally two methods for homework correction. One method involves teachers manually correcting homework directly. This method is labor-intensive and inefficient. The other method uses neural network models for intelligent correction. This method typically involves first photographing the question text, then using a neural network model to obtain the corresponding correct answer. Next, it photographs the handwritten portion of the answer area and uses the neural network model to compare the similarity between the handwritten portion and the correct answer to obtain the correction result. Similarly, this method is inefficient, and the accuracy needs improvement. Summary of the Invention

[0003] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method, apparatus, storage medium, and electronic device for making corrections.

[0004] In a first aspect, one embodiment of this application provides a grading method, including: obtaining a question stem text, and answer text and response text corresponding to the question stem text; based on the question stem text, and answer text and response text corresponding to the question stem text, using a grading model to generate grading text corresponding to the response text, wherein the grading model is trained based on a response text recognition model, and the response text recognition model is used to recognize the response text corresponding to the question stem text.

[0005] In conjunction with the first aspect, in some implementations of the first aspect, the grading model is trained based on the answer text recognition model, including: obtaining the network parameters of the answer text recognition model; constructing the network parameters of the grading model to be trained based on the network parameters of the answer text recognition model; obtaining a training sample set, which includes question stem samples, and answer samples, answer samples and grading samples corresponding to the question stem samples; and training the constructed grading model to be trained based on the training sample set to obtain the grading model.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the answer text recognition model includes a feature extraction network, the training grading model includes a feature extraction network, and the network parameters of the training grading model are constructed based on the network parameters of the answer text recognition model, including: constructing the network parameters of the feature extraction network included in the training grading model based on the network parameters of the feature extraction network included in the answer text recognition model, and the network parameters of the feature extraction network included in the constructed training grading model include the network parameters of the feature extraction network included in the answer text recognition model.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the constructed grading model is trained based on the training sample set to obtain the grading model, including: constructing the loss function of the grading model based on the loss function of the answer text recognition model; determining the text fusion features corresponding to the question stem sample, the answer sample, the answer sample, and the grading sample based on the training sample set; and training the grading model using the text fusion features and the loss function of the grading model to obtain the grading model.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, based on the training sample set, the text fusion features corresponding to the question stem sample, the answer sample, the response sample, and the grading sample are determined, including: based on the question stem sample, determining the semantic features corresponding to the question stem sample; based on the answer sample, determining the semantic features corresponding to the answer sample; based on the response sample, determining the semantic and visual features corresponding to the response sample; based on the grading sample, determining the semantic features corresponding to the grading sample; and performing feature fusion on the semantic features corresponding to the question stem sample, the answer sample, the response sample, and the grading sample, as well as the visual features corresponding to the response sample, to obtain the text fusion features.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, based on the question stem text, and the corresponding answer text and response text, a grading model is used to generate the grading text corresponding to the response text. This includes: based on the question stem text, and the corresponding answer text and response text, determining the semantic features corresponding to the question stem text, the semantic features corresponding to the answer text, the semantic features corresponding to the response text, and the visual features corresponding to the response text; and based on the semantic features corresponding to the question stem text, the answer text, and the response text, and the visual features corresponding to the response text, a grading model is used to generate the grading text corresponding to the response text.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, based on the semantic features corresponding to the question stem text, answer text, and response text, as well as the visual features corresponding to the response text, a grading model is used to generate the grading text corresponding to the response text. This includes: performing feature fusion on the semantic features corresponding to the question stem text, answer text, and response text, as well as the visual features corresponding to the response text, to obtain text fusion features; and using the grading model based on the text fusion features to generate the grading text corresponding to the response text.

[0011] Secondly, one embodiment of this application provides a grading device, including: an acquisition module for acquiring a question stem text, and answer text and response text corresponding to the question stem text; and a grading module for generating grading text corresponding to the response text based on the question stem text, and answer text and response text corresponding to the question stem text, using a grading model, wherein the grading model is trained based on a response text recognition model, and the response text recognition model is used to recognize the response text corresponding to the question stem text.

[0012] Thirdly, one embodiment of this application provides a computer-readable storage medium storing a computer program for performing the method described in the first aspect.

[0013] Fourthly, one embodiment of this application provides an electronic device, the electronic device comprising: a processor; a memory for storing processor-executable instructions; the processor being configured to perform the method described in the first aspect.

[0014] The training method for the correction model provided in this application has the following beneficial effects.

[0015] On the one hand, the grading model in this application is an end-to-end, one-stop, full-process model. Using this grading model, intelligent grading of response texts can be achieved in a simpler and more efficient way. On the other hand, the response text recognition model is mainly used to recognize response texts. Using the response text recognition model to train the grading model allows the model to focus more on the relevant features of the response samples, thereby improving the recognition accuracy of the model and further enhancing the grading accuracy of the grading model. Attached Figure Description

[0016] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 The diagram shown is a schematic representation of the implementation environment of the correction method provided in an exemplary embodiment of this application.

[0018] Figure 2 The above is a flowchart illustrating an exemplary embodiment of the correction method provided in this application.

[0019] Figure 3 The diagram shown is a flowchart illustrating the training method of the correction model provided in an exemplary embodiment of this application.

[0020] Figure 4 The diagram shown is a schematic diagram of training a model to be trained and graded according to an exemplary embodiment of this application.

[0021] Figure 5 The diagram shown is a schematic flowchart of obtaining a modified model provided by an exemplary embodiment of this application.

[0022] Figure 6 The diagram shown is a schematic diagram of obtaining text fusion features provided by an exemplary embodiment of this application.

[0023] Figure 7 The diagram shown is a structural schematic of an exemplary embodiment of the present application for a correction device.

[0024] Figure 8 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Exemplary scenario

[0027] Figure 1 The diagram shown illustrates the implementation environment of the correction method provided in an exemplary embodiment of this application. Figure 1 As shown, this implementation environment includes a user terminal 11 and a server 12, which are connected for communication. The server 12 can be a standalone physical server, a server cluster consisting of multiple servers, or a cloud server capable of cloud computing. Furthermore, the server can be considered a server for a specific business (e.g., homework grading). In addition, the server can be a physical machine or a virtual machine, and the number can be one or more. This application embodiment does not limit the type or number of servers.

[0028] Specifically, the user terminal first obtains the question stem text to be graded, as well as the corresponding answer text and response text, and inputs them into the user terminal 11. Further, the user terminal 11 sends the question stem text, answer text, and response text to the server 12. The server 12 is equipped with a grading model. Based on the grading model, the server grades the response text and answer text corresponding to the question stem text, obtains the grading result, and sends the grading result back to the user terminal 11 so that the user can visually view the responses of the relevant respondents through the user terminal 11.

[0029] For example, after a school's final exams, the workload of grading exam papers is large, and relying solely on teachers to grade them is inefficient. In this case, image acquisition devices (such as mobile phones, cameras, or scanners) can be used to capture images of the exam papers. Furthermore, the image acquisition device can wirelessly communicate with the user terminal 11, directly sending the captured exam paper images to the user terminal 11. Simultaneously, the user terminal 11 obtains the answer text corresponding to each question in the exam paper. Using a pre-installed preprocessing application, the user terminal 11 maps the question text, answer text, and reply text for each student's exam paper. Further, the user terminal 11 sends the question text, along with the corresponding answer text and reply text, for each student's exam paper to the server 12. The server 12, based on its deployed grading model, grades the students' answers and sends the grading results back to the user terminal 11. User terminal 11 can use a pre-installed post-processing application to match each question text and answer text for the same test paper, and finally display the grading results of each question text in the form of a test paper image.

[0030] Exemplary methods

[0031] Figure 2 The diagram shown is a flowchart illustrating an exemplary embodiment of the correction method provided in this application. Figure 2 As shown in the embodiments of this application, the correction method includes the following steps.

[0032] Step S210: Obtain the question stem text, as well as the corresponding answer text and response text.

[0033] Specifically, the system can acquire a single question stem text, along with the corresponding answer text and response text. Alternatively, it can acquire a captured image to be graded, containing at least one question stem text, the response text for the corresponding answer area, and the corresponding answer text. The image to be graded is then input into a preprocessing network model to obtain the response text and answer text for each question stem text. Furthermore, the question stem text can be from any subject; for example, it could be a multiple-choice question, fill-in-the-blank question, or problem-solving question in mathematics, or a word formation question or a classical poem recitation question in language arts.

[0034] Step S220: Based on the question stem text, the corresponding answer text, and the response text, generate the grading text corresponding to the response text using the grading model.

[0035] The grading model is trained based on the answer text recognition model, which is used to identify the answer text corresponding to the question stem.

[0036] For example, based on the question stem text, the corresponding answer text, and the response text, the semantic features corresponding to the question stem text, the semantic features corresponding to the answer text, the semantic features corresponding to the response text, and the visual features corresponding to the response text are determined respectively. Further, the semantic features corresponding to the question stem text, the answer text, and the response text, as well as the visual features corresponding to the response text, are fused to obtain text fusion features. Finally, based on the text fusion features, the grading model is used to generate the grading text corresponding to the response text.

[0037] The grading model in this application is an end-to-end, one-stop, full-process model. Using this grading model, intelligent grading of response texts can be achieved in a simpler and more efficient way. On the other hand, the response text recognition model is mainly used to recognize the response text. Using the response text recognition model to train the grading model allows the model to pay more attention to the relevant features of the response samples, thereby improving the recognition accuracy of the model and further enhancing the grading accuracy of the grading model.

[0038] Figure 3 The diagram shown is a flowchart illustrating a training method for a correction model provided in an exemplary embodiment of this application. Figure 3 As shown in the embodiments of this application, the training method for the correction model includes the following steps.

[0039] Step S310: Obtain the network parameters of the response text recognition model.

[0040] The answer text recognition model is used to identify the answer text corresponding to the question stem. For example, the network structure of the answer text recognition model includes an encoder and a decoder. The training process of the answer text recognition model includes: acquiring a large number of answer samples and setting the loss function of the answer text recognition model to be trained; inputting the answer samples into the answer text recognition model to be trained; adjusting the network parameters of the answer text recognition model based on the loss value obtained from the loss function of the answer text recognition model to be trained until the training conditions are met, thus obtaining the trained answer text recognition model; further, obtaining the network parameters of the answer text recognition model based on the trained answer text recognition model.

[0041] Step S320: Based on the network parameters of the response text recognition model, construct the network parameters of the model to be trained for grading.

[0042] Specifically, firstly, based on the network structure of the response text recognition model, the network structure of the model to be trained and graded is constructed. Then, according to the network parameters of the response text recognition model, the corresponding network parameters are set on the network structure of the model to be trained and graded.

[0043] Step S330: Obtain the training sample set, which includes question stem samples, as well as answer samples, response samples, and grading samples corresponding to the question stem samples.

[0044] The answer samples in this step and the answer samples corresponding to the text recognition model to be trained in step S310 can be the same or different. This application embodiment does not limit the type of question stem samples in the training sample set; they can be question stem samples from mathematics subjects or question stem samples from physics subjects.

[0045] Step S340: Based on the training sample set, train the constructed correction model to obtain the correction model.

[0046] Specifically, the grading model is used to grade the answer text corresponding to the question stem text. The grading model is an end-to-end model. The training sample set is input into the grading model to be trained, and the grading model to be trained after the network parameters are constructed is repeatedly trained to obtain the grading model.

[0047] In this application embodiment, on the one hand, the grading model trained using the training sample set is an end-to-end, one-stop, full-process model. Using the grading model in this application, intelligent grading of answer texts can be achieved in a simpler and more efficient way. On the other hand, the answer text recognition model is mainly used to recognize answer texts. By using the network parameters of the answer text recognition model to construct the network parameters of the grading model to be trained, the grading model to be trained can pay more attention to the relevant features of the answer samples, thereby improving the recognition accuracy of the answer samples and further improving the grading accuracy of the grading model.

[0048] In an exemplary embodiment of this application, the response text recognition model includes a feature extraction network, and the model to be trained for grading also includes a feature extraction network. The network parameters of the model to be trained for grading are constructed based on the network parameters of the response text recognition model, including: constructing the network parameters of the feature extraction network included in the model to be trained for grading based on the network parameters of the feature extraction network included in the response text recognition model. The network parameters of the feature extraction network included in the constructed model to be trained for grading include the network parameters of the feature extraction network included in the response text recognition model.

[0049] For example, the model structure of the response text recognition model and the model to be trained and graded is a graph neural network (transformer), and the feature extraction network is an encoder. In addition, the response text recognition model also includes a decoder.

[0050] The encoder mainly consists of convolutional layers, pooling layers, and normalization layers. Convolutional layers are responsible for acquiring local features of the image, pooling layers downsample the image and pass scale-invariant features to the next layer, and normalization layers mainly normalize the image distribution to accelerate learning. In other words, the encoder primarily classifies and analyzes low-level local pixel values ​​of the image to obtain high-order semantic features.

[0051] The decoder upsamples the scaled-down feature image and then performs convolution on the upsampled image to compensate for the loss of detail caused by the pooling layer in the encoder scaling down the image.

[0052] Figure 4 The diagram shown is a schematic representation of an exemplary embodiment of this application, illustrating the training of a model to be trained and graded. Figure 4As shown, the response text recognition model includes an Encoder and a Decoder, used to recognize handwritten response text. The network parameters of the Encoder in the response text recognition model are shared with the network parameters of the Encoder in the model to be trained for grading. That is, when constructing the network structure of the model to be graded, if the network structure of the Encoder in the model to be trained for grading is the same as that of the Encoder in the response text recognition model, then the network parameters of the Encoder in the model to be trained for grading are the same as those of the Encoder in the response text recognition model. If the network structure of the Encoder in the model to be trained for grading is different from that of the Encoder in the response text recognition model (i.e., different network depths), then the network parameters of the part of the network structure in the Encoder of the model to be trained for grading that is the same as that in the Encoder of the response text recognition model are set to be consistent with the network parameters of the Encoder in the response text recognition model.

[0053] In this application, to achieve end-to-end grading, the grading model receives features from many objects, such as features of the question text, the answer text, and the response text. Among these features, the answer text features are the most critical to the grading result. The answer text recognition model is mainly used to recognize the answer text; therefore, the network parameters of the answer text recognition model have a high response to the features of the answer text. This application's embodiments construct the network parameters of the grading model to be trained based on the network parameters of the feature extraction network of the answer text recognition model. This further enhances the grading model's focus on the features of the answer samples, thereby improving the grading accuracy.

[0054] Figure 5 The diagram shown is a schematic representation of the process for obtaining a modified model according to an exemplary embodiment of this application. Figure 3 Extending from the illustrated embodiment Figure 5 The illustrated embodiment will be described in detail below. Figure 5 The illustrated embodiments and Figure 3 The differences between the embodiments shown are not repeated here, and the similarities are not repeated here.

[0055] like Figure 5 As shown in the embodiment of this application, the constructed model to be trained is trained based on the training sample set to obtain the model, including the following steps.

[0056] Step S510: Based on the loss function of the response text recognition model, construct the loss function of the model to be trained for grading.

[0057] For example, in addition to the encoder, the model to be trained also includes a feature comparison network module. This module compares the similarity between the features of the answer sample and the features of the response sample. When the features of the answer sample and the response sample are consistent, the module outputs a correct grading result for the response sample; when the features of the answer sample and the response sample are inconsistent, the module outputs an incorrect grading result for the response sample. For example, a correct grading result can be represented by a "√" symbol, and an incorrect grading result can be represented by an "×" symbol.

[0058] Furthermore, since the network parameters of the model to be trained are constructed based on the network parameters of the response text recognition model, the loss function of the model to be trained includes the loss function of the response text recognition model. In addition, the loss function of the model to be trained also includes the loss function corresponding to the feature comparison network module. For example, if the loss function of the response text recognition model is denoted as LOSS1, and the loss function corresponding to the feature comparison network module is denoted as LOSS2, then the loss function corresponding to the model to be trained is LOSS = a × LOSS1 + b × LOSS2, where a represents the weight corresponding to LOSS1, and b represents the weight corresponding to LOSS2.

[0059] It should be noted that the network structure of the training and correction model provided in this application embodiment is only an example. In actual application, the network modules other than the encoder included in the training network model can be set according to the actual situation, and the final loss function corresponding to the training and correction model can be determined according to the loss function of the other network modules and the loss function of the answer text recognition model.

[0060] Step S520: Based on the training sample set, determine the text fusion features corresponding to the question stem sample, the answer sample, the response sample, and the grading sample.

[0061] Specifically, based on the question stem sample, the semantic features corresponding to the question stem sample are determined; based on the answer sample, the semantic features corresponding to the answer sample are determined; based on the answer sample, the semantic and visual features corresponding to the answer sample are determined; based on the graded sample, the semantic features corresponding to the graded sample are determined; the semantic features corresponding to the question stem sample, answer sample, answer sample, and graded sample, as well as the visual features corresponding to the answer sample, are fused to obtain the text fusion features.

[0062] For example, the question stem sample, answer sample, response sample, and grading sample are input into a Bidirectional Encoder Representation from Transformers (BERT) model to obtain the semantic features corresponding to the question stem sample, the answer sample, the response sample, and the grading sample. Further, the response sample is input into the ImageNet model to obtain the visual features corresponding to the response sample.

[0063] For example, the semantic features corresponding to the question stem sample, the semantic features corresponding to the answer sample, the semantic and visual features corresponding to the response sample, and the semantic features corresponding to the grading sample can be simply concatenated and passed through a fully connected network to obtain the text fusion features. For instance, if the semantic features corresponding to the question stem sample are T×160, the semantic features corresponding to the answer sample are T×160, the semantic features corresponding to the response sample are T×160, the visual features corresponding to the response sample are T×120, and the semantic features corresponding to the grading sample are T×120, then the text fusion features are T×720, where T represents the feature vector.

[0064] Alternatively, an attention mechanism can be used to initially filter the semantic features corresponding to the question stem sample, the semantic features corresponding to the answer sample, the semantic and visual features corresponding to the answer sample, and the semantic features corresponding to the grading sample. The semantic features corresponding to the filtered question stem sample, answer sample, answer sample, and grading sample, as well as the visual features corresponding to the answer sample, are then concatenated to obtain the text fusion features.

[0065] Step S530: Use the text fusion features and the loss function of the model to be trained to train the model to obtain the model.

[0066] When training a model to be trained and graded, there will be a difference between the model's predicted values ​​and the actual values. "Loss" means the penalty the model receives for failing to produce the expected results. The loss function determines the model's performance by comparing the predicted output of the model to be trained and graded with the expected output, and then finds the direction for optimization. If the deviation between the two is very large, the loss value corresponding to the loss function of the model to be trained and graded will be large; if the deviation is small, the loss value will be very low.

[0067] Specifically, the text fusion features are input into the model to be trained, the loss value is calculated based on the loss function of the model to be trained, and the parameters of the model to be trained are continuously adjusted until the training conditions are met, thus obtaining the model to be trained.

[0068] Constructing an appropriate loss function is crucial for successful network model training. In this embodiment, on the one hand, since the network parameters of the grading model to be trained are constructed based on the network parameters of the answer text recognition model, using the loss function of the answer text recognition model to construct the loss function of the network model to be trained ensures the consistency of the training process of the grading model and improves the overall training effect of the network model. On the other hand, this embodiment trains the grading model based on the text fusion features corresponding to the question stem samples, answer samples, and grading samples, which can improve the grading accuracy of the trained grading model.

[0069] Figure 6 The diagram shown is an exemplary embodiment of this application illustrating the acquisition of text fusion features. Figure 6 As shown, the question stem is "Assume the line y = x and…", the answer text is "y = 2x + 1", and the response text is "y = 2x". The question stem, answer text, and response text are input into a BERT model to obtain the semantic features corresponding to the question stem, answer text, and response text. The response text is then input into an ImageNet model to obtain its visual features. These semantic features are then fused with the visual features to obtain the text fusion features. Finally, the text fusion features are input into a grading model to obtain the grading results for the response text.

[0070] In this embodiment, the grading model is an end-to-end model. Furthermore, the network parameters of the grading model are constructed based on the network parameters of the answer text recognition model, allowing the grading model to focus more on the features of the answer text. This enables the grading result to be determined simply, quickly, and efficiently, while also improving the accuracy of the grading. In addition, this application uses multimodal features—namely, semantic features corresponding to the question stem, semantic features corresponding to the answer text, semantic features corresponding to the answer text, and visual features corresponding to the answer text—to obtain the final grading result for the answer text, further improving the grading accuracy.

[0071] The above text combined Figures 2 to 6 The method embodiments of this application are described in detail below, in conjunction with... Figure 7 The present application provides a detailed description of the apparatus embodiments. It should be understood that the descriptions of the method embodiments correspond to the descriptions of the apparatus embodiments; therefore, any parts not described in detail can be found in the foregoing method embodiments.

[0072] Figure 7 The diagram shown is a structural schematic of an exemplary embodiment of the present application for an alteration device. Figure 7As shown, the correction device 70 provided in this application embodiment includes:

[0073] The acquisition module 710 is used to acquire the question stem text, as well as the corresponding answer text and response text.

[0074] The grading module 720 is used to generate grading text corresponding to the answer text based on the question stem text, the corresponding answer text, and the answer text. The grading model is trained based on the answer text recognition model, which is used to recognize the answer text corresponding to the question stem text.

[0075] In one embodiment of this application, the acquisition module 710 is further configured to: acquire the network parameters of the answer text recognition model; construct the network parameters of the grading model to be trained based on the network parameters of the answer text recognition model; acquire a training sample set, which includes question stem samples, and answer samples, answer samples and grading samples corresponding to the question stem samples; and train the constructed grading model to be trained based on the training sample set to obtain a grading model, which is used to grade the answer text corresponding to the question stem text.

[0076] In one embodiment of this application, the acquisition module 710 is further configured to construct, based on the network parameters of the feature extraction network included in the answer text recognition model, the network parameters of the feature extraction network included in the training and grading model, wherein the network parameters of the feature extraction network included in the training and grading model include the network parameters of the feature extraction network included in the answer text recognition model.

[0077] In one embodiment of this application, the acquisition module 710 is further configured to: construct a loss function for the grading model to be trained based on the loss function of the answer text recognition model; determine the text fusion features corresponding to the question stem sample, the answer sample, the answer sample, and the grading sample based on the training sample set; and train the grading model to be trained using the text fusion features and the loss function of the grading model to be trained to obtain the grading model.

[0078] In one embodiment of this application, the acquisition module 710 is further configured to: determine the semantic features corresponding to the question stem sample based on the question stem sample; determine the semantic features corresponding to the answer sample based on the answer sample; determine the semantic features and visual features corresponding to the answer sample based on the answer sample; determine the semantic features corresponding to the grading sample based on the grading sample; and perform feature fusion on the semantic features corresponding to the question stem sample, answer sample, answer sample, and grading sample, as well as the visual features corresponding to the answer sample, to obtain text fusion features.

[0079] In one embodiment of this application, the grading module 720 is further configured to, based on the question stem text and the corresponding answer text and response text, determine the semantic features corresponding to the question stem text, the semantic features corresponding to the answer text, the semantic features corresponding to the response text, and the visual features corresponding to the response text; and, based on the semantic features corresponding to the question stem text, the answer text, and the response text, and the visual features corresponding to the response text, generate the grading text corresponding to the response text using the grading model.

[0080] In one embodiment of this application, the grading module 720 is further configured to perform feature fusion on the semantic features corresponding to the question stem text, the answer text, and the response text, as well as the visual features corresponding to the response text, to obtain text fusion features; and based on the text fusion features, generate the grading text corresponding to the response text using the grading model.

[0081] Below, for reference Figure 8 This describes an electronic device according to embodiments of the present application. Figure 8 The diagram shown is a structural schematic of an electronic device provided in an exemplary embodiment of this application.

[0082] like Figure 8 As shown, the electronic device 80 includes one or more processors 801 and memory 802.

[0083] The processor 801 may be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device 80 to perform desired functions.

[0084] The memory 802 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 801 may execute the program instructions to implement the methods of the various embodiments of this application described above and / or other desired functions. The computer-readable storage medium may also store various contents such as network parameters of the response text recognition model, network parameters of the training and grading model, training sample set, grading model, text fusion features, etc.

[0085] In one example, the electronic device 80 may also include an input device 803 and an output device 804, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0086] The input device 803 may include, for example, a keyboard, a mouse, etc.

[0087] The output device 804 can output various information to the outside, including network parameters of the text recognition model, network parameters of the model to be trained and graded, training sample set, grading model, text fusion features, etc. The output device 804 may include, for example, a display, speaker, printer, and communication network and its connected remote output devices, etc.

[0088] Of course, for the sake of simplicity, Figure 8 Only some of the components of the electronic device 80 relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device 80 may include any other suitable components depending on the specific application.

[0089] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps of the methods described above according to various embodiments of this application.

[0090] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0091] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform the steps of the methods described above according to various embodiments of this application.

[0092] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0093] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0094] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0095] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0096] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0097] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for grading, characterized in that, include: Obtain the question stem text, as well as the corresponding answer text and response text; Based on the question stem text, the corresponding answer text and the response text, a grading model is used to generate the grading text corresponding to the response text. The correction model is obtained through the following processing: Obtain the network parameters of the answer text recognition model, which is used to recognize the answer text corresponding to the question stem text; Based on the network parameters of the answer text recognition model, construct the network parameters of the grading model to be trained; Obtain a training sample set, which includes question stem samples, and corresponding answer samples, response samples, and graded samples. Based on the loss function of the response text recognition model, the loss function of the model to be trained and graded is constructed. Based on the training sample set, determine the text fusion features corresponding to the question stem sample, the answer sample, the response sample, and the grading sample; The text fusion features and the loss function of the training correction model are used to train the training correction model to obtain the correction model.

2. The method according to claim 1, characterized in that, The response text recognition model includes a feature extraction network, and the model to be trained for grading also includes a feature extraction network. The process of constructing the network parameters of the model to be trained for grading based on the network parameters of the response text recognition model includes: Based on the network parameters of the feature extraction network included in the answer text recognition model, the network parameters of the feature extraction network included in the training and grading model are constructed. The network parameters of the feature extraction network included in the constructed training and grading model are the same as the network parameters of the feature extraction network included in the answer text recognition model.

3. The method according to claim 1, characterized in that, The step of determining the text fusion features corresponding to the question stem sample, the answer sample, the response sample, and the grading sample based on the training sample set includes: Based on the question stem sample, determine the semantic features corresponding to the question stem sample; Based on the answer sample, determine the semantic features corresponding to the answer sample; Based on the response sample, determine the semantic and visual features corresponding to the response sample; Based on the corrected sample, determine the semantic features corresponding to the corrected sample; The semantic features corresponding to the question stem sample, the answer sample, the response sample, and the grading sample, as well as the visual features corresponding to the response sample, are fused to obtain the text fusion features.

4. The method according to claim 1, characterized in that, The step of generating graded text corresponding to the answer and response text based on the question stem text, the corresponding answer text, and the response text using a grading model includes: Based on the question stem text, the corresponding answer text, and the response text, the semantic features corresponding to the question stem text, the semantic features corresponding to the answer text, the semantic features corresponding to the response text, and the visual features corresponding to the response text are determined respectively. Based on the semantic features corresponding to the question stem text, the answer text, and the response text, as well as the visual features corresponding to the response text, the grading model is used to generate the grading text corresponding to the response text.

5. The method according to claim 4, characterized in that, The step of generating graded text corresponding to the answer text using the grading model, based on the semantic features corresponding to the question stem text, the answer text, and the answer text, as well as the visual features corresponding to the answer text, includes: The semantic features corresponding to the question stem text, the answer text, and the response text, as well as the visual features corresponding to the response text, are fused to obtain text fusion features; Based on the text fusion features, the correction model is used to generate the correction text corresponding to the answer text.

6. A correction device, characterized in that, include: The acquisition module is used to acquire the question stem text, as well as the corresponding answer text and response text; The grading module is used to generate graded text corresponding to the answer text based on the question stem text, the corresponding answer text, and the corresponding answer text, using a grading model. The grading model is obtained through the following processes: obtaining the network parameters of the answer text recognition model, which is used to recognize the answer text corresponding to the question stem text; constructing the network parameters of the grading model to be trained based on the network parameters of the answer text recognition model; obtaining a training sample set, which includes question stem samples, and the corresponding answer samples, answer samples, and graded samples; constructing the loss function of the grading model to be trained based on the loss function of the answer text recognition model; and determining the text fusion features corresponding to the question stem samples, and the corresponding answer samples, answer samples, and graded samples based on the training sample set. The text fusion features and the loss function of the training correction model are used to train the training correction model to obtain the correction model.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program for performing the method described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to perform the method described in any one of claims 1 to 5.

Images