Test question label prediction method and related device, equipment and storage medium

Abstract

The present application discloses a method for predicting test question labels and related devices, equipment and storage media, wherein the method for predicting test question labels includes: obtaining a label feature set and obtaining a test question to be identified; wherein the label feature set includes label features of several knowledge labels, and the label features of any knowledge label are extracted by the knowledge label and the reference test question under it through the label tower sub-model in the double tower model, the double tower model also includes a test question tower sub-model, and the label tower sub-model and the test question tower sub-model are decoupled for training during the training of the double tower model; extracting the test question features of the test question to be identified based on the test question tower sub-model; selecting the knowledge label as the target label of the test question to be identified based on the similarity between the test question features and the label features of each knowledge label in the label feature set. The above scheme can realize the real-time prediction of knowledge labels and reduce the consumption of computing resources and storage resources, especially when there may be new knowledge labels.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a test question label prediction method and related devices, equipment and storage media. Background Art

[0002] In the field of education, educational data such as test questions are usually formed in a targeted manner around one or certain knowledge points, that is, knowledge tags are the core and most basic information of educational data.

[0003] In existing research and applications at home and abroad, the test question label prediction model is usually incrementally trained after the educational resources are updated to meet the prediction needs of the newly added knowledge labels. However, this method cannot support the real-time prediction of the newly added knowledge labels. Since incremental training is required once the new knowledge labels are added, the consumption of computing resources is also large. In order to mitigate the problem of catastrophic forgetting, the relevant training data of the original knowledge labels must be retained, so the consumption of storage resources is also large. In view of this, how to achieve real-time prediction of knowledge labels and reduce the consumption of computing resources and storage resources, especially when there may be new knowledge labels, has become an urgent problem to be solved. Summary of the invention

[0004] The main technical problem solved by this application is to provide a test question label prediction method and related devices, equipment and storage media, which can realize real-time prediction of knowledge labels and reduce the consumption of computing resources and storage resources, especially when there may be new knowledge labels.

[0005] In order to solve the above technical problems, the first aspect of the present application provides a test question label prediction method, including: obtaining a label feature set and obtaining a test question to be identified; wherein the label feature set includes label features of several knowledge tags, and the label features of any knowledge label are extracted by the knowledge label and the reference test question attached to it through the label tower sub-model in the double-tower model, the double-tower model also includes a test question tower sub-model, and during the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training; based on the test question tower sub-model, the test question features of the test question to be identified are extracted; based on the similarity between the test question features and the label features of each knowledge label in the label feature set, the knowledge label is selected as the target label of the test question to be identified.

[0006] In order to solve the above-mentioned technical problems, the second aspect of the present application provides a test question label prediction device, including: a test question collection acquisition module, a test question feature extraction module and a knowledge label selection module, the test question collection acquisition module is used to obtain a label feature set and obtain test questions to be identified; wherein, the label feature set includes label features of several knowledge labels, and the label features of any knowledge label are extracted by the knowledge label and the reference test questions under it through the label tower sub-model in the double-tower model, the double-tower model also includes a test question tower sub-model, and the label tower sub-model and the test question tower sub-model are decoupled for training during the training of the double-tower model; the test question feature extraction module is used to extract the test question features of the test question to be identified based on the test question tower sub-model; the knowledge label selection module is used to select the knowledge label as the target label of the test question to be identified based on the similarity between the test question features and the label features of each knowledge label in the label feature set.

[0007] In order to solve the above technical problems, the third aspect of the present application provides an electronic device, which at least includes a memory and a processor coupled to each other, wherein the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the test question label prediction method in the above first aspect.

[0008] In order to solve the above technical problems, the fourth aspect of the present application provides a computer-readable storage medium storing program instructions that can be executed by a processor, and the program instructions are used to implement the test question label prediction method of the first aspect.

[0009] The above scheme obtains a label feature set and a test question to be identified, and the label feature set includes label features of several indicative labels. The label features of any knowledge label are extracted by the knowledge label and the reference test question under it through the label tower sub-model in the double-tower model. The double-tower model also includes a test question tower sub-model, and during the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training. The test question features of the test question to be identified are then extracted based on the test question tower sub-model, and based on the similarity between the test question features and the label features of each knowledge label in the label feature set, the knowledge label is selected as the target label of the test question to be identified. Therefore, due to the decoupled training of the label tower sub-model and the test question tower sub-model in the dual-tower model, it can ensure that various knowledge labels can be represented in the same feature space when the label tower sub-model performs feature extraction, thereby improving the generalization representation ability of the model, and helping to ensure the unbiased representation of various knowledge labels as much as possible when new knowledge labels are added. When new knowledge labels are added, only the label tower sub-model needs to extract its label features to ensure real-time prediction of the new knowledge labels without the need for incremental training, thereby avoiding the consumption of a large amount of computing resources and storage resources caused by incremental training as much as possible. Therefore, real-time prediction of knowledge labels can be achieved, and the consumption of computing resources and storage resources can be reduced, especially when there may be new knowledge labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flowchart of an embodiment of the test question label prediction method of the present application; Figure 2a is a schematic diagram of a process of extracting tag features according to an embodiment; Figure 2b is a process schematic diagram of an embodiment of decoupling training; Figure 2c is a schematic diagram of a process of an embodiment of real-time prediction of incremental test questions; Figure 3 It is a schematic diagram of the framework of an embodiment of the test question label prediction device of the present application; Figure 4 It is a schematic diagram of the framework of an embodiment of the electronic device of the present application; Figure 5 It is a schematic diagram of a framework of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0011] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0012] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0013] The terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the fragment " / " in this article generally indicates that the associated objects before and after are in an "or" relationship. In addition, "many" in this article means two or more than two.

[0014] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of the test question label prediction method of the present application. Specifically, it may include the following steps: Step S11: Obtain a label feature set and obtain the test questions to be identified.

[0015] In the disclosed embodiment, the label feature set may include label features of several knowledge labels. It should be noted that the knowledge label represents a knowledge point. For example, if the knowledge label of a test question is "trigonometric function", it means that the knowledge point examined by the test question involves the knowledge label "trigonometric function"; or, if the knowledge label of a test question is "present perfect tense", it means that the knowledge point examined by the test question involves the knowledge label "present perfect tense". Of course, the above examples are only a few possible examples in the actual application process, and the specific knowledge points involved in the knowledge labels will not be given one by one. In addition, the label feature can be represented by a feature vector, such as a feature vector with feature dimensions of 128 dimensions, 256 dimensions, etc., and the feature dimensions of the feature vector are not limited here.

[0016] In the disclosed embodiments, the label features of any knowledge label can be extracted from the knowledge label and the reference test questions attached to it through the label tower sub-model in the dual-tower model. The dual-tower model can also include a test question tower sub-model, and the label tower sub-model and the test question tower sub-model can be decoupled for training during the dual-tower model training process. It should be noted that the label tower sub-model can be used to model label representations, while the test question tower sub-model can be used to model test question representations. In addition, the decoupled training of the label tower sub-model and the test question tower sub-model during the training process of the dual-tower model means that the network parameters of the label tower sub-model and the test question tower sub-model are independent of each other during the training process, rather than being adjusted together. For the specific process of decoupling training, please refer to the following relevant description, which will not be repeated here.

[0017] In one implementation scenario, in order to construct a dual-tower model, a first language model and a second language model can be obtained based on pre-training of education field data, and then a label tower sub-model is constructed based on the first language model and the first neural network layer, and a test tower sub-model is constructed based on the second and second neural network layers, and the first neural network layer and the second neural network layer are used for feature dimensionality reduction. The above method, combining the language model obtained by pre-training based on education field data and the neural network layer for feature dimensionality reduction to construct a dual-tower model, can enrich the language comprehension ability of the dual-tower model, especially the language comprehension ability of data related to the education field, better understand the professional terms and concepts in the education field, and help provide more accurate prediction results in the application scenario of knowledge label prediction.

[0018] In a specific implementation scenario, the first language model and the second language model may adopt the same network structure. For example, taking the case where both the first language model and the second language model adopt BERT (Bidirectional Encoder Representation from Transformers, bidirectional encoding representation based on Transformer) as an example, a BERT model can be pre-trained using education field data as the first language model, and at the same time, another BERT model can be pre-trained using education field data as the second language model. Of course, the above example is only a possible example in the actual application process, and the specific structure of the language model is not limited here. For example, the language model can also adopt PaLM, etc., and examples are not given one by one here. Alternatively, the first language model and the second language model can adopt different network structures. For example, a BERT model can be pre-trained using education field data as the first language model, and at the same time, a PaLM model can be pre-trained using education field data as the second language model. Of course, the above example is only a possible example in the actual application process, and the specific structure of the first language model and the second language model is not limited here. For example, the first language model can also adopt the LLAMA model, and the second language model can also adopt the Galactica model, and examples are not given one by one here. It should be noted that for the specific process of pre-training, please refer to the technical details of pre-training of language models such as BERT, which will not be repeated here.

[0019] In a specific implementation scenario, the first language model and the second language model may be the same neural network model. Exemplarily, the BERT model may be pre-trained using education field data, and the pre-trained BERT model and its copy model (i.e., the network structure is exactly the same as the pre-trained BERT model and the network parameters are exactly the same as the pre-trained BERT model) may be used as the first language model and the second language model, respectively. Of course, the above example is only a possible example of the first language model and the second language model, and the network structure of the first language model and the second language model may not be limited here. For example, the first language model and the second language model may also adopt PaLM, etc., which will not be given one by one here. Alternatively, the first language model and the second language model may also be different neural network models. Exemplarily, the first language model and the second language model may adopt the same network structure or different network structures. For details, please refer to the aforementioned related description, which will not be repeated here.

[0020] In a specific implementation scenario, similar to the above description of the first language model and the second language model, the first neural network layer and the second neural network layer may also adopt the same network structure or different network structures. For example, the first neural network layer and the second neural network layer may both include M convolutional layers and N fully connected layers, or the first neural network layer may include M1 convolutional layers and N1 fully connected layers, and the second neural network layer may include M2 ​​convolutional layers and N2 fully connected layers.

[0021] In a specific implementation scenario, after obtaining the first language model and the first neural network layer, the first neural network layer can be connected to the first language model to obtain a label tower sub-model; similarly, after obtaining the second language model and the second neural network layer, the second neural network layer can be connected to the second language model to obtain a test question tower sub-model.

[0022] In an implementation scenario, after constructing the dual-tower model, you can first use the label tower sub-model to construct a label feature set. Specifically, you can extract features from the knowledge label based on the label sub-model to obtain the first feature, and extract features from each reference question under the knowledge label based on the label tower sub-model to obtain the second feature of the reference question. On this basis, you can fuse the first feature of the knowledge label and the second feature of each reference question under the knowledge label to obtain the label feature of the knowledge label. The above method, which fuses the first feature of the knowledge label itself and the second feature of the reference question it is attached to as a label feature, can make the label feature include both the implicit question representation encoded by the language model at the semantic level, and the text representation vector that explicitly highlights and focuses on the important representations in the label.

[0023] In a specific implementation scenario, for knowledge tags, the text of the knowledge tag itself (such as the aforementioned "trigonometric function", "present perfect tense", etc.) can be sent to the first language model in the label tower sub-model for encoding to obtain the corresponding representation, and the first neural network layer in the label tower sub-model reduces the dimension of the representation to obtain the first feature of the knowledge tag.

[0024] In a specific implementation scenario, for reference test questions, you can first use a word segmenter such as WordPiece to split the reference test questions into word lists, add corresponding special tags, and then map them into word vectors. Finally, you can use the first language model in the label tower sub-model to encode the word vectors to obtain the corresponding representation, and then use the first neural network layer in the label tower sub-model to reduce the dimension of the representation to obtain the second feature of the reference test questions.

[0025] In a specific implementation scenario, the first feature of the knowledge label and the second features of each reference test question under the knowledge label can be fused by concatenation, summing and averaging, weighting, etc. to obtain the label feature of the knowledge label.

[0026] In a specific implementation scenario, please refer to Figure 2a , Figure 2a FIG. 1 is a schematic diagram of a process of extracting tag features according to an embodiment of the present invention. Figure 2a As shown, taking the kth knowledge tag as an example, the reference questions under it can be recorded as ,in, represents the i-th reference question. , you can first use a word segmenter such as WordPiece to convert the reference test questions Split into word lists, add corresponding special tags, and then map them into word vectors Finally, we can use the first language model in the label tower sub-model to classify the word vector Encode and get the corresponding representation , and the first neural network layer in the label tower sub-model represents this Perform dimensionality reduction to obtain the second feature of the reference test questions At the same time, the text of the kth knowledge tag itself It can be sent to the first language model in the label tower sub-model for encoding to obtain the corresponding representation , and the first neural network layer in the label tower sub-model represents this Perform dimensionality reduction to obtain the first feature of the knowledge label On this basis, the first feature and the second characteristic of each reference test question Fusion is performed to obtain the label features of the kth knowledge label : …… (1) In the above formula (1), It represents a fusion operation, such as concatenation, summing and averaging, weighting, etc. The specific method of the fusion operation is not limited here. Through the above process, the label feature of the kth knowledge label can be extracted, and other knowledge labels can be deduced by analogy to obtain the corresponding label features, and then a label feature set can be constructed.

[0027] In an implementation scenario, after constructing the label feature set, the dual-tower model can be decoupled for training. Figure 2b , Figure 2b Schematic diagram of the process of decoupling training embodiment. Figure 2bAs shown, we can first obtain sample questions, and the sample questions can be marked with real labels, which represent the knowledge labels to which the sample questions actually belong. On this basis, we can extract the sample question features of the sample questions based on the question tower sub-model, and then use the similarity between the sample question features and the label features of each knowledge label in the label feature set to extract the sample question features. and the true label, and measure the training loss. Then, based on the training loss, we can adjust the network parameters of the question tower sub-model and fix the network parameters of the label tower sub-model. Indicates the similarity between the label feature of the kth knowledge label and the sample test question feature. In the above method, in the decoupled training process, the training loss only acts on the test question tower sub-model, so the test question tower sub-model can be constrained by the real label and the training loss measured by the similarity to model the feature information related to the knowledge point in the test question text as accurately as possible with the label feature as the reference object.

[0028] In a specific implementation scenario, in the process of extracting sample test question features by the test question tower sub-model, you can first use a word segmenter such as WordPiece to segment the sample test question into a word list, add corresponding special tags, and then map it to a word vector. Finally, you can use the second language model in the test question tower sub-model to encode the word vector to obtain the corresponding representation, and the second neural network layer in the test question tower sub-model reduces the dimension of the representation to obtain the sample test question features of the sample test question. It should be noted that the specific process of extracting sample test question features of sample test questions by the test question tower sub-model can refer to the related description of the aforementioned label tower sub-model extracting the second feature of the reference test question, which will not be repeated here.

[0029] In a specific implementation scenario, the similarity between the sample test question features and the label features of each knowledge label in the label feature set can be measured by cosine similarity and other measurement methods. For specific technical details, please refer to cosine similarity and other technical details, which will not be repeated here.

[0030] In a specific implementation scenario, the training loss can be measured by a loss function such as cross entropy. Exemplarily, the similarity between the sample test question feature and the positive example label in the label feature set and the training loss can be negatively correlated, and the positive example label is the same knowledge label as the true label. That is, if the similarity between the sample test question feature and the label feature with the same knowledge label is higher, the training loss is smaller, and conversely, if the similarity between the sample test question feature and the label feature with the same knowledge label is lower, the training loss is greater; similarly, the similarity between the sample test question feature and the negative example label in the label feature set and the training loss can be positively correlated, and the negative example label is a knowledge label different from the true label. That is, if the similarity between the sample test question feature and the label feature with different knowledge labels is higher, the training loss is greater, and conversely, if the similarity between the sample test question feature and the label feature with different knowledge labels is lower, the training loss is smaller. On this basis, by minimizing the training loss during the training process, the test question tower model can be forced to use label features with the same knowledge labels as the sample test questions as positive references and label features with different knowledge labels from the sample test questions as negative references, so as to model the feature information related to the knowledge points in the test question text as accurately as possible.

[0031] Step S12: Extracting the test feature of the test question to be identified based on the test question tower sub-model.

[0032] Specifically, after the decoupling training of the label tower sub-model and the test question tower sub-model in the double tower model is completed, the test question features of the test question to be identified can be extracted based on the test question tower sub-model. Exemplarily, a word segmenter such as WordPiece can be used to segment the test question to be identified into a word list, and corresponding special tags can be added, and then mapped to a word vector, and finally the second language model in the test question tower sub-model can be used to encode the word vector to obtain the corresponding representation, and the second neural network layer in the test question tower sub-model reduces the dimension of the representation to obtain the test question features of the test question to be identified. It should be noted that the specific process of the test question tower sub-model extracting the test question features of the test question to be identified can refer to the related description of the second feature of the reference test question extracted by the aforementioned label tower sub-model, which will not be repeated here.

[0033] Step S13: Based on the similarity between the test question feature and the label features of each knowledge label in the label feature set, a knowledge label is selected as the target label of the test question to be identified.

[0034] In an implementation scenario, each knowledge tag can be sorted based on the order of similarity from high to low between the test question features and the label features of each knowledge tag in the label feature set, and then among the knowledge tags after sorting, the knowledge tag that is located before the preset rank is selected as the target tag of the test question to be identified. It should be noted that the preset rank can be set according to the actual application needs. Exemplarily, the preset rank can be set to the top 5, top 4, top 3, etc.; or, the preset rank can be set to the top 5%, top 10%, etc., and the specific settings of the preset rank will not be given one by one here.

[0035] In a specific implementation scenario, as a special example, the preset order can be directly set to 1, that is, only the knowledge tag with the highest similarity is selected as the target tag. The two right-angled sides and The side lengths are 3 and 4 respectively. and Taking "angle" as an example, if the label feature set includes the knowledge labels "trigonometric function", "plane geometry", and "present perfect tense", the similarities calculated by the above process steps are ranked from high to low as: "trigonometric function", "plane geometry", and "present perfect tense". The knowledge label "trigonometric function" with the highest similarity can be taken as the target label of the test question to be identified. Of course, the above example is only a possible example of selecting the target label, and it does not limit other possible situations. No more examples will be given here.

[0036] In another specific implementation scenario, as another special example, a similarity threshold can be preset, and a knowledge tag with a similarity higher than the similarity threshold among the knowledge tags before the preset order is selected as the target tag. As a possible example, the test question to be identified is "Right triangle The two right-angled sides and The side lengths are 3 and 4 respectively. and Taking "angle" as an example, if the label feature set includes the knowledge labels "trigonometric function", "plane geometry", and "present perfect tense", the similarities calculated through the above process steps are ranked from high to low as follows: "trigonometric function" (similarity is 0.95), "plane geometry" (similarity is 0.93), and "present perfect tense" (similarity is 0.1). When the similarity threshold is set to 0.9, the knowledge label "trigonometric function" with the highest similarity and the knowledge label "plane geometry" with the second highest similarity can be taken as the target labels for the test questions to be identified. Of course, the above example is only a possible example of selecting the target label, and does not limit other possible situations. No examples will be given here one by one.

[0037] In an implementation scenario, in the actual application process of knowledge label prediction, there may be situations where new knowledge labels are added. Exemplarily, still taking the label feature set containing the knowledge labels "trigonometric functions", "plane geometry", and "present perfect tense" as an example, in the actual application process of knowledge label prediction, new knowledge labels such as "arithmetic operations" are added. In this case, in response to detecting that a new knowledge label is marked on at least one test question, the above-mentioned at least one test question can be selected as a reference test question under the new knowledge label, and the new knowledge label and the reference test question under it can be feature extracted based on the label tower sub-model to obtain the label features of the new knowledge label, and then based on the label features of the new knowledge label, the label feature set is updated to provide real-time prediction for new test questions that may be to be identified and should actually be marked as new knowledge labels. In the above method, when a new knowledge label is detected on at least one test question, the label feature set is updated based on the label tower sub-model, which can meet the real-time prediction of incremental test questions.

[0038] In a specific implementation scenario, for a new knowledge label and the reference questions under it, the new knowledge label can be specifically extracted based on the label tower sub-model to obtain the first feature, and the reference questions under the new knowledge label can be extracted based on the label tower sub-model to obtain the second feature of each reference question, and then the first feature of the new knowledge label and the second feature of each reference question under the new knowledge label are fused to obtain the label feature of the new knowledge label. For the specific extraction process, please refer to the aforementioned description of the label features of the knowledge label, which will not be repeated here.

[0039] In a specific implementation scenario, see Figure 2c , Figure 2c 2 is a schematic diagram of the process of an embodiment of real-time prediction of incremental test questions. Figure 2c As shown, after updating the label feature set based on the label features of the new knowledge label, in response to obtaining a new test question to be identified, the new test question to be identified can be based on the new label feature set, and the step of extracting the test feature of the test question to be identified based on the test question tower sub-model can be returned to determine the target label of the new test question to be identified. In other words, the test feature of the new test question to be identified can be extracted based on the test question tower sub-model, and then based on the similarity between the test feature and the label features of each knowledge label in the label feature set, the knowledge label is selected as the target label of the new test question to be identified. In the above method, after updating the label feature set based on the label features of the new knowledge label, the new label feature set is used for the new test question to be identified, and the aforementioned process steps are returned to determine the target label of the new test question to be identified, without the need for incremental training, and real-time prediction of incremental test questions can be achieved.

[0040] The above scheme obtains a label feature set and a test question to be identified, and the label feature set includes label features of several indicative labels. The label features of any knowledge label are extracted by the knowledge label and the reference test question under it through the label tower sub-model in the double-tower model. The double-tower model also includes a test question tower sub-model, and during the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training. The test question features of the test question to be identified are then extracted based on the test question tower sub-model, and based on the similarity between the test question features and the label features of each knowledge label in the label feature set, the knowledge label is selected as the target label of the test question to be identified. Therefore, due to the decoupled training of the label tower sub-model and the test question tower sub-model in the dual-tower model, it can ensure that various knowledge labels can be represented in the same feature space when the label tower sub-model performs feature extraction, thereby improving the generalization representation ability of the model, and helping to ensure the unbiased representation of various knowledge labels as much as possible when new knowledge labels are added. When new knowledge labels are added, only the label tower sub-model needs to extract its label features to ensure real-time prediction of the new knowledge labels without the need for incremental training, thereby avoiding the consumption of a large amount of computing resources and storage resources caused by incremental training as much as possible. Therefore, real-time prediction of knowledge labels can be achieved, and the consumption of computing resources and storage resources can be reduced, especially when there may be new knowledge labels.

[0041] See also Figure 3 , Figure 3 It is a schematic diagram of the framework of an embodiment of the test question label prediction device of the present application. The test question label prediction device 30 includes: a test question collection acquisition module 31, a test question feature extraction module 32 and a knowledge label selection module 33. The test question collection acquisition module 31 is used to obtain a label feature set and obtain test questions to be identified; wherein, the label feature set includes label features of several knowledge labels, and the label features of any knowledge label are extracted by the knowledge label and the reference test questions under it through the label tower sub-model in the double tower model. The double tower model also includes a test question tower sub-model, and the label tower sub-model and the test question tower sub-model are decoupled for training during the training of the double tower model; the test question feature extraction module 32 is used to extract the test question features of the test question to be identified based on the test question tower sub-model; the knowledge label selection module 33 is used to select the knowledge label as the target label of the test question to be identified based on the similarity between the test question features and the label features of each knowledge label in the label feature set.

[0042] In the above scheme, the test question label prediction device 30 obtains a label feature set and obtains the test question to be identified, and the label feature set includes label features of several indication labels, and the label features of any knowledge label are extracted by the knowledge label and the reference test question attached thereto through the label tower sub-model in the double-tower model, and the double-tower model also includes a test question tower sub-model, and in the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training, and then the test question features of the test question to be identified are extracted based on the test question tower sub-model, and based on the similarity between the test question features and the label features of each knowledge label in the label feature set, the knowledge label is selected as the test question to be identified. The target label of the question, therefore, due to the decoupled training of the label tower sub-model and the question tower sub-model in the dual-tower model, it can ensure that various knowledge labels can be represented in the same feature space when the label tower sub-model performs feature extraction, improving the generalization representation ability of the model, and helping to ensure the unbiased representation of various knowledge labels as much as possible when adding new knowledge labels. When adding new knowledge labels, only the label tower sub-model needs to extract its label features to ensure the real-time prediction of the new knowledge labels, without the need for incremental training, thereby avoiding the consumption of a large amount of computing resources and storage resources caused by incremental training as much as possible. Therefore, it is possible to achieve real-time prediction of knowledge labels and reduce the consumption of computing resources and storage resources, especially when there may be new knowledge labels.

[0043] In some disclosed embodiments, the test question label prediction device 30 includes a sample acquisition module for acquiring sample test questions; wherein the sample test questions are marked with real labels, and the real labels represent the knowledge labels to which the sample test questions actually belong; the test question label prediction device 30 includes a sample extraction module for extracting sample test question features of sample test questions based on the test question tower sub-model; the test question label prediction device 30 includes a loss measurement module for measuring the training loss based on the similarity between the sample test question features and the label features of each knowledge label in the label feature set and the real labels; the test question label prediction device 30 includes a parameter optimization module for adjusting the network parameters of the test question tower sub-model based on the training loss, and fixing the network parameters of the label tower sub-model.

[0044] In some disclosed embodiments, the similarity between the sample test question features and the positive example labels in the label feature set and the training loss are negatively correlated, and the positive example labels are knowledge labels that are the same as the true labels; and / or, the similarity between the sample test question features and the negative example labels in the label feature set and the training loss are positively correlated, and the negative example labels are knowledge labels that are different from the true labels.

[0045] In some disclosed embodiments, the test question label prediction device 30 includes a pre-training module for pre-training a first language model and a second language model based on education field data; the test question label prediction device 30 includes a model construction module for constructing a label tower sub-model based on the first language model and the first neural network layer, and constructing a test question tower sub-model based on the second language model and the second neural network layer; wherein the first neural network layer and the second neural network layer are both used for feature dimensionality reduction.

[0046] In some disclosed embodiments, the network structures of the first language model and the second language model are the same or different; or, the first language model and the second language model are the same neural network model or different neural network models.

[0047] In some disclosed embodiments, the test question label prediction device 30 includes a first extraction module for performing feature extraction on the knowledge label based on the label tower sub-model to obtain a first feature; the test question label prediction device 30 includes a second extraction module for performing feature extraction on each reference test question under the knowledge label based on the label tower sub-model to obtain a second feature of each reference test question; the test question label prediction device 30 includes a feature fusion module for performing fusion based on the first feature of the knowledge label and the second features of each reference test question under the knowledge label to obtain a label feature of the knowledge label.

[0048] In some disclosed embodiments, the test question label prediction device 30 includes a test question selection module for selecting at least one test question as a reference test question under the new knowledge label in response to detecting a new knowledge label on at least one test question; the test question label prediction device 30 includes a label feature extraction module for extracting features of the new knowledge label and the reference test questions under it based on the label tower sub-model to obtain label features of the new knowledge label; the test question label prediction device 30 includes a feature set updating module for updating the label feature set based on the label features of the new knowledge label.

[0049] In some disclosed embodiments, the test question label prediction device 30 includes a real-time prediction module for, in response to obtaining a new test question to be identified, based on a new label feature set, returning to execute the step of extracting the test question features of the test question to be identified based on the test question tower sub-model for the new test question to be identified, so as to determine the target label of the new test question to be identified.

[0050] In some disclosed embodiments, the knowledge tag selection module 33 includes a tag sorting submodule, which is used to sort each knowledge tag based on the order of similarity between the test question features and the tag features of each knowledge tag in the tag feature set from high to low; the knowledge tag selection module 33 includes a tag selection submodule, which is used to select the knowledge tag located before the preset sequence position among the knowledge tags after sorting as the target tag of the test question to be identified.

[0051] See also Figure 4 , Figure 4 : is a schematic diagram of the framework of an embodiment of an electronic device of the present application. The electronic device 40 includes at least a memory 41 and a processor 42 coupled to each other, the memory 41 stores at least program instructions, and the processor 42 is used to execute the program instructions to implement the steps in any of the above-mentioned test question label prediction method embodiments. For details, please refer to the aforementioned disclosed embodiments, which will not be repeated here. As a possible example, the electronic device 40 may include but is not limited to a server, a camera, a smart phone, a tablet computer, etc., and the specific type of the electronic device 40 is not limited here.

[0052] Specifically, the processor 42 is used to control itself and the memory 41 to implement the steps in any of the above-mentioned test question label prediction method embodiments. The processor 42 can also be called a CPU (Central Processing Unit). The processor 42 may be an integrated circuit chip with signal processing capabilities. The processor 42 can also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. In addition, the processor 42 can be implemented by an integrated circuit chip.

[0053] In the above scheme, the electronic device 40 obtains a label feature set and obtains a test question to be identified, and the label feature set includes label features of several indication labels, and the label features of any knowledge label are extracted by the knowledge label and the reference test question attached thereto through the label tower sub-model in the double-tower model, and the double-tower model also includes a test question tower sub-model, and in the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training, and then the test question features of the test question to be identified are extracted based on the test question tower sub-model, and based on the similarity between the test question features and the label features of each knowledge label in the label feature set, the knowledge label is selected as the test question to be identified. Target label, therefore, due to the decoupled training of the label tower sub-model and the test question tower sub-model in the dual-tower model, it can ensure that various knowledge labels can be represented in the same feature space when the label tower sub-model performs feature extraction, improve the generalization representation ability of the model, and help to ensure the unbiased representation of various knowledge labels as much as possible when adding new knowledge labels. When adding new knowledge labels, only the label tower sub-model needs to extract its label features to ensure the real-time prediction of the new knowledge labels, without the need for incremental training, thereby avoiding the consumption of a large amount of computing resources and storage resources caused by incremental training as much as possible. Therefore, it is possible to achieve real-time prediction of knowledge labels and reduce the consumption of computing resources and storage resources, especially when there may be new knowledge labels.

[0054] See also Figure 5 , Figure 5 1 is a schematic diagram of a framework of an embodiment of a computer-readable storage medium 50 of the present application. The computer-readable storage medium 50 stores program instructions 51 that can be executed by a processor, and the program instructions 51 are used to implement the steps in any of the above-mentioned test question label prediction method embodiments.

[0055] In the above scheme, the computer-readable storage medium 50 obtains a label feature set and obtains a test question to be identified, and the label feature set includes label features of several indicating labels, and the label features of any knowledge label are extracted by the knowledge label and the reference test question attached thereto through the label tower sub-model in the double-tower model, and the double-tower model also includes a test question tower sub-model, and in the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training, and then the test question features of the test question to be identified are extracted based on the test question tower sub-model, and based on the similarity between the test question features and the label features of each knowledge label in the label feature set, the knowledge label is selected as the test question to be identified. The target label of the question, therefore, due to the decoupled training of the label tower sub-model and the question tower sub-model in the dual-tower model, it can ensure that various knowledge labels can be represented in the same feature space when the label tower sub-model performs feature extraction, improving the generalization representation ability of the model, and helping to ensure the unbiased representation of various knowledge labels as much as possible when adding new knowledge labels. When adding new knowledge labels, only the label tower sub-model needs to extract its label features to ensure the real-time prediction of the new knowledge labels, without the need for incremental training, thereby avoiding the consumption of a large amount of computing resources and storage resources caused by incremental training as much as possible. Therefore, it is possible to achieve real-time prediction of knowledge labels and reduce the consumption of computing resources and storage resources, especially when there may be new knowledge labels.

[0056] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0057] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0058] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0059] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0060] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0061] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0062] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A test question label prediction method, characterized in that: include: Obtain a label feature set and obtain a test question to be identified; wherein the label feature set includes label features of a number of knowledge labels, and the label features of any of the knowledge labels are extracted by the label tower sub-model in the double-tower model from the knowledge label and the reference test question attached thereto, and the double-tower model also includes a test question tower sub-model, and during the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training; Extracting the test feature of the test question to be identified based on the test question tower sub-model; Based on the similarity between the test question feature and the label features of each of the knowledge labels in the label feature set, the knowledge label is selected as the target label of the test question to be identified.

2. The method according to claim 1, characterized in that: The decoupling training includes: Obtaining sample test questions; wherein the sample test questions are marked with real labels, and the real labels represent the knowledge labels to which the sample test questions actually belong; Extracting sample test question features of the sample test question based on the test question tower sub-model; Based on the similarity between the sample test question feature and the label feature of each of the knowledge labels in the label feature set and the true label, the training loss is measured; Based on the training loss, the network parameters of the question tower sub-model are adjusted, and the network parameters of the label tower sub-model are fixed.

3. The method according to claim 2, characterized in that The similarity between the sample test question feature and the positive example label in the label feature set and the training loss are negatively correlated, and the positive example label is a knowledge label that is the same as the true label; And / or, there is a positive correlation between the similarity between the sample test question feature and the negative example label in the label feature set and the training loss, and the negative example label is a knowledge label different from the true label.

4. The method according to claim 1, characterized in that: The steps of constructing the twin tower model include: The first language model and the second language model are obtained based on pre-training of education field data; The label tower sub-model is constructed based on the first language model and the first neural network layer, and the question tower sub-model is constructed based on the second language model and the second neural network layer; wherein the first neural network layer and the second neural network layer are both used for feature dimensionality reduction.

5. The method according to claim 4, characterized in that The network structures of the first language model and the second language model are the same or different; Alternatively, the first language model and the second language model are the same neural network model or different neural network models.

6. The method according to claim 1, characterized in that The step of extracting the label features comprises: Based on the label tower sub-model, feature extraction is performed on the knowledge label to obtain a first feature, and based on the label tower sub-model, feature extraction is performed on each of the reference test questions under the knowledge label to obtain a second feature of each of the reference test questions; The label feature of the knowledge label is obtained by fusing the first feature of the knowledge label and the second features of each of the reference test questions under the knowledge label.

7. The method according to claim 1, characterized in that The method further comprises: In response to detecting that a new knowledge label is marked on at least one test question, selecting the at least one test question as a reference test question under the new knowledge label; Based on the label tower sub-model, feature extraction is performed on the new knowledge label and the reference test questions attached thereto to obtain label features of the new knowledge label; The tag feature set is updated based on the tag feature of the new knowledge tag.

8. The method according to claim 7, characterized in that After updating the tag feature set based on the tag feature of the new knowledge tag, the method further includes: In response to obtaining a new test question to be identified, based on a new label feature set, for the new test question to be identified, the step of extracting the test feature of the test question to be identified based on the test question tower sub-model is returned to determine the target label of the new test question to be identified.

9. The method according to claim 1, characterized in that: The selecting the knowledge tag as the target tag of the test question to be identified based on the similarity between the test question feature and the tag features of each of the knowledge tags in the tag feature set includes: Sort each of the knowledge tags based on the order of similarity from high to low between the test question feature and the tag feature of each of the knowledge tags in the tag feature set; Among the knowledge tags after sorting, the knowledge tag located before the preset order is selected as the target tag of the test question to be identified.

10. A test question label prediction device, characterized in that: include: A set of test questions acquisition module, used to acquire a label feature set and to acquire test questions to be identified; wherein the label feature set includes label features of a number of knowledge labels, and the label features of any of the knowledge labels are extracted by the label tower sub-model in the double-tower model from the knowledge label and the reference test questions attached thereto, the double-tower model also includes a test question tower sub-model, and during the training process of the double-tower model, the label tower sub-model and the test question tower sub-model are decoupled for training; A test question feature extraction module, used for extracting the test question features of the test question to be identified based on the test question tower sub-model; The knowledge tag selection module is used to select the knowledge tag as the target tag of the test question to be identified based on the similarity between the test question feature and the tag features of each of the knowledge tags in the tag feature set.

11. An electronic device, characterized in that: The method comprises at least a memory and a processor coupled to each other, wherein the memory at least stores program instructions, and the processor is used to execute the program instructions to implement the test question label prediction method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that: Program instructions that can be executed by a processor are stored, and the program instructions are used to implement the test question label prediction method described in any one of claims 1 to 9.

Images