A method and device for training a question and answer model, a computer device and a storage medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses a training method and device of a question and answer model, computer equipment and a storage medium, the method comprising the following steps: acquiring an initial model and L pieces of server statements; replacing at least one target word in a plurality of words in the server statements with a preset identifier to generate new L pieces of server statements to form a first sample set; extracting a plurality of groups of two continuous server statements to obtain a plurality of groups of positive samples, sequentially exchanging the two server statements to obtain a plurality of groups of negative samples, and combining the plurality of groups of positive samples to form a second sample set; generating a server question according to the grammatical structure and content of the server statements, generating a server answer based on the dependency relationship of the words in each server statement, and combining the server question and the server answer to form a third sample set. The initial model is trained based on the first sample set, the second sample set and the third sample set to obtain a target question and answer model. The application can improve the efficiency of a customer inquiring about the corresponding answer of a server question.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a training method and device of a question and answer model, a computer device and a storage medium. BACKGROUND

[0002] With the rapid development of server technology, the functions of servers cover more and more aspects. Consequently, the related professional knowledge of servers is becoming more and more complicated. After a customer purchases a server, the manufacturer will provide a series of server-related manuals for the customer to query and refer to in the later operation and maintenance process, and the manufacturer will also provide an FAE team to manually answer the customer's questions. However, due to the difficulty of understanding and memorizing some professional knowledge, the content of the manuals is complicated and most of them are in English, which leads to a very low efficiency of the user in extracting information, and further makes many users skip the self-querying stage and directly ask the FAE team, which finally leads to an increase in the labor cost of the manufacturer.

[0003] Therefore, how to improve the efficiency of the customer in querying the corresponding answers to the server problems, reduce the labor cost, and provide good after-sales service has become a practical problem that needs to be solved by the server manufacturer. SUMMARY

[0004] Therefore, the present application provides a training method and device of a question and answer model, a computer device and a storage medium to solve the problems of low efficiency, high labor cost and poor user experience in querying the corresponding answers to the server problems by the customer.

[0005] In a first aspect, the present application provides a training method of a question and answer model, which comprises: obtaining an initial model and L pieces of server statements, L is an integer greater than or equal to 1; analyzing the L pieces of server statements to obtain a plurality of words in each piece of server statement, a syntax structure and content of each piece of server statement, and a dependency relationship of the words included in each piece of server statement; extracting at least one target word included in the plurality of words and replacing the at least one target word with a preset identifier to generate new L pieces of server statements, the new L pieces of server statements forming a first sample set; extracting a plurality of groups of two pieces of server statements in sequence from the L pieces of server statements to obtain a plurality of groups of positive samples, and exchanging the order of the two pieces of server statements included in each group of positive samples to obtain a plurality of groups of negative samples, the plurality of groups of positive samples and the plurality of groups of negative samples forming a second sample set; generating a plurality of server questions according to the syntax structure and content of the L pieces of server statements, and generating a plurality of server answers corresponding to the plurality of server questions based on the dependency relationship of the words included in each piece of server statement, the plurality of server questions and the plurality of server answers forming a third sample set.

[0006] The initial model is trained based on the first sample set, the second sample set and the third sample set to obtain a target question and answer model, and the target question and answer model is sent to the baseboard management controller, and the target question and answer model has the answering ability in response to the server question.

[0007] Based on the method of the first aspect, the initial model and the L server statements can be obtained, and the L data can be processed differently to obtain the first sample set, the second sample set and the third sample set, and the initial model is trained based on the first sample set, the second sample set and the third sample set to obtain a target question and answer model. Since the first number of target words included in the first sample set is replaced by a preset identifier, the trained initial model has the ability to predict words; since the second sample set is determined by the sequence of the server statements, the trained initial model has the ability to predict the sequence of the statements; after training the initial model based on the first sample set and the second sample set, the third sample set is further trained, and since the third sample set is divided into server questions and server answers, the target question and answer model obtained by training has the answering ability in response to the server question, and then the target question and answer model is sent to the baseboard management controller, so that the customer can directly query the corresponding answer of the server question on the client side of the server, thereby improving the efficiency of the customer to query the corresponding answer of the server question, reducing the labor cost, and ensuring the customer experience.

[0008] In an optional implementation, the L server statements are obtained by: obtaining at least one server document, and each first server statement is expressed in a first language type; translating each first server statement expressed in the first language type into a second server statement expressed in a second language type to obtain N second server statements, and one second server statement expressed in the second language type can be translated into at least one server statement expressed in the first language type; and translating each second server statement expressed in the second language type into a server statement expressed in the first language type to obtain L server statements, and L≥N.

[0009] Based on the above method, since the number of at least one server document is small, the server statements are expanded by the translation and back-translation method, for example, the Chinese document of the server statement is translated into English or the English document is translated into Chinese, and then the server statement is translated into the original language, so that the same server statement has different expression methods quickly and efficiently, the purpose of expanding the server statement is achieved, and the model training is more accurate.

[0010] In an optional implementation, the initial model is trained based on the first sample set, the second sample set and the third sample set to obtain the target question and answer model, including: inputting the first sample set into the initial model to perform a prediction word task training to obtain a first prediction model, the first prediction model having the ability to predict the word replaced by the preset identifier; inputting the second sample set into the first prediction model to perform a prediction sentence order training to obtain a second prediction model, the second prediction model having the ability to predict whether the order of the two server sentences is correct; inputting the third sample set into the second prediction model to perform a question and answer training to obtain a third prediction model, the third prediction model having the ability to respond to the server question; and performing a pruning processing on the third prediction model to obtain the target question and answer model.

[0011] Based on the above method, the initial model can be trained based on the first sample set, the second sample set and the third sample set to obtain the trained third prediction model, and the pruning processing is performed on the third prediction model to obtain the target question and answer model, so as to realize the model lightening and enable the target question and answer model to be embedded in the baseboard management controller.

[0012] In an optional implementation, the number of the at least one target word is a first number; the inputting the first sample set into the initial model to perform the prediction word task training to obtain the first prediction model includes: taking the first number of target words as first number of first real labels; and inputting the first sample set into the initial model to perform the prediction word task training to obtain a first number of first prediction labels, the first prediction label being used to indicate a prediction word corresponding to the target word.

[0013] The first evaluation result of the initial model is obtained by comparing the first number of first prediction labels with the first number of first real labels, the first evaluation result being used to indicate a ratio of a number of times that the first prediction label is consistent with the first real label to the first number; it is determined whether the first evaluation result is greater than or equal to a first threshold value; if yes, the trained initial model is determined as the first prediction model; if no, the model parameter of the trained initial model is adjusted, and the training is performed again based on the first sample set until the first evaluation result is greater than or equal to the first threshold value.

[0014] Based on the above method, the initial model can be trained based on the first sample set to obtain the first prediction model, so that the first prediction model has the ability to predict the word replaced by the preset identifier.

[0015] In an optional implementation, the second sample set includes a second number of groups of positive samples; the second sample set is input into the first prediction model for prediction sentence order training, to obtain a second prediction model, including: recording, as a second number of second real labels, an order of two server statements included in each group of positive samples in the second sample set; inputting the second sample set into the first prediction model, to obtain a second number of second prediction labels, the second prediction labels being used to indicate a predicted order of the two server statements included in each group of positive samples or each group of negative samples.

[0016] comparing the second number of second prediction labels with the second number of second real labels, to obtain a second evaluation result of the first prediction model, the second evaluation result being used to indicate a ratio of a number of times that the second prediction labels are consistent with the second real labels to the second number; determining whether the second evaluation result is greater than or equal to a second threshold value; if yes, determining the trained first prediction model as the second prediction model; if no, adjusting model parameters of the trained first prediction model, and retraining based on the second sample set, until the second evaluation result is greater than or equal to the second threshold value.

[0017] Based on the above method, the first prediction model can be trained based on the second sample set to obtain the second prediction model, so that the second prediction model has the ability to predict whether the order of the two server statements is correct.

[0018] In an optional implementation, the third sample set includes a third number of server questions, each server question corresponding to a server answer; the third sample set is input into the second prediction model for question and answer training, to obtain a third prediction model, including: inputting each server question in the third number of server questions and a server answer corresponding to each server question as a third number of third real labels; inputting the third sample set into the second prediction model, to obtain a third number of third prediction labels, the third prediction labels being used to indicate predicted server answers corresponding to each server question.

[0019] comparing the third number of third prediction labels with the third number of third real labels, to obtain a third evaluation result of the second prediction model, the third evaluation result being used to indicate a ratio of a number of times that the third prediction labels are consistent with the third real labels to the third number; determining whether the third evaluation result is greater than or equal to a third threshold value; if yes, determining the trained second prediction model as the third prediction model; if no, adjusting model parameters of the trained second prediction model, and retraining based on the third sample set, until the third evaluation result is greater than or equal to the third threshold value.

[0020] Based on the above method, the second prediction model can be trained based on the third sample set to obtain the third prediction model, so that the third prediction model has the ability to answer the server question.

[0021] In an optional implementation, the third prediction model comprises an embedding layer, an attention layer and a fully connected layer; the third prediction model is pruned to obtain the target question and answer model, comprising: reducing the dimension of the embedding vector in the embedding layer, the number of attention heads and the dimension of each attention head in the attention layer, or the number of hidden units in the fully connected layer included in the third prediction model to obtain the target question and answer model after pruning processing, the memory resource required by the target question and answer model is less than or equal to a fourth threshold, and the central processing unit resource required is less than or equal to a fifth threshold.

[0022] Based on the above method, the third prediction model can be pruned to make the memory resource and central processing unit resource required by the obtained target question and answer model meet the conditions of embedding the baseboard management controller.

[0023] In a second aspect, the present application provides a question and answer model training device, an acquisition module is used to acquire an initial model and L pieces of server sentences, L≥1 and is an integer; the acquisition module is also used to analyze the L pieces of server sentences to acquire a plurality of words corresponding to each piece of server sentence, the syntax structure and content of each piece of server sentence, and the dependency relationship of the words included in each piece of server sentence.

[0024] A processing module is used to extract at least one target word included in the plurality of words and replace the at least one target word with a preset identifier to generate new L pieces of server sentences, the new L pieces of server sentences forming a first sample set; the processing module is also used to extract a plurality of groups of two pieces of server sentences from the L pieces of server sentences to obtain a plurality of groups of positive samples, and exchange the order of the two pieces of server sentences included in each group of positive samples to obtain a plurality of groups of negative samples, and combine the plurality of groups of positive samples and the plurality of groups of negative samples to form a second sample set; the processing module is also used to generate a plurality of server questions according to the syntax structure and content of the L pieces of server sentences, and generate a plurality of server answers corresponding to the plurality of server questions based on the dependency relationship of the words included in each piece of server sentence, and combine the plurality of server questions and the plurality of server answers to form a third sample set; the processing module is also used to train the initial model based on the first sample set, the second sample set and the third sample set to obtain a target question and answer model, and send the target question and answer model to the baseboard management controller, the target question and answer model having the ability to respond to server questions.

[0025] In a third aspect, the present application provides a computer device, comprising: a memory and a processor, the memory and the processor are in communication connection with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the question and answer model training method of the first aspect or any one of the corresponding embodiments thereof.

[0026] In a fourth aspect, the present application provides a computer readable storage medium, having stored thereon computer instructions for causing a computer to execute the question and answer model training method of the first aspect or any of the corresponding embodiments thereof.

[0027] In a fifth aspect, the present application provides a computer program product comprising computer instructions for causing a computer to execute the question and answer model training method of the first aspect or any of the corresponding embodiments thereof. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings required to be used in the specific embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0029] Figure 1 is an architecture diagram of a BERT model according to an embodiment of the present application;

[0030] Figure 2 is a topological diagram of a question and answer model training system according to an embodiment of the present application;

[0031] Figure 3 is a flowchart of a question and answer model training method according to an embodiment of the present application;

[0032] Figure 4 is a structural block diagram of a question and answer model training device according to an embodiment of the present application;

[0033] Figure 5 is a hardware structure diagram of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0035] Before introducing the specific embodiments of the present application, the related technical terms of the present application will be introduced.

[0036] BERT model: BERT model is a natural language processing model based on deep learning. As shown in Figure 1 ,Figure 1 is an architecture diagram of a BERT model of an embodiment of the present application. Figure 1 In the BERT model, the BERT model includes an Input layer, an embedding layer, a Network structure of BERT layer and an Output layer. The embedding layer is divided into three subparts, namely Token embedding, Segment embedding and Position embedding.

[0037] Low-Rank Adaptation (LoRA) method: In deep learning, the parameters of a model are usually represented as a high-dimensional weight matrix W. The core idea of the LoRA method is to add a small, low-rank matrix AW to these weight matrices, resulting in a new weight matrix W+AW. This low-rank matrix AW can be obtained by training without changing the original pre-trained weights W.

[0038] Specifically, assuming that the dimension of the original weight matrix W is m x n, LoRA introduces a low-rank matrix AW, which can be decomposed into the product of two smaller matrices A and B, i.e. AW=AB, where the dimension of A is m x r, the dimension of B is r x n, and r<<min(m,n). In this way, by adjusting the values of A and B, we can indirectly adjust AW, and then adjust the weights of the entire model. Weight matrix update: W_new=W+AW=W+AB. Forward propagation: In the forward propagation process of the model, the updated weight matrix W_new is used for calculation. Backpropagation: In the backpropagation process, the gradient with respect to AW (i.e. A and B) is calculated, and these parameters are updated to minimize the loss function. The gradient can be calculated by the chain rule and applied to A and B. Gradient descent or other optimization algorithms are used to update the values of A and B until the stopping conditions are met (such as reaching the maximum number of iterations, the loss function value no longer significantly decreases (converges), etc.).

[0039] The present application is applied to the field of server operation and maintenance. The target question and answer model trained by the present application needs to be embedded in the baseboard management controller (BMC). The BMC is an independent system, which does not depend on other hardware (such as central processing unit, memory, etc.) on the system, nor on the input-output system, operating system. Moreover, the program server manufacturer in the BMC can be developed by itself, therefore, it is feasible for the manufacturer to implant the target question and answer model in the BMC, and the customer can use it by accessing the web page of the BMC after purchasing the server.

[0040] As described in the background, due to the low efficiency, high labor cost and poor user experience of the answers corresponding to the customer query server problems, the embodiment of the application provides a training method of a question and answer model, which trains the question and answer model to improve the efficiency of the answers corresponding to the customer query server problems and reduce the labor cost.

[0041] The following describes the method provided by the embodiment of the application by taking the training system 200 of the question and answer model shown in Figure 2 The following describes the method provided by the embodiment of the application by taking the training system 200 of the question and answer model shown in Figure 2 The above is only a schematic diagram and does not constitute a limitation on the applicable scenarios of the technical solutions provided by the application.

[0042] As shown in Figure 2 As shown in Figure 2 is a topological graph of the training system of the question and answer model according to the embodiment of the application. Figure 2 In the embodiment, the training system 200 of the question and answer model can include a training device 201 of the question and answer model and a baseboard management controller 202.

[0043] The training device 201 of the question and answer model in the embodiment of the application can be any device with computing and communication capabilities, such as a processor or a central processing unit.

[0044] The baseboard management controller 202 in the embodiment of the application can be a hardware manager integrated in a server, a network device and other computer systems. The baseboard management controller is used to monitor the hardware state of the device, perform remote management operations and provide monitoring and control functions for the device.

[0045] Figure 2 The training system 200 of the question and answer model shown in the above is only used for example and is not used to limit the technical solutions of the application. Those skilled in the art should understand that the training system 200 of the question and answer model can also include other devices in the specific implementation process, which is not limited.

[0046] According to the embodiment of the application, a training method of a question and answer model is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0047] In the embodiment, a training method of a question and answer model is provided, which can be used for the training device of the question and answer model described above, Figure 3 is a flowchart of the training method of the question and answer model according to the embodiment of the application, as shown in Figure 3 The flowchart includes the following steps:

[0048] S301: Obtain an initial model and L server statements.

[0049] The initial model can be a pre-trained model. For example, the initial model can be an ALBERT model.

[0050] It can be understood that the ALBERT model is a lightweight variant of the BERT model, which significantly reduces the number of model parameters through parameter sharing and embedding parameterization decomposition techniques, while maintaining comparable performance to the BERT model. The ALBERT model can reduce the number of parameters and computing cost while retaining the basic architecture of the BERT model, so it can be further pre-trained and fine-tuned like the BERT model.

[0051] In some optional embodiments, the question and answer model training device obtains at least one server document, each server document including N first server statements; translates each first server statement expressed in a first language type into a second server statement expressed in a second language type to obtain N second server statements; and translates each second server statement expressed in the second language type into a server statement expressed in the first language type to obtain L server statements.

[0052] L≥N.

[0053] In the embodiments of the application, the at least one server document can be a professional document of any one or more server fields. For example, the server document can be a Redfish manual, an IPMI manual, a BMC document, a BIOS document, and a server operation and maintenance document.

[0054] In the embodiments of the application, the language type of each first server statement is a first language type. A second server statement expressed in a second language type can be translated into at least one server statement expressed in the first language type. For example, the first language type can be Chinese, and the second language type can be English.

[0055] Optionally, the question and answer model training device can further perform a deduplication process on the N first server statements before translating each first server statement expressed in the first language type into a second server statement expressed in the second language type.

[0056] It can be understood that the question and answer model training device performs a deduplication process on the N server statements to delete the repeated content in the N server statements, which can avoid the model learning repeated patterns during the training process and affect the generalization ability.

[0057] S302: Analyze the L server statements to obtain a plurality of words corresponding to each server statement in the L server statements, the syntax structure and content of each server statement, and the dependency relationship of the words included in each server statement.

[0058] For example, the training device of the question and answer model can parse the L server statements through a word segmentation tool (for example, a jieba tool or a HanLP tool), perform word segmentation processing on the L server statements, divide the L server statements into multiple sentences or shorter text segments (commonly referred to as tokens), and extract the syntax structure and content of each server statement and the dependency relationship of the words included in each server statement.

[0059] Optionally, the training device of the question and answer model can also mark specific words or entities included in each server statement, improve the accuracy of word segmentation, or facilitate the initial model to identify and learn important information during the training process.

[0060] S303: Extract at least one target word included in the plurality of words, and replace the at least one target word with a preset identifier to generate new L server statements, and the new L server statements form a first sample set.

[0061] The target word can be a randomly selected word from the plurality of words. The number of the at least one target word can be a first number. For example, the first number can be a preset proportion of the words in the plurality of words. The preset proportion can be set according to actual needs. For example, the preset proportion can be 15%.

[0062] The preset identifier can be any kind of identifier or symbol.

[0063] For example, the training device of the question and answer model extracts at least one target word included in the plurality of words, and replaces the at least one target word with a preset identifier to generate new L server statements, and saves the new L server statements as a first sample set in a preset format.

[0064] The preset format can be a TensorFlow TFRecord file format or a PyTorch Dataset object format.

[0065] S304: Extract multiple groups of consecutive two server statements from the L server statements to obtain multiple groups of positive samples, and exchange the order of the two server statements included in each group of positive samples to obtain multiple groups of negative samples, and combine the multiple groups of positive samples and the multiple groups of negative samples to form a second sample set.

[0066] The positive sample includes two consecutive server statements, and the negative sample includes two server statements with exchanged order. It can be understood that the negative sample is similar to the positive sample in vocabulary and theme, but the order between the sentences is wrong.

[0067] The second sample set includes a second number of groups of positive samples.

[0068] S305: According to the syntax structure and content of the L server statements, a plurality of server questions are generated, and a plurality of server answers corresponding to the plurality of server questions are generated based on the dependency relationship of the words included in each server statement, and the plurality of server questions and the plurality of server answers are combined to form a third sample set.

[0069] For example, the server statement is "when the model of the server is the first model, the operating system of the server is the first operating system by default". The training device of the question and answer model generates the server question "what is the operating system of the first model server" according to the syntax structure (subject-predicate-object structure) and content of the server statement; based on the dependency relationship of the words included in the server statement, the server answer corresponding to the server question is "the first operating system", and "what is the operating system of the first model server" and "the first operating system" are saved in the third sample set.

[0070] The third sample set includes a third number of server questions.

[0071] S306: Training the initial model based on the first sample set, the second sample set, and the third sample set to obtain a target question and answer model, and sending the target question and answer model to the baseboard management controller.

[0072] The target question and answer model has the ability to respond to server questions.

[0073] The specific process of the training device of the question and answer model training the initial model based on the first sample set, the second sample set, and the third sample set to obtain the target question and answer model is introduced below.

[0074] In one example, the training device of the question and answer model inputs the first sample set into the initial model for prediction word task training to obtain a first prediction model; inputs the second sample set into the first prediction model for prediction sentence order training to obtain a second prediction model; inputs the third sample set into the second prediction model for question and answer training to obtain a third prediction model; and prunes the third prediction model to obtain the target question and answer model.

[0075] The first prediction model has the ability to predict the words replaced by the preset identifier.

[0076] The second prediction model has the ability to predict whether the order of the two server statements is correct.

[0077] The third prediction model has the ability to respond to server questions.

[0078] The training process of the first prediction model is as follows:

[0079] In an example, the training device of the question and answer model takes the first quantity of target words as the first quantity of first real labels; inputs the first sample set into the initial model to perform prediction word task training, to obtain the first quantity of first prediction labels; compares the first quantity of first prediction labels with the first quantity of first real labels, to obtain a first evaluation result of the initial model; judges whether the first evaluation result is greater than or equal to a first threshold value; if yes, the trained initial model is determined as a first prediction model; if no, the model parameters of the trained initial model are adjusted, and training is performed again based on the first sample set until the first evaluation result is greater than or equal to the first threshold value.

[0080] The first prediction label is used to indicate a predicted word corresponding to the target word.

[0081] The first evaluation result is used to indicate a ratio of a number of times that the first prediction label is consistent with the first real label to the first quantity.

[0082] The first threshold value can be set according to actual needs, and is not limited, for example, the first threshold value can be 90%.

[0083] The training process of the second prediction model is as follows:

[0084] In an example, the training device of the question and answer model records an order of two server sentences included in each group of positive samples in the second sample set as a second quantity of second real labels; inputs the second sample set into the first prediction model, to obtain a second quantity of second prediction labels; compares the second quantity of second prediction labels with the second quantity of second real labels, to obtain a second evaluation result of the first prediction model; judges whether the second evaluation result is greater than or equal to a second threshold value; if yes, the trained first prediction model is determined as a second prediction model; if no, the model parameters of the trained first prediction model are adjusted, and training is performed again based on the second sample set until the second evaluation result is greater than or equal to the second threshold value.

[0085] The second prediction label is used to indicate a predicted order of two server sentences included in each group of positive samples or each group of negative samples.

[0086] The second evaluation result is used to indicate a ratio of a number of times that the second prediction label is consistent with the second real label to the second quantity.

[0087] The second threshold value can be set according to actual needs, and is not limited, for example, the second threshold value can be 95%.

[0088] Optionally, the training device of the question and answer model can compare the second number of second predicted labels with the second number of second real labels by using a cross-entropy loss function to obtain a second evaluation result of the first prediction model. And when the second evaluation result is less than a second threshold, the model parameters of the trained first prediction model are adjusted by a back propagation algorithm until the second evaluation result is greater than or equal to the second threshold.

[0089] The training process of the third prediction model is as follows:

[0090] In an example, the training device of the question and answer model takes each of the third number of server questions and the server answer corresponding to each of the third number of server questions as a third number of third real labels; inputs the third sample set into the second prediction model to obtain a third number of third predicted labels; compares the third number of third predicted labels with the third number of third real labels to obtain a third evaluation result of the second prediction model; judges whether the third evaluation result is greater than or equal to a third threshold; if yes, the trained second prediction model is determined as the third prediction model; if no, the model parameters of the trained second prediction model are adjusted and retrained based on the third sample set until the third evaluation result is greater than or equal to the third threshold.

[0091] The third predicted label is used to indicate the predicted server answer corresponding to each service question.

[0092] The third evaluation result is used to indicate the ratio of the number of times that the third predicted label is consistent with the third real label to the third number.

[0093] Optionally, before the training device of the question and answer model trains the third prediction model, the parameters of the second prediction model can also be fine-tuned by a parameter efficient fine-tuning algorithm.

[0094] It can be understood that the parameter efficient fine-tuning can include an additive method, a selective method or a reparameterization method. Considering the limitation of the size of the question and answer model and the best model training effect, the training device of the question and answer model fine-tunes by a low-rank adaptation (LoRA) method under the reparameterization category.

[0095] The LoRA method reduces the number of parameters that need to be trained by introducing two trainable low-rank matrices (A and B). During the training process, the second prediction model parameters are frozen, and only these two low-rank matrices are trained, thereby achieving a quick adjustment of the model behavior without training the entire model from scratch. During the inference process, these two low-rank matrices are merged back into the second prediction model parameters to generate the final third prediction model.

[0096] Further, the third prediction model includes an embedding layer, an attention layer, and a fully connected layer. Among them, the embedding layer (InputEmbedding) contains token, segment, and position embedding methods. The attention layer (Multi-HeadAttention) includes multi-head attention. The fully connected layer (Feed Forward) is used to further map the output vector of the attention layer.

[0097] In an example, the training device of the question and answer model reduces the dimension of the embedding vector in the embedding layer included in the third prediction model, the number of attention heads and the dimension of each attention head in the attention layer, or the number of hidden units in the fully connected layer, to obtain a target question and answer model after pruning processing.

[0098] Among them, the required memory resources of the target question and answer model are less than or equal to a fourth threshold value, and the required central processing unit resources are less than or equal to a fifth threshold value.

[0099] The fourth threshold value can be a part of the memory amount of the baseboard management controller. For example, the fourth threshold value can be 50% of the total memory of the baseboard management controller.

[0100] The fifth threshold value can be a part of the central manager resource amount of the baseboard management controller. For example, the fifth threshold value can be 30% of the central manager resources of the baseboard management controller.

[0101] It can be understood that the embedding layer is usually used to convert the input text into a fixed-size vector representation. In the third prediction model, the training device of the question and answer model can reduce the parameter amount and the calculation amount of the model by reducing the dimension of the embedding vector. For example, the third prediction model uses a 768-dimensional embedding vector, which is reduced to 384 dimensions or lower through dimension pruning.

[0102] The attention layer (especially the multi-head attention layer) is a computationally intensive part. In the third prediction model, the training device of the question and answer model reduces the number of attention heads or the dimension of each head, which can significantly reduce the calculation amount and the parameter amount. For example, the third prediction model has 12 attention heads, each with a dimension of 64, which can be reduced to 6 heads or a dimension of 32 per head through dimension pruning.

[0103] Fully connected layers (especially those in the middle layer) typically have a large number of parameters. In the third prediction model, the question-answering model training apparatus reduces the number of hidden units in the fully connected layers, thereby reducing the number of model parameters. For example, the fully connected layer of the third prediction model has 3072 hidden units, which can be reduced to 1536 or less (768 dimensions) through dimensionality pruning.

[0104] The target question answering model performs dimension pruning to reduce the dimensions of each model layer to achieve model lightweighting. In the third prediction model, dimension pruning can be performed on different components such as the embedding layer, attention layer, and fully connected layer.

[0105] In one example, after the training device of the question-answering model performs dimension pruning on the third prediction model, the dimensionally pruned third prediction model is trained based on the third sample set to obtain a target question-answering model.

[0106] It can be understood that after the training device of the question-answering model performs dimension pruning on the third prediction model, it can train again to restore the model performance and avoid the problem of reduced prediction accuracy of the target question-answering model due to dimension pruning.

[0107] Furthermore, the training device of the question-answering model converts the format of the finally generated target question-answering model into a .gguf format file, and sends the .gguf format file corresponding to the target question-answering model to the baseboard management controller.

[0108] It is understandable that since the original image of the baseboard management controller does not have the operating environment of the large model, the training device of the question-and-answer model needs to use a cross-compilation tool chain to transplant the environment required for the model operation into the BMC, and then import the model file into the baseboard management controller to run and realize the question-and-answer function.

[0109] Based on the above Figure 2 According to the method, the training device of the question-answering model can obtain the initial model and L server statements, and perform different processing on the L data to obtain a first sample set, a second sample set, and a third sample set, and train the initial model based on the first sample set, the second sample set, and the third sample set to obtain a target question-answering model.

[0110] The first number of target words included in the first sample set is replaced by a preset identifier, so that the trained initial model has the ability to predict words; the second sample set is determined in the order of the server sentences, so that the trained initial model has the ability to predict the order of the sentences; after training the initial model based on the first sample set and the second sample set, further training is performed based on the third sample set; since the third sample set is to split L server sentences into server questions and server answers, the target question and answer model obtained by training has the ability to respond to the answers to the server questions, and then the target question and answer model is sent to the baseboard management controller, so that the customer can directly query the answers corresponding to the server questions on the client of the server, thereby improving the efficiency of the customer to query the answers corresponding to the server questions, reducing the labor cost, and ensuring the customer experience.

[0111] In the embodiment, a training device of a question and answer model is also provided, which is used to implement the above-mentioned embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware is also possible and is contemplated.

[0112] The embodiment provides a training device of a question and answer model, as shown in Figure 4 Figure 4 is a structural block diagram of the training device of the question and answer model according to the embodiment of the application; the device comprises:

[0113] The acquisition module 401 is configured to acquire an initial model and L server sentences, L≥1 and L is an integer.

[0114] The acquisition module 401 is further configured to analyze the L server sentences, acquire a plurality of words corresponding to each server sentence in the L server sentences, a syntax structure and content of each server sentence, and a dependency relationship of the words included in each server sentence.

[0115] The processing module 402 is configured to extract at least one target word included in the plurality of words, replace the at least one target word with a preset identifier, generate new L server sentences, and generate a first sample set composed of the new L server sentences.

[0116] The processing module 402 is further configured to extract a plurality of groups of two consecutive server sentences from the L server sentences to obtain a plurality of groups of positive samples, exchange the order of the two server sentences included in each group of positive samples to obtain a plurality of groups of negative samples, and group the plurality of groups of positive samples and the plurality of groups of negative samples to obtain a second sample set.

[0117] ​The processing module 402 is further configured to generate a plurality of server questions according to the syntax structure and content of the L server statements, and generate a plurality of server answers corresponding to the plurality of server questions based on the dependency relationship of the words included in each server statement, and combine the plurality of server questions and the plurality of server answers to form a third sample set.

[0118] The processing module 402 is further configured to train the initial model based on the first sample set, the second sample set, and the third sample set to obtain a target question and answer model, and send the target question and answer model to the baseboard management controller, the target question and answer model having the ability to answer server questions.

[0119] In some optional embodiments, the acquisition module 401 is specifically configured to acquire at least one server document, each server document including N first server statements, and each first server statement being of a first language type; the acquisition module 401 is further specifically configured to translate each first server statement expressed in the first language type into a second server statement expressed in a second language type to obtain N second server statements, one second server statement expressed in the second language type being translatable into at least one server statement expressed in the first language type; and the acquisition module 401 is further specifically configured to translate each second server statement expressed in the second language type into a server statement expressed in the first language type to obtain L server statements, L≥N.

[0120] In some optional embodiments, the processing module 402 is further specifically configured to input the first sample set into an initial model to perform a prediction word task training to obtain a first prediction model, the first prediction model having the ability to predict words replaced by preset identifiers; the processing module 402 is further specifically configured to input the second sample set into the first prediction model to perform a prediction sentence order training to obtain a second prediction model, the second prediction model having the ability to predict whether the order of two server statements is correct; the processing module 402 is further specifically configured to input the third sample set into the second prediction model to perform a question and answer training to obtain a third prediction model, the third prediction model having the ability to answer server questions; and the processing module 402 is further specifically configured to perform a pruning processing on the third prediction model to obtain a target question and answer model.

[0121] In some optional embodiments, the number of the at least one target word is a first number; the processing module 402 further specifically takes the first number of target words as a first number of first true labels; inputs the first sample set into the initial model to perform a predicted word task training, to obtain a first number of first predicted labels, the first predicted label being used to indicate a predicted word corresponding to the target word; compares the first number of first predicted labels with the first number of first true labels, to obtain a first evaluation result of the initial model, the first evaluation result being used to indicate a ratio of a number of times that the first predicted label is consistent with the first true label to the first number; judges whether the first evaluation result is greater than or equal to a first threshold value; if yes, determines the trained initial model as a first prediction model; if no, adjusts model parameters of the trained initial model, and re-performs training based on the first sample set, until the first evaluation result is greater than or equal to the first threshold value.

[0122] In some optional embodiments, the second sample set includes a second number of groups of positive samples; the processing module 402 further specifically records an order of two server statements included in each group of positive samples in the second sample set as a second number of second true labels; inputs the second sample set into the first prediction model, to obtain a second number of second predicted labels, the second predicted label being used to indicate a predicted order of two server statements included in each group of positive samples or each group of negative samples; compares the second number of second predicted labels with the second number of second true labels, to obtain a second evaluation result of the first prediction model, the second evaluation result being used to indicate a ratio of a number of times that the second predicted label is consistent with the second true label to the second number; judges whether the second evaluation result is greater than or equal to a second threshold value; if yes, determines the trained first prediction model as a second prediction model; if no, adjusts model parameters of the trained first prediction model, and re-performs training based on the second sample set, until the second evaluation result is greater than or equal to the second threshold value.

[0123] In some optional embodiments, the third sample set includes a third number of server questions; the processing module 402 is further specifically configured to take each of the third number of server questions and a server answer corresponding to each of the server questions as a third number of third real labels; input the third sample set into the second prediction model to obtain a third number of third predicted labels, the third predicted labels being used to indicate predicted server answers corresponding to each of the server questions; compare the third number of third predicted labels with the third number of third real labels to obtain a third evaluation result of the second prediction model, the third evaluation result being used to indicate a ratio of a number of times that the third predicted labels are consistent with the third real labels to the third number; determine whether the third evaluation result is greater than or equal to a third threshold value; if yes, determine the trained second prediction model as the third prediction model; and if no, adjust model parameters of the trained second prediction model and retrain based on the third sample set until the third evaluation result is greater than or equal to the third threshold value.

[0124] In some optional embodiments, the third prediction model includes an embedding layer, an attention layer and a fully connected layer; the processing module 402 is further specifically configured to reduce dimensions of embedding vectors in the embedding layer, a number of attention heads and dimensions of each of the attention heads in the attention layer or a number of hidden units in the fully connected layer included in the third prediction model to obtain a target question and answer model after pruning processing, the target question and answer model requiring memory resources less than or equal to a fourth threshold value and requiring central processing unit resources less than or equal to a fifth threshold value.

[0125] Further function descriptions of the above-mentioned various modules and units are the same as those of the above-mentioned corresponding embodiments, which will not be repeated here.

[0126] The training device of the question and answer model in the embodiment is presented in the form of a functional unit. The unit herein refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and a memory executing one or more software or fixed programs, and / or other devices that can provide the above-mentioned functions.

[0127] The embodiment of the present application also provides a computer device having the above-mentioned Figure 4 training device of the question and answer model.

[0128] Please refer to Figure 5 , Figure 5 is a structural schematic diagram of a computer device provided by an optional embodiment of the present application, as Figure 5As shown, the computer device includes one or more processors 10, memory 20, and interfaces 30 for the various components to communicate with one another. The various components communicate through the use of the various buses, and can be mounted on a common motherboard or in other manners as appropriate. The processor 10 can process instructions for execution within the computer device, including instructions stored in the memory 20 or elsewhere to implement routines for displaying graphical information, such as a GUI on an external input / output device, such as a display device coupled to the interface 30. In some embodiments, multiple processors and / or multiple buses can be employed as appropriate, as will be appreciated by those skilled in the art. Additionally, various components of the computer device can be used for other purposes, such as storage of data to implement routines, as will also be appreciated by those skilled in the art. Figure 5 The processor 10 is taken as an example in the embodiments.

[0129] The processor 10 can be a central processing unit, a network processor, or a combination thereof. The processor 10 can further include a hardware chip. The hardware chip can be an application specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device can be a complex programmable logic device, a field programmable logic device, a general array logic, or any combination thereof.

[0130] The memory 20 stores instructions that are executable by the at least one processor 10, so as to enable the at least one processor 10 to perform the method shown in the above embodiments.

[0131] The memory 20 can include a program region and a data region. The program region can store an operating system and an application program required by at least one function. The data region can store data created according to the use of the computer device, and the like. In addition, the memory 20 can include a high-speed random access memory, and can further include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some alternative embodiments, the memory 20 can optionally include a memory that is remotely arranged with respect to the processor 10, and these remote memories can be connected to the computer device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0132] The memory 20 can include a volatile memory, such as a random access memory, and can also include a non-volatile memory, such as a flash memory, a hard disk, or a solid state disk. The memory 20 can further include a combination of the above-mentioned kinds of memories.

[0133] The computer device further includes a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0134] The embodiments of the present application further provide a computer readable storage medium, and the method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or stored in a remote storage medium or a non-transitory machine readable storage medium and downloaded to a local storage medium through network, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special hardware. The storage medium can be a magnetic disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid state disk, etc. Further, the storage medium can also include a combination of the above-mentioned memories. It can be understood that the computer, the processor, the microprocessor controller, or the programmable hardware includes a storage component that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the method shown in the above embodiments is implemented.

[0135] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, the method and / or technical solutions according to the present application can be called or provided. Those skilled in the art should understand that the form of computer program instructions in a computer readable medium includes but is not limited to source files, executable files, installation package files, etc. Correspondingly, the way of executing computer program instructions by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0136] Although the embodiments of the present application are described in conjunction with the accompanying drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A training method for a question-answering model, characterized in that: The method comprises: Get the initial model and L server statements, where L is an integer and is ≥ 1. Parsing the L server statements to obtain a plurality of words corresponding to each of the L server statements, a grammatical structure and content of each server statement, and dependency relationships among the words included in each server statement; Extracting at least one target word from the plurality of words, and replacing the at least one target word with a preset identifier to generate L new server statements, wherein the L new server statements constitute a first sample set; Extracting multiple groups of two consecutive server statements from the L server statements to obtain multiple groups of positive samples, and sequentially swapping the two server statements included in each group of positive samples to obtain multiple groups of negative samples, and combining the multiple groups of positive samples and the multiple groups of negative samples into a second sample set, wherein the second sample set includes a second number of groups of positive samples; generating a plurality of server questions based on the grammatical structure and content of the L server statements, and generating a plurality of server answers corresponding to the plurality of server questions based on dependency relationships between terms included in each of the server statements, and forming a third sample set from the plurality of server questions and the plurality of server answers; Training the initial model based on the first sample set, the second sample set, and the third sample set to obtain a target question-answering model, and sending the target question-answering model to the baseboard management controller, wherein the target question-answering model has the ability to answer questions in response to the server; The training of the initial model based on the first sample set, the second sample set, and the third sample set to obtain a target question-answering model includes: Inputting the first sample set into the initial model for word prediction task training to obtain a first prediction model, wherein the first prediction model has the ability to predict the word replaced by the preset identifier; Inputting the second sample set into the first prediction model to perform statement sequence prediction training to obtain a second prediction model, wherein the second prediction model has the ability to predict whether the sequence of the two server statements is correct; Inputting the third sample set into the second prediction model for question-answering training to obtain a third prediction model, wherein the third prediction model has the ability to answer the server question; Pruning the third prediction model to obtain the target question-answering model; The step of inputting the second sample set into the first prediction model to perform prediction sentence sequence training to obtain a second prediction model includes: Recording the order of the two server statements included in each group of positive samples in the second sample set as the second number of second true labels; Inputting the second sample set into the first prediction model to obtain a second number of second prediction labels, where the second prediction labels are used to indicate a prediction order of the two server statements included in each group of positive samples or each group of negative samples; Comparing the second number of second predicted labels with the second number of second true labels to obtain a second evaluation result of the first prediction model, where the second evaluation result is used to indicate a ratio of the number of times the second predicted labels are consistent with the second true labels to the second number; Determining whether the second evaluation result is greater than or equal to a second threshold; If yes, the trained first prediction model is determined as the second prediction model; If not, adjust the model parameters of the trained first prediction model and retrain based on the second sample set until the second evaluation result is greater than or equal to the second threshold.

2. The method according to claim 1, characterized in that The obtaining of L server statements includes: Obtain at least one server document, each server document including N first server statements, and each first server statement being in a first language type; translating each of the first server statements expressed in the first language type into a second server statement expressed in a second language type, thereby obtaining N second server statements, wherein one of the second server statements expressed in the second language type can be translated into at least one server statement expressed in the first language type; Each of the second server statements expressed in the second language type is translated into a server statement expressed in the first language type to obtain the L server statements, where L≥N.

3. The method according to claim 2, characterized in that The number of the at least one target word is a first number; and the inputting the first sample set into the initial model for word prediction task training to obtain a first prediction model includes: using the first number of target words as the first number of first true labels; Inputting the first sample set into the initial model for word prediction task training to obtain a first number of first prediction labels, where the first prediction labels are used to indicate predicted words corresponding to the target words; Comparing the first number of first predicted labels with the first number of first true labels to obtain a first evaluation result of the initial model, where the first evaluation result is used to indicate a ratio of the number of times the first predicted label is consistent with the first true label to the first number; Determining whether the first evaluation result is greater than or equal to a first threshold; If yes, the trained initial model is determined as the first prediction model; If not, adjust the model parameters of the trained initial model and retrain based on the first sample set until the first evaluation result is greater than or equal to the first threshold.

4. The method according to claim 3, characterized in that The third sample set includes a third number of server questions; each server question corresponds to a server answer; and inputting the third sample set into the second prediction model for question-answering training to obtain a third prediction model, including: using each server question in the third number of server questions and the server answer corresponding to each server question as a third number of third true labels; Inputting the third sample set into the second prediction model to obtain a third number of third prediction labels, wherein the third prediction labels are used to indicate a predicted server answer corresponding to each service question; Comparing the third number of third predicted labels with the third number of third true labels to obtain a third evaluation result of the second prediction model, wherein the third evaluation result is used to indicate a ratio of the number of times the third predicted labels are consistent with the third true labels to the third number; Determining whether the third evaluation result is greater than or equal to a third threshold; If yes, the trained second prediction model is determined as the third prediction model; If not, the model parameters of the trained second prediction model are adjusted, and retraining is performed based on the third sample set until the third evaluation result is greater than or equal to the third threshold.

5. The method according to claim 3, characterized in that The third prediction model includes an embedding layer, an attention layer, and a fully connected layer; and the pruning process of the third prediction model to obtain the target question-answering model includes: Reduce the dimension of the embedding vector in the embedding layer, the number of attention heads in the attention layer and the dimension of each attention head, or the number of hidden units in the fully connected layer included in the third prediction model to obtain the target question-answering model after pruning, wherein the memory resources required by the target question-answering model are less than or equal to the fourth threshold, and the required central processing unit resources are less than or equal to the fifth threshold.

6. A training device for a question-answering model, characterized in that: The device comprises: An acquisition module is used to obtain an initial model and L server statements, where L is an integer and is greater than or equal to 1. The acquisition module is further configured to parse the L server statements to obtain a plurality of words corresponding to each of the L server statements, a grammatical structure and content of each server statement, and dependency relationships among the words included in each server statement; a processing module, configured to extract at least one target word from the plurality of words, and replace the at least one target word with a preset identifier to generate L new server statements, wherein the L new server statements constitute a first sample set; The processing module is further configured to extract multiple groups of two consecutive server statements from the L server statements to obtain multiple groups of positive samples, and sequentially swap the two server statements included in each group of positive samples to obtain multiple groups of negative samples, and combine the multiple groups of positive samples and the multiple groups of negative samples into a second sample set, wherein the second sample set includes a second number of groups of positive samples; The processing module is further configured to generate a plurality of server questions based on the grammatical structure and content of the L server statements, and generate a plurality of server answers corresponding to the plurality of server questions based on dependency relationships between terms included in each of the server statements, and to form a third sample set with the plurality of server questions and the plurality of server answers; The processing module is further configured to train the initial model based on the first sample set, the second sample set, and the third sample set to obtain a target question-answering model, and send the target question-answering model to the baseboard management controller, wherein the target question-answering model has the ability to answer questions in response to the server; The processing module is specifically used to: Inputting the first sample set into the initial model for word prediction task training to obtain a first prediction model, wherein the first prediction model has the ability to predict the word replaced by the preset identifier; Inputting the second sample set into the first prediction model to perform statement sequence prediction training to obtain a second prediction model, wherein the second prediction model has the ability to predict whether the sequence of the two server statements is correct; Inputting the third sample set into the second prediction model for question-answering training to obtain a third prediction model, wherein the third prediction model has the ability to answer the server question; Pruning the third prediction model to obtain the target question-answering model; Recording the order of the two server statements included in each group of positive samples in the second sample set as the second number of second true labels; Inputting the second sample set into the first prediction model to obtain a second number of second prediction labels, where the second prediction labels are used to indicate a prediction order of the two server statements included in each group of positive samples or each group of negative samples; Comparing the second number of second predicted labels with the second number of second true labels to obtain a second evaluation result of the first prediction model, where the second evaluation result is used to indicate a ratio of the number of times the second predicted labels are consistent with the second true labels to the second number; Determining whether the second evaluation result is greater than or equal to a second threshold; If yes, the trained first prediction model is determined as the second prediction model; If not, adjust the model parameters of the trained first prediction model and retrain based on the second sample set until the second evaluation result is greater than or equal to the second threshold.

7. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the training method of the question-answering model according to any one of claims 1 to 5 by executing the computer instructions.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, which are used to enable a computer to execute the training method of the question-answering model according to any one of claims 1 to 5.

Images