Theme mining model training method and device and storage medium

Abstract

The embodiment of the invention provides a topic mining model training method and device and a storage medium, and can be applied to the technical field of natural language process.The method comprises the steps that for any word segmentation category in multiple obtained word segmentation categories, whether the word segmentation category meets semantic consistency or not is determined through a large language model; determining corresponding topics and word counting matrixes of the plurality of sample documents under the word segmentation categories with the semantic consistency, and determining global weights of the plurality of topics; each segmented word category corresponds to one theme, and the number of times of occurrence of the same segmented word in the plurality of sample documents is used as the word count of the segmented word under the corresponding theme; taking the topic and word counting matrix as initial values of topic and word distribution in a topic mining model; the topic mining model is trained through the plurality of sample documents and the global weight until the topic mining model with the target value and the word distribution is obtained, so that through semantic consistency judgment, the accuracy of topic mining is improved, and the training time of the model is shortened.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of natural language processing, and particularly relate to a topic mining model training method, device and storage medium. BACKGROUND

[0002] In the field of text mining on the Internet, topic mining of text refers to mining the potential topics of at least one document, which is often applied in news recommendation scenarios or content classification scenarios.

[0003] In related technologies, a Latent Dirichlet Allocation (LDA) model is used to mine the topics of text, and through multiple rounds of iterative updating and training of the distribution of each word in the text and the topic and the distribution of the text and the topic, the topics of the text are finally determined until convergence.

[0004] However, in the initialization stage of the traditional LDA model, each word in the text is randomly grouped, and such random allocation not only reduces the accuracy of topic mining, but also leads to the need for more rounds of iterative training to make the model converge, thereby increasing the time cost of model training. SUMMARY

[0005] Embodiments of the present application provide a topic mining model training method, device and storage medium, which are used to improve the accuracy of topic mining and reduce the time cost of model training.

[0006] In one aspect, the present application provides a topic mining model training method, which comprises:

[0007] clustering the word segmentation set obtained through the plurality of sample documents to obtain a plurality of word segmentation categories;

[0008] For any word segmentation category, determining whether each word in the word segmentation category meets semantic consistency through a large language model;

[0009] determining a topic and word count matrix corresponding to the plurality of sample documents under the word segmentation category of semantic consistency, and determining global weights of a plurality of topics through the word count of the words in each topic; wherein each word segmentation category corresponds to a topic, and the number of occurrences of the same word in the plurality of sample documents is taken as the word count of the word in the corresponding topic;

[0010] taking the topic and word count matrix as the initial value of the topic and word distribution in the topic mining model;

[0011] training the topic mining model through the plurality of sample documents and the global weights until a topic mining model with a target value of topic and word distribution is obtained.

[0012] Optionally, after determining whether each word segment in the segmentation category satisfies semantic consistency through a large language model, and before determining the topic and word count matrix corresponding to the multiple sample documents under the semantically consistent segmentation category, the method further includes:

[0013] Re-cluster multiple words contained in a word segmentation category that does not have semantic consistency to obtain at least one updated word segmentation category;

[0014] The large language model is used to determine whether each word in the at least one updated word segmentation category satisfies semantic consistency, until a word segmentation category that meets the requirements is obtained.

[0015] Optionally, after obtaining the topic mining model with target values ​​and word distribution, the method further includes:

[0016] For any topic in the topic and word distribution with the target value, the co-occurrence probability of the word segment with other words is obtained through the large language model, and the word segments with co-occurrence probabilities under the topic that are all lower than the first threshold are removed;

[0017] Then, using the large language model, other word segments under the topic are summarized to obtain the topic tags corresponding to the topic.

[0018] Optionally, training the topic mining model using the multiple sample documents and the global weights until a topic mining model with target values ​​and word distributions is obtained includes:

[0019] For the topic and word count matrix, based on the global weight, the probability of each word in the multiple sample documents belonging to each topic is calculated to obtain the multiple topic probabilities of the word;

[0020] The word segmentation is updated to the topic corresponding to the topic with the highest topic probability, and the topic and word counting matrix is ​​updated until a topic mining model with the target value and word distribution is obtained.

[0021] Optionally, determining the global weight of multiple topics by counting the words in the word segmentation under each topic includes:

[0022] For any word segment under the topic, the conditional probability of the word segment is obtained by the word count of the word segment, the word count of all words under the topic, and the word count of all words in the multiple sample documents.

[0023] The global weights of the multiple topics are obtained based on the conditional probabilities of each word segment under multiple topics.

[0024] Optionally, it also includes:

[0025] The large language model is used to sample the documents to be processed in the inference phase of the topic mining model;

[0026] Determine the coverage of the sampled word segments under the multiple topic tags;

[0027] If the coverage rate is less than the second threshold, the topic mining model is then trained and updated.

[0028] Optionally, training and updating the topic mining model includes:

[0029] The document to be processed is identified where the new word segment is located in the sampled word segmentation; the new word segmentation is one that cannot be categorized under the multiple topic tags;

[0030] The topic mining model is retrained using the document containing the new word and the multiple sample documents.

[0031] On one hand, embodiments of this application provide a topic mining model training apparatus, the apparatus comprising:

[0032] The clustering module is used to cluster the word segmentation set obtained from multiple sample documents to obtain multiple word segmentation categories;

[0033] The judgment module is used to determine, for any word segmentation category, whether each word in the segmentation category satisfies semantic consistency through a large language model; determine the topic and word count matrix corresponding to the multiple sample documents under the semantically consistent word segmentation category, and determine the global weight of multiple topics through the word count of the words under each topic; wherein, each word segmentation category corresponds to a topic, and the number of times the same word appears in the multiple sample documents is used as the word count of the word under the corresponding topic;

[0034] The training module is used to use the topic and word count matrix as the initial values ​​of topic and word distribution in the topic mining model; and to train the topic mining model using the multiple sample documents and the global weights until a topic mining model with target topic and word distribution values ​​is obtained.

[0035] Optionally, the clustering module is further configured to:

[0036] Re-cluster multiple words contained in a word segmentation category that does not have semantic consistency to obtain at least one updated word segmentation category;

[0037] The large language model is used to determine whether each word in the at least one updated word segmentation category satisfies semantic consistency, until a word segmentation category that meets the requirements is obtained.

[0038] Optionally, the determination module is further configured to:

[0039] For any topic in the topic and word distribution with the target value, the co-occurrence probability of the word segment with other words is obtained through the large language model, and the word segments with co-occurrence probabilities under the topic that are all lower than the first threshold are removed;

[0040] Then, using the large language model, other word segments under the topic are summarized to obtain the topic tags corresponding to the topic.

[0041] Optionally, the training module is specifically used for:

[0042] For the topic and word count matrix, based on the global weight, the probability of each word in the multiple sample documents belonging to each topic is calculated to obtain the multiple topic probabilities of the word;

[0043] The word segmentation is updated to the topic corresponding to the topic with the highest topic probability, and the topic and word counting matrix is ​​updated until a topic mining model with the target value and word distribution is obtained.

[0044] Optionally, the determination module is specifically used for:

[0045] For any word segment under the topic, the conditional probability of the word segment is obtained by the word count of the word segment, the word count of all words under the topic, and the word count of all words in the multiple sample documents.

[0046] The global weights of the multiple topics are obtained based on the conditional probabilities of each word segment under multiple topics.

[0047] Optionally, the determination module is further configured to:

[0048] The large language model is used to sample the documents to be processed in the inference phase of the topic mining model;

[0049] Determine the coverage of the sampled word segments under the multiple topic tags;

[0050] If the coverage rate is less than the second threshold, the topic mining model is then trained and updated.

[0051] Optionally, the training module is specifically used for:

[0052] The document to be processed is identified where the new word segment is located in the sampled word segmentation; the new word segmentation is one that cannot be categorized under the multiple topic tags;

[0053] The topic mining model is retrained using the document containing the new word and the multiple sample documents.

[0054] On one hand, embodiments of this application provide a computer device, including:

[0055] Memory, used to store program instructions;

[0056] The processor is used to call the program instructions stored in the memory and execute the steps of the topic mining model training method described above according to the obtained program.

[0057] On one hand, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer to perform the steps of the topic mining model training method described above.

[0058] On one hand, embodiments of this application provide a computer program product, including a computer program stored on a computer-readable storage medium, the computer program including program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the topic mining model training method described above.

[0059] In this embodiment, the word segmentation set obtained from multiple sample documents is clustered to obtain multiple word segmentation categories. For any word segmentation category, a large language model is used to determine whether each word in the category satisfies semantic consistency. The topic and word count matrix corresponding to the semantically consistent word segmentation categories of multiple sample documents is determined, and the global weight of multiple topics is determined by the word count of the words under each topic. Each word segmentation category corresponds to a topic, and the number of times the same word appears in multiple sample documents is used as the word count of the word under the corresponding topic. The topic and word count matrix is ​​used as the initial value of the topic and word distribution in the topic mining model. Initial values; the topic mining model is trained using multiple sample documents and global weights until a topic mining model with target values ​​and word distribution is obtained. This utilizes the semantic understanding capabilities of the large language model to perform semantic consistency judgment on word segments in multiple segmentation categories, replacing the random grouping of each word segment in existing technologies. This ensures that the initial topic and word count matrix has a certain semantic relevance, improving the accuracy of topic mining. Furthermore, the global weights calculated by the large language model provide prior knowledge for subsequent topic mining model training, reducing the number of iterations in the topic mining model training and thus lowering the time cost of model training. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1A schematic diagram of a system architecture provided in an embodiment of this application;

[0062] Figure 2 A flowchart illustrating a topic mining model training method provided in an embodiment of this application;

[0063] Figure 3 This is a schematic diagram of the structure of a topic mining model training method provided in an embodiment of this application;

[0064] Figure 4 A schematic diagram of the structure of a topic mining model training device provided in an embodiment of this application;

[0065] Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0066] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0067] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0068] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0069] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0070] The term "module" refers to any known or subsequently developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware and / or software code that is capable of performing the functions associated with that element.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0072] The technical terms used in this invention will be explained below.

[0073] Large Language Model (LLM): refers to a natural language processing model based on large-scale pre-training.

[0074] Latent Dirichlet Allocation (LDA): An unsupervised machine learning algorithm based on probabilistic graphical models, mainly used for topic modeling of discrete data (such as text, image features, etc.).

[0075] Gibbs sampling: A Markov chain Monte Carlo method for approximate inference of topic assignments in LDA models.

[0076] In existing technologies, each word in a sample document is randomly assigned a topic number. This random assignment can lead to multiple words in the same category being completely unrelated. For example, in a news document, "football" and "artificial intelligence" might be randomly assigned to category 1, while "maternal and infant care" and "technology" might be randomly assigned to category 2. However, the correlation between "football" and "artificial intelligence," and between "maternal and infant care" and "technology" is weak. This negatively impacts the accuracy of subsequent topic mining. Furthermore, based on the randomly assigned word categories and their multiple words, training the topic mining model requires more rounds of iterative training to cluster related words into the same category, significantly increasing the time cost of model training.

[0077] The following is a brief introduction to the system architecture diagram applicable to the technical solutions of the embodiments of this application. It should be noted that the process described below is only for illustrating the embodiments of this application and is not intended to limit the scope of the application.

[0078] refer to Figure 1 This is a system architecture diagram applicable to the embodiments of this application. The system architecture includes at least terminal device 101 and server 102. The number of terminal devices 101 can be one or more, and the number of servers 102 can also be one or more. This application does not specifically limit the number of terminal devices 101 and servers 102.

[0079] The terminal device 101 pre-installs an application with model training functions. This application can be a client application, a web application, a mini-program application, etc. The terminal device 101 can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart home appliance, a smart voice interaction device, a smart vehicle-mounted device, etc., but is not limited thereto.

[0080] The server 102 is a server for deploying a topic mining model and a large language model. The server 102 can be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms, but is not limited thereto.

[0081] It should be noted that the method in the embodiments of this application can be executed independently by the terminal device 101 or the server 102, or jointly executed by the terminal device 101 and the server 102.

[0082] In the embodiments of this application, the terminal device 101 and the server 102 can be directly or indirectly communicatively connected through one or more networks. The network can be a wired network or a wireless network. For example, the wireless network can be a mobile cellular network or a Wireless-Fidelity (WIFI) network. Of course, it can also be other possible networks, and the embodiments of this application do not limit this.

[0083] Next, based on Figure 1 the system architecture diagram shown, the embodiments of this application provide a process of a topic mining model training method. The process of this method can be executed by Figure 1 the terminal device 101 shown, or by the server 102, or by the interaction between the terminal device 101 and the server 102. As Figure 2 shown, it includes the following steps:

[0084] Step 201: Cluster the token set obtained from multiple sample documents to obtain multiple token categories.

[0085] Specifically, perform preprocessing operations on the collected sample texts. First, tokenize the sample texts to obtain multiple tokens, remove duplicates from the obtained multiple tokens, and剔除 stop words and low-frequency words, such as stop words "的", "得", "是", etc., and tokens that appear less than 2 times. Generate multiple token categories from the de-duplicated results through a clustering algorithm. Among them, the number of sample texts is not specifically limited in this application.

[0086] For example, multiple word segments include: technology, sports, football, maternal and infant, quantum, basketball, artificial intelligence, university, and airplane. Clustering these multiple word segments yields three categories: [technology, quantum, artificial intelligence], [sports, football, basketball], and [maternal and infant, university, airplane].

[0087] Step 202: For any word segmentation category, determine whether each word segmentation in the segmentation category satisfies semantic consistency through a large language model, determine the topic and word count matrix corresponding to multiple sample documents under the semantically consistent word segmentation category, and determine the global weight of multiple topics through the word count of each word segmentation under each topic.

[0088] Specifically, the large language model evaluates the semantic consistency of each word in a word segmentation category. When there is semantic consistency between any two words, then the words in that word segmentation category satisfy semantic consistency.

[0089] For example, for any word segmentation category, the large language model can determine whether the word segmentation category has semantic consistency using the following calculation formula (1).

[0090]

[0091] Where, sim avg The average semantic consistency of a word segmentation category is represented by k; the total number of words in a word segmentation category is represented by e. i This represents the semantic vector of the i-th word segment; e j This represents the semantic vector of the j-th word segment; cos(e i ,e j ) represents the cosine similarity between the i-th and j-th word segments.

[0092] In addition, large language models can also determine the semantic consistency of word segmentation categories by calculating the Euclidean distance between the i-th and j-th word segments under each segmentation category.

[0093] Furthermore, the large language model intuitively determines the semantically consistent word segmentation categories among multiple word segmentation categories using the following formula (2).

[0094] g n =[M(C1),M(C2),…,M(C n (2)

[0095] Where n represents the number of word segmentation categories; g n An array representing the semantic consistency of multiple word segmentation categories; C1, C2, ..., C n Indicates word segmentation category; M(C1), M(C2), ..., M(Cn () is a two-dimensional label used to characterize the semantic consistency of word segmentation categories.

[0096] When a certain word segmentation category C is determined by calculation formula (1) n sim avg If the value exceeds the set threshold, and if the set threshold is 0.6, then set its corresponding two-dimensional label M(C). n M(C) is 1; otherwise M(C) is 1. n The value is 0.

[0097] In this embodiment of the application, each word segmentation category corresponds to a topic, and the number of times the same word appears in multiple sample documents is used as the word count of the word under the corresponding topic.

[0098] Specifically, for word segmentation categories with semantic consistency that have been determined, multiple sample documents are traversed, and the count of each word is obtained from the multiple sample documents. The word segments of each document are assigned to the corresponding word segmentation category. Based on the multiple word segmentation categories and the count of each word, the topic and word count matrix corresponding to the word segmentation category is obtained. In the topic and word count matrix, the word count of each word under the topic and the total count of all words under the topic are determined to obtain the conditional probability of each word. Based on this, the global weight of multiple topics is calculated.

[0099] For example, if the word segmentation categories [technology, quantum, artificial intelligence] and [sports, football, basketball] have semantic consistency, then the count of each word segment is determined from multiple training documents, such as document 1 [technology, quantum, artificial intelligence, football, sports], document 2 [sports, basketball, technology, artificial intelligence], and document 3 [quantum, artificial intelligence, technology, basketball, football, sports]. Based on the count of each word segment in these two word segmentation categories, two word count matrices are obtained, namely topic 1 [technology: 3; quantum: 2; artificial intelligence: 3; sports: 0; football: 0; basketball: 0] and topic 2 [sports: 3; football: 2; basketball: 2; technology: 0; quantum: 0; artificial intelligence: 0].

[0100] In this embodiment, the semantic understanding capability of the large language model is used to determine whether the word segmentation categories have semantic consistency, thereby obtaining a high-quality topic and word count matrix, providing semantic priors for topic mining modeling, replacing the random allocation in the prior art, and improving the accuracy of topic mining.

[0101] In some embodiments, multiple words contained in a segmentation category that does not have semantic consistency are re-clustered to obtain at least one updated segmentation category; a large language model is used to determine whether each word in the at least one updated segmentation category satisfies semantic consistency, until a segmentation category that meets the requirements is obtained.

[0102] Specifically, when one or more word segmentation categories do not meet semantic consistency, the words in these word segmentation categories are re-clustered to obtain at least one updated word segmentation category. The semantic consistency of the at least one updated word segmentation category is then evaluated through the large language model. If there are still word segmentation categories that do not meet semantic consistency in the at least one updated word segmentation category, the re-clustering is continued and the semantic consistency of the clustering results is evaluated until the word segmentation categories meet the semantic consistency judgment of the large language model.

[0103] If, when the number of re-aggregation exceeds a finite number, there are still multiple individual word segments that are not semantically consistent with the word segments in each word segmentation category, and are not semantically consistent with each other, and cannot be clustered into a word segmentation category, then the individual word segments are deleted.

[0104] For example, if the large language model determines that the word segmentation category [mother and baby, university, airplane, graduate student] lacks semantic consistency, it will re-cluster the four word segments it contains, resulting in two new word segmentation categories: [university, graduate student] and [mother and baby, airplane]. If the large language model determines that the new word segmentation category [university, graduate student] has semantic consistency, but determines that the new word segmentation category [mother and baby, airplane] does not, it will re-cluster the word segments [mother and baby] and [airplane] again, continuing to use the large language model to determine semantic consistency. Ultimately, it will be determined that [mother and baby] and [airplane] lack semantic consistency with other word segments in other categories, such as [university, graduate student], [sports, football, basketball], etc., and also lack semantic consistency with each other. Therefore, the word segments [mother and baby] and [airplane] will be removed.

[0105] In this embodiment, the semantic understanding capability of the large language model is used to evaluate the semantic consistency of word segmentation categories, providing prior knowledge for subsequent topic mining models, generating high-quality topic and word count matrices, avoiding the blindness of random initialization in traditional topic mining models, improving the accuracy of topic mining, and reducing the number of iterations in topic mining model training, thereby reducing the time cost of model training.

[0106] In some embodiments, for any word segment under a topic, the conditional probability of the word segment is obtained by the word count of the word segment, the word count of all words under the topic, and the number of times the word segment appears in multiple sample documents; based on the conditional probabilities of each word segment under multiple topics, the global weight of the multiple topics is obtained.

[0107] Specifically, for any word segment under a topic, firstly, the first probability of the word segment under the topic is obtained, then the second probability of the word segment under multiple sample documents is determined, and based on the first and second probabilities, the conditional probability of the word segment is obtained; the conditional probabilities of each word segment under multiple topics are summed to obtain the summation result, and then the summation result is averaged to obtain the global weight of the topic. The specific calculation formula is shown in the following formula (3):

[0108]

[0109] Among them, W s P(w) represents the semantic weight of a topic; K represents the number of topics; k |T k ) indicates the participle w k In topic T k The first probability; P(w) k ) indicates the participle w k The second probability across all word segments in multiple sample documents.

[0110] For example, if multiple sample documents contain Topic 1 [Football: 10; Goals: 15; Referees: 5] and Topic 2 [Code: 8; Algorithms: 7; Chips: 5], then... Therefore, the conditional probability of the word segmentation "football" is: but Similarly, the conditional probabilities of word segmentation for goals and referees under Topic 1, and the conditional probabilities of word segmentation code, algorithm, and chip under Topic 2 are calculated using the same method, ultimately yielding the global weight W for Topic 1 and Topic 2. s =1.12.

[0111] In this embodiment, the global weights of multiple topics are calculated using a large language model, providing prior knowledge for the training of the subsequent topic mining model. This allows the topic mining model to converge in fewer training rounds, reducing the training time.

[0112] Step 203: Use the topic and word count matrix as the initial value of the topic and word distribution in the topic mining model; train the topic mining model with multiple sample documents and global weights until a topic mining model with the target value of topic and word distribution is obtained.

[0113] Specifically, the topic and word count matrix is ​​input into the topic mining model as the initial values ​​for the topic and word distribution. Based on the initial topic and word distribution, each word segment is reassigned a corresponding topic. During the training process of the topic mining model, when the difference between the training result of round Q and the training result of round Q-1 is extremely small, or when the topic mining model reaches a certain preset number of training rounds, the topic mining model converges, and a topic mining model with the target value is obtained.

[0114] In this embodiment, the word segmentation set obtained from multiple sample documents is clustered to obtain multiple word segmentation categories. For any word segmentation category, the semantic consistency of each word in the category is determined using a large language model. The topic and word count matrix corresponding to the semantically consistent word segmentation categories of multiple sample documents is determined, and the global weight of multiple topics is determined by the word count of the words under each topic. Each word segmentation category corresponds to a topic, and the number of times the same word appears in multiple sample documents is used as the word count of the word under the corresponding topic. The topic and word count matrix is ​​used as the initial value of the topic and word distribution in the topic mining model. The topic mining model is trained using multiple sample documents and global weights until a topic mining model with target values ​​and word distributions is obtained. This leverages the semantic understanding capabilities of the large language model to perform semantic consistency judgment on word segments in multiple segmentation categories, replacing the random grouping of each word segment in existing technologies. This ensures that the initial topic and word count matrix has a certain semantic relevance, improving the accuracy of topic mining. Furthermore, the global weights calculated by the large language model provide prior knowledge for subsequent topic mining model training, reducing the number of iterations in the topic mining model training and thus lowering the time cost of model training.

[0115] In some embodiments, for the topic and word count matrix, based on global weights, the probability of each word segment belonging to each topic in multiple sample documents is calculated to obtain multiple topic probabilities of the word segment; the word segment is updated to the topic corresponding to the topic with the highest topic probability and the topic and word count matrix is ​​updated until a topic mining model with topic and word distribution with target value is obtained.

[0116] Specifically, when calculating the topic probability of a word being assigned to various topics, the count of a word under its initial topic is first decremented by 1. Then, the global weights are substituted to calculate the topic probabilities of that word under its initial topic and other topics. Under the topic corresponding to the highest topic probability, the count of a word is incremented by 1. In each round of training iteration, the topic probability calculation for the corresponding count of each word is performed.

[0117] The topic mining model can be Probabilistic Latent Semantic Analysis (pLSA), Hierarchical Latent Dirichlet Allocation (hLDA), or Latent Dirichlet Allocation (LDA), and this application does not impose any specific limitations on it.

[0118] The training process of topic mining models will be introduced below, taking the most widely used topic mining model (LDA model) as an example.

[0119] During the training of the LDA model, based on the initial distribution of topics and words, the topic assignment of each word is updated according to the Gibbs sampling formula. The calculation formula is shown in the following formula (4):

[0120]

[0121] Wherein, P(z) i =t|w i ,z -i ,w -i ) indicates that the current word w is known. i The respective themes of all other word segments z -i All other word segmentations w -i In the case of the current word w i The probability of being assigned to topic t; n wt Indicates the current word w i The count of those assigned to topic t; n t β represents the count of all word segments under topic t; β represents the prior hyperparameters of topic and word distribution; W s Wβ represents the global weights of multiple topics in the topic and word distribution matrix; W in Wβ represents the number of words in all training texts; n td n represents the count of topic t in a training document d; d Tα represents the total count of all words in the training document; α represents the prior hyperparameter of the document and topic distribution; T in Tα represents the number of topics.

[0122] Suppose there are two topic and word distributions: Topic 1 [Football: 10; Goal: 15; Referee: 5] and Topic 2 [Code: 8; Algorithm: 7; Chip: 5]. Assume the count for Topic 1 in training document d is 25 (i.e., the three word segments under Topic 1 appear a total of 25 times in document d, of which "goal" appears 15 times), and the count for Topic 2 in training document d is 10. Calculate the probability that the word segment "goal" is assigned to Topic 1 and Topic 2, setting α and β to 0.5 respectively.

[0123] First, subtract 1 from the number of goals in theme 1 [Football: 10; Goals: 15; Referee: 5] to get theme 1 [Football: 10; Goals: 14; Referee: 5].

[0124] The probability of "scoring a goal" being assigned to Topic 1:

[0125] The probability of "scoring a goal" being assigned to Topic 2:

[0126] It is evident that the probability of the word segment "goal" being assigned to topic 2 is much lower than that of being assigned to topic 1. Therefore, the number of goals in topic 1 [football: 10; goals: 14; referee: 5] is increased by 1 to obtain topic 1 [football: 10; goals: 15; referee: 5].

[0127] When the probability obtained by direct calculation is small, the probability can be normalized to obtain a normalized probability. By comparing the normalized probabilities, it is possible to more clearly determine which topic the word segmentation is assigned to.

[0128] For example, the normalized probability of "scoring a goal" being assigned to Topic 1:

[0129] The normalized probability of "scoring a goal" being assigned to Topic 1:

[0130] Since 0.02 is much smaller than 0.98, "goal" will be assigned to Topic 1. Simultaneously, the same topic probability calculation steps are performed for the remaining 14 "goals". Similarly, for the word segment "football" with a count of 10 and the word segment "referee" with a count of 5 in Topic 1, the corresponding topic probability calculation is performed; for the word segment "code" with a count of 8, the word segment "algorithm" with a count of 7, and the word segment "chip" with a count of 5 in Topic 2, the corresponding number of counts of topic probability calculation are performed.

[0131] In this embodiment, the global weights obtained by the large language model are incorporated into the sampling formula to train the topic mining model, thereby improving the accuracy of topic mining.

[0132] In some embodiments, for any topic in the topic and word distribution with a target value, the co-occurrence probability of the word segment with other words is obtained through a large language model, and the word segments with co-occurrence probabilities under the topic that are all lower than a first threshold are removed; and the other word segments under the topic are summarized through the large language model to obtain the topic tags corresponding to the topic.

[0133] Specifically, for the topic and word distribution with target values ​​output at the end of the topic mining model training, for any topic, the large language model is used again to obtain the co-occurrence frequency of every two words in multiple word segments. If there is a word whose co-occurrence frequency with other words is lower than the first threshold, then that word is removed from the topic.

[0134] The specific calculation formula is shown in the following formula (5):

[0135]

[0136] Among them, C t Indicates the participle w i With the word segment w jThe co-occurrence probability between them; λ represents the decay factor; D(w) i ,w j ) indicates the participle w i With the word segment w j Semantic distance.

[0137] For example, if there is a topic and its segments in the topic and word distribution with target values: Topic 1 [baby, mother, constellation], the first threshold is set to 0.5, and the large language model calculates and determines NPMI(baby, mother) = 0.95, NPMI(baby, constellation) = 0.3, NPMI(mother, constellation) = 0.3 through formula (4). It can be seen that the co-occurrence probability between the segment constellation and the segments baby and mother is less than 0.5. Then, the segment constellation is removed under topic 1, and based on the remaining segments baby and mother, the topic label corresponding to topic 1 is summarized as the maternal and infant topic.

[0138] It should be noted that in this example, each word segment is counted. The actual topic mining model outputs topic and word distributions that include word segments and their counts. The topic and word distributions are used to determine the document and topic distributions of any training document.

[0139] For example, the topic and word distribution with target values ​​includes: Topic 1 [Football: 10; Goals: 15; Referee: 5], Topic 2 [Code: 8; Algorithm: 7; Chip: 5]. By traversing the training documents, we know: Document 1 [Football: 10; Goals: 15; Algorithm: 4], Document 2 [Referee: 5; Code: 8; Algorithm 3; Chip: 5]. Combining the topic and word distribution with target values, we can determine the topic distribution in Document 1 and Document 2, i.e., Document 1 [Topic 1: 25; Topic 2: 4], Document 2 [Topic 1: 5; Topic 2: 16].

[0140] In this embodiment, after training the topic mining model with multiple sample documents until a topic mining model with target values ​​and word distributions is obtained, the co-occurrence probability of the topic and word distributions with target values ​​is judged to eliminate the erroneous allocation in the iterative sampling process of the topic mining model. Based on the remaining more accurate word segmentation, the topic tags corresponding to the topic are summarized to ensure the interpretability and accuracy of the topic, make up for the semantic drift problem in the later stage of the topic mining model, and thus improve the accuracy of topic mining.

[0141] In some embodiments, the document to be processed by the topic mining model during the inference phase is sampled using a large language model; the coverage of the sampled word segmentation under the multiple topic tags is determined; if the coverage is less than a second threshold, the topic mining model is trained and updated.

[0142] Specifically, the large language model monitors the inference phase of the topic mining model and periodically collects the documents to be processed during the inference phase. Multiple new word segments are obtained from these documents, and the large language model evaluates whether these new word segments satisfy semantic consistency with existing topic tags and their segments. The semantic consistency between each pair of word segments can be judged using cosine similarity in formula (1), or by calculating Euclidean distance, etc., thus determining whether the new word segments and existing topic tags have semantic consistency.

[0143] The formula for determining whether a new word segment can be categorized into topic tags using a large language model is shown in the following formula (6):

[0144]

[0145] Among them, w i Indicates the current new word segmentation; L indicates topic tag; LLM score The semantic score represents the judgment of the large language model, when LLM score =1 indicates that there is semantic consistency between the new word segmentation and the topic tag, that is, the new word segmentation can be covered by the topic tag; when LLM score =0 indicates that the new word segment and the topic tag do not have semantic consistency, that is, the new word segment cannot be covered by the topic tag.

[0146] Furthermore, after obtaining the number of new word segments that can be covered by existing topic tags, the coverage rate of existing topic tags over multiple new word segments is determined based on the number of new word segments that can be covered by existing topic tags. The formula for calculating the coverage rate of existing topic tags over multiple new word segments is shown in the following formula (7):

[0147]

[0148] Among them, P covered This indicates the coverage of existing topic tags for multiple new word segments; W sample Indicates the total number of new word segments; This represents the number of new word segments that can be incorporated into existing topic tags, i.e., the number of new word segments with a semantic consistency of 1 obtained from formula (6).

[0149] For example, existing topic tags and their corresponding word segments are: technology topic [technology, quantum, artificial intelligence], maternal and infant topic [maternal and infant, milk powder, diapers], and environmental protection topic [environmental protection, energy saving, recycling], with a preset coverage threshold of 0.8. New word segments obtained from the document to be processed include: cloud computing, marathon, and blockchain. LLM is obtained through a large language model. score (Blockchain, Technology Theme) = 1. LLM score (Blockchain, Technology Theme) = 1. LLMscore (Marathon, Technology Theme) = 0, LLM score (Marathon, Mother and Baby Theme) = 0, LLM score (Marathon, Environmental Protection Theme) = 0, meaning the new word segment "marathon" cannot be covered by existing themes. Therefore, the coverage rate of these three new word segments is... If the calculated coverage is less than a preset threshold, then the topic mining model needs to be trained and updated.

[0150] In this embodiment, the topic mining model is monitored by a large language model, and the coverage between the new word segment and the existing topic tags is determined in a timely manner. Based on the coverage, it is determined whether the topic mining model needs to be trained and updated. This forms a closed-loop improvement and enhances the robustness of the topic mining model.

[0151] In some embodiments, the document to be processed containing the new word in the sampled word segmentation is determined; the new word cannot be categorized under multiple topic tags; the document to be processed containing the new word and the multiple sample documents are retrained to obtain the topic mining model.

[0152] Specifically, if the new word segmentation in the document to be processed cannot be categorized under multiple topic tags, the document to be processed and multiple sample documents are put into preprocessing together. The topic mining model is retrained based on the document to be processed and multiple sample documents using a feedback mechanism, so as to obtain a topic mining model with new target values ​​and word distribution.

[0153] In this embodiment of the application, based on the actual application of the pre-trained topic mining model in the inference stage, the coverage rate of existing topic tags to new word segmentation is calculated to determine whether the quality of the pre-trained topic mining model has deteriorated. Combined with the feedback mechanism, the topic mining model with deteriorated quality is automatically retrained to continuously optimize the topic mining model.

[0154] To better explain the embodiments of this application, the following describes a method for training a topic mining model provided by the embodiments of this application, in conjunction with the LDA model. (Refer to...) Figure 3 The process is mainly divided into three stages: initialization stage, LDA modeling stage, and post-correction stage, with the operation of each stage coordinated by a scheduling center deployed on the server.

[0155] In the initialization phase, the training documents are first preprocessed using a large language model. Preprocessing operations include clustering and deduplication to obtain multiple word segmentation categories. The large language model then performs semantic consistency checks on each of these categories. For categories lacking semantic consistency, the words are re-clustered and semantic consistency checks are repeated until all final word segmentation categories meet the semantic consistency requirements. Each word is counted within the training document. Based on the word category and the count of each word, a topic and word count matrix is ​​obtained. The global weight of each topic is determined based on the word count of the words under each topic. Here, topics are represented by topic numbers.

[0156] In the LDA modeling phase, the topic and word count matrices obtained in the initialization phase are used as the initial values ​​for the topic and word distribution of the LDA model. Based on global weights, the LDA model is trained. The training process involves multiple rounds of Gibbs iteration sampling to adjust the counts of each word segment under each topic in the topic and word count matrices, thereby obtaining an LDA model with the target topic and word distribution.

[0157] In the post-correction phase, the topic and word distribution output from the LDA modeling phase are obtained. A large language model is used to determine the co-occurrence probability of each word segment under each topic, removing multiple words with excessively low co-occurrence probabilities. Based on the remaining words under each topic, topic tags corresponding to that topic are summarized. The document to be processed, input to the pre-trained LDA model during the inference phase, is obtained, and new word segments are extracted from it. The large language model is used to calculate the coverage rate between existing topic tags and new word segments. If the coverage rate is less than a second threshold, the document is preprocessed to obtain a new topic and word count matrix. This new topic and word count matrix is ​​used as the initial values ​​for the topic and word distribution of the LDA model for training.

[0158] When the amount of data is small, the documents to be processed and the training documents can be preprocessed together, and the LDA model can be retrained using the documents to be processed and the training documents.

[0159] In this embodiment, the large language model and the topic mining model are integrated in stages during the initialization phase, the post-correction phase, and the training phase of the topic mining model. This integrates the strong semantic understanding capability of the large language model with the probabilistic modeling advantage of the topic mining model, solving the problem of random grouping of words during initialization in traditional topic mining models, reducing computational overhead. It also overcomes the input length limitation and illusion problem that may occur when generating topics using only the large language model. It can stably generate accurate and comprehensive topic and word distributions, thereby obtaining the corresponding document and topic distributions, and adapting to diverse text topic mining tasks, such as news recommendation or public opinion analysis.

[0160] Based on the same technical concept, this application provides a schematic diagram of the structure of a topic mining model training device, such as... Figure 4 As shown, the topic mining model training device 400 includes:

[0161] Clustering module 401 is used to cluster the word segmentation set obtained from multiple sample documents to obtain multiple word segmentation categories;

[0162] The judgment module 402 is used to determine, for any word segmentation category, whether each word in the word segmentation category satisfies semantic consistency through a large language model; determine the topic and word count matrix corresponding to the multiple sample documents under the semantically consistent word segmentation category, and determine the global weight of each topic through the word count of the word segmentation under each topic; wherein, each word segmentation category corresponds to a topic, and the number of times the same word appears in the multiple sample documents is used as the word count of the word segmentation under the corresponding topic;

[0163] The training module 403 is used to use the topic and word count matrix as the initial value of the topic and word distribution in the topic mining model; and to train the topic mining model using the multiple sample documents and the global weights until a topic mining model with the target value of topic and word distribution is obtained.

[0164] Optionally, the clustering module 401 is further configured to:

[0165] Re-cluster multiple words contained in a word segmentation category that does not have semantic consistency to obtain at least one updated word segmentation category;

[0166] The large language model is used to determine whether each word in the at least one updated word segmentation category satisfies semantic consistency, until a word segmentation category that meets the requirements is obtained.

[0167] Optionally, the determination module 402 is further configured to:

[0168] For any topic in the topic and word distribution with the target value, the co-occurrence probability of the word segment with other words is obtained through the large language model, and the word segments with co-occurrence probabilities under the topic that are all lower than the first threshold are removed;

[0169] Then, using the large language model, other word segments under the topic are summarized to obtain the topic tags corresponding to the topic.

[0170] Optionally, the training module 403 is specifically used for:

[0171] For the topic and word count matrix, based on the global weight, the probability of each word in the multiple sample documents belonging to each topic is calculated to obtain the multiple topic probabilities of the word;

[0172] The word segmentation is updated to the topic corresponding to the topic with the highest topic probability, and the topic and word counting matrix is ​​updated until a topic mining model with the target value and word distribution is obtained.

[0173] Optionally, the determination module 402 is specifically used for:

[0174] For any word segment under the topic, the conditional probability of the word segment is obtained by the word count of the word segment, the word count of all words under the topic, and the word count of all words in the multiple sample documents.

[0175] The global weights of the multiple topics are obtained based on the conditional probabilities of each word segment under multiple topics.

[0176] Optionally, the determination module 402 is further configured to:

[0177] The large language model is used to sample the documents to be processed in the inference phase of the topic mining model;

[0178] Determine the coverage of the sampled word segments under the multiple topic tags;

[0179] If the coverage rate is less than the second threshold, the topic mining model is then trained and updated.

[0180] Optionally, the training module 403 is specifically used for:

[0181] The document to be processed is identified where the new word segment is located in the sampled word segmentation; the new word segmentation is one that cannot be categorized under the multiple topic tags;

[0182] The topic mining model is retrained using the document containing the new word and the multiple sample documents.

[0183] In this embodiment, the word segmentation set obtained from multiple sample documents is clustered to obtain multiple word segmentation categories. For any word segmentation category, the semantic consistency of each word segmentation in the category is determined using a large language model. The topic and word count matrix corresponding to the semantically consistent word segmentation categories of multiple sample documents is determined, and the global weight of each topic is determined by the word count of the words under each topic. Each word segmentation category corresponds to a topic, and the number of times the same word appears in multiple sample documents is used as the word count of the word under the corresponding topic. The topic and word count matrix is ​​used as the initial value of the topic and word distribution in the topic mining model. The topic mining model is trained using multiple sample documents and global weights until a topic mining model with target values ​​and word distributions is obtained. This leverages the semantic understanding capabilities of the large language model to perform semantic consistency judgment on word segments in multiple segmentation categories, replacing the random grouping of each word segment in existing technologies. This ensures that the initial topic and word count matrix has a certain semantic relevance, improving the accuracy of topic mining. Furthermore, the global weights calculated by the large language model provide prior knowledge for subsequent topic mining model training, reducing the number of iterations in the topic mining model training and thus lowering the time cost of model training.

[0184] Based on the same technical concept, embodiments of this application provide a computer device, which can be... Figure 1 The server shown is as follows: Figure 5 As shown, it includes at least one processor 501 and a memory 502 connected to at least one processor. In this embodiment, the specific linking medium between the processor 501 and the memory 502 is not limited. Figure 5 Taking the connection between processor 501 and memory 502 via a bus as an example, the bus can be divided into address bus, data bus, control bus, etc.

[0185] In this embodiment of the application, the memory 502 stores instructions executed by at least one processor 501. By executing the instructions stored in the memory 502, at least one processor 501 can perform the steps of the topic mining model training method described above.

[0186] The processor 501 is the control center of the computer device, capable of connecting to various parts of the computer device via various interfaces and lines. It trains the topic mining model by running or executing instructions stored in the memory 502 and accessing data stored in the memory 502. Optionally, the processor 501 may include one or more processing modules. The processor 501 may integrate an application processor and a modem processor. The application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may not be integrated into the processor 501. In some embodiments, the processor 501 and the memory 502 may be implemented on the same chip; in other embodiments, they may be implemented on separate chips.

[0187] Processor 501 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0188] Memory 502, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 502 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 502 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer device, but is not limited thereto. In the embodiments of this application, memory 502 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0189] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the above-described topic mining model training method.

[0190] Based on the same inventive concept, embodiments of this application provide a computer program product, including a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer device, cause the computer device to perform the steps of the above-described topic mining model training method.

[0191] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0193] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0194] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0195] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for training a topic mining model, characterized in that, include: Cluster the word segmentation sets obtained from multiple sample documents to obtain multiple word segmentation categories; For any word segmentation category, a large language model is used to determine whether each word segmentation in the segmentation category satisfies semantic consistency. The topic and word count matrix corresponding to the multiple sample documents under the semantically consistent word segmentation category are determined, and the global weight of the multiple topics is determined by the word count of the word segmentation under each topic; wherein, each word segmentation category corresponds to a topic, and the number of times the same word appears in the multiple sample documents is used as the word count of the word segmentation under the corresponding topic; The topic and word count matrix is ​​used as the initial value for the topic and word distribution in the topic mining model; The topic mining model is trained using the multiple sample documents and the global weights until a topic mining model with target values ​​and word distribution is obtained.

2. The method as described in claim 1, characterized in that, After determining whether each word segment in the segmentation category satisfies semantic consistency through a large language model, and before determining the topic and word count matrix corresponding to the multiple sample documents under the semantically consistent segmentation category, the method further includes: Re-cluster multiple words contained in a word segmentation category that does not have semantic consistency to obtain at least one updated word segmentation category; The large language model is used to determine whether each word in the at least one updated word segmentation category satisfies semantic consistency, until a word segmentation category that meets the requirements is obtained.

3. The method as described in claim 1, characterized in that, After obtaining the topic mining model with target values ​​and word distribution, the following is also included: For any topic in the topic and word distribution with the target value, the co-occurrence probability of the word segment with other words is obtained through the large language model, and the word segments with co-occurrence probabilities under the topic that are all lower than the first threshold are removed; Then, using the large language model, other word segments under the topic are summarized to obtain the topic tags corresponding to the topic.

4. The method as described in claim 1, characterized in that, The process of training the topic mining model using the multiple sample documents and the global weights until a topic mining model with target values ​​and word distributions is obtained includes: For the topic and word count matrix, based on the global weight, the probability of each word in the multiple sample documents belonging to each topic is calculated to obtain the multiple topic probabilities of the word; The word segmentation is updated to the topic corresponding to the topic with the highest topic probability, and the topic and word counting matrix is ​​updated until a topic mining model with the target value and word distribution is obtained.

5. The method as described in claim 1, characterized in that, The process of determining the global weight of multiple topics by counting the words segmented under each topic includes: For any word segment under the topic, the conditional probability of the word segment is obtained by the word count of the word segment, the word count of all words under the topic, and the word count of all words in the multiple sample documents. The global weights of the multiple topics are obtained based on the conditional probabilities of each word segment under multiple topics.

6. The method according to any one of claims 1 to 5, characterized in that, Also includes: The large language model is used to sample the documents to be processed in the inference phase of the topic mining model; Determine the coverage of the sampled word segments under the multiple topic tags; If the coverage rate is less than the second threshold, the topic mining model is then trained and updated.

7. The method as described in claim 6, characterized in that, The training and updating of the topic mining model includes: The document to be processed is identified where the new word segment is located in the sampled word segmentation; the new word segmentation is one that cannot be categorized under the multiple topic tags; The topic mining model is retrained using the document containing the new word and the multiple sample documents.

8. A computer device, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke program instructions stored in the memory and execute the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a computer-readable storage medium, the computer program including program instructions that, when executed by a computer device, cause the computer device to perform the steps of the method as described in any one of claims 1 to 7.

Images