Artificial intelligence scoring training method and device

Abstract

This application provides an artificial intelligence scoring training method and device, including: obtaining raw data for training artificial intelligence, including comparison data and iteration data; cleaning the raw data according to preset data cleaning rules to generate sample data; extracting classification-regression model samples corresponding to the sample data attributes based on the sample data attributes; constructing a similarity classification-regression model based on the classification-regression model samples; obtaining user data and analyzing the similarity of the user data based on the similarity classification-regression model. This method combines data collection, data cleaning and augmentation, model architecture design, model training, and model launch processes, and uses multiple natural language processing technologies such as attention mechanisms, twin networks, and pre-trained models to generate semantic similarity scores between students and training standards. This enables the robot to understand the student's training level and recognize the student's subsequent intentions, thereby achieving intelligent human-machine training.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and more specifically, to an artificial intelligence scoring training method and device. Background Art

[0002] In the insurance industry, traditional training typically relies on face-to-face instruction, requiring both students and instructors to be present for instruction and exams. However, the high turnover of trainees, such as sales and customer service personnel, makes traditional training inefficient and assessment more difficult. Furthermore, each instructor has their own unique training methods, resulting in low retention and reusability. Summary of the Invention

[0003] The purpose of the embodiments of the present application is to provide an artificial intelligence scoring training method and device to solve the technical problems existing in the prior art.

[0004] In a first aspect, an embodiment of the present application provides an artificial intelligence scoring training method, comprising: obtaining original data for training artificial intelligence, the original data including comparison data and iteration data; cleaning the original data according to preset data cleaning rules to generate sample data; extracting classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data; constructing a similarity classification-regression model based on the classification-regression model samples; obtaining user data, and analyzing the similarity of the user data according to the similarity classification-regression model.

[0005] In an optional embodiment, the raw data used to train artificial intelligence includes: speech recognition text and phoneme samples corresponding to the speech recognition text; according to preset data cleaning rules, the raw data is cleaned to generate sample data, including: regularizing the speech recognition text and phoneme samples into two fields; extracting characters in the field and deleting invalid characters in the field; matching punctuation in the field according to a preset character comparison table; extracting non-Chinese characters from the characters and matching phonemes to the non-Chinese characters.

[0006] In an optional embodiment, the attributes of the sample data include text samples; based on the attributes of the sample data, a classification-regression model sample corresponding to the attributes of the sample data is extracted, including: mixing general corpus data with standard text to generate a pre-trained model sample.

[0007] In an optional embodiment, the attributes of the sample data include text samples; based on the attributes of the sample data, extracting classification-regression model samples corresponding to the attributes of the sample data, also includes: extracting two text samples from the same source as positive samples; extracting two text samples from different sources as negative samples with replacement; judging whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discarding a number of elements to generate a text classification model sample.

[0008] In an optional embodiment, a similarity classification-regression model is constructed based on the classification-regression model sample, including: constructing a pre-trained twin network similarity classification model based on the pre-trained model sample and the text classification model sample.

[0009] In an optional embodiment, the attributes of the sample data include text samples; based on the attributes of the sample data, extracting classification-regression model samples corresponding to the attributes of the sample data, also includes: discarding a preset length of the text sample at a random position to generate a first processed sample; repeatedly extracting the first processed sample and any text sample in the sample data with replacement to form a second processed sample with target similarity; judging whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discarding a number of elements to generate a text regression model sample.

[0010] In an optional embodiment, a similarity classification-regression model is constructed based on the classification-regression model samples, including: constructing a pre-trained embedded dual-tower structure similarity regression model based on the pre-trained model samples and the text regression model samples.

[0011] In an optional embodiment, the attributes of the sample data include phoneme samples; based on the attributes of the sample data, extracting classification-regression model samples corresponding to the attributes of the sample data also includes: extracting two phoneme samples from the same source as positive samples; extracting two phoneme samples from different sources as negative samples with replacement; judging whether the number of sentence elements in the comparison data and the iterative data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discarding a number of elements to generate phoneme classification model samples.

[0012] In an optional embodiment, a similarity classification-regression model is constructed based on the classification-regression model sample, including: constructing an attention-convolutional twin network similarity classification model based on the phoneme classification model sample.

[0013] In the second aspect, an embodiment of the present application also provides a corpus synthesis device, including: a first acquisition module, used to obtain original data for training artificial intelligence, the original data including comparison data and iterative data; a first generation module, used to clean the original data according to preset data cleaning rules and generate sample data; a first extraction module, used to extract classification-regression model samples corresponding to the attributes of the sample data based on the attributes of the sample data; a first construction module, used to construct a similarity classification-regression model based on the classification-regression model samples; a first analysis module, used to obtain user data and analyze the similarity of the user data based on the similarity classification-regression model.

[0014] In an optional embodiment, the original data used to train artificial intelligence includes: speech recognition text and phoneme samples corresponding to the speech recognition text; the first generation module is also used to: organize the speech recognition text and phoneme samples into two fields; extract characters in the field and delete invalid characters in the field; match punctuation in the field according to a preset character comparison table; extract non-Chinese characters from the characters and match phonemes to the non-Chinese characters.

[0015] In an optional embodiment, the attributes of the sample data include text samples; the first extraction module is further used to: mix general corpus data with the standard text to generate a pre-trained model sample.

[0016] In an optional embodiment, the attributes of the sample data include text samples; the first extraction module is further used to: extract two text samples from the same source as positive samples; extract two text samples from different sources with replacement as negative samples; determine whether the number of sentence elements in the comparison data and the iterative data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discard a number of elements to generate text classification-regression model samples.

[0017] In an optional embodiment, the first construction module is further used to: construct a pre-trained twin network similarity classification model based on the pre-trained model samples and the text classification model samples.

[0018] In an optional embodiment, the attributes of the sample data include text samples; the first extraction module is further used to: discard a preset length of the text sample at a random position to generate a first processed sample; repeatedly extract the first processed sample and any text sample in the sample data with replacement to form a second processed sample with target similarity; determine whether the number of sentence elements in the comparison data and the iterative data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discard a number of elements to generate a text regression model sample.

[0019] In an optional implementation, the first construction module is further used to: construct a pre-trained embedded dual-tower structure similarity regression model based on the pre-trained model samples and the text regression model samples.

[0020] In an optional embodiment, the attributes of the sample data include phoneme samples; the first extraction module is also used to: extract two phoneme samples from the same source as positive samples; extract two phoneme samples from different sources as negative samples with replacement; determine whether the number of sentence elements in the comparison data and the iterative data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discard a number of elements to generate a phoneme classification model sample.

[0021] In an optional embodiment, the first construction module is further used to: construct an attention-convolutional twin network similarity classification model based on the phoneme classification model sample.

[0022] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing a method as described in any one of the aforementioned embodiments.

[0023] The embodiments of the present application provide an artificial intelligence scoring training method and apparatus, which have the following advantages during the corpus synthesis process:

[0024] Based on AI technology, the system replaces instructors with robots, simulates personnel training models in actual scenarios, realizes human-machine intelligent training applications, and helps improve training efficiency.

[0025] The attention-based scoring mechanism and method proposed in this application are used for the core functions of intelligent training robots. They combine process mechanisms such as data collection, data cleaning and augmentation, model architecture design, model training, and model online deployment. Based on multiple natural language processing technologies such as attention mechanism, twin network, and pre-training model, a semantic similarity score between students and training standards is generated. On the one hand, this score enables the robot to have the ability to understand the training level of students, and on the other hand, it has the intelligence to recognize the subsequent intentions of students, thereby achieving intelligent human-computer training. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0027] Figure 1 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0028] Figure 2 A flowchart of an artificial intelligence scoring training method provided in an embodiment of the present application;

[0029] Figure 3 A schematic diagram of the structure of an artificial intelligence scoring training device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0031] like Figure 1 As shown, this embodiment provides an electronic device 1, including: at least one processor 11 and a memory 12, Figure 1 In the example of FIG. 1 , a processor 11 and a memory 12 are connected via a bus 10 . The memory 12 stores instructions that can be executed by the processor 11 , and the instructions are executed by the processor 11 .

[0032] In one embodiment, the electronic device 1 can be a mobile phone, a laptop computer, or other device. Cost-reducing and efficiency-enhancing data collection can be achieved through data collection, combined with the collection of text data in the field, manual annotation and expansion of speech, standard sample entry of trainers, and user sample entry. A sample augmentation and generation method is adopted, and positive and negative samples are generated by extracting them from the same node and across nodes, and words are randomly removed to enhance robustness as classification-regression model samples, and text is retained by percentage to generate regression model samples. Phoneme data is introduced for similarity calculation of intelligent training robots to expand semantic information outside of Chinese characters. Hybrid BM25 algorithm, Attention-CNN twin network structure similarity classification model, pre-trained twin network structure similarity classification model, pre-trained Embedding (embedding) dual-tower structure similarity regression model score, and pre-storage technology is used to speed up model use to achieve real-time call.

[0033] like Figure 2 As shown, it is a flow chart of an artificial intelligence scoring training method provided by an embodiment of the present application, which can be Figure 1 The method is performed by the electronic device 1 shown. The method includes the following steps:

[0034] Step 201: Obtain raw data for training artificial intelligence.

[0035] In this step, the raw data includes comparison data and iteration data. In order for the robot to acquire judgment capabilities, it is necessary to collect data in advance for the training of the intelligent robot. In one embodiment, the comparison data collects the following data, which are applied in different stages:

[0036] 1. Long-term collection of sales and customer service script data in the insurance sector. This is used to enhance the pre-trained model's understanding of domain text and to serve as standard training scripts for selected scripts. Multiple interfaces are used to continuously collect text data from the insurance sector, including introduction documents for various insurance products (including those no longer available), basic terms and conditions for various insurance policies, accumulated sales scripts, telemarketing scripts, basic customer service process terms, and incident handling scripts. These scripts are stored by source interface name, category name, text content, and source date. Scripts expected to go live are stored separately as "golden scripts" in the script and category management system, including information such as the category ID, script ID, script content, and update time. Experts develop extended scripts in two steps: one protects the keywords in the gold-medal scripts and modifies only the non-protected parts. The other allows for free expression, while maintaining semantic consistency, to maximize the scope for correct scripting.

[0037] 2. For specific training subjects, before the training subjects go online, we hire annotation personnel to provide extended scripts of standard scripts and generate text + phoneme samples through conversion for similarity model training.

[0038] 3. For specific training subjects, before the training subjects go live, recruit trainers to provide standard ASR (Automatic Speech Recognition) text and phoneme samples for similarity model training. Trainers will input ASR data based on the speech scripts, obtain the standard ASR text and phoneme data of the speech scripts, and perform simple calculations to match the words and characters with the standard and expanded speech scripts. Samples with low matching scores will be deleted. If the cost is high, this module can be omitted.

[0039] In one embodiment, after the specific training subject system is launched, iterative data can be continuously collected from users whose ASR text and phoneme samples meet similarity requirements for iterative training of the similarity model. After the system is launched, ASR text and phoneme data are continuously collected during user use. Qualified text and phoneme data are initially screened using the existing scoring model, and unqualified samples are removed based on rules such as length requirements and word matching requirements for the final collection.

[0040] Step 202: Clean the original data according to the preset data cleaning rules to generate sample data.

[0041] In this step, the speech recognition text and phoneme samples are organized into two fields, and then the characters in the field are extracted, and the invalid characters in the field are deleted, and the punctuation in the field is matched according to the preset character comparison table. Finally, the non-Chinese characters in the characters are extracted, and the non-Chinese characters are matched with phonemes. Specifically, text-phoneme data separation: some data collection end text and phoneme data are mixed together, and they are cleaned and organized into two fields. Text format cleaning: invalid characters such as line breaks and spaces are removed. Text punctuation cleaning: common punctuation marks are normalized, and wildcards are set for punctuation marks that are not in the common character table. Text number cleaning: the pronunciation of numbers, percent signs, addition, subtraction, multiplication and division, etc. that have Chinese pronunciations are converted.

[0042] Step 203: According to the attributes of the sample data, extract the classification-regression model samples corresponding to the attributes of the sample data.

[0043] In this step, after the collected data is cleaned, some augmentation work is performed to expand and enrich our samples, and finally divided into three parts of samples for use.

[0044] In an optional implementation, the attributes of the sample data include text samples. Sales and customer service speech corpus data in the insurance field are collected over a long period of time. On the one hand, this is used to enhance the pre-training model's ability to understand texts in the field. On the other hand, some golden speech scripts are directly used as standard texts for training subjects. General corpus data is mixed into these texts to generate pre-training model samples, which are then continued to be used for training in the pre-training model.

[0045] In an optional embodiment, two text samples from the same source are extracted as positive samples, and two text samples from different sources are extracted with replacement as negative samples. A determination is made as to whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit. If so, a number of elements are randomly discarded to generate a text classification model sample. Specifically, for each specific node in a specific training subject, a text sample of standard speech, a text sample of extended speech, an ASR standard text sample, and a text sample qualified by an ASR user are used as the text sample for that node.

[0046] Extract two text samples A and B in the same node as positive samples, and then extract them with replacement. Then extract two text samples across nodes A and B as negative samples. The construction ratio of positive and negative samples is γ∈[0.5,0.9] and can be adjusted.

[0047] For sentences with more than 20 elements on either side of A or B, 0-3 elements are randomly discarded to enhance robustness. For sentences exceeding the set length L∈[256,512] (adjustable), they are truncated.

[0048] In an optional embodiment, a text sample is discarded at a random position for a preset length to generate a first processed sample. The first processed sample and any text sample in the sample data are repeatedly extracted with replacement to form a second processed sample with target similarity. It is determined whether the number of sentence elements in the comparison data and the iterative data exceeds a preset upper limit. If the number exceeds the preset upper limit, a number of elements are randomly discarded to generate a text regression model sample. Specifically, for a specific node in each specific training subject, a text sample of standard speech, a text sample of extended speech, an ASR standard text sample, and a text sample qualified by an ASR user are used as the text sample of the node.

[0049] For any text sample A in a certain node, randomly discard 20% of the length of text sample A' and any sample B in the node, and repeat the replacement extraction to form a sample with 80% similarity. If the random position is in the last 20%, continue to discard towards the beginning of the text; extract any text sample A in a certain node, randomly discard 40% of the length of text sample A' and any sample B in the node, and repeat the replacement extraction to form a sample with 60% similarity. If the discarded length L exceeds 20, randomly choose to discard L / 2, and then discard L / 2; and so on to construct samples with similarity of 40% and 20%, and at the same time construct samples with scores of 1 and 0.

[0050] For sentences with more than 20 elements on either side, 0-3 elements are randomly discarded to enhance robustness. For sentences exceeding the set length L∈[256,512] (adjustable), they are truncated.

[0051] In an optional embodiment, two phoneme samples from the same source are extracted as positive samples, and two phoneme samples from different sources are extracted with replacement as negative samples. A determination is made as to whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit. If so, a number of elements are randomly discarded to generate a phoneme classification model sample. Specifically, for each specific node in a specific training subject, the phoneme samples of the standard speech, the phoneme samples of the extended speech, the ASR standard phoneme samples, and the phoneme samples qualified by the ASR user are used as the phoneme samples of that node.

[0052] Two phoneme samples, A and B, within the same node are extracted as positive samples. Two phoneme samples across nodes A and B are extracted with replacement as negative samples. The positive-negative sample ratio is γ∈[0.5,0.9], which is adjustable. For sentences with more than 20 elements on either side of A or B, 0-3 elements are randomly discarded to enhance robustness. For sentences exceeding a set length L∈[256,512], which is adjustable, truncation is performed.

[0053] Step 204: Construct a similarity classification-regression model based on the classification-regression model samples.

[0054] In this step, a pre-trained twin network similarity classification model is constructed based on the pre-trained model samples and the text classification-regression model samples. Specifically, the model includes: (1) Embedding + sentence information extraction layer: the pre-trained model after pre-training is used, and the first three layers of the 12-layer block structure containing self-attention (if there are sufficient computing resources, other layers or layer combinations can also be flexibly selected). The weights on both sides are shared and participate in parameter adjustment. (2) Additive splicing layer: the result of the expression of the two sentences + the absolute value of the difference between the expressions of the two sentences. (3) Fully connected layer: 2 fully connected layers extract interactive information and finally map it to the classification problem: (4) Classification softmax, the loss uses the cross entropy loss function.

[0055] In an optional embodiment, a pre-trained embedding dual-tower structural similarity regression model is constructed based on the pre-trained model samples and the text regression model samples. Specifically, the model includes: (1) Embedding + sentence information extraction layer: directly use the pre-trained model after pre-training, and average the first three layers of tensors in the 12-layer block structure containing self-attention. It is fixed and no longer participates in subsequent parameter training. If the computing resources are sufficient, other layers or layer combinations can be flexibly selected. (2) Additive splicing layer: the result of the expression of two sentences + the absolute value of the difference between the expressions of the two sentences. (3) Fully connected layer: 2 fully connected layers extract interactive information. Finally, it is mapped to the regression problem: (4) Regression sigmoid, and the loss uses the mean square error function.

[0056] In an optional embodiment, an attention-convolution twin network similarity classification model is constructed based on the phoneme classification model sample. Specifically, the model includes: (1) Embedding layer: each sentence is embedded through the Embedding layer. (2) Sentence information extraction layer: 1 self-Attention layer understands the information within the sentence, and 2 CNN layers extract the information within the sentence to obtain the sentence expression. The embedding stage and the sentence information extraction stage share parameters to form a twin. Then enter the interaction stage of the two sentences: (3) Additive splicing layer: the result of the two sentence expressions + the interactive attention between the two sentence expressions + the absolute value of the difference between the two sentence expressions. (4) Fully connected layer: 2 fully connected layers extract interactive information. Finally, it is mapped to the classification problem: (5) Classification softmax, and the loss uses the cross entropy loss function.

[0057] In an optional implementation, an unsupervised similarity evaluation model, BM25, can be used. This model requires no training and primarily compares the similarity of words within two sentences. The main principle is to segment sentence A, calculate the correlation between each morpheme or word and another sentence B, and then calculate the weighted sum.

[0058] Step 205: Obtain user data, and analyze the similarity of the user data based on a similarity classification-regression model.

[0059] In this step, as students participate in training on specific subjects, the robot and the students engage in human-machine dialogue. The invented mechanism receives the student's input and calculates the similarity between the input and all standard scripts in the subject space according to the set model. The highest score is the user intention determined by the mechanism, and the score is also used as the student's performance effect under this standard script. Different business parties such as sales and customer service will have different timeliness requirements and hardware facilities, so the online method can be flexibly adjusted:

[0060] 1. The various structures in the online model mixing step 204 can flexibly adjust the weights [ω_1,ω_2,ω_3,ω_4] of Model1_Score, Model2_Score, Model3_Score, and Model4_Score, where ω_1+ω_2+ω_3+ω_4=1, ω_i,i∈1,2,3,4in[0,1].

[0061] 2. The results of any model embedding + sentence information extraction layer of standard dialogue can be pre-stored. When called in real time, only the model embedding + sentence information extraction layer of user input and subsequent interactions are calculated, which can significantly speed up the process.

[0062] See Figure 3 , which is an artificial intelligence scoring training device 300 provided in an embodiment of the present application, the artificial intelligence scoring training device 300 can be applied to Figure 1 The electronic device 1 shown achieves cost-saving and efficiency-enhancing data collection through data collection, combined with text data collection in the field, manual annotation and expansion of speech, trainer standard sample entry, and user sample entry. The sample augmentation and generation method is adopted, and the positive and negative samples are extracted from the same node and across nodes to generate samples and randomly remove words to enhance robustness as classification-regression model samples, and the text is retained according to the percentage to generate regression model samples. Phoneme data is introduced for similarity calculation of intelligent training robots to expand semantic information outside Chinese words. The BM25 algorithm, Attention-CNN twin network structure similarity classification model, pre-trained twin network structure similarity classification model, pre-trained Embedding (embedding) dual tower structure similarity regression model score are mixed, and pre-storage technology is used to speed up the use of the model to achieve real-time call. The device includes: a first acquisition module 301, a first generation module 302, a first extraction module 303, a first construction module 304 and a first analysis module 305. The specific principle relationship is as follows:

[0063] The first acquisition module 301 is used to acquire raw data for training artificial intelligence, the raw data including comparison data and iteration data. For details, please refer to the description of step 201 in the above embodiment.

[0064] The first generating module 302 is used to clean the original data according to the preset data cleaning rules to generate sample data.

[0065] In an optional embodiment, the raw data used to train the artificial intelligence includes: speech recognition text and phoneme samples corresponding to the speech recognition text; the first generation module 302 is further configured to: organize the speech recognition text and phoneme samples into two fields; extract characters from the fields and delete invalid characters in the fields; match punctuation in the fields according to a preset character comparison table; extract non-Chinese characters from the characters and match phonemes to the non-Chinese characters. For details, please refer to the description of step 202 in the above embodiment.

[0066] The first extraction module 303 is used to extract classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data.

[0067] In an optional embodiment, the attributes of the sample data include text samples; the first extraction module 303 is further used to: mix general corpus data with the standard text to generate a pre-trained model sample.

[0068] In an optional embodiment, the attributes of the sample data include text samples; the first extraction module 303 is further used to: extract two text samples from the same source as positive samples; extract two text samples from different sources with replacement as negative samples; determine whether the number of sentence elements in the comparison data and the iterative data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discard a number of elements to generate text classification-regression model samples.

[0069] In an optional embodiment, the attributes of the sample data include text samples; the first extraction module 303 is further used to: discard a preset length of the text sample at a random position to generate a first processed sample; repeatedly extract the first processed sample and any text sample in the sample data with replacement to form a second processed sample with target similarity; determine whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; if it exceeds the preset upper limit, randomly discard a number of elements to generate a text regression model sample.

[0070] In an optional embodiment, the attributes of the sample data include phoneme samples; the first extraction module 303 is further configured to: extract two phoneme samples from the same source as positive samples; extract two phoneme samples from different sources with replacement as negative samples; determine whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; and if so, randomly discard a number of elements to generate a phoneme classification model sample. For details, please refer to the description of step 203 in the above embodiment.

[0071] The first construction module 304 is used to construct a similarity classification-regression model based on the classification-regression model samples.

[0072] In an optional embodiment, the first construction module 304 is further used to: construct a pre-trained twin network similarity classification model based on the pre-trained model samples and the text classification model samples.

[0073] In an optional implementation, the first construction module 304 is further configured to construct a pre-trained embedded dual-tower structure similarity regression model based on the pre-trained model samples and the text regression model samples.

[0074] In an optional embodiment, the first construction module 304 is further configured to construct an attention-convolutional twin network similarity classification model based on the phoneme classification model sample. For details, please refer to the description of step 204 in the above embodiment.

[0075] The first analysis module 305 is used to obtain user data and analyze the similarity of the user data based on the similarity classification-regression model. For details, please refer to the description of step 205 in the above embodiment.

[0076] In the embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interface, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0077] In addition, the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0078] Furthermore, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0079] It should be noted that if the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0080] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any actual relationship or order between these entities or operations.

[0081] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. An artificial intelligence scoring training method, characterized in that: include: Acquiring raw data for training artificial intelligence, wherein the raw data includes comparison data and iteration data; Cleaning the raw data according to preset data cleaning rules to generate sample data; Extracting classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data; Constructing a similarity classification-regression model based on the classification-regression model sample; Acquire user data, and analyze the similarity of the user data according to the similarity classification-regression model; The attributes of the sample data include text samples; and extracting classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data includes: Mixing general corpus data with standard text to generate pre-trained model samples; The attributes of the sample data include text samples; and extracting classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data further includes: Extracting two text samples from the same source as positive samples; Extract two text samples from different sources as negative samples with replacement; Determining whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; If the number exceeds the preset upper limit, some elements are randomly discarded to generate text classification model samples; The constructing of a similarity classification-regression model based on the classification-regression model sample includes: Constructing a pre-trained twin network similarity classification model based on the pre-trained model sample and the text classification-regression model sample; The attributes of the sample data include text samples; and extracting classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data further includes: discarding a preset length of the text sample at a random position to generate a first processed sample; The first processed sample and any text sample in the sample data are repeatedly extracted with replacement to form a second processed sample with target similarity; Determining whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; If the number exceeds the preset upper limit, a number of elements are randomly discarded to generate text regression model samples; The constructing of a similarity classification-regression model based on the classification-regression model sample includes: Constructing a pre-trained embedded dual-tower structure similarity regression model based on the pre-trained model sample and the text regression model sample; The attributes of the sample data include phoneme samples; and extracting classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data further includes: Extracting two phoneme samples from the same source as positive samples; Extracting two phoneme samples from different sources as negative samples with replacement; Determining whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; If the number exceeds the preset upper limit, a number of elements are randomly discarded to generate a phoneme classification model sample; The constructing of a similarity classification-regression model based on the classification-regression model sample includes: Based on the phoneme classification model sample, an attention-convolutional twin network similarity classification model is constructed.

2. The method according to claim 1, characterized in that The raw data used to train artificial intelligence includes: speech recognition text and phoneme samples corresponding to the speech recognition text; the raw data is cleaned according to the preset data cleaning rules to generate sample data, including: Regularizing the speech recognition text and the phoneme sample into two fields; Extract characters from the field and delete invalid characters from the field; Matching punctuation in the field according to a preset character comparison table; Non-Chinese characters are extracted from the characters, and phonemes are matched to the non-Chinese characters.

3. An artificial intelligence scoring training device, characterized in that: include: A first acquisition module is used to acquire raw data for training artificial intelligence, wherein the raw data includes comparison data and iteration data; A first generating module, configured to clean the raw data according to preset data cleaning rules to generate sample data; A first extraction module is used to extract classification-regression model samples corresponding to the attributes of the sample data according to the attributes of the sample data; A first construction module is used to construct a similarity classification-regression model based on the classification-regression model sample; A first analysis module, configured to obtain user data and analyze the similarity of the user data according to the similarity classification-regression model; The attributes of the sample data include text samples; the first extraction module is further configured to: Mixing general corpus data with standard text to generate pre-trained model samples; The attributes of the sample data include text samples; the first extraction module is further configured to: Extracting two text samples from the same source as positive samples; Extract two text samples from different sources as negative samples with replacement; Determining whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; If the number exceeds the preset upper limit, some elements are randomly discarded to generate text classification model samples; The first building block is further configured to: Constructing a pre-trained twin network similarity classification model based on the pre-trained model sample and the text classification-regression model sample; The attributes of the sample data include text samples; the first extraction module is further configured to: discarding a preset length of the text sample at a random position to generate a first processed sample; The first processed sample and any text sample in the sample data are repeatedly extracted with replacement to form a second processed sample with target similarity; Determining whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; If the number exceeds the preset upper limit, a number of elements are randomly discarded to generate text regression model samples; The first building block is further configured to: Constructing a pre-trained embedded dual-tower structure similarity regression model based on the pre-trained model sample and the text regression model sample; The attributes of the sample data include phoneme samples; the first extraction module is further configured to: Extracting two phoneme samples from the same source as positive samples; Extracting two phoneme samples from different sources as negative samples with replacement; Determining whether the number of sentence elements in the comparison data and the iteration data exceeds a preset upper limit; If the number exceeds the preset upper limit, a number of elements are randomly discarded to generate a phoneme classification model sample; The first building block is further configured to: Based on the phoneme classification model sample, an attention-convolutional twin network similarity classification model is constructed.

Images