Intelligent auxiliary review method and system

Abstract

The invention discloses an intelligent auxiliary review method and system, and the method comprises the four steps: model design, data set construction, model training and prediction execution, and specifically comprises the steps: constructing a two-stage model comprising a compression stage and a classification stage; defining a classification label of the trigger rule, and determining positive and negative example data of the classification label; setting training times of the data set, and updating model parameters according to the loss function; performing prediction on the trained model; the method comprises the steps of obtaining to-be-processed text data, and analyzing all entries in a text based on a regular expression to obtain an entry list; and inputting the item list into the trained model to obtain a prediction result of the corresponding rule. According to the method, whether the text content triggers the preset rule or not can be automatically identified, and meanwhile, the method has efficient processing capability and good generalization capability so as to meet requirements in different scenes.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an intelligent assisted review method and system. Background Art

[0002] With the acceleration of digital transformation, enterprises and institutions are facing many challenges in processing large amounts of text data. This data may come from various sources, such as customer feedback, market research reports, technical documents, etc. In these scenarios, quickly and accurately identifying whether the text content complies with preset rules (such as data compliance rules or data quality rules) is the key to improving work efficiency and quality.

[0003] Most existing text processing methods rely on rule matching, such as extracting text content with specific patterns through regular expressions. However, these methods have obvious limitations. First, rule matching usually requires a large number of rules to be manually defined, and it is difficult to adapt to different types of text data and lacks generalization capabilities. Second, for complex text content, rule matching is prone to misjudgment or omission. In addition, existing methods are inefficient when processing large-scale data and are difficult to meet real-time requirements. Summary of the invention

[0004] The present application provides an intelligent assisted review method and system that can automatically identify whether the text content triggers preset rules, and at the same time has efficient processing capabilities and good generalization capabilities to meet the needs of different scenarios.

[0005] In the first aspect, an intelligent assisted review method is provided, the method comprising four steps: model design, data set construction, model training, and prediction execution, specifically including: Model design, build a two-stage model including compression stage and classification stage; in the compression stage, first split the text into multiple blocks through the block splitting algorithm, use the BERT model to encode the split text blocks to capture the bidirectional semantic information of the text, and further process the BERT output through the multi-layer perceptron to extract key information and compress the text; in the classification stage, use the BERT model and multi-layer perceptron to classify the compressed text to predict whether it triggers the preset rules; Dataset construction, including defining classification labels that trigger rules, and determining positive and negative example data of classification labels; wherein positive example data is composed of texts that trigger corresponding rules, and negative examples are composed of texts in the patents where positive examples are located that do not trigger corresponding rules, and lists of items extracted from other data that do not trigger corresponding rules; Model training, including setting the number of training times for the dataset and updating model parameters based on the loss function; Execute prediction and use the trained model to perform prediction; wherein, obtain the text data to be processed, parse all entries in the text based on regular expressions, and obtain an entry list; then input the entry list into the trained model to obtain the prediction result of the corresponding rule.

[0006] Optionally, the model design uses a block splitting algorithm to split the text into multiple blocks, specifically including: Set the block splitting parameters, including defining the cost dictionary end_tokens of the block separator, where different symbols including line break, period, semicolon, question mark, exclamation mark and comma are assigned different cost values, and setting the maximum block length BLOCK_MAX to 63 characters and the minimum block length BLOCK_MIN to 10 characters; Use the block separator with cost value less than 2 to preliminarily split the text and get the initial block list init_blocks; Declare an empty final block list final_blocks to store the final blocks; Traverse the initial block list init_blocks, for each block, if its length does not exceed the maximum block length BLOCK_MAX, directly add it to the final block list final_blocks; if the block length exceeds BLOCK_MAX, continue to split it based on the block separator with a cost value greater than 1. If the length of the sub-block after splitting still exceeds BLOCK_MAX, further split it into multiple blocks not exceeding BLOCK_MAX, and add these blocks to final_blocks; Returns the final block list final_blocks for subsequent processing and analysis.

[0007] Optionally, it includes defining the classification labels for triggering rules and determining the positive and negative example data for the classification labels, specifically including: Extract text from the dataset containing positive and negative examples and split the items into items; The list of entries that trigger the preset rules is taken as positive examples; The entries in the dataset where the positive examples are located that do not trigger the preset rules are regarded as negative examples; If the number of negative examples is insufficient, some entries are extracted from the data set that has not triggered the preset rules to supplement the negative examples, so that the positive and negative examples reach the requirement of 1:1; The training set, validation set, and test set are divided into the ratio of 0.7:0.15:0.15.

[0008] Optionally, the preset rules are specifically data compliance rules and data quality rules.

[0009] Optionally, model training includes setting the number of training times for the data set and updating model parameters according to the loss function, specifically including: Set the number of traversals of the dataset to 10 rounds, which means that the entire training process will perform 10 complete traversals of the dataset; Start a training loop for each round. For each round, iterate over each sample pair (x, y) in the training set, where x is the text and y is the corresponding label; For each sample x, define an empty key block list z to store the key blocks in x. The key blocks are screened by deleting any block from x and judging the impact of the deletion on the loss. Continue to traverse each sample pair (x, y) in the training set, obtain the key block list z of each sample x, and define the loss function for calculation; among them, the cross entropy loss is used, involving the category label, the key block list corresponding to the input item list x, and the classification component of the model; Update the model parameters according to the calculated loss to optimize the model's predictive performance. After 10 rounds of training, save and return the trained model.

[0010] Optionally, the loss function is defined as: , in, loss Indicates loss, CrossEntropy () calculates the cross entropy loss, x _ represents the remaining block sequence after deleting any block, judge () represents the relevant probability value of the compressed output, y _ means from y Delete d The tag sequence after the block corresponds to the tag, relv_label () indicates the correlation label calculated based on the manually defined correlation features.

[0011] In the second aspect, an intelligent auxiliary review system includes a model design module, a data set construction module, a model training module, and an execution prediction module, which specifically includes: The model design module is used to build a two-stage model including a compression stage and a classification stage. In the compression stage, the text is first split into multiple blocks by a block splitting algorithm, and the split text blocks are encoded using the BERT model to capture the bidirectional semantic information of the text. The output of BERT is further processed by a multi-layer perceptron to extract key information and compress the text. In the classification stage, the compressed text is classified using the BERT model and the multi-layer perceptron to predict whether it triggers the preset rules. A data set construction module, including a classification label that defines a triggering rule, and positive and negative example data that determine the classification label; wherein the positive example data is composed of text that triggers the corresponding rule, and the negative example is composed of text in the patent where the positive example is located that does not trigger the corresponding rule and a list of items extracted from other data that do not trigger the corresponding rule; Model training module, including setting the number of training times for the data set and updating the model parameters according to the loss function; The prediction module is executed to perform prediction on the trained model; wherein, the text data to be processed is obtained, all entries in the text are parsed based on regular expressions to obtain an entry list; then the entry list is input into the trained model to obtain the prediction result of the corresponding rule.

[0012] In a third aspect, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the intelligent assisted review method described in any one of the first aspects is implemented.

[0013] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the intelligent assisted review method described in any one of the first aspects is implemented.

[0014] In a fifth aspect, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the intelligent assisted review method described in any one of the first aspects.

[0015] It can be seen from the above that the present invention provides an intelligent auxiliary review method and system, which can automatically identify whether the text content triggers a preset rule, and has efficient processing capabilities and good generalization capabilities to meet the needs of different scenarios.

[0016] The invention has good generalization ability based on the model implementation. The invention is based on the two-stage model implementation of compression and classification and is applicable to texts of any length. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0018] Figure 1 The main flow chart provided for the embodiment of the present application; Figure 2 A block diagram of the module architecture of an intelligent assisted review system provided in one embodiment of the present application; Figure 3 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0020] In the description of the present application: the terms "including", "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units that have been explicitly listed, but may also include other steps or units that are not explicitly listed but are inherent to these processes, methods, products or apparatuses, or steps or units added based on further optimization schemes conceived by the present invention.

[0021] In one embodiment, Figure 1 As shown, an intelligent assisted review method is provided, which can be applied to the server. The method includes four steps: model design, data set construction, model training, and prediction execution, specifically including: Model design, build a two-stage model including compression stage and classification stage; in the compression stage, first split the text into multiple blocks through the block splitting algorithm, use the BERT model to encode the split text blocks to capture the bidirectional semantic information of the text, and further process the BERT output through the multi-layer perceptron to extract key information and compress the text; in the classification stage, use the BERT model and multi-layer perceptron to classify the compressed text to predict whether it triggers the preset rules; Dataset construction, including defining classification labels that trigger rules, and determining positive and negative example data of classification labels; wherein positive example data is composed of texts that trigger corresponding rules, and negative examples are composed of texts in the patents where positive examples are located that do not trigger corresponding rules, and lists of items extracted from other data that do not trigger corresponding rules; Model training, including setting the number of training times for the dataset and updating model parameters based on the loss function; Execute prediction and use the trained model to perform prediction; wherein, obtain the text data to be processed, parse all entries in the text based on regular expressions, and obtain an entry list; then input the entry list into the trained model to obtain the prediction result of the corresponding rule.

[0022] In the embodiment of the present application, the two-stage model specifically includes: compression and classification. The two-stage model includes: (1) model architecture; (2) model training; (3) data set.

[0023] Model architectures include compression and classification, such as " Figure 2 The entry list triggers the second and twenty-fifth prediction model architectures. The compression is characterized by including: deblocking, BERT (Bidirectional Encoder Representation from Transformers) and multi-layer perceptron. The deblocking is implemented by "Algorithm 1 Deblocking Algorithm". Algorithm 1 Deblocking Algorithm is specifically as follows: Input: Text Output: List of blocks step: S101: Set the deblocking parameters: S1011: Block separator cost dictionary end_tokens, pseudo code is as follows: end_tokens = { :0, '.':1, ';':1, '?':1, '!':1, ',':2} S1012: Maximum block length: BLOCK_MAX=63, minimum block length: BLOCK_MIN=10.

[0024] S102: Decompose the text into blocks based on block separators with a cost less than 2 to obtain an initial block list init_blocks.

[0025] S103: Declare the final block list final_blocks and initialize it to an empty list.

[0026] S104: Traverse the initial block list init_blocks: S1041: If the length of the current block curr_block is not longer than BLOCK_MAX, add it to the final block list final_blocks; S1042: Otherwise, the current block is further split based on the block separator with a cost greater than 1. If a sub-block after the split is still larger than BLOCK_MAX, it is split into multiple blocks based on BLOCK_MAX and added to the final block list final_blocks.

[0027] S105: Return the final block list final_blocks.

[0028] BERT is a pre-trained language model proposed by Google AI Research in 2018. It is a deep bidirectional pre-trained model based on Transformers that can utilize information in both the forward and backward directions at the same time.

[0029] Multilayer Perceptron (MLP) is a feedforward artificial neural network model that maps multiple input data sets to a single output data set. Classifications include: BERT and Multilayer Perceptron.

[0030] Model training includes: corresponding rules triggering the training of prediction models. The training process of corresponding rules triggering prediction models differs only in the data sets. The algorithms are implemented using "Algorithm 2 Model Training". Algorithm 2 Model Training is as follows: Input: training set ,in, Represents the first piece of data, x is the text, and y is the label.

[0031] Output: The trained model.

[0032] step: S201: Set the number of data set traversal epochs=10.

[0033] S202: for epoch from 1 to epochs do S2021: for x, y in D do S20211: Define a key block list z=[] to store the key blocks in x.

[0034] S20212: Delete a block d (any block) from x and determine the impact of the deletion on the loss. If the loss does not decrease significantly or even increases after the deletion of block d, add the block d to the block list z. The definition of loss is as follows: , in, loss Indicates loss, CrossEntropy () calculates the cross entropy loss, x _ means to delete a block d The remaining block sequence after judge () represents the relevant probability value of the compressed output, y _ means from y Delete d The tag sequence after the block corresponds to the tag, relv_label () indicates the correlation label calculated based on the manually defined correlation features.

[0035] S2022: for x, y in D do S20221: Get the key block list z of x.

[0036] S20222: Define loss and calculate: , in, loss Indicates loss, CrossEntropy () calculates the cross entropy loss, y is the category label, z is the corresponding input entry list x A list of key blocks, Reasoner () is the classification component of the model.

[0037] S20223: Update model parameters.

[0038] S203: Save and return the trained model.

[0039] The model training and application of the present invention involves the steps of constructing a data set, including: (1) defining classification labels; (2) determining positive and negative example data; and (3) constructing a data set and dividing it into a training set, a validation set, and a test set.

[0040] Define classification labels: for example, label 0 means the second rule is triggered, and label 1 means the second rule is not triggered.

[0041] Determine positive and negative data: The model training and application of the present invention involve data sets, including the second prediction data set and the twenty-fifth prediction data set.

[0042] The second prediction data set consists of positive examples consisting of a list of items that trigger the second item, and negative examples consisting of patents in the positive example item list that do not trigger the second item list, and other item lists extracted from patents that do not trigger the second item. The ratio of positive and negative examples is 1:1.

[0043] The Article 25 prediction data set consists of positive examples consisting of a list of items that trigger Article 25, and negative examples consisting of patents in the positive example item list that do not trigger the Article 25 item list, and other item lists extracted from patents that do not trigger Article 25. The ratio of positive to negative examples is 1:1.

[0044] Construct a dataset and divide it into training set, validation set, and test set.

[0045] Constructing a data set includes: constructing a second prediction data set, and constructing a twenty-fifth prediction data set.

[0046] The steps to construct the second prediction data set include: Step 1: Extract text from the dataset containing positive and negative examples and split the items into items; Step 2: Take the list of items that trigger the second item as positive examples; Step 3: The entries in the dataset where the positive examples are located that do not trigger the preset rules are taken as negative examples; Step 4: If the number of negative examples in the third step is insufficient, some entries are extracted from the data set that has not triggered the preset rules to supplement the negative examples, so that the positive and negative examples reach the requirement of 1:1; Step 5: Divide the training set, validation set, and test set into a ratio of 0.7:0.15:0.15.

[0047] The steps of constructing the prediction data set of Article 25 include: Step 1: Extract text from the dataset containing positive and negative examples and split the items into items; Step 2: List the items that trigger Article 25 as positive examples; Step 3: Take the list of entries in the patents containing the positive examples that do not trigger Article 25 as negative examples; Step 4: If the number of negative examples in step 3 is insufficient, some item lists are extracted from patents that do not trigger Article 25 in any item list to supplement the negative examples, so that the positive and negative examples meet the requirement of 1:1; Step 5: Divide the training set, validation set, and test set into a ratio of 0.7:0.15:0.15.

[0048] Legal provision prediction method and system: Data compliance rules and data quality rules trigger prediction methods and systems, such as " Figure 3 The item list triggers the second and twenty-fifth prediction method and system diagram as shown in the "schematic diagram of the prediction method and system", which is characterized by comprising the steps of: Step 1: Obtain the claims, parse the claims from the patent XML text according to the XML tags, or directly receive the text data to be processed input by the user; Step 2: Parse the item list. Parse all the item lists in the claims based on regular expressions. The algorithm involved is shown in "Algorithm 3 Parsing the Item List". Algorithm 3 Parsing the Item List specifically includes: Input: claims Output: List of entries claim_lst step: S301: cleaning all spaces and horizontal tabs in the claims text claims; S302: split the claims based on line breaks and periods to obtain an entry list init_claim_lst; S303: Cleaning, traversing the entry list: if it does not start with a number, merge it into the previous entry list; if the first entry list does not start with a number, end it with the entry list that starts with a number, or, end it. The entire content is taken as the first priority.

[0049] S3031: If the current item list item is empty, do nothing and continue to traverse the next item list; S3032: If the current entry list item starts with a number, a new entry list is encountered and added to the entry list claim_lst; otherwise, the current entry list content item is concatenated to the end of the last entry list in the entry list; S304: Return the entry list claim_lst.

[0050] Step 3: Apply model prediction, input the item lists into the corresponding rule trigger prediction model, and output the second trigger prediction result {item list 1: [v0, v1], item list 2: [v0, v1], …, item list n: [v0.v1]}, and the twenty-fifth trigger prediction result {item list 1: [v0, v1], item list 2: [v0, v1], …, item list n: [v0.v1]}; where v0 is predicted to be the label output value of 0, and v1 is predicted to be the label output value of 1.

[0051] Step 4: Calculate the trigger probability, the formula is as follows: , in, prob represents the probability value, Indicates that the calculation is based on e v0 is the calculation result of the power exponent, It represents the result of calculating the power v1 with base e as the exponent.

[0052] Step 5: Arrange the result format, the second trigger prediction result {item list 1: prob, item list 2: prob, ..., item list n: prob}, and the twenty-fifth trigger prediction result {item list 1: prob, item list 2: prob, ..., item list n: prob}; where prob is the probability value of predicting the triggering of the second or twenty-fifth item.

[0053] In one embodiment, Figure 2 As shown, an intelligent assisted review system is provided.

[0054] The system includes a model design module, a data set construction module, a model training module, and an execution prediction module, which specifically includes: a model design module, which is used to build a two-stage model including a compression stage and a classification stage; wherein, in the compression stage, the text is first split into multiple blocks by a block splitting algorithm, and the split text blocks are encoded using a BERT model to capture the bidirectional semantic information of the text, and the output of BERT is further processed by a multi-layer perceptron to extract key information and compress the text; in the classification stage, the compressed text is classified using a BERT model and a multi-layer perceptron to predict whether it triggers a preset rule; A data set construction module, including a classification label that defines a triggering rule, and positive and negative example data that determine the classification label; wherein the positive example data is composed of text that triggers the corresponding rule, and the negative example is composed of text in the patent where the positive example is located that does not trigger the corresponding rule and a list of items extracted from other data that do not trigger the corresponding rule; Model training module, including setting the number of training times for the data set and updating the model parameters according to the loss function; The prediction module is executed to perform prediction on the trained model; wherein, the text data to be processed is obtained, all entries in the text are parsed based on regular expressions to obtain an entry list; then the entry list is input into the trained model to obtain the prediction result of the corresponding rule.

[0055] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 3 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. The processor of the computer device is used to provide computing and control capabilities, and the network interface is used to communicate with an external terminal through a network connection. The computer device implements the above-mentioned intelligent auxiliary review method by loading and running a computer program.

[0056] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0057] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which involves all or part of the processes in the above-mentioned embodiment method.

[0058] In one embodiment, a computer program product is also provided, including a computer program / instruction, which involves all or part of the process in the above embodiment method.

[0059] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in M ​​forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (SyMchliMk) DRAM (SLDRAM), memory bus (RaMbus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0060] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. An intelligent assisted review method, characterized in that: The method includes four steps: model design, data set construction, model training, and prediction execution, specifically including: Model design, build a two-stage model including compression stage and classification stage; in the compression stage, first split the text into multiple blocks through the block splitting algorithm, use the BERT model to encode the split text blocks to capture the bidirectional semantic information of the text, and further process the BERT output through the multi-layer perceptron to extract key information and compress the text; in the classification stage, use the BERT model and multi-layer perceptron to classify the compressed text to predict whether it triggers the preset rules; Dataset construction, including defining classification labels that trigger rules, and determining positive and negative example data of classification labels; wherein positive example data is composed of texts that trigger corresponding rules, and negative examples are composed of texts in the patents where positive examples are located that do not trigger corresponding rules, and lists of items extracted from other data that do not trigger corresponding rules; Model training, including setting the number of training times for the dataset and updating model parameters based on the loss function; Execute prediction and use the trained model to perform prediction; wherein, obtain the text data to be processed, parse all entries in the text based on regular expressions, and obtain an entry list; then input the entry list into the trained model to obtain the prediction result of the corresponding rule.

2. The method according to claim 1, characterized in that: In the model design, the text is split into multiple blocks through the block splitting algorithm, which includes: Set the block splitting parameters, including defining the cost dictionary end_tokens of the block separator, where different symbols including line break, period, semicolon, question mark, exclamation mark and comma are assigned different cost values, and setting the maximum block length BLOCK_MAX to 63 characters and the minimum block length BLOCK_MIN to 10 characters; Use the block separator with cost value less than 2 to preliminarily split the text and get the initial block list init_blocks; Declare an empty final block list final_blocks to store the final blocks; Traverse the initial block list init_blocks, for each block, if its length does not exceed the maximum block length BLOCK_MAX, directly add it to the final block list final_blocks; if the block length exceeds BLOCK_MAX, continue to split it based on the block separator with a cost value greater than 1. If the length of the sub-block after splitting still exceeds BLOCK_MAX, further split it into multiple blocks not exceeding BLOCK_MAX, and add these blocks to final_blocks; Returns the final block list final_blocks for subsequent processing and analysis.

3. The method according to claim 1, characterized in that This includes defining the classification labels that trigger the rules and determining the positive and negative example data for the classification labels, including: Extract text from the dataset containing positive and negative examples and split the items into items; The list of entries that trigger the preset rules is taken as positive examples; The entries in the dataset where the positive examples are located that do not trigger the preset rules are regarded as negative examples; If the number of negative examples is insufficient, some entries are extracted from the data set that has not triggered the preset rules to supplement the negative examples, so that the positive and negative examples reach the requirement of 1:1; The training set, validation set, and test set are divided into the ratio of 0.7:0.15:0.

15.

4. The method according to claim 3, characterized in that The preset rules specifically include data compliance rules and data quality rules.

5. The method according to claim 1, characterized in that Model training includes setting the number of training times for the dataset and updating the model parameters based on the loss function, including: Set the number of traversals of the dataset to 10 rounds, which means that the entire training process will perform 10 complete traversals of the dataset; Start a training loop for each round. For each round, iterate over each sample pair (x, y) in the training set, where x is the text and y is the corresponding label; For each sample x, define an empty key block list z to store the key blocks in x. The key blocks are screened by deleting any block from x and judging the impact of the deletion on the loss. Continue to traverse each sample pair (x, y) in the training set, obtain the key block list z of each sample x, and define the loss function for calculation; among them, the cross entropy loss is used, involving the category label, the key block list corresponding to the input item list x, and the classification component of the model; Update the model parameters according to the calculated loss to optimize the model's predictive performance. After 10 rounds of training, save and return the trained model.

6. The method according to claim 5, characterized in that The loss function is defined as: ， in, loss Indicates loss, CrossEntropy () calculates the cross entropy loss, x _ represents the remaining block sequence after deleting any block, judge () represents the relevant probability value of the compressed output, y _ means from y Delete d The label sequence after the block corresponds to the label, relv_label () indicates the correlation label calculated based on the manually defined correlation features.

7. An intelligent auxiliary review system, characterized in that: The system includes a model design module, a data set construction module, a model training module, and an execution prediction module, which specifically include: The model design module is used to build a two-stage model including a compression stage and a classification stage. In the compression stage, the text is first split into multiple blocks by a block splitting algorithm, and the split text blocks are encoded using the BERT model to capture the bidirectional semantic information of the text. The output of BERT is further processed by a multi-layer perceptron to extract key information and compress the text. In the classification stage, the compressed text is classified using the BERT model and the multi-layer perceptron to predict whether it triggers the preset rules. A data set construction module, including a classification label that defines a triggering rule, and positive and negative example data that determine the classification label; wherein the positive example data is composed of text that triggers the corresponding rule, and the negative example is composed of text in the patent where the positive example is located that does not trigger the corresponding rule and a list of items extracted from other data that do not trigger the corresponding rule; Model training module, including setting the number of training times of the data set and updating the model parameters according to the loss function; The prediction module is executed to perform prediction on the trained model; wherein, the text data to be processed is obtained, all entries in the text are parsed based on regular expressions to obtain an entry list; then the entry list is input into the trained model to obtain the prediction result of the corresponding rule.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images