TWI882378B - Contract assisted review method and device - Google Patents

Abstract

A contract-assisted review method is applicable to marking multiple paragraphs of a contract document as at least one of multiple contract structure categories, and is implemented by a computer device, comprising: (A) obtaining multiple contract paragraphs to be classified based on a contract document to be classified using a natural language document parsing algorithm; (B) obtaining a first sentence meaning vector for each contract paragraph to be classified using a first sentence conversion vector model; (C) obtaining a structure category vector to be analyzed based on each first sentence meaning vector; (D) determining whether the value of each element of each structure category vector to be analyzed is greater than a preset threshold; (E) when it is determined that the value of the element is greater than the preset threshold, using the contract structure category corresponding to the element as a target contract structure category.

Description

Contract assisted review method and device

The present invention relates to a natural language document marking and classification method and device, and in particular to a natural language document marking and classification method and device applied to contract documents.

In modern society, contracts have become indispensable to ensure the rights and interests of both parties in transaction cooperation and to avoid possible legal disputes.

In the past, in order to classify contracts, confirm the content of contracts and clarify the matters that should be paid attention to in contract clauses, a lot of manpower must be invested in studying the contents of many contracts, finding the key words or descriptions of each clause, and summarizing their description writing methods, and then writing relevant natural language rules to determine whether the contract clause to be analyzed meets specific conditions, thereby assisting reviewers in reviewing contracts. However, the rules written will affect the accuracy of the clause summary due to the diversity of contract clauses (same meaning but different words, clause wording errors and omissions).

In view of this, it is necessary to find a solution to overcome the previous problems of excessive labor costs and inaccuracy of contract markings.

Therefore, the purpose of the present invention is to provide a contract assisted review method for assisting reviewers in reviewing contracts.

Therefore, the contract-assisted review method of the present invention is applicable to marking multiple paragraphs of a natural language contract document as at least one of multiple contract structure categories, and is implemented by a computer device that stores a first sentence conversion vector model for converting a sentence to be converted into a sentence meaning vector to be classified. The contract-assisted review method includes step (A), step (B), step (C), step (D), and step (E).

Step (A) is to use a natural language document parsing algorithm based on a received contract document to be classified related to a natural language to obtain a plurality of contract paragraphs to be classified corresponding to the contract document to be classified.

Step (B) is to obtain a first sentence vector corresponding to each contract paragraph to be classified by using the first sentence conversion vector model according to the contract paragraph to be classified.

Step (C) is to obtain, for each first sentence meaning vector, a structure category vector to be analyzed corresponding to the first sentence meaning vector according to the first sentence meaning vector, wherein each element of the structure category vector to be analyzed corresponds to one of the contract structure categories.

Step (D) is to determine, for each element of each to-be-analyzed framework category vector, whether the value of the element is greater than a preset threshold.

Step (E) is that for each element of each architecture category vector to be analyzed, when it is determined that the value of the element is greater than the preset threshold, the contract paragraph to be classified corresponding to the architecture category vector to be analyzed is treated as a classified contract paragraph, and the contract architecture category corresponding to the element is used as a target contract architecture category of the classified contract paragraph.

Another object of the present invention is to provide a contract auxiliary review device for assisting reviewers in reviewing contracts.

Therefore, the contract assisted review device of the present invention includes a storage module, an input module, and a processing module electrically connected to the storage module and the input module.

The storage module stores a first sentence-to-vector model for converting a sentence to be converted into a sentence meaning vector to be classified.

The input module is used for receiving a contract document to be classified in natural language.

The processing module uses a natural language document parsing algorithm to obtain a plurality of contract paragraphs to be classified corresponding to the contract document to be classified according to the contract document to be classified. For each contract paragraph to be classified, the processing module uses the first sentence conversion vector model to obtain a first sentence meaning vector corresponding to the contract paragraph to be classified according to the contract paragraph to be classified. For each first sentence meaning vector, the processing module obtains a corresponding structure category vector to be analyzed according to the first sentence meaning vector. An element corresponds to one of the contract structure categories respectively. For each element of each structure category vector to be analyzed, the processing module determines whether the value of the element is greater than a preset threshold. For each element of each structure category vector to be analyzed, when the processing module determines that the value of the element is greater than the preset threshold, the processing module regards the contract segment to be classified corresponding to the structure category vector to be analyzed as a classified contract segment, and regards the contract structure category corresponding to the element as a target contract structure category of the classified contract segment.

The effect of the present invention is that: each contract paragraph to be classified is converted into the first sentence meaning vector by the computer device using the first sentence conversion vector model, and then the corresponding structure category vector to be analyzed is obtained, and finally the element value of each structure category vector to be analyzed is analyzed, and the corresponding contract paragraph to be classified is marked with the target contract structure category, thereby marking all the contract paragraphs to be classified in the contract document to be classified, and when reviewing the contract document to be classified, the reviewer can refer to the marking results of the contract paragraphs to be classified to increase the review efficiency and accuracy.

Before the present invention is described in detail, it should be noted that similar components are represented by the same reference numerals in the following description.

Referring to FIG. 1 , an embodiment of the contract assisted review method of the present invention is applicable to marking multiple paragraphs of a natural language contract document as at least one of multiple contract structure categories, and is implemented by a computer device 1, which includes a storage module 11, an input module 12, a display module 13, and a processing module 14 electrically connecting the storage module 11, the input module 12, and the display module 13.

The storage module 11 stores a first sentence conversion vector model for converting a sentence to be converted into a sentence vector to be classified, a second sentence conversion vector model, a sentence conversion word vector model, a feature extraction classification model for converting the sentence vector to be classified into a vector of a structure category to be analyzed with a dimension equal to the number of the contract structure categories, a contract key prompt corresponding table, a default relevance result sentence for indicating the relevance result according to a default sentence, and a default relevance result sentence for converting the sentence vector to be classified into a feature extraction classification model for converting the sentence vector to be classified into a vector of a structure category to be analyzed with a dimension equal to the number of the contract structure categories, and a default relevance result sentence for indicating the relevance result according to a default sentence. A correlation prediction model for obtaining the degree of correlation among the three paragraphs based on another preset sentence about the description of the penalty for breach of contract, a corresponding table of contract penalty reminders, and a paragraph extraction model for obtaining, based on another preset sentence about the description of the penalty for breach of contract and one of the paragraphs, a starting probability value corresponding to the first word of the description related to the other preset sentence and an ending probability value corresponding to the last word of the description related to the other preset sentence for each word in the paragraph.

The contract key prompt corresponding table includes a plurality of relevant contract structure categories, at least one relevant test sentence corresponding to each relevant contract structure category, and a plurality of relevant result sentences corresponding to each relevant contract structure category. Each relevant contract structure category is one of the contract structure categories.

The contract penalty reminder corresponding table includes a plurality of penalty contract structure categories and at least one penalty test sentence corresponding to each penalty contract structure category. Each penalty contract structure category is one of the contract structure categories.

The storage module 11 also stores a plurality of contract structure training data sets, a plurality of contract correlation training data sets, and a plurality of contract default penalty training data sets.

Each contract structure training data set includes a training contract paragraph and at least one contract structure category answer corresponding to the training contract paragraph. The at least one contract structure category answer is one of the contract structure categories and is generated by manual marking.

Each contract-related training data includes a training-related contract paragraph, the framework category corresponding to the training-related contract paragraph, and a contract-related answer corresponding to the training-related contract paragraph. The framework category corresponding to the training-related contract paragraph is one of the related contract framework categories. The contract-related answer is one of all related results corresponding to the framework category corresponding to the training-related contract paragraph and is generated by manual marking.

Each contract default penalty training data set includes a training default penalty contract paragraph, the framework category corresponding to the training default penalty contract paragraph, and a contract default penalty answer corresponding to the training default penalty contract paragraph, which includes a starting position and an ending position. The framework category corresponding to the training default penalty contract paragraph is one of the default penalty contract framework categories. The starting position of the contract breach penalty answer is used to indicate the position of a word in the corresponding training contract paragraph, and the ending position of the contract breach penalty answer is used to indicate the position of another word in the corresponding training contract paragraph, so a section of content selected from the corresponding training contract paragraph can be defined by the starting position and the ending position. The contract breach penalty answer is generated by manual annotation.

It is worth noting that in this embodiment, the contract document includes a confidentiality contract document, but is not limited thereto. The contract structure categories include a contract subject category, a contract purpose category, a confidential information category, a confidentiality obligation category, a confidentiality obligation exclusion category, a confidentiality period and effectiveness category, a rights ownership and guarantee category, a confidential information return category, a breach of contract liability damages category, a breach of contract liability liquidated damages category, a governing law and jurisdictional court category, and other categories, but are not limited to the above.

The computer device 1 may be a personal computer, a tablet computer or a notebook computer, but is not limited thereto.

The following will be used in conjunction with the embodiment of the contract assisted review method of the present invention to explain the operating details of each component in the computer device 1. The embodiment of the contract assisted review method includes a framework training program, a relevance training program, a penalty training program, a framework classification program, a relevance determination program, and a penalty paragraph capture program.

Referring to FIG. 2 , the framework training process is trained based on all contract framework training data sets and includes steps 41 to 44.

In step 41, for each contract framework training data set, the processing module 14 obtains a training first sentence meaning vector corresponding to the training contract paragraph based on the training contract paragraph of the contract framework training data set and using the first sentence conversion vector model to be trained.

It is worth noting that in this embodiment, the first sentence-to-vector model is a transformer-based bidirectional encoder representation technology (BERT, Bidirectional Encoder Representations from Transformers) encoder, but not limited to this. And this embodiment uses the [CLS] vector output by the first sentence-to-vector model as the training first sentence meaning vector, but not limited to this.

In step 42, for each contract framework training data set, the processing module 14 obtains a training framework category vector to be processed corresponding to the training first sentence meaning vector according to the training first sentence meaning vector corresponding to the contract framework training data set and using the feature extraction classification model to be trained.

It is worth mentioning that in this embodiment, the feature extraction classification model is a neural network (NN) or a convolutional neural network model (CNN), but is not limited thereto.

In step 43, for each contract framework training data set, the processing module 14 uses a sigmoid function (S-type function) to map and transform each element in the training framework class vector to be processed corresponding to the contract framework training data set, so as to obtain a training framework class vector to be analyzed corresponding to the training framework class vector to be processed.

In step 44, for each contract framework training data set, the processing module 14 compares the training framework category vector to be analyzed corresponding to the contract framework training data set with the at least one contract framework category answer corresponding to the contract framework training data set, and adjusts the parameters of the first sentence conversion vector model to be trained (fine tuning) and the feature extraction classification model to be trained, thereby obtaining the trained first sentence conversion vector model and the feature extraction classification model.

Referring to FIG. 3 , the relevance training procedure is trained based on all contract relevance training data sets, and processes each contract relevance training data set, and includes steps 51 to 55.

In step 51, the processing module 14 obtains all relevance test sentences and all relevance result sentences corresponding to the training relevance contract paragraph of the contract relevance training data set according to the architecture category corresponding to the training relevance contract paragraph of the contract relevance training data set and the contract key prompt corresponding table.

In step 52, for each relevance test sentence corresponding to the training relevance contract paragraph of the contract relevance training data set, the processing module 14 uses the second sentence conversion vector model to be trained to obtain multiple training second sentence meaning vectors about the relevance test sentence to be tested and the relevance contract paragraph to be analyzed for each relevance result sentence corresponding to the training relevance contract paragraph of the contract relevance training data set based on all relevance result sentences corresponding to the training relevance contract paragraph of the contract relevance training data set, the relevance test sentence and the training relevance contract paragraph of the contract relevance training data set.

It is worth noting that in this embodiment, the second sentence-to-vector model is a transformer-based bidirectional encoder representation technology (BERT, Bidirectional Encoder Representations from Transformers) encoder, but not limited to this. And this embodiment uses the [CLS] vector output by the second sentence-to-vector model as the training second sentence meaning vector, but not limited to this.

In step 53, for each relevance test sentence corresponding to the training relevance contract paragraph of the contract relevance training data set, the processing module 14 obtains a to-be-processed training relevance vector corresponding to the training second sentence meaning vector based on each training second sentence meaning vector corresponding to the relevance test sentence using a first linear transformation matrix to be trained.

In step 54, for each relevance test sentence corresponding to the training relevance contract paragraph of the contract relevance training data set, the processing module 14 uses a normalized exponential function (Softmax function) to perform mapping transformation according to each training relevance vector to be processed corresponding to the relevance test sentence to obtain a training relevance vector to be analyzed.

In step 55, for each relevance test sentence corresponding to the training relevance contract paragraph of the contract relevance training data set, the processing module 14 compares each training relevance vector to be analyzed corresponding to the relevance test sentence with the contract relevance answer of the contract relevance training data set, and adjusts the parameters of the second sentence transformation vector model to be trained (fine tuning) and the first linear transformation matrix to be trained, thereby obtaining the trained relevance prediction model.

Referring to FIG. 4 , the default penalty training procedure is trained based on all contract default penalty training data sets, and processes each contract default penalty training data set, and includes steps 61 to 67.

In step 61, the processing module 14 obtains at least one breach penalty test sentence corresponding to the training breach penalty contract paragraph of the contract breach penalty training data set according to the contract breach penalty prompt corresponding table.

In step 62, for each breach of contract fee test sentence corresponding to the training breach of contract fee paragraph of the contract breach of contract training data set, the processing module 14 obtains a plurality of to-be-processed training word meaning vectors for each word of the training breach of contract paragraph based on the breach of contract fee test sentence and the training breach of contract paragraph of the contract breach of contract training data set, using the sentence-to-word vector model to be trained.

It is worth noting that in this embodiment, the sentence-to-word vector model is a transformer-based bidirectional encoder representation technology (BERT, Bidirectional Encoder Representations from Transformers) encoder, but not limited to this. In this embodiment, each word in the training default contract paragraph is converted by the sentence-to-word vector model to output the word vector as the waiting-for-processing training word meaning vector, but not limited to this.

In step 63, for each breach of contract test sentence corresponding to the training breach of contract paragraph of the breach of contract training data set, the processing module 14 obtains a plurality of to-be-processed training starting point probability values respectively corresponding to the to-be-processed training word meaning vectors according to the to-be-processed training word meaning vectors of each word in the training breach of contract paragraph of the breach of contract test sentence using a second linear transformation matrix to be trained.

In step 64, for each penalty test sentence corresponding to the training penalty contract paragraph of the contract penalty training data set, the processing module 14 uses the normalized exponential function to perform mapping transformation according to all the to-be-processed training starting point probability values corresponding to the penalty test sentence to obtain multiple training starting point probability values of each word of the penalty test sentence with respect to the training penalty contract paragraph.

In step 65, for each breach of contract test sentence corresponding to the training breach of contract paragraph of the contract breach of contract training data set, the processing module 14 obtains a plurality of pending training endpoint probability values respectively corresponding to the pending training word meaning vectors according to the pending training word meaning vectors of each word in the training breach of contract paragraph in the breach of contract test sentence, using a third linear transformation matrix to be completed for training.

In step 66, for each penalty test sentence corresponding to the training penalty contract paragraph of the contract penalty training data set, the processing module 14 uses the normalized exponential function to perform mapping transformation according to all the pending training endpoint probability values corresponding to the penalty test sentence to obtain multiple training endpoint probability values of each word of the penalty test sentence with respect to the training penalty contract paragraph.

In step 67, for each breach of contract test sentence corresponding to the training breach of contract paragraph of the breach of contract training data set, all training starting point probability values and all training ending point probability values corresponding to the breach of contract test sentence are compared with the breach of contract answer of the breach of contract training data set, and the sentence-to-word vector model to be trained is adjusted (fine tuning is performed), the second linear transformation matrix to be trained, and the parameters of the third linear transformation matrix to be trained are adjusted to obtain the trained paragraph capture model. Among them, the position of the word with the largest probability value of the training starting point and the position of the word with the largest probability value of the training ending point are the answers to the contract penalty.

Referring to FIG. 5 , the framework classification process includes steps 71 to 75.

In step 71, the processing module 14 uses a natural language document parsing algorithm according to a contract document to be classified related to a natural language received by the input module 12 to obtain a plurality of contract paragraphs to be classified corresponding to the contract document to be classified.

It is worth mentioning that in this embodiment, the natural language document parsing algorithm is Docx Parser software, but not limited to this.

In step 72, for each contract paragraph to be classified, the processing module 14 uses the first sentence conversion vector model according to the contract paragraph to be classified to obtain a first sentence meaning vector corresponding to the contract paragraph to be classified. In this embodiment, the processing module 14 uses the [CLS] vector output by the first sentence conversion vector model as the first sentence meaning vector, but is not limited thereto.

In step 73, for each first sentence meaning vector, the processing module 14 uses the feature extraction classification model according to the first sentence meaning vector to obtain a to-be-analyzed structure category vector corresponding to the first sentence meaning vector, wherein each element of the to-be-analyzed structure category vector corresponds to one of the contract structure categories.

Referring to FIG. 6 , step 73 further includes sub-steps 731 and 732 .

In sub-step 731, for each first sentence meaning vector, the processing module 14 uses the feature extraction classification model according to the first sentence meaning vector to obtain a to-be-processed framework category vector corresponding to the first sentence meaning vector.

In sub-step 732, for each architecture class vector to be processed, the processing module 14 uses the sigmoid function to map and transform each element in the architecture class vector to be processed according to the architecture class vector to be processed, so as to obtain an architecture class vector to be analyzed corresponding to the architecture class vector to be processed.

In step 74, for each element of each to-be-analyzed framework class vector, the processing module 14 determines whether the value of the element is greater than a preset threshold. When the processing module 14 determines that the value of the element is greater than the preset threshold, the process proceeds to step 75; when the processing module 14 determines that the value of the element is not greater than the preset threshold, no program is executed.

In step 75, for each element of each architecture category vector to be analyzed, the processing module 14 treats the contract paragraph to be classified corresponding to the architecture category vector to be analyzed as a classified contract paragraph, and treats the contract architecture category corresponding to the element as a target contract architecture category of the classified contract paragraph, and displays the classified contract paragraph and the target contract architecture category in the display module 13.

Referring to Figures 7, 8 and Table 1, the relevance determination process is executed after the contract structure classification process, and processes each classified contract paragraph. The relevance determination process includes steps 81-84.

In step 81, the processing module 14 determines whether there is at least one relevant contract structure category to be analyzed that is the same as any of the relevant contract structure categories in the contract key prompt corresponding table in the at least one target contract structure category corresponding to the classified contract paragraph. When the processing module 14 determines that there is at least one relevant contract structure category to be analyzed, the process step 82 is performed; when the processing module 14 determines that there is no relevant contract structure category to be analyzed, the relevance determination procedure is terminated.

In step 82, for each relevant contract structure category to be analyzed, the processing module 14, based on the contract key prompt corresponding table, takes the at least one relevant test sentence corresponding to the relevant contract structure category that is the same as the relevant contract structure category to be analyzed as at least one relevant test sentence to be tested, and takes all relevant result statements corresponding to the relevant contract structure category that is the same as the relevant contract structure category to be analyzed as multiple relevant result statements to be tested, and takes the classified contract paragraph as a relevant contract paragraph to be analyzed.

In step 83, for each relevance contract structure category to be analyzed in the relevance contract paragraph to be analyzed, the processing module 14 uses the relevance prediction model to obtain multiple relevance prediction results to be determined based on all relevance result sentences to be detected corresponding to the relevance contract structure category to be analyzed, all relevance test sentences to be detected corresponding to the relevance contract structure category to be analyzed, and the relevance contract paragraph to be analyzed. Among them, each relevance prediction result to be determined includes multiple relevance probability values to be determined between each relevance result sentence to be detected corresponding to the relevance contract structure category to be analyzed, each relevance test sentence to be detected corresponding to the relevance contract structure category to be analyzed, and the relevance contract paragraph to be analyzed regarding the relevance results corresponding to the relevance contract structure category to be analyzed.

Referring to FIG. 8 , for each relevant contract structure category to be analyzed in the relevant contract paragraph to be analyzed, the processing module performs a relevance probability value acquisition procedure to obtain the at least one relevance prediction result to be determined as step 83, and step 83 also includes sub-steps 831~833.

In sub-step 831, for each relevance test sentence to be detected of the relevance contract framework category to be analyzed, the processing module 14 uses the second sentence conversion vector model to obtain multiple second sentence meaning vectors about all relevance result sentences to be detected, the relevance test sentences to be detected, and the relevance contract paragraph to be analyzed based on all relevance result sentences to be detected of the relevance contract framework category to be analyzed, the relevance test sentences to be detected, and the relevance contract paragraph to be analyzed. In this embodiment, the processing module 14 uses the [CLS] vector output by the second sentence conversion vector model as the second sentence meaning vector, but is not limited to this.

In sub-step 832, for each second sentence meaning vector of the relevance contract structure category to be analyzed, the processing module 14 uses the first linear transformation matrix according to the second sentence meaning vector to obtain a relevance vector to be processed corresponding to the second sentence meaning vector.

In sub-step 833, for each correlation vector to be processed of the correlation contract structure category to be analyzed, the processing module 14 uses the normalized index function to perform mapping transformation according to all elements of the correlation vector to be processed to obtain a correlation vector to be analyzed as the prediction result of the correlation to be determined. Among them, all elements of the correlation vector to be analyzed are the correlation probability values to be determined.

In step 84, for each relevant contract structure category to be analyzed of the relevant contract paragraph to be analyzed, the processing module 14 obtains a target relevance marking result indicating that the relevant contract paragraph to be analyzed belongs to one of all the relevance results corresponding to the relevant contract structure category to be analyzed based on all the to-be-determined relevance prediction results corresponding to the relevant contract structure category to be analyzed, and displays the target relevance marking result on the display module 13.

In particular, in step 84, for each relevant contract structure category to be analyzed in the relevant contract paragraph to be analyzed, the processing module 14 uses the correlation result corresponding to the largest relevant probability value to be determined among all relevant probability values to be determined corresponding to the relevant contract structure category to be analyzed as the target relevance marking result, but is not limited to this.

Refer to Table 1. Specifically, in steps 81 and 82, when the processing module 14 determines that the relevant contract paragraph to be analyzed belongs to the "confidentiality period and effectiveness category", it can be known that the at least one relevance test sentence to be tested corresponding to the relevant contract paragraph to be analyzed is "How long is the confidentiality period" and "The period of confidentiality required after receiving the data is" and all the relevance result sentences to be tested corresponding to the relevant contract paragraph to be analyzed are "within one year (inclusive)", "more than one year" and "not specified", and the classified contract paragraph is used as the relevant contract paragraph to be analyzed. In step 83, the processing module 14 obtains a first relevance vector to be analyzed according to "how long is the confidentiality period", "within one year (inclusive)" and the relevant contract paragraph to be analyzed; obtains a second relevance vector to be analyzed according to "how long is the confidentiality period", "more than one year" and the relevant contract paragraph to be analyzed; obtains a third relevance vector to be analyzed according to "how long is the confidentiality period", "unspecified" and the relevant contract paragraph to be analyzed; "The period of confidentiality required after receiving the data is", "within one year (inclusive)" and the relevant contract paragraph to be analyzed, a fourth relevance vector to be analyzed is obtained; based on "The period of confidentiality required after receiving the data is", "more than one year" and the relevant contract paragraph to be analyzed, a fifth relevance vector to be analyzed is obtained; based on "The period of confidentiality required after receiving the data is", "unspecified" and the relevant contract paragraph to be analyzed, a sixth relevance vector to be analyzed is obtained. In step 84, if it is determined that the relevance result statement corresponding to the maximum relevance probability value to be determined for all elements in the above six relevance vectors to be analyzed is "within one year (inclusive)", then "within one year (inclusive)" is used as the target relevance marking result, and the target relevance marking result is displayed on the display module 13. In particular, the at least one correlation test sentence and the correlation result sentences are not limited to the examples in Table 1, and can be freely added or modified by the designer. Related Contract Structure Class Relevance Test Sentences Relevance Result Statement Confidential Information Category 1. What should be marked when confidential information is disclosed? 2. What should be stated when confidential information is disclosed? 3. What should be done when disclosing? 1. Scope of non-classified confidential data 2. Scope of classified confidential data Types of confidentiality obligations 1. Who has to keep the other party's secrets confidential? 2. What are the responsibilities of both parties? 3. Who has to keep the other party's secrets confidential? 1. Unilateral 2. Bilateral Confidentiality period and effectiveness category 1. How long is the confidentiality period? 2. How long is the confidentiality period after receiving the information? 1. Within one year (inclusive) 2. More than one year 3. Not specified Breach of contract damages and penalty categories 1. What is the stipulation for damages for breach of contract? 2. Scope of damages 1. Compensation for direct losses 2. Compensation for losses exceeding direct losses 3. Not agreed upon Breach of contract penalty category 1. What is the agreement on liquidated damages? 2. What is liquidated damages or punitive liquidated damages? 3. What is the description of liquidated damages or punitive liquidated damages? 1. With penalty 2. Without penalty Governing Law and Type of Jurisdiction 1. What is the applicable law? 2. What is the jurisdiction of the court? 1. Taiwan 2. China 3. Arbitration: Taiwan 4. Arbitration: China 5. Unable to determine Table 1

Referring to Figures 9-11 and Table 2, the penalty paragraph extraction process is executed after the contract structure classification process, and processes each classified contract paragraph, and includes steps 91-97.

In step 91, the processing module 14 determines whether there is at least one target contract structure category corresponding to the classified contract paragraph that is the same as any of the default penalty contract structure categories in the contract default penalty prompt corresponding table. When the processing module 14 determines that there is at least one default penalty contract structure category to be analyzed, the process step 92 is performed; when the processing module 14 determines that there is no default penalty contract structure category to be analyzed, the default penalty paragraph capture procedure is terminated.

In step 92, for each penalty contract structure category to be analyzed, the processing module 14 uses the at least one penalty test sentence corresponding to the penalty contract structure category that is the same as the penalty contract structure category to be analyzed as at least one penalty test sentence to be tested, and uses the classified contract paragraph as a penalty contract paragraph to be analyzed according to the contract penalty prompt corresponding table.

In step 93, for each of the to-be-analyzed breach of contract structure categories of the to-be-analyzed breach of contract paragraph, the processing module 14 uses the paragraph capture model to obtain at least one to-be-determined breach of contract capture result according to each to-be-detected breach of contract test sentence of the to-be-analyzed breach of contract structure category and the to-be-analyzed breach of contract paragraph. Each to-be-determined breach of contract capture result includes a breach of contract starting point probability value corresponding to the first word of the description of the corresponding to-be-detected breach of contract test sentence and a breach of contract ending point probability value corresponding to the last word of the description of the corresponding to-be-detected breach of contract test sentence.

Referring to FIG. 10 , for each default penalty contract structure category to be analyzed in the default penalty contract section to be analyzed, the processing module performs an endpoint probability value acquisition procedure to obtain at least one default penalty capture result to be determined as step 93, and step 93 also includes sub-steps 931 to 936.

In sub-step 931, for each test sentence of the breach of contract to be analyzed of the category of the breach of contract structure to be analyzed, the processing module 14 uses the sentence-to-word vector model to obtain multiple word vectors to be processed for each word of the test sentence of the breach of contract to be analyzed with respect to the paragraph of the breach of contract to be analyzed. In this embodiment, the processing module 14 uses the word vector output by the sentence-to-word vector model for each word of the paragraph of the breach of contract to be analyzed as the word vector to be processed, but is not limited thereto.

In sub-step 932, for each penalty test sentence to be detected in the penalty contract structure category to be analyzed, the processing module 14 uses the second linear transformation matrix to obtain multiple penalty starting point probability values to be processed that correspond to the word meaning vectors to be processed according to all the word meaning vectors to be processed of each word in the penalty test sentence to be detected in the penalty contract paragraph to be analyzed.

In sub-step 933, for each penalty test sentence to be detected in the penalty contract structure category to be analyzed, the processing module 14 uses the normalized index function to perform mapping transformation according to all penalty starting point probability values to be processed corresponding to the penalty test sentence to be detected, so as to obtain multiple penalty starting point probability values of each word in the penalty test sentence to be detected with respect to the penalty contract paragraph to be analyzed.

In sub-step 934, for each penalty test sentence to be tested in the penalty contract structure category to be analyzed, the processing module 14 uses the third linear transformation matrix to obtain multiple penalty endpoint probability values to be processed that respectively correspond to the word meaning vectors to be processed based on all the word meaning vectors to be processed of each word in the penalty contract paragraph to be analyzed in the penalty test sentence to be detected.

In sub-step 935, for each penalty test sentence to be tested in the penalty contract structure category to be analyzed, the processing module 14 uses the normalized exponential function to perform mapping transformation according to all penalty endpoint probability values to be processed corresponding to the penalty test sentence to be detected, so as to obtain multiple penalty endpoint probability values of each word in the penalty test sentence to be detected with respect to the penalty contract paragraph to be analyzed.

In sub-step 936, for each penalty test sentence to be detected in the penalty contract structure category to be analyzed, the processing module 14 uses all penalty starting point probability values and all penalty ending point probability values corresponding to the penalty test sentence to be detected as the penalty capture result to be determined.

In step 94, for each penalty contract structure category to be analyzed in the penalty contract paragraph to be analyzed, the processing module 14 obtains a target penalty capture result for the penalty contract structure category to be analyzed based on all pending penalty capture results for the penalty contract structure category to be analyzed, and indicates a target penalty capture result for a penalty paragraph defined by a target penalty starting point word and a target penalty ending point word in the penalty contract paragraph to be analyzed, and displays the target penalty capture result on the display module 13.

Referring to FIG. 11 , step 94 further includes sub-steps 941 to 943.

In sub-step 941, for each penalty contract structure category to be analyzed in the penalty contract paragraph to be analyzed, the processing module 14 obtains a word with the largest penalty starting probability value from all the words in the penalty contract paragraph to be analyzed according to all the penalty starting probability values corresponding to the penalty contract structure category to be analyzed, and uses it as the target penalty starting word.

In sub-step 942, for each penalty contract structure category to be analyzed in the penalty contract paragraph to be analyzed, the processing module 14 obtains a word with the largest penalty endpoint probability value from all the words in the penalty contract paragraph to be analyzed according to all the penalty endpoint probability values corresponding to the penalty contract structure category to be analyzed, and uses it as the target penalty endpoint word.

In sub-step 943, for each liquidated damages contract structure category to be analyzed of the liquidated damages contract paragraph to be analyzed, the processing module 14 defines the liquidated damages paragraph as the target liquidated damages capture result according to the target liquidated damages starting point word and the target liquidated damages ending point word corresponding to the liquidated damages contract structure category to be analyzed.

Referring to Table 2, specifically, in steps 91 and 92, when the processing module 14 determines that the liquidated damages contract paragraph to be analyzed belongs to the "liquidation damages category", it can be known that the at least one liquidated damages test sentence to be tested corresponding to the liquidated damages contract paragraph to be analyzed is "what is the liquidated damages agreement" and "what is the liquidated damages or punitive liquidated damages"; in step 93, the liquidated damages contract paragraph to be analyzed is the "liquidation damages" and "what is the liquidated damages or punitive liquidated damages"; The processing module 14 obtains a first penalty starting point probability value and a first penalty ending point probability value for each word of the penalty contract paragraph to be analyzed with respect to "what is the penalty agreement" according to "what is the penalty agreement" and the penalty contract paragraph to be analyzed; the processing module 14 then obtains the penalty or punitive penalty according to "what is the penalty agreement" and the relevant contract paragraph to be analyzed. Analyze each word in the penalty contract paragraph for a second penalty starting probability value and a second penalty ending probability value regarding "what is the penalty or punitive penalty"; and in step 94, based on all the first penalty starting probability values and all the second penalty starting probability values, use the word with the largest penalty starting probability value as the target penalty starting word, and based on all the first penalty starting probability values and all the second penalty starting probability values The first default penalty endpoint probability value and all the second default penalty endpoint probability values are used, and the word with the largest default penalty endpoint probability value is used as the target default penalty endpoint word, and the target default penalty extraction result about the "default liability damages default penalty category" and including the default penalty paragraph (for example: the default penalty is 200,000 US dollars) is obtained, and the target default penalty extraction result is displayed on the display module 13. In particular, the default penalty test sentences are not limited to the examples in Table 2, and can be freely added or modified according to the designer. Types of penalty contract structures Liquidated Damage Test Sentence Breach of contract damages and penalty categories 1. What is the agreement on liquidated damages? 2. What is liquidated damages or punitive liquidated damages? Breach of contract penalty category Table 2

In step 95, for each penalty contract structure category to be analyzed in the penalty contract paragraph to be analyzed, the processing module 14 obtains a penalty value in the penalty paragraph according to the penalty paragraph of the target penalty capture result corresponding to the penalty contract structure category to be analyzed, using a known natural language technology.

In step 96, for the penalty value corresponding to each penalty contract structure category to be analyzed in the penalty contract segment to be analyzed, the processing module 14 determines whether the penalty value is greater than a warning amount. When the processing module 14 determines that it is greater than the warning amount, the process proceeds to step 97; when the processing module 14 determines that it is not greater than the warning amount, the penalty segment capture procedure is terminated.

In step 97 , the processing module 14 generates a warning signal indicating that the default penalty section corresponding to the default penalty contract structure type to be analyzed needs to be warned and displays it on the display module 13 .

In summary, the contract assisted review method of the present invention, through the processing module 14 executing the framework classification procedure, marks all the contract paragraphs to be classified in the contract document to be classified as the target contract framework category to which they belong, so that the review personnel who review the contract document to be classified can clearly understand the subject of each contract paragraph to be classified; and through the processing module 14 executing the relevance training procedure, some of the contract paragraphs to be classified in the contract document to be classified are additionally marked with key points. The prompt explanation is provided to assist the reviewer to understand the core content described in the contract paragraph to be classified; finally, the processing module 14 executes the penalty paragraph extraction procedure, in addition to extracting all the contents related to the penalty in the contract document to be classified for the reviewer to judge its risk, it also generates the warning signal for the penalty paragraph with too high penalty, so as to help the reviewer reduce the risk of the review, increase the efficiency and accuracy of the review, so that the purpose of the present invention can be achieved.

However, the above is only an embodiment of the present invention and should not be used to limit the scope of implementation of the present invention. All simple equivalent changes and modifications made according to the scope of the patent application of the present invention and the content of the patent specification are still within the scope of the present patent.

1: Computer device 11: Storage module 12: Input module 13: Display module 14: Processing module 41~44: Steps 51~55: Steps 61~67: Steps 71~75: Steps 731~732: Sub-steps 81~84: Steps 831~833: Sub-steps 91~97: Steps 931~936: Sub-steps 941~943: Sub-steps

Other features and functions of the present invention will be clearly presented in the implementation method with reference to the drawings, in which: FIG. 1 is a block diagram illustrating a computer device for executing an embodiment of the contract assisted review method of the present invention; FIG. 2 is a flow chart illustrating a framework training procedure of the embodiment of the contract assisted review method of the present invention; FIG. 3 is a flow chart illustrating a relevance training procedure of the embodiment; FIG. 4 is a flow chart illustrating a default fee training procedure of the embodiment; FIG. 5 is a flow chart illustrating a framework classification procedure of the embodiment; FIG. 6 is a flow chart illustrating how the framework classification procedure of the embodiment obtains a framework category vector to be analyzed; FIG. 7 is a flow chart illustrating a relevance determination procedure of the embodiment; FIG. 8 is a flow chart illustrating how the relevance determination procedure of the embodiment obtains a plurality of relevance prediction results to be determined; FIG. 9 is a flow chart illustrating a penalty paragraph extraction procedure of the embodiment; FIG. 10 is a flow chart illustrating how the penalty paragraph extraction procedure of the embodiment obtains at least one penalty extraction result to be determined; and FIG. 11 is a flow chart illustrating how the penalty paragraph extraction procedure of the embodiment obtains a target penalty extraction result.

71~75······· Steps

Claims (18)

A contract assisted review method is used to mark multiple paragraphs of a natural language contract document as at least one of multiple contract structure categories, and is implemented by a computer device, the computer device storing a first sentence conversion vector model based on a transformer-based bidirectional encoder representation technology for converting a sentence to be converted into a sentence vector to be classified, a contract key prompt corresponding table, and a transformer-based bidirectional encoder representation technology for indicating a correlation result according to a preset sentence. A correlation prediction model of the correlation degree between the three is obtained by presetting a correlation result statement and one of the paragraphs. The contract key prompt correspondence table includes multiple correlation contract structure categories, at least one correlation test sentence corresponding to each correlation contract structure category, and multiple correlation result statements corresponding to multiple correlation results corresponding to each correlation contract structure category. Each correlation contract structure category is one of the contract structure categories. The contract assisted review method includes the following steps: (A) Based on a received contract document to be classified related to a natural language, a natural language document parsing algorithm is used to obtain a plurality of contract paragraphs to be classified corresponding to the contract document to be classified; (B) For each contract paragraph to be classified, based on the contract paragraph to be classified, a first sentence vector corresponding to the contract paragraph to be classified is obtained using the first sentence conversion vector model; (C) For each first sentence vector, based on the first sentence vector, a structure category vector to be analyzed corresponding to the first sentence vector is obtained, each element of the structure category vector to be analyzed corresponds to one of the contract structure categories; (D) For each element of each structure category vector to be analyzed, whether the value of the element is greater than a preset threshold is determined; (E) For each element of each to-be-analyzed framework category vector, when it is determined that the value of the element is greater than the preset threshold, the to-be-classified contract paragraph corresponding to the to-be-analyzed framework category vector is treated as a classified contract paragraph, and the contract framework category corresponding to the element is treated as a target contract framework category of the classified contract paragraph; (F) For each classified contract paragraph, the computer device performs a relevance determination procedure on the classified contract paragraph, and the relevance determination procedure includes sub-steps (F-1), (F-2), (F-3) and (F-4): (F-1) Determine whether there is at least one to-be-analyzed relevant contract framework category that is the same as any of the relevant contract framework categories in the contract key prompt corresponding table in the at least one target contract framework category corresponding to the classified contract paragraph; (F-2) When it is determined that there is at least one relevant contract structure category to be analyzed, for each relevant contract structure category to be analyzed, according to the contract key prompt corresponding table, the at least one relevant test sentence corresponding to the relevant contract structure category that is the same as the relevant contract structure category to be analyzed is used as at least one relevant test sentence to be tested, and all relevant result sentences corresponding to the relevant contract structure category that is the same as the relevant contract structure category to be analyzed are used as multiple relevant result sentences to be tested, and the classified contract paragraph is used as a relevant contract paragraph to be analyzed; (F-3) For each relevant contract structure category to be analyzed of the relevant contract paragraph to be analyzed, based on all relevant result statements to be tested corresponding to the relevant contract structure category to be analyzed, all relevant test sentences to be tested corresponding to the relevant contract structure category to be analyzed, and the relevant contract paragraph to be analyzed, the relevant prediction model is used to obtain multiple relevant prediction results to be determined, each relevant prediction result to be determined includes multiple relevant probability values to be determined between each relevant result statement to be tested corresponding to the relevant contract structure category to be analyzed, each relevant test sentence to be tested corresponding to the relevant contract structure category to be analyzed, and the relevant contract paragraph to be analyzed, regarding the relevant results corresponding to the relevant contract structure category to be analyzed; and (F-4) For each relevant contract structure category to be analyzed of the relevant contract paragraph to be analyzed, according to all the to-be-determined relevant prediction results corresponding to the relevant contract structure category to be analyzed, a target relevance marking result indicating that the relevant contract paragraph to be analyzed belongs to one of all the relevant results corresponding to the relevant contract structure category to be analyzed is obtained. The contract assisted review method as described in claim 1, wherein the computer device further stores a feature extraction classification model for converting the sentence meaning vector to be classified into a structure category vector to be analyzed having the same dimension as the number of the contract structure categories, and step (C) further includes the following steps: (C-1) for each first sentence meaning vector, based on the first sentence meaning vector, using the feature extraction classification model to obtain a structure category vector to be processed corresponding to the first sentence meaning vector; and (C-2) for each structure category vector to be processed, based on the structure category vector to be processed, using a sigmoid function to map and transform each element in the structure category vector to be processed, so as to obtain a structure category vector to be analyzed corresponding to the structure category vector to be processed. The contract assisted review method as described in claim 1, wherein, in sub-step (F-3), for each relevant contract structure category to be analyzed of the relevant contract paragraph to be analyzed, the computer device further performs the following steps: (F-3-1) For each relevant test sentence to be detected of the relevant contract structure category to be analyzed, based on all relevant result sentences to be detected of the relevant contract structure category to be analyzed, the relevant test sentence to be detected and the relevant contract paragraph to be analyzed, a second sentence conversion vector model is used to obtain multiple second sentence meaning vectors of all relevant result sentences to be detected, the relevant test sentence to be detected and the relevant contract paragraph to be analyzed; (F-3-2) For each second semantic vector of the relevance contract structure category to be analyzed, a first linear transformation matrix is used according to the second semantic vector to obtain a relevance vector to be processed corresponding to the second semantic vector; and (F-3-3) For each relevance vector to be processed of the relevance contract structure category to be analyzed, a normalized exponential function is used to perform mapping transformation according to all elements of the relevance vector to be processed to obtain a relevance vector to be analyzed as the prediction result of the relevance to be determined, and all elements of the relevance vector to be analyzed are the relevance probability values to be determined. A contract assisted review method as described in claim 1, wherein in step (F-4), for each relevant contract structure category to be analyzed in the relevant contract paragraph to be analyzed, the computer device uses the relevance result corresponding to the largest relevant probability value to be determined among all relevant probability values to be determined corresponding to the relevant contract structure category to be analyzed as the target relevance marking result. The contract assisted review method as described in claim 1, wherein the computer device further stores a contract breach penalty prompt correspondence table and a paragraph extraction model, the paragraph extraction model is used to obtain a starting probability value corresponding to the first word of the description related to the other default sentence and an ending probability value corresponding to the last word of the description related to the other default sentence for each word in the one of the paragraphs according to another default sentence about the breach penalty description and one of the paragraphs, The contract penalty reminder corresponding table includes multiple penalty contract structure categories, and at least one penalty test sentence corresponding to each penalty contract structure category, and each penalty contract structure category is one of the contract structure categories. After step (E), the contract auxiliary review method further includes a step (G), for each classified contract paragraph, the computer device performs a penalty paragraph extraction process on the classified contract paragraph, wherein the penalty paragraph extraction process includes the following sub-steps: (G-1) Determine whether there is at least one target contract structure category to be analyzed that is identical to any of the contract structure categories in the contract default penalty prompt corresponding table in the at least one target contract structure category corresponding to the classified contract paragraph; (G-2) When it is determined that there is at least one contract structure category to be analyzed, for each contract structure category to be analyzed, according to the contract default penalty prompt corresponding table, the at least one default penalty test sentence corresponding to the contract structure category identical to the contract structure category to be analyzed is used as at least one default penalty test sentence to be tested, and the classified contract paragraph is used as a default penalty contract paragraph to be analyzed; (G-3) For each breach of contract structure category to be analyzed of the breach of contract paragraph to be analyzed, based on each breach of contract test sentence to be tested of the breach of contract structure category to be analyzed and the breach of contract paragraph to be analyzed, the paragraph extraction model is used to obtain at least one breach of contract extraction result to be determined, each breach of contract extraction result to be determined includes a breach of contract starting point probability value corresponding to the first word of the description of the corresponding breach of contract test sentence to be tested and a breach of contract ending point probability value corresponding to the last word of the description of the corresponding breach of contract test sentence to be tested for each word of the breach of contract paragraph to be analyzed; and (G-4) For each of the to-be-analyzed penalty contract structure categories of the to-be-analyzed penalty contract paragraph, according to all the to-be-determined penalty capture results of the to-be-analyzed penalty contract structure category, a target penalty capture result of a penalty paragraph defined by a target penalty starting point word and a target penalty ending point word in the to-be-analyzed penalty contract paragraph is obtained for the to-be-analyzed penalty contract structure category, and indicated. The contract assisted review method as described in claim 5, wherein, in step (G-3), for each breach of contract structure category to be analyzed in the breach of contract paragraph to be analyzed, the computer device further performs the following steps: (G-3-1) For each breach of contract test sentence to be tested in the breach of contract structure category to be analyzed, based on the breach of contract test sentence to be tested and the breach of contract paragraph to be analyzed, a sentence-to-word vector model is used to obtain multiple to-be-processed word meaning vectors for each word in the breach of contract paragraph to be analyzed in the breach of contract test sentence to be tested; (G-3-2) For each penalty test sentence to be tested in the penalty contract structure category to be analyzed, a second linear transformation matrix is used to obtain multiple penalty starting point probability values to be processed corresponding to the word meaning vectors to be processed according to all the penalty starting point probability values to be processed corresponding to the penalty test sentence to be analyzed, so as to obtain multiple penalty starting point probability values of the penalty test sentence to be tested for each word in the penalty contract paragraph to be analyzed; (G-3-3) For each penalty test sentence to be tested in the penalty contract structure category to be analyzed, a normalized index function is used to perform mapping transformation according to all the penalty starting point probability values to be processed corresponding to the penalty test sentence to be tested, so as to obtain multiple penalty starting point probability values of the penalty test sentence to be tested for each word in the penalty contract paragraph to be analyzed; (G-3-4) For each pending penalty test sentence of the pending penalty contract structure category, based on all pending word meaning vectors of each word in the pending penalty contract paragraph of the pending penalty test sentence, a third linear transformation matrix is used to obtain a plurality of pending penalty endpoint probability values corresponding to the pending word meaning vectors; (G-3-5) For each penalty test sentence to be tested of the penalty contract structure category to be analyzed, the normalized index function is used to perform mapping transformation according to all pending penalty endpoint probability values corresponding to the penalty test sentence to be tested, so as to obtain multiple penalty endpoint probability values of each word of the penalty contract paragraph to be analyzed for the penalty test sentence to be tested; and (G-3-6) For each penalty test sentence to be tested of the penalty contract structure category to be analyzed, all penalty starting probability values and all penalty endpoint probability values corresponding to the penalty test sentence to be tested are used as the penalty extraction result to be determined. The contract assisted review method as described in claim 5, wherein step (G-4) further comprises the following steps: (G-4-1) For each breach of contract structure category to be analyzed in the breach of contract paragraph to be analyzed, according to all breach of contract starting point probability values corresponding to the breach of contract structure category to be analyzed, a word with the largest breach of contract starting point probability value is obtained from all words in the breach of contract paragraph to be analyzed and used as the target breach of contract starting point word; (G-4-2) For each of the penalty contract structure categories to be analyzed in the penalty contract paragraph to be analyzed, according to all the penalty end point probability values corresponding to the penalty contract structure category to be analyzed, a word with the largest penalty end point probability value is obtained from all the words in the penalty contract paragraph to be analyzed and used as the target penalty end point word; and (G-4-3) For each of the penalty contract structure categories to be analyzed in the penalty contract paragraph to be analyzed, according to the target penalty starting point word and the target penalty end point word corresponding to the penalty contract structure category to be analyzed, the penalty paragraph is defined and used as the target penalty capture result. The contract assisted review method as described in claim 5, wherein after step (G-4) in the penalty paragraph extraction procedure, the following sub-steps are further included: (G-5) For each penalty contract structure category to be analyzed in the penalty contract paragraph to be analyzed, according to the penalty paragraph of the target penalty extraction result corresponding to the penalty contract structure category to be analyzed, using a natural language technology, a penalty value in the penalty paragraph is obtained; (G-6) For the penalty value corresponding to each penalty contract structure category to be analyzed in the penalty contract paragraph to be analyzed, determine whether the penalty value is greater than a warning amount; and (G-7) For the penalty value corresponding to each penalty contract structure type to be analyzed in the penalty contract segment to be analyzed, when it is determined to be greater than the warning amount, a warning signal indicating that the penalty segment corresponding to the penalty contract structure type to be analyzed needs to be warned is generated and displayed on a screen. A contract assisted review method as described in claim 1, wherein, in step (A), the contract document to be classified includes a confidential contract document, and the contract structure categories include a contract subject category, a contract purpose category, a confidential information category, a confidentiality obligation category, a confidentiality obligation exclusion category, a confidentiality period and effectiveness category, a rights ownership and guarantee category, a confidential information return category, a breach of contract liability damages category, a breach of contract liability penalty category, a governing law and jurisdiction category, and an other category. A contract review-assisted device is used to mark multiple paragraphs of a natural language contract document as at least one of multiple contract structure categories, including: A storage module stores a first sentence conversion vector model based on a transformer-based bidirectional encoder representation technology and used to convert a sentence to be converted into a sentence semantic vector to be classified, a contract key prompt corresponding table, and a transformer-based bidirectional encoder representation technology and used to obtain a correlation prediction model based on a preset sentence and a preset correlation result sentence indicating a correlation result and one of the paragraphs to obtain the correlation degree between the three. The contract key prompt corresponding table includes multiple correlation contract structure categories, at least one correlation test sentence corresponding to each correlation contract structure category, and multiple correlation result sentences corresponding to each correlation contract structure category corresponding to multiple correlation results, and each correlation contract structure category is one of the contract structure categories; An input module for receiving a contract document to be classified related to a natural language; A processing module electrically connected to the storage module and the input module; and wherein the processing module uses a natural language document parsing algorithm based on the contract document to be classified to obtain a plurality of contract paragraphs to be classified corresponding to the contract document to be classified, for each contract paragraph to be classified, the processing module uses the first sentence conversion vector model based on the contract paragraph to be classified to obtain a first sentence meaning vector corresponding to the contract paragraph to be classified, for each first sentence meaning vector, the processing module obtains a structure category vector to be analyzed corresponding to the first sentence meaning vector based on the first sentence meaning vector, and each element of the structure category vector to be analyzed corresponds to one of the contract structure categories, For each element of each architecture category vector to be analyzed, the processing module determines whether the value of the element is greater than a preset threshold. For each element of each architecture category vector to be analyzed, when the processing module determines that the value of the element is greater than the preset threshold, the processing module regards the contract paragraph to be classified corresponding to the architecture category vector to be analyzed as a classified contract paragraph, and regards the contract architecture category corresponding to the element as a target contract architecture category of the classified contract paragraph. For each classified contract paragraph, the processing module performs a relevance determination procedure on the classified contract paragraph, wherein the relevance determination procedure includes the processing module determining whether there is at least one target contract structure category to be analyzed that is the same as any of the relevant contract structure categories in the contract key prompt corresponding table in the at least one target contract structure category corresponding to the classified contract paragraph. When the processing module determines that there is at least one relevant contract structure category to be analyzed, for each relevant contract structure category to be analyzed, The processing module uses the at least one relevance test sentence corresponding to the relevance contract structure category that is the same as the relevance contract structure category to be analyzed as at least one relevance test sentence to be tested according to the contract key prompt corresponding table, and uses all relevance result sentences corresponding to the relevance contract structure category that is the same as the relevance contract structure category to be analyzed as multiple relevance result sentences to be tested, and uses the classified contract paragraph as a relevance contract paragraph to be analyzed, and for each relevance contract sentence to be analyzed of the relevance contract paragraph to be analyzed, The processing module obtains a plurality of correlation prediction results to be determined by using the correlation prediction model according to all correlation result sentences to be detected corresponding to the correlation contract structure category to be analyzed, all correlation test sentences to be detected corresponding to the correlation contract structure category to be analyzed, and the correlation contract paragraph to be analyzed. Each correlation prediction result to be determined includes each correlation result sentence to be detected corresponding to the correlation contract structure category to be analyzed, each correlation test sentence to be detected corresponding to the correlation contract structure category to be analyzed, and each correlation prediction result to be determined. The test sentence and the relevant contract paragraph to be analyzed have multiple to-be-determined correlation probability values for the correlation results corresponding to the relevant contract structure category to be analyzed. For each to-be-analyzed relevant contract structure category of the relevant contract paragraph to be analyzed, the processing module obtains a target correlation marking result indicating that the relevant contract paragraph to be analyzed belongs to one of all the correlation results corresponding to the relevant contract structure category to be analyzed based on all the to-be-determined correlation prediction results corresponding to the relevant contract structure category to be analyzed. The contract assisted review device as described in claim 10, wherein the storage module further stores a feature extraction classification model for converting the sentence meaning vector to be classified into a structure category vector to be analyzed with the same dimension as the number of the contract structure categories. For each first sentence meaning vector, the processing module uses the feature extraction classification model according to the first sentence meaning vector to obtain a structure category vector to be processed corresponding to the first sentence meaning vector. For each structure category vector to be processed, the processing module uses a sigmoid function according to the structure category vector to be processed to map and transform each element in the structure category vector to be processed to obtain a structure category vector to be analyzed corresponding to the structure category vector to be processed. The contract assisted review device as described in claim 10, wherein, when the processing module obtains the at least one relevance prediction result to be determined, for each relevance contract structure category to be analyzed in the relevance contract paragraph to be analyzed, the processing module performs a relevance probability value acquisition procedure to obtain the at least one relevance prediction result to be determined, and the relevance probability value acquisition procedure includes, For each relevance test sentence to be detected of the relevance contract structure category to be analyzed, the processing module uses a second sentence conversion vector model to obtain multiple second sentence meaning vectors about all relevance result sentences to be detected, the relevance test sentence to be detected, and the relevance contract paragraph to be analyzed according to all relevance result sentences to be detected of the relevance contract structure category to be analyzed, the relevance test sentence to be detected, and the relevance contract paragraph to be analyzed. For each second sentence meaning vector of the relevance contract structure category to be analyzed, the processing module uses a first linear conversion matrix according to the second sentence meaning vector to obtain a relevance vector to be processed corresponding to the second sentence meaning vector. For each correlation vector to be processed of the correlation contract structure category to be analyzed, the processing module uses a normalized index function to perform mapping transformation according to all elements of the correlation vector to be processed to obtain a correlation vector to be analyzed as the prediction result of the correlation to be determined, and all elements of the correlation vector to be analyzed are the correlation probability values to be determined. A contract assisted review device as described in claim 10, wherein, for each relevant contract structure category to be analyzed in the relevant contract paragraph to be analyzed, the processing module uses the relevance result corresponding to the largest relevant probability value to be determined among all relevant probability values to be determined corresponding to the relevant contract structure category to be analyzed as the target relevance marking result. The contract assisted review device as described in claim 10, wherein the storage module further stores a contract breach penalty prompt correspondence table and a paragraph extraction model, the paragraph extraction model is used to obtain a starting point probability value corresponding to the first word of the description related to the other default sentence and a probability value of belonging to the description related to the other default sentence for each word in the paragraph according to another default sentence about the breach penalty description and one of the paragraphs. The last word corresponds to a terminal probability value, the contract penalty prompt corresponding table includes a plurality of penalty contract structure categories, and at least one penalty test sentence corresponding to each penalty contract structure category, each penalty contract structure category is one of the contract structure categories, and for each classified contract paragraph, the processing module performs a penalty paragraph extraction procedure on the classified contract paragraph, wherein the penalty paragraph extraction procedure includes, The processing module determines whether there is at least one target contract structure category to be analyzed that is the same as any of the default penalty contract structure categories in the contract default penalty prompt corresponding table in the at least one target contract structure category corresponding to the classified contract paragraph. When the processing module determines that there is at least one default penalty contract structure category to be analyzed, for each default penalty contract structure category to be analyzed, the processing module uses the at least one default penalty test sentence corresponding to the default penalty contract structure category that is the same as the default penalty contract structure category to be analyzed as at least one default penalty test sentence to be tested, and uses the classified contract paragraph as a default penalty contract paragraph to be analyzed. For each breach of contract structure category to be analyzed of the breach of contract paragraph to be analyzed, the processing module obtains at least one breach of contract capture result to be determined using the paragraph capture model according to each breach of contract test sentence to be detected of the breach of contract structure category to be analyzed and the breach of contract paragraph to be analyzed. Each breach of contract capture result to be determined includes a breach of contract starting point probability value corresponding to the first word of the description of the corresponding breach of contract test sentence to be detected and a breach of contract ending point probability value corresponding to the last word of the description of the corresponding breach of contract test sentence to be detected for each word of the breach of contract paragraph to be analyzed. For each of the to-be-analyzed penalty contract structure categories of the to-be-analyzed penalty contract paragraph, the processing module obtains a target penalty capture result for the to-be-analyzed penalty contract structure category according to all the to-be-determined penalty capture results of the to-be-analyzed penalty contract structure category, and indicates a target penalty capture result for a penalty paragraph defined by a target penalty starting point word and a target penalty ending point word in the to-be-analyzed penalty contract paragraph. The contract assisted review device as described in claim 14, wherein, when the processing module obtains the at least one pending penalty capture result, for each pending penalty contract structure category of the pending penalty contract paragraph, the processing module performs an end point probability value acquisition procedure to obtain the at least one pending penalty capture result, and the start and end point probability value acquisition procedure includes, For each test sentence of the contract structure category of the contract to be analyzed, the processing module uses a sentence-to-word vector model to obtain multiple word meaning vectors to be processed for each word of the contract to be analyzed paragraph of the contract. For each test sentence of the contract to be analyzed, the processing module uses a second linear transformation matrix to obtain multiple starting probability values of the contract to be processed corresponding to the word meaning vectors to be processed, according to the test sentence of the contract to be analyzed and the contract paragraph of the contract to be analyzed. For each penalty test sentence to be tested of the penalty contract structure category to be analyzed, the processing module uses a normalized index function to perform mapping conversion according to all penalty starting point probability values to be processed corresponding to the penalty test sentence to be detected, so as to obtain multiple penalty starting point probability values of each word of the penalty test sentence to be detected with respect to the penalty contract paragraph to be analyzed, For each pending penalty test sentence of the pending penalty contract structure category, the processing module uses a third linear transformation matrix to obtain multiple pending penalty endpoint probability values corresponding to the pending penalty word meaning vectors according to all pending word meaning vectors of each word in the pending penalty contract paragraph of the pending penalty test sentence. For each penalty test sentence to be detected in the penalty contract structure category to be analyzed, the processing module uses the normalized index function to perform mapping conversion according to all penalty endpoint probability values to be processed corresponding to the penalty test sentence to be detected, so as to obtain multiple penalty endpoint probability values of each word of the penalty contract paragraph to be analyzed in the penalty test sentence to be detected. For each penalty test sentence to be detected in the penalty contract structure category to be analyzed, the processing module uses all penalty starting point probability values and all penalty endpoint probability values corresponding to the penalty test sentence to be detected as the penalty capture result to be determined. The contract assisted review device as described in claim 14, wherein, for each breach of contract structure category to be analyzed in the breach of contract paragraph to be analyzed, the processing module obtains a word with the largest breach of contract starting point probability value from all words in the breach of contract paragraph to be analyzed according to all breach of contract starting point probability values corresponding to the breach of contract structure category to be analyzed and uses it as the target breach of contract starting point word, For each of the penalty contract structure categories to be analyzed in the penalty contract paragraph to be analyzed, the processing module obtains a word with the largest penalty end point probability value from all the words in the penalty contract paragraph to be analyzed according to all the penalty end point probability values corresponding to the penalty contract structure category to be analyzed and uses it as the target penalty end point word. For each of the penalty contract structure categories to be analyzed in the penalty contract paragraph to be analyzed, the processing module defines the penalty paragraph according to the target penalty starting point word and the target penalty end point word corresponding to the penalty contract structure category to be analyzed and uses it as the target penalty capture result. The contract auxiliary review device as described in claim 14 further comprises a display module electrically connected to the processing module, wherein, for each default penalty contract structure category to be analyzed in the default penalty contract paragraph to be analyzed, the processing module obtains a default penalty value in the default penalty paragraph according to the default penalty paragraph of the target default penalty capture result corresponding to the default penalty contract structure category to be analyzed, for the default penalty value corresponding to each default penalty contract structure category to be analyzed in the default penalty contract paragraph to be analyzed, the processing module determines whether the default penalty value is greater than a warning amount, For the default penalty value corresponding to each default penalty contract structure type to be analyzed in the default penalty contract segment to be analyzed, when the processing module determines that it is greater than the warning amount, the processing module generates a warning signal indicating that the default penalty segment corresponding to the default penalty contract structure type to be analyzed needs to be warned and displays it on the display module. A contract-assisted review device as described in claim 10, wherein the contract document to be classified includes a confidential contract document, and the contract structure categories include a contract subject category, a contract purpose category, a confidential information category, a confidentiality obligation category, a confidentiality obligation exclusion category, a confidentiality period and effectiveness category, a rights attribution and guarantee category, a confidential information return category, a breach of contract liability damages category, a breach of contract liability penalty category, a governing law and jurisdiction category, and an other category.

Images