Control automation intelligent system based on large model fine tuning

Abstract

The invention discloses a control automation intelligent system based on large model fine tuning, and the system comprises a receiving unit which is used for receiving a first letter, carrying out the OCR recognition of the first letter, converting the first letter into a corresponding electronic document, and extracting the basic information of the letter from the electronic document through an information extraction model; the processing unit is used for introducing a fine tuning model to judge the cause of action of the first letter according to the basic information of the letter, classifying the cause of action and outputting cause classification and corresponding acceptance suggestions; and the problem feedback unit is used for judging whether the first letter is related to the public security law according to the cause classification and the corresponding acceptance suggestions, and monitoring the public security law when judging that the first letter is related to the public security law. Therefore, after the letter is received, case handling can be realized without participation of personnel with very rich case handling experience, and acceptance suggestions are given, so that the waste of human resources is greatly reduced, and meanwhile, the case handling efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent case management and control, and in particular to an automated intelligent system for complaints and appeals based on large model fine-tuning. Background Art

[0002] In the relevant technology, the case handling process of the procuratorate is complicated and changeable. After receiving the letter, it needs to be manually processed to give the acceptance opinion, which not only consumes a lot of human resources, but also requires personnel with rich case handling experience to handle it, and the case handling efficiency is low. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides an automated intelligent complaint and petition system based on large model fine-tuning. After receiving the first letter, the system can handle the case and give acceptance opinions without the participation of personnel with rich case handling experience, which greatly reduces the waste of human resources and improves case handling efficiency.

[0004] The technical solution adopted by the present invention is as follows: An automated intelligent complaint and petition system based on large model fine-tuning includes: a receiving unit, the receiving unit is used to receive a first letter, and perform OCR recognition on the first letter to convert the first letter into a corresponding electronic document, and extract basic information of the letter from the electronic document through an information extraction model, wherein the first letter is uploaded by the public; a processing unit, the processing unit is used to introduce a fine-tuning model to determine the case cause of the first letter according to the basic information of the letter, and to classify the case cause, and output the case cause classification and corresponding acceptance opinion; a problem feedback unit, the problem feedback unit is used to determine whether the first letter is related to the public security, procuratorial, judicial and judicial organs according to the case cause classification and the corresponding acceptance opinion, and to supervise the public security, procuratorial, judicial and judicial organs when it is determined that the first letter is related to the public security, procuratorial, judicial and judicial organs.

[0005] In one embodiment of the present invention, the automated intelligent system for complaint and application based on large model fine-tuning also includes a model training unit, which is used to: query and extract files related to previous letter monitoring work from a database to obtain a corresponding set of documents to be learned; construct a basic information type set and a document excerpt type set according to the file characteristics of the letter monitoring work; build an associated vocabulary of the basic information type set and the document excerpt type set according to the basic information type set and the document excerpt type set; identify all documents to be learned in the set of documents to be learned according to the associated vocabulary to obtain an identification queue set corresponding to each document to be learned; and train and obtain the information extraction model according to the identification queue set.

[0006] In one embodiment of the present invention, the model training unit identifies all the documents to be learned in the document set to be learned according to the associated vocabulary to obtain an identification queue set corresponding to each document to be learned, specifically including: dividing the corresponding excerpt types of the documents to be learned according to the document excerpt type set and the excerpt feature pairs of each excerpt of each of the documents to be learned; obtaining the basic information type corresponding to each type of document excerpt in each of the documents to be learned according to the associated vocabulary; identifying each type of document excerpt according to the basic information type corresponding to each type of document excerpt in each of the documents to be learned to obtain an identification queue of each type of document excerpt; obtaining the corresponding identification queue set of the document to be learned according to the recognition queue of each type of document excerpt.

[0007] In one embodiment of the present invention, the model training unit is also used to: extract basic information of letters from each document to be learned in the document set to be learned; rewrite the corresponding document to be learned according to the basic information of the letters; perform data augmentation on the document set to be learned according to the documents to be learned before and after the rewriting; and generate the fine-tuning model based on the training of the document set to be learned after the data augmentation.

[0008] In one embodiment of the present invention, the automated intelligent system for complaint and petition based on large model fine-tuning also includes: a situation analysis unit, which is used to perform situation analysis based on the current letter; and a report generation unit, which is used to generate an analysis report according to the situation analysis results.

[0009] In one embodiment of the present invention, the situation analysis unit is specifically used to: obtain the case type of the current letter, and query the database for letters with the same case type as the current letter to generate a set of candidate letters; perform case coding on the current letter and each letter in the candidate letter set; calculate the Jaccard distance between the case code corresponding to the current letter and the case code corresponding to each letter in the candidate letter set; select letters with a Jaccard distance less than a preset value to generate a target letter set; sort each letter in the target letter set in order of upload time, and perform situation analysis on the sorted letters.

[0010] In one embodiment of the present invention, the situation analysis unit performs case coding on the current letter and each letter in the candidate letter set, specifically including: extracting case features in the electronic document corresponding to each letter, wherein the case features include: public positions, public emotions, events, core words and monetary transaction amounts; except for the public positions and public emotions, the remaining case features are represented in chronological order, wherein events that do not occur simultaneously with other case features are represented as independent events, and events that occur simultaneously with other case features are represented in combination; wherein the event features of the event occurring at the kth moment are represented by the following formula: , Among them, m is the number of case features occurring simultaneously at the kth moment, represents the event characteristics of the event occurring at the kth moment, The numerical representation of the j-th case feature that occurs at the same time at the k-th time. ; The event occurring at the kth moment is expressed by the following formula: , in, , ,……, are the weight coefficients corresponding to each element in the event feature representation of the event occurring at the kth moment, , is the vector representation of the event occurring at the kth moment; The event representations of each time node are connected in series, and the case code is generated by the following formula: , Where n is the total number of independent case features and merged case features extracted from the electronic document corresponding to the letter.

[0011] Beneficial effects of the present invention: After receiving the letter, the present invention can handle the case and give acceptance opinions without the participation of personnel with rich case handling experience, which greatly reduces the waste of human resources and improves the case handling efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a block diagram of an intelligent control and petition automation system based on large model fine-tuning according to an embodiment of the present invention; Figure 2 The block diagram of an intelligent control and petition automation system based on large model fine-tuning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] like Figure 1 As shown, the automated intelligent system for complaint and application based on large model fine-tuning of an embodiment of the present invention may include: a receiving unit 100, a processing unit 200 and a problem feedback unit 300.

[0015] Among them, the receiving unit 100 is used to receive a first letter, and perform OCR (Optical Character Recognition) recognition on the first letter to convert the first letter into a corresponding electronic document, and extract basic information of the letter from the electronic document through an information extraction model, and the first letter is uploaded by the public; the processing unit 200 is connected to the receiving unit 100 through an electrical signal, and the processing unit 200 is used to receive the basic information of the letter sent by the receiving unit 100, and introduce a fine-tuning model to determine the case of the first letter according to the basic information of the letter, and classify the case, and output the case classification and corresponding acceptance opinions; the problem feedback unit 300 is connected to the processing unit 200 through an electrical signal, and the problem feedback unit 300 is used to determine whether the first letter is related to the public security, procuratorial, judicial and judicial organs according to the case classification and the corresponding acceptance opinions, and when it is determined that the first letter is related to the public security, procuratorial, judicial and judicial organs, the judgment result is fed back to the monitoring terminal, wherein the monitoring terminal is used to monitor the public security, procuratorial, judicial and judicial organs.

[0016] It is understandable that before the receiving unit 100 extracts the basic information of the letter from the electronic document through the information extraction model, the information extraction model may be pre-trained.

[0017] In one embodiment of the present invention, Figure 2 As shown, the control and petition automation intelligent system based on large model fine-tuning also includes a model training unit 400.

[0018] Among them, the model training unit 400 is used to: query and extract relevant files of previous letter monitoring work from the database to obtain the corresponding document set to be learned; construct a basic information type set and a document excerpt type set according to the file characteristics of the letter monitoring work; build an associated vocabulary of the basic information type set and the document excerpt type set according to the basic information type set and the document excerpt type set; identify all the documents to be learned in the document set to be learned according to the associated vocabulary to obtain the identification queue set corresponding to each document to be learned; train and obtain the information extraction model according to the identification queue set.

[0019] Specifically, the database stores documents related to previous letter supervision work, which may include letters from previous years, etc. After obtaining letters from previous years, they can be converted into corresponding document sets to be learned through OCR recognition. Then, according to the file characteristics of letter supervision work, a basic information type set and a document excerpt type set are constructed, and a related word library of the basic information type set and the document excerpt type set is constructed based on the basic information type set and the document excerpt type set. Specifically, taking letters as an example, the electronic documents corresponding to letters can be divided into document excerpts of types such as basic information of the people, case description, and purpose of uploading by the people. The corresponding basic information types may include: people's names, people's education, people's positions, time of case occurrence, location of case occurrence, subject of problem, persons involved in the case, and main demands. Among them, when building the related word library of the basic information type set and the document excerpt type set, in the document excerpt of the basic information type of the people, the associated basic information types are people's names, people's education and people's positions; in the document excerpt of the case description type, the associated basic information types are case time, case location, subject of problem and persons involved in the case; in the document excerpt of the purpose of uploading by the people, the associated basic information type is the main demand.

[0020] In one embodiment of the present invention, the model training unit 400 identifies all documents to be learned in the document set to be learned according to the associated vocabulary to obtain an identification queue set corresponding to each document to be learned, specifically including: dividing the corresponding excerpt types of the documents to be learned according to the document excerpt type set and the excerpt feature pairs of each excerpt of each document to be learned; obtaining the basic information type corresponding to each type of document excerpt in each document to be learned according to the associated vocabulary; identifying each type of document excerpt according to the basic information type corresponding to each type of document excerpt in each document to be learned to obtain an identification queue of each type of document excerpt; obtaining the corresponding identification queue set of the document to be learned according to the recognition queue of each type of document excerpt.

[0021] Specifically, when marking each document to be studied, the corresponding excerpt type of the document to be studied can be divided according to the document excerpt type set and the excerpt feature of each excerpt of each document to be studied. For example, the electronic document corresponding to the letter can be divided into document excerpt types such as basic information, case description, and purpose of public upload. Then, the basic information type corresponding to each type of document excerpt in each document to be studied is obtained according to the associated vocabulary, and each type of document excerpt is marked to obtain the identification queue of each type of document excerpt. Finally, the identification queue of each type of document excerpt is integrated.

[0022] In one embodiment of the present invention, after extracting the basic information of the first letter from the electronic document corresponding to the first letter, the processing unit 200 can introduce a fine-tuning model to determine the case of the first letter according to the basic information of the letter, classify the case, and output the case classification and the corresponding acceptance opinion. Of course, before calling the fine-tuning model, it is also necessary to obtain the fine-tuning model through training based on previous letters through the model training unit 400.

[0023] However, due to the particularity of the procuratorial system, the data of the regional procuratorate is one-sided and difficult to cover the complete complaint situation, and the generalization ability is poor. Therefore, it is necessary to augment the data stored in the database and retrain the fine-tuning model based on the augmented data. It is understandable that the information extraction model can also be retrained based on the augmented data.

[0024] In one embodiment of the present invention, the model training unit 400 is also used to: extract basic information of letters from each document to be learned in the document set to be learned; rewrite the corresponding document to be learned according to the basic information of the letters; perform data augmentation on the document set to be learned according to the documents to be learned before and after the rewriting; and generate a fine-tuning model based on the training of the document set to be learned after data augmentation.

[0025] Among them, the basic information of the letter may include people's educational background, age, emotions, and whether they have lawyers. Specifically, people of different educational backgrounds and age groups have different styles of explaining cases, and the emotions of most people will also affect the wording in the final letter. Whether there is a lawyer will also affect the professionalism of the letter. These can all be used as dynamic parameters for data augmentation. Combining these conditions as random items can increase the number of cases that were originally few. In addition, the desensitization or random generation of information such as names / related case numbers can also improve the generalization ability of the fine-tuning model.

[0026] In one embodiment of the present invention, Figure 2 As shown, the automated intelligent system for complaint based on large model fine-tuning also includes: a situation analysis unit 500 and a report generation unit 600. The situation analysis unit 500 is used to perform situation analysis based on the current letter; the report generation unit 600 is used to generate an analysis report based on the situation analysis results.

[0027] In one embodiment of the present invention, the situation analysis unit 500 is specifically used to: obtain the case type of the current letter, and query the database for letters with the same case type as the current letter to generate a set of candidate letters; perform case coding on the current letter and each letter in the candidate letter set; calculate the Jaccard distance between the case code corresponding to the current letter and the case code corresponding to each letter in the candidate letter set; select letters with a Jaccard distance less than a preset value to generate a target letter set; sort each letter in the target letter set according to the order of upload time by the public, and perform situation analysis on the sorted letters.

[0028] Specifically, when the situation analysis unit 500 performs the situation analysis, it is necessary to perform a comprehensive analysis based on the current letter and in combination with previous letters. However, in the actual analysis, if the type of the previous letter is quite different from that of the current letter, not only will the situation analysis result be quite different from the actual result, but also the amount of calculation of the situation analysis unit 500 for situation analysis will be greatly increased, affecting the analysis speed. Therefore, in the present invention, the situation analysis unit 500 first obtains the case type of the current letter, and queries the letters of the same case type as the current letter from the database to generate a candidate letter set. Then, the case is coded for each letter in the current letter and the candidate letter set, and the Jaccard distance between the case code corresponding to the current letter and the case code corresponding to each letter in the candidate letter set is calculated, and the letters with a Jaccard distance less than a preset value are selected to generate a target letter set. Finally, each letter in the target letter set is sorted in the order of upload time, and the sorted letters are situationally analyzed.

[0029] In one embodiment of the present invention, the situation analysis unit 500 performs case coding on each letter in the current letter and the candidate letter set, specifically including: extracting case features in the electronic documents corresponding to each letter, wherein the case features include: public positions, public emotions, events, core words and monetary transaction amounts; except for public positions and public emotions, the remaining case features are represented in chronological order, wherein events that do not occur simultaneously with other case features are represented as independent events, and events that occur simultaneously with other case features are represented in combination; wherein the event features of the event occurring at the kth moment are represented by the following formula: , Among them, m is the number of case features occurring simultaneously at the kth moment, represents the event characteristics of the event occurring at the kth moment, The numerical representation of the j-th case feature that occurs at the same time at the k-th time. ; The event occurring at the kth moment is expressed by the following formula: , in, , ,……, are the weight coefficients corresponding to each element in the event feature representation of the event occurring at the kth moment, , is the vector representation of the event occurring at the kth moment; The event representations of each time node are linked together and the case code is generated using the following formula: , Where n is the total number of independent case features and merged case features extracted from the electronic document corresponding to the letter.

[0030] In summary, according to the large model fine-tuning based automated intelligent complaint and petition system of the embodiment of the present invention, the first letter is received by the receiving unit, and the first letter is subjected to OCR recognition to convert the first letter into a corresponding electronic document, and the basic information of the letter is extracted from the electronic document by the information extraction model, and the case of the first letter is determined according to the basic information of the letter by introducing the fine-tuning model through the processing unit, and the case is classified, and the case classification and corresponding acceptance opinions are output, and the problem feedback unit is used to determine whether the first letter is related to the public security, procuratorial, judicial and judicial organs according to the case classification and the corresponding acceptance opinions, and the public security, procuratorial, judicial and judicial organs are monitored when it is determined to be related. Therefore, after receiving the letter, the case can be handled and the acceptance opinions can be given without the participation of personnel with rich case handling experience, which greatly reduces the waste of human resources and improves the efficiency of case handling.

[0031] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. "Multiple" means two or more, unless otherwise clearly and specifically defined.

[0032] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0033] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0034] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0035] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0036] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. An automated intelligent control and complaint system based on large model fine-tuning, characterized in that: include: a receiving unit, the receiving unit being used to receive a first letter, and perform OCR recognition on the first letter to convert the first letter into a corresponding electronic document, and extract basic information of the letter from the electronic document through an information extraction model, wherein the first letter is uploaded by a citizen; A processing unit, wherein the processing unit is connected to the receiving unit via an electrical signal, and the processing unit is used to receive the basic information of the letter sent by the receiving unit, and introduce a fine-tuning model to determine the case cause of the first letter according to the basic information of the letter, and classify the case cause, and output the case cause classification and corresponding acceptance opinion; A problem feedback unit, wherein the problem feedback unit is connected to the processing unit via an electrical signal, and the problem feedback unit is used to determine whether the first letter is related to the public security, procuratorial, judicial and judicial organs according to the case classification and the corresponding acceptance opinion, and when it is determined to be related to the public security, procuratorial and judicial organs, the judgment result is fed back to the monitoring terminal, wherein the monitoring terminal is used to monitor the public security, procuratorial and judicial organs.

2. The automated intelligent control and petition system based on large model fine-tuning according to claim 1 is characterized in that: It also includes a model training unit, which is used to: Query and extract documents related to previous letter monitoring work from the database to obtain the corresponding document set to be studied; Construct basic information type set and document excerpt type set according to the document characteristics of letter monitoring work; Building a related word library of the basic information type set and the document excerpt type set according to the basic information type set and the document excerpt type set; Identify all the documents to be learned in the set of documents to be learned according to the associated vocabulary to obtain an identification queue set corresponding to each document to be learned; The information extraction model is trained and obtained according to the identification queue set.

3. The automated intelligent control and petition system based on large model fine-tuning according to claim 2 is characterized in that: The model training unit identifies all the documents to be learned in the set of documents to be learned according to the associated vocabulary to obtain an identification queue set corresponding to each document to be learned, specifically including: Dividing the corresponding excerpt types of the documents to be learned according to the document excerpt type set and the excerpt feature pairs of each excerpt of each of the documents to be learned; Acquire the basic information type corresponding to each type of document excerpt in each of the documents to be learned according to the associated vocabulary; According to the basic information type corresponding to each type of document excerpt in each of the documents to be learned, each type of document excerpt is identified to obtain an identification queue of each type of document excerpt; According to the recognition queue of each type of document excerpt, the corresponding identification queue set of the document to be learned is obtained.

4. The automated intelligent control and petition system based on large model fine-tuning according to claim 2 is characterized in that: The model training unit is also used for: Extracting basic information of letters from each document to be studied in the document set to be studied; Rewrite the corresponding document to be studied according to the basic information of the letter; Performing data augmentation on the set of documents to be learned according to the documents to be learned before and after the rewriting; The fine-tuning model is generated by training the document set to be learned after data augmentation.

5. The automated intelligent control and petition system based on large model fine-tuning according to claim 2 is characterized in that: Also includes: A situation analysis unit, the situation analysis unit is used to perform situation analysis based on the current letter; A report generating unit, wherein the report generating unit is used to generate an analysis report according to the situation analysis result.

6. The automated intelligent control and petition system based on large model fine-tuning according to claim 5 is characterized in that: The situation analysis unit is specifically used for: Obtaining the case type of the current letter, and searching the database for letters of the same case type as the current letter to generate a candidate letter set; Performing case coding on the current letter and each letter in the candidate letter set; Calculate the Jaccard distance between the case code corresponding to the current letter and the case code corresponding to each letter in the candidate letter set; Select the letters whose Jaccard distance is less than the preset value to generate the target letter set; Each letter in the target letter set is sorted according to the order of upload time, and a situation analysis is performed on the sorted letters.

7. The automated intelligent control and petition system based on large model fine-tuning according to claim 6 is characterized in that: The situation analysis unit performs case coding on the current letter and each letter in the candidate letter set, specifically including: Extracting case features from the electronic documents corresponding to each letter, wherein the case features include: public positions, public emotions, events, core words and monetary transaction amounts; Except for the public positions and public emotions, the remaining case features are represented in chronological order, where events that do not occur simultaneously with other case features are represented as independent events, and events that occur simultaneously with other case features are represented together; where the event feature of the event occurring at the kth moment is represented by the following formula: ， Among them, m is the number of case features occurring simultaneously at the kth moment, represents the event characteristics of the event occurring at the kth moment, The numerical representation of the j-th case feature that occurs at the same time at the k-th time. ; The event occurring at the kth moment is expressed by the following formula: ， in, , ,……, are the weight coefficients corresponding to each element in the event feature representation of the event occurring at the kth moment, , is the vector representation of the event occurring at the kth moment; The event representations of each time node are connected in series, and the case code is generated by the following formula: ， Where n is the total number of independent case features and merged case features extracted from the electronic document corresponding to the letter.

Images