CN113946649A

CN113946649A - Providing method of mediation plan, training method, related device and storage medium

Info

Publication number: CN113946649A
Application number: CN202010690804.4A
Authority: CN
Inventors: 周鑫; 孙常龙; 张雅婷
Original assignee: Alibaba Cloud Computing Ltd
Current assignee: Alibaba Cloud Computing Ltd
Priority date: 2020-07-17
Filing date: 2020-07-17
Publication date: 2022-01-18

Abstract

Providing method, training method, related device and storage medium of the mediation plan, wherein the providing method of the mediation plan comprises the following steps: obtaining input information related to a case to be mediated; and inputting the input information into a trained mediation plan providing model to obtain one or more output mediation plans corresponding to the case to be mediated. By adopting the method for providing the mediation plan in the embodiment of the specification, the mediation plan can be directly provided with high efficiency, and the mediation efficiency can be improved by assisting mediation personnel; in addition, more mediation reference information can be provided when the mediation plan content is selected, which is helpful for improving the mediation success rate.

Description

Providing method of mediation plan, training method, related device and storage medium

Technical Field

The embodiment of the specification relates to the technical field of data processing, in particular to a providing method, a training method, a related device and a storage medium of a mediation plan.

Background

With the rapid development of commercial economy, the number of various commercial disputes has also increased dramatically. Particularly, the internet economy is closely combined with the clothes and food residents of people, so that the rapid and accurate online shopping service is brought; however, there are a lot of legal disputes, such as network transaction disputes.

When the dispute is resolved, the dispute is resolved by the dispute parties in a relatively efficient manner. However, at present, no suitable intelligent tool is provided, and assistance for forming a mediation scheme is provided when a mediator, a judge and the like solve a dispute, which is not beneficial to the resolution efficiency of the dispute.

Disclosure of Invention

In view of this, embodiments of the present disclosure provide a providing method, a training method, a related apparatus, and a storage medium for a mediation plan, which can efficiently generate an accurate mediation plan.

First, an embodiment of the present specification provides a method for providing a mediation plan, including:

acquiring case information of a case to be mediated;

obtaining mediation reference information of the case to be mediated according to the case information;

and inputting the case information and the mediation reference information into a mediation plan providing model to obtain one or more mediation plans of the case to be mediated.

Optionally, the mediation reference information includes class case information; the class information is obtained by the following method:

extracting a first case from a preset referee document library;

calculating the similarity between the case to be mediated and the first case according to the case information;

and taking the first case with the similarity exceeding a preset threshold value as the case to be mediated, and obtaining the case information.

Optionally, the referee library is obtained by:

and classifying the historical referee documents according to judicial factors to obtain a referee document library.

Optionally, the method further includes:

and structuring the referee text library.

Optionally, the extracting the first case from the preset referee library includes:

and extracting a first case with judicial elements matched with the case to be mediated from the preset referee text library.

Optionally, the calculating, according to the case information, the similarity between the case to be mediated and the first case includes:

extracting case texts from case information, and carrying out vector coding on the case texts of the cases to be mediated to obtain first text semantic features;

extracting a case text of a first case, and carrying out vector coding on the case text of the first case to obtain a second text semantic feature;

and calculating the cosine distance between the first text semantic feature and the second text semantic feature to obtain the similarity.

Optionally, the class case information includes statistical results of litigation results for the class case, the method further comprising:

counting the statistical results of the litigation results of the multiple classes to obtain statistical results;

and the statistical result is used for being input into the mediation plan providing model together with the case information and the class information to obtain one or more mediation plans of the case to be mediated.

Optionally, the mediation reference information further includes party characteristic information, where the party characteristic information includes: and presentation information of one or more combinations of the party's complaint records, executed records, and risk preferences.

Optionally, the mediation reference information further includes: appeal result prediction information; the appeal result prediction information is obtained through the following method:

calculating to obtain legal provision prediction information of a legal provision applicable to the case to be mediated according to the case information;

and predicting support of the appeal of the party in the case information according to the legal forecast information to obtain the appeal result forecast information.

The embodiment of the present specification provides a training method of a mediation plan providing model suitable for mediation plan provision, including:

obtaining a sample data set; wherein the sample data set comprises a training data set comprising: case information of each historical case, mediation reference information obtained according to the case information and a mediation result;

and inputting the sample data set into a mediation plan providing model so that the mediation plan providing model is trained according to the loss between the mediation plan obtained by the mediation plan providing model and the mediation result.

extracting a first case from a preset referee document library;

calculating the similarity between the historical case and the first case according to the case information;

and taking the first case with the similarity exceeding a preset threshold value as the historical case, and acquiring the information of the case.

Optionally, the referee library is obtained by:

Optionally, the method further includes:

and structuring the referee text library.

and extracting a first case with judicial elements matched with the historical cases from the preset referee text library.

Optionally, the calculating the similarity between the historical case and the first case according to the case information includes:

extracting case texts from case information, and carrying out vector coding on the case texts of the historical cases to obtain first text semantic features;

and the statistical result is used for being input into the mediation plan providing model together with the case information and the class information to obtain one or more mediation plans of the historical cases.

calculating to obtain applicable legal rules of historical cases according to the case information;

A further embodiment of the present specification provides a provision apparatus for a mediation plan, including:

the data acquisition unit is used for acquiring case information of a case to be mediated;

the information obtaining unit is used for obtaining mediation reference information of the case to be mediated according to the case information;

and the input unit is used for inputting the case information and the mediation reference information into a mediation plan providing model to obtain one or more mediation plans of the case to be mediated.

An embodiment of the present specification further provides a training device suitable for a mediation plan provision model provided by a mediation plan, where the training device includes:

a data obtaining unit, configured to obtain a sample data set; wherein the sample data set comprises a training data set comprising: case information about each historical case, mediation reference information obtained according to the case information, and a mediation result;

and the training unit is used for inputting the sample data set into the mediation plan providing model so as to enable the mediation plan providing model to obtain training according to the loss between the mediation plan obtained by the mediation plan providing model and the mediation result. The present specification further provides an electronic device, including a memory and a processor, where the memory stores computer instructions executable on the processor, and the processor executes the computer instructions to perform the steps of the method according to any one of the foregoing embodiments.

The present specification also provides a computer readable storage medium, on which computer instructions are stored, and the computer instructions execute the steps of the method of any one of the foregoing embodiments when executed.

Compared with the prior art, the technical scheme of the embodiment of the specification has the following beneficial effects:

on one hand, by adopting the method for providing the mediation plan in the embodiment of the specification, case information and mediation reference information related to a case to be mediated are obtained and input into a mediation plan providing model, so that one or more output mediation plans corresponding to the case to be mediated are obtained; the whole process is automatically completed by providing a model based on a trained mediation plan, and the method is efficient and direct; moreover, the dispute resolution method is directly provided for a moderator, a legal officer and the like in a mode of mediating a plan, and compared with an indirect mode of providing legal provisions and similar cases, the dispute resolution method is higher in efficiency.

On the other hand, by adopting the method for providing the mediation plan in the embodiment of the present specification, the mediation reference information may include laws, appeal support, characteristic information of the parties, and the like, and the mediation plan generated by referring to the information makes the content of the mediation plan legally and more acceptable to the parties, thereby improving the success rate of mediation.

Drawings

Fig. 1A is a schematic diagram of an example of an application scenario in an embodiment of this specification.

Fig. 1B is a schematic diagram of another application scenario example in this embodiment.

Fig. 1C is a schematic diagram of another application scenario example in the embodiment of the present specification.

Fig. 2 is a schematic flow chart of a mediation plan providing method in the embodiment of the present specification.

Fig. 3 is a schematic flow chart of obtaining class case information in the embodiment of the present specification.

Fig. 4 is a schematic flow chart of calculating similarity in the flow of obtaining the class case information in the embodiment of the present specification.

Fig. 5 is a schematic structural diagram of a class information acquisition system in an embodiment of the present specification.

Fig. 6 is a flowchart illustrating obtaining appeal result prediction information in an embodiment of the present disclosure.

Fig. 7 is a schematic structural diagram of a legal issue prediction model in an embodiment of the present specification.

Fig. 8 is a schematic structural diagram of a mediation plan provision model in the embodiment of the present specification.

Fig. 9 is a schematic structural diagram of a part of functional units corresponding to case information processing in the mediation plan providing model of fig. 8.

Fig. 10 is a schematic diagram of the residual structure in the embodiment of the present specification.

Fig. 11 is a flowchart of a training method of a mediation plan providing model suitable for mediation plan provision in an embodiment of the present specification.

Fig. 12 is a schematic structural diagram of a provision device of the mediation plan in the embodiment of the present specification.

Fig. 13 is a schematic structural diagram of a training device suitable for a mediation plan provision model of mediation plan provision in the embodiment of the present specification.

Fig. 14 is a schematic structural diagram of an electronic device in an embodiment of the present specification.

Detailed Description

The number of legal dispute events, especially network transaction disputes, is increasing in enormous quantities. More targets for network transaction disputes are smaller, and the method is more suitable for mediation and resolution. In order to improve the mediation efficiency, the mediation mode is gradually changed from off-line mediation to on-line mediation.

However, the existing online mediation basically depends on the experience of the mediator, so that the online mediation of the target just gets through the space barrier and does not provide corresponding assistance for the mediator in dealing with dispute events.

In view of this, in the embodiments of the present specification, an artificial intelligence technique is applied to the field of legal dispute resolution, so that a resolution plan can be output intelligently and automatically, so as to better assist a moderator in processing dispute resolution transactions quickly, efficiently, and accurately.

It should be noted that the application of the embodiment of the present disclosure may be to perform various legal dispute resolution scenarios, for example, to assist a moderator in network transaction dispute, or to assist a judge in legal action to efficiently handle dispute resolution affairs.

Reference is made to fig. 1A, which is a schematic diagram of an application scenario example in the embodiment of the present specification. In this application scenario, a mediation plan interface 11 is presented on the screen of the user terminal 10. Although the user terminal is a desktop computer in the example of fig. 1A, in other examples, the user terminal 10 may also be a smart phone, a notebook computer, a tablet computer, a smart band, a smart watch, or other smart devices, which is not limited to the illustration.

The mediation plan interface is exemplary. Optionally, in the mediation plan interface 11, a plurality of mediation plans are displayed for selection, such as "mediation plan one", "mediation plan two", and "mediation plan three". The mediation plan may be defined as a plan agreed upon between parties to the dispute, generally a support for the original appeal, and a constraint on the execution means.

For example, FIG. 1A also shows "original claim," original claim A claiming "hope triple offset, return good" and claim B hoping "return good" with the content of the proposed reconciliation plan tending to reconcile A, B. For example, the content of the first mediation plan is "refund payment, triple compensation", the content of the second mediation plan is "refund payment, double compensation", and the content of the third mediation plan is "refund payment".

Optionally, the mediation plan interface 11 also provides success rate for each mediation plan, for example, 60%, 20%, 15% in the figure, and may be sorted for selection. The success rate may be predicted by considering one or a combination of factors such as case, law, party, etc.

After referring to the interface 11 for displaying the mediation plan, a mediator (or a mediation person such as a judge) can choose one of the mediation plans to perform mediation work between the parties. The higher the success rate of the selection of a mediation plan, the better the consensus may be reached by the parties when the mediation is performed.

Refer to fig. 1B for a schematic diagram of another example of an application scenario in an embodiment of this specification.

The scene in this example may be a scene of dispute resolution performed offline, such as a court or a judicial bureau. In these scenarios, the moderator 12 may use, for example, the user terminal 10 in the embodiment of fig. 1A to assist in work, and perform the moderation work with reference to the information of each moderation plan provided by the user terminal 10 in the process of communicating with the party 13.

Refer again to fig. 1C, which is a schematic diagram of another application scenario example in the embodiment of this specification.

In the communication system provided in the figure, parties 13 and a moderator 12 access a network 16 through

user terminals

10, 14 and 15 to perform online communication, and each of the

user terminals

10, 14 and 15 needs to have a network communication capability; each of the

user terminals

10, 14, 15 may be, for example, a smart phone, a tablet computer, a desktop computer, a notebook computer, etc., and is equipped with a display screen and a wired or wireless communication module; the network 16 may be a wired or wireless internet or the like.

In this scenario, the moderator 12 may communicate with the principal 13 online, and the user terminal 10 of the moderator 12 may display a mediation plan interface as the user terminal 10 in fig. 1A, and the moderator 12 may perform mediation work with reference to the mediation plan therein.

Optionally, a service terminal 17 (e.g. a server or a server group) running an intelligent mediation platform may be provided in the scenario, and both the mediator 12 and the party 13 may communicate with the service terminal 17 through the

respective user terminals

10, 14, 15, and log in the intelligent mediation platform by using the held identity ID, so as to complete the mediation work through online communication. The online communication mode comprises the following steps: online multiparty video, multiparty teleconference, multiparty online conversation, online text exchange, and the like.

Alternatively, the intelligent mediation platform may be used to generate each of the mediation plans and send them to the user terminal 10 held by the mediator 12 for display in the form of a graphical user interface, for example, in fig. 1A. Optionally, the service terminal 17 and each

user terminal

10, 14, 15 may interact based on a Browser/Server (B/S) architecture, that is, the data transmitted by the intelligent mediation platform to each

user terminal

10, 14, 15 may be displayed in a page Browser (Browser) of the

user terminal

10, 14, 15, or may be displayed in a page provided in a platform program (e.g., a small program of a pay pal); alternatively, the service terminal 17 and each of the

user terminals

10, 14, 15 may also interact based on a Client/Server (C/S) architecture, and the mediation plan interface 11 may be displayed on the screen of the user terminal 10 through Client software communicating with the service software.

The flow and principle of the method for providing the mediation plan in the embodiments of the present specification are described below by using a plurality of embodiments.

Referring to a flow diagram of a mediation plan providing method in the embodiment of the present specification shown in fig. 2, an execution flow includes the following steps:

and S21, obtaining case information of the case to be mediated.

In particular implementations, the obtaining refers to receiving from outside or generating locally.

The case information of the case to be mediated may be from text information in a paper or electronic document of the case to be mediated, such as text information recorded in front of a court and describing the case of the case to be mediated, where the content may include: party information (e.g., original, advertised name, identity, etc.), party complaints (e.g., original complaints), case process descriptions, etc.; optionally, the recording mode includes: by handwriting a record on a paper document, inputting characters or voice to an electronic device through an input device (such as a keyboard, a microphone), and the like; wherein the material of the handwritten record (which may be made by, for example, a moderator, a bookmarker, etc.) may form an image by, for example, scanning, photographing, and further recognized as text information in the form of electronic data by, for example, Optical Character Recognition (OCR) Character Recognition technology; alternatively, data of the voice input may be converted into text information in the form of electronic data by a voice recognition technique.

And S22, obtaining the mediation reference information of the case to be mediated according to the case information.

In an alternative example, the mediation reference information may include: and (4) class case information.

As shown in fig. 3, the manner of acquiring the class information in the embodiment of this specification may include the following steps:

step S31: and extracting the first case from a preset referee document library.

In a specific implementation, the official document library is obtained by: and classifying the historical referee documents according to judicial factors to obtain a referee document library.

Specifically, the judicial factors include: and presentation information of text information related to legal concepts involved in case-approval. Judicial elements are concepts that are legally abstracted. Taking the case of dispute loan as an example, it is assumed that text information such as "whether there is an electronic agreement", "agreeable loan", "road overtaking", "payment" recorded in the case information and the actual dispute case auditing process need to be identified as two elements: the text information may be encoded as a judicial element in the case where there is a consensus of a civil loan between the parties (i.e., "lending consensus") and the case of the civil loan requires the lender to pay (i.e., "delivery").

Optionally, the official document library may be structured, for example, by using a structured database to store official document data marked with judicial elements in an associated manner.

Step S32: and calculating the similarity between the case to be mediated and the first case according to the case information.

As shown in fig. 4, the calculation manner of the similarity in step S32 may include:

step S41: and extracting case texts from the case information, and carrying out vector coding on the case texts of the cases to be mediated to obtain first text semantic features.

Step S42: extracting a case text of a first case, and carrying out vector coding on the case text of the first case to obtain a second text semantic feature;

step S43: and calculating the cosine distance between the first text semantic feature and the second text semantic feature to obtain the similarity.

The vector coding refers to converting case texts in the case information of the first case and the case to be mediated into vector-form semantic features, and measuring and obtaining text semantic similarity between the case texts of each first case and the case to be mediated by calculating cosine distances and the like between the semantic features. Alternatively, the vector conversion method may be, for example, a one-hot (one-hot) method, or a Word Embedding (Word Embedding) method.

Optionally, a Bi-directional Long Short-Term Memory (Bi-directional Long Short-Term Memory, BiLSTM) model may be used to measure the similarity of the text. The BilSTM is formed by combining a forward long-short memory model (LSTM) and a backward LSTM, and the LSTM is very suitable for natural language processing; the LSTM can learn to memorize certain information and forget certain information through a training process, and the dependence relationship of a longer distance can be better captured by using an LSTM model; however, the LSTM model is deficient in the information coding from the back to the front of the sentence, so that the problem can be well compensated by the BilTM. However, it should be noted that the above is only an example of the implementation manner of the similarity calculation, and is not limited.

It can be understood that if the text semantic features are used for the first matching, more noise is introduced; therefore, in the above example, the first case similar to the case to be mediated in judicial elements is screened out first, and then the class case is screened out from the first case according to the semantic similarity of the text, so that the efficiency of acquiring class case information can be effectively improved.

Step S33: and taking the first case with the similarity exceeding a preset threshold value as the case to be mediated, and obtaining the case information.

For example, the similarity may be measured by a probability value between [0,1], the preset threshold may be defined as 0.45, 0.5, 0.55, 0.6, 0.65, 0.7, etc., and when the preset threshold is reached, the corresponding first case is considered as a class, and the class information is obtained; or sorting the similarity, and selecting a plurality of cases with the highest similarity as the class.

In a specific embodiment, the class information includes: statistical results of litigation results of various cases are obtained, such as distribution information. For example, for 100 similar cases to be mediated, 10 cases are determined as a result, such as "win, refund and 2-time claim for reimbursement", and 20 cases are determined as B result, such as "win, refund and 1-time claim for reimbursement" … …, and after statistics, a relevant statistical result is obtained, and the statistical result is used for being input to the mediation plan providing model to obtain one or more mediation plans of the cases to be mediated. The mediation plan provides a model to learn the statistical results, and is helpful for accurately predicting the content tendency of the mediation plan of the case to be mediated to be consistent with the actual case.

In specific implementation, when the first case is extracted from the official document library according to the judicial element matching, the matching degree of the corresponding judicial element can be calculated and obtained, and the matching degree can be used for being stored/output to be used as an explanation basis for obtaining the official document of the case type. Optionally, the judicial element matching degree may be expressed as statistical results of case information and case referee documents of the case to be mediated on dimensions of the same or similar judicial elements in number, word frequency, and the like, for example, statistical results of the number of characters or word frequency, such as "food poisoning", "dining in xx food shop", and the like, exist in both the case information of the case to be mediated and the output case referee documents of the category; it should be noted that the above is only some example implementations of the judicial element matching degree, and does not limit other implementations.

Refer to fig. 5, which is a schematic structural diagram of a system for obtaining a class and class information in an embodiment of this specification.

In a specific implementation, a structured referee's text library 54 can be constructed, for example, a judicial elements library 52 can be constructed by manually labeling part of referee's text data in the existing referee's text library 51 and extracting judicial elements in the referee's text. A sample set (which may contain a training set and a test set) is generated from the labeled partial referee document data and corresponding judicial element labels for use in training a judicial element analysis model 53 implemented based on a machine learning model (e.g., a deep neural network). After the judicial element analysis model 53 is trained, the judicial element analysis model 53 can be used to input various other referee documents to be classified into judicial labels by the judicial element analysis model 53, so as to construct a structured referee document library 54.

In the process of obtaining the class, a class matching model 55 constructed in advance may be used to complete the process; when the case and the case information related to the case to be mediated need to be obtained, the case information of the case to be mediated is input into the case matching model 55, so that the first case is extracted by matching judicial factors in the referee document library 54, and then the case is determined by matching the similarity based on text semantics of the first case and the case to be mediated, and further the case information, such as referee document data of the case, statistical results of litigation results of the case, and the like, is obtained.

In a specific implementation, judicial elements can be extracted from case information of a case to be mediated based on Natural Language Processing (NLP) technology for the matching. Natural Language Processing (NLP) techniques include: sentence segmentation, word segmentation, syntactic analysis, entity identification, extraction of entity relationships, pattern identification, information extraction methods and the like.

Alternatively, the similarity matching process performed by the class matching model 55 may be as shown in the flow of fig. 3.

It should be noted that, in practical implementation, the process of constructing the structured official document library 54 in fig. 5 and the process of obtaining the class information may be performed by different implementers, for example, the enterprise a constructs the structured official document library, and the enterprise B uses the official document library to perform the process of obtaining the class information in fig. 3, for example, without limitation, the process is performed by the same implementer, and without limitation, the process must be performed continuously in the same process.

In addition, the schematic architecture of FIG. 5 is merely exemplary and may be varied according to implementation requirements. For example, the acquisition mode of the structured document library can be changed, and the acquisition mode can be completed by utilizing the existing database or manually classifying and indexing, and the like; further alternatively, algorithms such as judicial element matching and similarity calculation of the pattern matching model 55 may be changed, and the above description is not intended to be limiting. Furthermore, the embodiments of fig. 3, 4, 5 are not limited to be applied in the provision method of the mediation plan.

In an alternative example, the mediation reference information may include: the characteristic information of the party. The characteristic information of the party is the characteristic information of the party of the case to be mediated, which is inclined to treat the legal problem. Optionally, the characteristic information of the party includes: and presentation information of one or more combinations of the party's complaint records, executed records, and risk preferences. Wherein the risk preference refers to the ability of the party to tolerate the degree of risk, such as investment risk, legal risk, etc.

Optionally, the characteristic information of the party may be a result of directly encoding (e.g. by feature extraction) the complaint record, the executed record, the risk preference, and the like of the party, or may be a result of encoding statistical information of these information. For example, statistical information obtained by counting the data in the complaint records, such as the number of debts and the number of complaints; the risk preferences include, for example: for low legal risk, information that tends to be paid out and settled as soon as possible so as not to affect the reputation, etc., can be represented by encoded data, such as in vector form.

The character information of the concerned person is input into the mediation plan providing model, so that the mediation plan providing model can learn the portrait data of the concerned person in the aspect of legal dispute when predicting the mediation plan, and the predicted mediation plan is closer to the legal dispute, thereby being beneficial to improving the mediation success rate. For example, if the complaint is preceded by a large number of outstanding complaint records, the content of the predicted mediation plan may be closer to the original complaint than if the complaint had not been filed.

In an alternative example, the mediation reference information may include: appeal result prediction information. As shown in fig. 6, the complaint result prediction information is obtained as follows:

step S61: calculating to obtain legal provision prediction information of a legal provision applicable to the case to be mediated according to the case information;

step S62: and predicting support of the appeal of the party in the case information according to the legal forecast information to obtain the appeal result forecast information.

The appeal result prediction information can be used for representing the support situation of the appeal of the original report of the case to be mediated on the legal terms. For example, if the prose complaint is not legally supported, it can be characterized by the complaint outcome prediction information, e.g., the complaint outcome prediction information is represented as a supported or unsupported two-class 0 or 1 value; or the appeal result prediction information is represented as a probability value and the like in [0,1], and reflects the probability of the legal support original appeal; alternatively, the prediction information of the result may be expressed in a specific encoding format.

In the specific implementation, case information of a case to be mediated can be processed through a legal problem prediction model, and complaint result prediction information is output. Logically, the legal problem prediction model can predict related legal terms according to the case information of the case to be mediated, and the adopted technical means is similarity matching based on text semantics (the above judicial elements can also be provided); and judging whether the appeal is supported according to the predicted legal terms so as to obtain the appeal result prediction information.

In specific implementation, reference may be made to a schematic structural diagram of the legal issue prediction model 70 in the embodiment illustrated in fig. 7. The legal issue prediction model may be a Multi-Task Learning (MTL) model. And taking the acquisition of the legal provision prediction result as a first task and taking the acquisition of the appeal prediction result as a second task. In particular implementations, the legal issue prediction model may be implemented by a neural network.

The legal issue prediction model 70 includes: an input layer 71, a hidden layer 72 and an output layer 73.

In a specific implementation, the input information of the legal issue prediction model 70 includes: legal provision text information, case information, and case discrete characteristic information. Optionally, the legal provision text information may include an original text of a legal provision, and the like; the case information is text information in paper or electronic documents of cases to be mediated, such as text information recorded in front of a court and describing cases of the cases to be mediated, appetitive text information and the like; the case discrete characteristic information can be discrete information of case arrangement (such as extraction or statistics) of cases manually or mechanically, such as 'number of words containing criminals', 'advertised age group', 'stolen amount', 'weight of drugs', 'alcohol concentration in blood', and the like.

In alternative examples, the legal provision textual information, case information, and case discrete feature information may be converted into vector form in the input layer 71, before being input to the input layer 71, or in the hidden layer 72; meaning that the input layer 71 may have a vector conversion unit integrated therein, or the vector conversion unit may be provided outside the legal issue prediction model 70, or the hidden layer 72 may have a vector conversion unit integrated therein, such as the first layer of the hidden layer. In an alternative example, the vector form may include an embedded vector (Embedding) form, by which high-dimensional sparse (or discrete) data can be converted into a low-dimensional vector that preserves semantic relationships. For example, for text processing, the corresponding embedding vector may be a word embedding vector, a sentence embedding vector, a paragraph embedding vector, an article embedding vector, and the like.

Optionally, in consideration of semantic complexity of case information, a text semantic vector representation model such as context may be set before the input layer or the input layer to process the case information to obtain a case information vector, such as an ELMo model pre-trained by a related field text, and the like, where the ELMo model considers context information and word vectors obtained under different context semantics are different, and can solve problems such as ambiguous words.

Alternatively, the case discrete feature information may be converted into a sparse vector by discretization when input into the layer, and then converted into an embedded vector when input into the hidden layer 72.

In the example of fig. 7, the hidden layer includes a plurality of coding layers corresponding to input information processing and output information. The plurality of encoding layers includes: a french information coding layer 721, a case text coding layer 722, a discrete feature coding layer 723, a french prediction coding layer 724, and a complaint result coding layer 725. The law information coding layer 721, the case text coding layer 722 and the discrete feature coding layer 723 may respectively correspond to the vectors of the law information, the case information and the case discrete feature information input by the input layer and further code the vectors to obtain vectors capable of more accurately reflecting the semantic features of the upper and lower contexts.

For example, corresponding to processing legal text information, the french information encoding layer 721 and the case text encoding layer 722 may have an Attention calculation unit, for example, using a Self-Attention (Self-Attention) algorithm to give higher weight to key information in the french information, and further encode a vector of the french information by the Attention calculation unit to obtain a text embedding vector that can reflect more accurate upper and lower semantics.

In response to processing the case discrete feature information, the discrete feature encoding layer 723 may have, for example, a factorization machine, a linear regression model, etc. to convert the case discrete feature information into a continuous vector, such as an embedded vector; the advantage of using the factorization machine is that it also associatively combines each discrete feature in the case discrete feature information when the case discrete feature information is subjected to the vector conversion processing.

The vectors output by the law information coding layer 721, the case text coding layer 722 and the discrete feature coding layer 723 are input to the law prediction coding layer 724 and the complaint result coding layer 725, respectively. Logically, the case information and the case discrete feature information learned in advance and the related law are used by the law prediction encoding layer 724 to predict the related law according to the vector of the input case information and case discrete feature information, encode and generate law prediction information (for example, in a vector form) representing the related law, and output the law prediction information to the output layer; the law enforcement prediction information may be input to the law enforcement result coding layer 725, and the law enforcement prediction information coding layer 725 encodes and generates a vector of law enforcement request result prediction information indicating a support result of a related law on an litigation request in case information according to a vector of the input case information and case discrete feature information and a vector of the law enforcement prediction information by using a relationship between the law enforcement prediction information, the case discrete feature information, and the law enforcement prediction information learned in advance, and outputs the vector to the output layer.

The output layer 73 generates corresponding appeal result prediction information according to the vector and outputs the appeal result prediction information; optionally, the output layer 73 may further generate a law enforcement prediction result according to the law enforcement prediction information, and output the law enforcement prediction result so as to provide predicted information applicable to the legal provision of the case to be mediated.

In a specific implementation, the first task of the law prediction is actually a multi-label classification task of a related law, and a cross entropy function can be adopted as a loss function; the second task of calling for outcome prediction is to ask for legally supported or unsupported dichotomy problems, and cross-entropy functions can also be used as loss functions. For multitasking loss functions, their combined loss may be calculated by weighted sum or the like.

In some examples, when Training the legal problem prediction model 70, the legal problem prediction model 70 may be trained by, for example, an alternative Training mode, that is, a loss alternative optimization that is respectively output to the loss function of the first task and the loss function of the second task, or a Joint Training (Joint Training) mode, that is, a combined loss obtained by combining (e.g., weighting and summing, etc.) the losses of the two first tasks and the second task is optimized.

And S23, inputting the case information and the mediation reference information into a mediation plan providing model to obtain one or more mediation plans of the case to be mediated.

In the specific implementation, the content of each mediation plan is taken as a label, and the content of various mediation results in the collected historical cases is taken as a label. The mediation plan model is a multi-classification label model for obtaining labels of mediation plan contents according to the input information in a classification mode.

Referring to fig. 8, a schematic diagram of the structure of the mediation plan providing model 80 is shown. The mediation plan provisioning model 80 may be implemented based on a deep neural network, including: an input layer 81, a hidden layer 82 and an output layer 83.

The input layer 81 is used for receiving case information of a case to be mediated and mediation reference information. Optionally, the mediation reference information includes: at least one of the category information, the party characteristic information, and the appeal prediction result information.

Although in the embodiment of fig. 8 it is exemplarily shown that the mediation reference information includes appeal result prediction information, class case information, and party characteristic information; it will be appreciated, however, that variations in implementation can be made depending on the accuracy requirements of the actual mediation plan, and not limiting the implementation. For example, various information in the mediation reference information may be changed according to the demand, such as addition or deletion, and the like.

Optionally, the mediation plan providing model 80 includes: at least one vector conversion unit for converting the input information into a form of an embedded vector for transmission to a rear layer. In the case where there are a plurality of input information, there may be a plurality of vector conversion units. For example, in the example of fig. 8, each of the case information, the complaint result prediction information, the class information, and the party characteristic information is provided with a vector conversion unit, that is,

vector conversion units

821, 822, 823, 824.

Optionally, the case information, the complaint result prediction information, the category information, and the party characteristic information may be converted into a reduced-dimension continuous vector, such as an embedded vector, by the

vector conversion units

821, 822, 823, 824.

For example, the case information may be text information, such as text information describing case situations of the case to be mediated from pre-court records. In consideration of semantic complexity of context of case information, in the example of fig. 9, the vector conversion unit 821 includes: the text semantic vector representation model 8211 is used for converting case information into a first text semantic vector which is related to context.

In a specific implementation, the text semantic vector representation model 8211 may include a text pre-training model, i.e., a model with parameters that have been pre-trained using a text data set, such as an ELMo model, a bert (bidirectional Encoder retrieval from transformations) model, and the like, which are trained by massive legal texts. The ELMo model is realized based on a BilM (bidirectional language model), and can learn the complexity of vocabulary usage, such as grammar and semantics; the vocabulary ambiguity under different context conditions can be learned; compared with fixed word vectors, the ELMo considers context information, and different context word vectors are different, so that the problems of ambiguous words and the like can be solved, and the method is more suitable for feature extraction of case information with complex semantics. The BERT model learns feature representations for words by running an auto-supervised learning method on the basis of massive corpora.

Taking the ELMo model as an example, case information in the form of text using a case to be mediated can be converted into an embedding (embedding) vector as the first text semantic vector by inputting the ELMo model after preprocessing (such as word segmentation and the like).

Optionally, in fig. 9, the mediation plan providing model 80 may further include: the attention calculating unit 84 is adapted to calculate attention weights of the words corresponding to the first text semantic vector to obtain a second text semantic vector. Wherein, Attention refers to an Attention (Attention) mechanism in deep learning, and refers to a feature of simulating uneven distribution of Attention weights of a person on elements (such as words in a sentence or blocks in an image) in an object, and distinguishing key information from non-key information (such as represented by weights) in information to be processed.

In a specific implementation, the attention mechanism includes: additive Attention (Additive Attention), Multiplicative Attention (Multiplicative Attention), Self-Attention (Self-Attention), and key-value Attention (key-value Attention). The attention calculation unit 84 may be implemented by an attention model based on one or more attention mechanisms therein.

The text pre-training model is taken as an example of the ELMo model, and a vector output by the ELMo model is input into, for example, a self-attention computing unit for processing, so that the characteristic semantic information of the text can be retained to the maximum extent, and an embedded vector capable of reflecting more accurate upper and lower semantics can be obtained.

Optionally, the text pre-training model and the attention calculation unit may also be integrated. For example, if the text pre-training model is a BERT model, a Multi-Head Attention (Multi-Head Attention) calculation unit is integrated into an Encoder (Encoder) and a Decoder (Decoder).

Alternatively, the input complaint result prediction information, the class information, or the party characteristic information may include a plurality of discrete features, so that the input complaint result prediction information, the class information, or the party characteristic information may be converted into a continuous vector by a regression model as the corresponding

vector conversion units

822, 823, 824, such as a factor decomposition Machine (FM), a linear regression model, or the like. Considering that the appeal result prediction information, the class case information and the characteristic information of the party are possibly high-dimensional sparse discrete characteristics, the connection parameters of the neurons of the neural network are too many, and the factorization machine is more suitable for solving the problem; the processing of the factorization machine is equivalent to the low-dimensional embedding of the discrete features with high-dimensional sparsity, on one hand, the discrete features can be converted into continuous dense vectors, and the calculation of a subsequent neural network layer is facilitated; on the other hand, the factorization machine may automatically perform feature combining on discrete features.

In a specific implementation, each of the

vector converting units

821, 822, 823, 824 may be implemented in the hidden layer 82, and configured to convert input information into a vector (e.g., an embedded vector) for transmission to a subsequent layer.

In a specific implementation, the hidden layer 82 includes a nonlinear transformation layer 825 adapted to transform at least a portion of the features from the input layer 81 to a nonlinear transformation and output the transformed features to the output layer 83. The nonlinear conversion layer 825 is located at a later layer of each of the

vector conversion units

821, 822, 823, 824, and is configured to receive the vector output by each of the

vector conversion units

821, 822, 823, 824.

Optionally, the nonlinear transformation layer 825 may be implemented based on a Highway neural Network (Highway Network) or a residual neural Network (ResNet). The required depth of the deep neural network is increased due to the need of processing complex tasks, the problems of gradient disappearance and the like exist during training, and the high-speed path neural network and the residual neural network appearing later have the capability of transmitting the original input across layers, so that the problem of difficulty in training the deep neural network is solved. The network depth that the residual neural network can realize is greater than the highway neural network, so when the nonlinear transformation layer 825 is realized, the network depth can be selected according to the actual neural network depth requirement.

The forward propagation of the highway network can be simplified as expressed by the following formula:

y＝H(x,W_H)·T(x,W_T)+x·(1-T(x,W_T))； (1)

wherein H, T is a non-linear transformation function; when T is 0, y is x; when T is 1, y is H (x, W)_H) (ii) a From the second half of the formula, it can be known that the original feature x of the highway structure is controlled by the function T without nonlinear changeInstead, the portion that passes directly.

Reference is made to the principle schematic of the residual structure shown in fig. 10. The residual structure (residual block) is obtained by modifying the above highway structure, and is expressed by formula (2):

y＝F(x,W_i)+x； (2)

where F is a residual function, x represents the input, and y represents the desired output; x is weighted by weight layer (weight parameter is W)_i) Partial output F (x, W) after ReLu function_i) And summing the original input x to obtain y.

The Residual structure of ResNet includes two mappings, one is Identity Mapping (Identity Mapping) and the other is Residual Mapping (Residual Mapping), where Identity Mapping refers to itself, i.e. x in the formula, and Residual Mapping refers to "difference", i.e. y-x, i.e. part f (x) in the formula.

Aiming at the phenomenon that the accuracy of a training set is reduced along with the deepening of a network, ResNet provides two selection modes, namely identity mapping and residual mapping, if the network reaches a target state, the network is deepened continuously, the residual mapping approaches to 0, only the identity mapping is left, and therefore the network is in the target state all the time theoretically, and the performance of the network cannot be reduced along with the increase of the depth.

In practical application, the models of common depth residual error networks have 50 layers, 101 layers, 152 layers and the like; however, in the scenario of the multi-class model of the mediation plan in the embodiment of the present specification, if the actual network depth does not need to reach the number of layers of the model of the depth residual network, the highway neural network may be preferred.

Optionally, the attention calculation unit in the previous example may be integrated in the non-linear transformation layer. For example, the non-linear transformation layer is a highway neural network in which the attention calculation unit may be implemented.

Optionally, the mediation plan providing model further outputs mediation success rates corresponding to each mediation plan, i.e. the success rates as exemplarily shown in fig. 1A. For example, the mediation success rate may be obtained according to a prediction value of the mediation plan by the mediation plan providing model. For example, the output layer of the mediation plan providing model includes a Sigmod function, and is used to output the predicted value of the tag corresponding to the content of each mediation plan, and the success rate of the mediation plan can be obtained according to the predicted value.

The predicted value output by the Sigmod function and corresponding to each label is between 0 and 1, and the predicted value can be converted into a corresponding success rate and the like, for example, the first success rate of the regulation plan is 80%, the second success rate of the regulation plan is 50%, and the third success rate of the regulation plan is 20%; or, the success rate is obtained by recalculating with the predicted value as one of the reference factors. Of course, other functions, such as softmax function, etc., are also possible, but not limited thereto.

Referring to fig. 11, an embodiment of the present specification further provides a flowchart of a training method of a mediation plan providing model suitable for mediation plan provision, where the training process may specifically include the following steps:

s111: a sample data set is obtained.

Wherein the sample data set comprises a training data set comprising: case information of each historical case, mediation reference information obtained according to the case information, and a mediation result. In some examples, the sample data set may also include a test data set, and the test data set and the training data set are simply the result of dividing the same data set into two parts for different purposes.

Optionally, the mediation reference information includes class case information; the class information is obtained by the following method: extracting a first case from a preset referee document library; calculating the similarity between the historical case and the first case according to the case information; and taking the first case with the similarity exceeding a preset threshold value as the historical case, and acquiring the information of the case.

Optionally, the referee library is obtained by: and classifying the historical referee documents according to judicial factors to obtain a referee document library. Optionally, the method further includes: and structuring the referee text library.

Optionally, the extracting the first case from the preset referee library includes: and extracting a first case with judicial elements matched with the historical cases from the preset referee text library.

Optionally, the calculating the similarity between the historical case and the first case according to the case information includes: extracting case texts from case information, and carrying out vector coding on the case texts of the historical cases to obtain first text semantic features; extracting a case text of a first case, and carrying out vector coding on the case text of the first case to obtain a second text semantic feature; and calculating the cosine distance between the first text semantic feature and the second text semantic feature to obtain the similarity.

Optionally, the category information includes statistical results of litigation results of the category, and the statistical results are used for being input into the mediation plan providing model to obtain one or more mediation plans of the historical case.

It should be noted that, the specific principle of obtaining the class information may refer to the previous embodiment shown in fig. 3 to 5, and is not repeated herein.

Optionally, the mediation reference information further includes: appeal result prediction information; the appeal result prediction information is obtained through the following method: calculating to obtain applicable legal rules of historical cases according to the case information; and predicting support of the appeal of the party in the case information according to the legal forecast information to obtain the appeal result forecast information.

It should be noted that, the specific principle of obtaining the prediction information of the appeal result can refer to the embodiment of fig. 6 to 7, and the description thereof is not repeated herein.

S112: and inputting the sample data set into a mediation plan providing model so that the mediation plan providing model is trained according to the loss between the mediation plan obtained by the mediation plan providing model and the mediation result.

Optionally, the mediation plan model is a multi-classification label model, and is used for classifying according to the input information to obtain a label about mediation plan content.

Optionally, the mediation plan providing model is implemented based on a neural network, and includes: an input layer, a hidden layer and an output layer; the hidden layer comprises a nonlinear transformation layer and is suitable for outputting at least part of the characteristics from the input layer direction to the output layer direction after nonlinear transformation. Optionally, the nonlinear transformation layer is implemented based on a highway neural network or a residual neural network.

Optionally, the mediation plan providing model includes: at least one vector conversion unit for converting the input information into a form of an embedded vector for transmission to a rear layer.

Optionally, the vector is a continuous dense vector.

Optionally, the mediation reference information includes: case information of historical cases, wherein the case information is in a text form; the vector conversion unit includes: and the text semantic vector representation model is used for converting case information into a first text semantic vector which is related with the context.

Optionally, the text semantic vector representation model includes: a text pre-training model pre-trained with legal text data.

Optionally, the mediation plan providing model further includes: and the attention calculating unit is suitable for calculating attention weights of all words corresponding to the first text semantic vector to obtain a second text semantic vector.

Optionally, the at least one vector conversion unit includes: the regression model is used for converting at least one of the class case information, the party characteristic information and the appeal prediction result information in the mediation reference information into a vector form; wherein the regression model is capable of performing associative combination on discrete features in the processed information.

In practical implementation, the mediation plan providing model may be implemented by a deep neural network, and part or all of various sub-networks in a hidden layer of the mediation plan providing model, such as a nonlinear conversion unit (e.g. a highway neural network or a residual neural network, etc.), various vector conversion units (e.g. a factorization machine, a text semantic vector representation model), and an attention calculation unit, etc., may be trained when training the mediation plan providing model through the sample data set.

Because the prediction mediation plan in the embodiment of the specification can be a multi-classification label task model, a cross entropy loss function can be selected correspondingly as a loss function of the mediation plan providing model, and the cross entropy loss function is used for measuring the difference between the real probability distribution and the prediction value distribution as loss (loss); the real probability distribution is from the input mediation result of the historical case, and the predicted value distribution is from each of the predicted mediation plans output by the mediation plan providing model.

Inputting a regulation and regulation plan providing model through the sample data set, and regulating parameters of a hidden layer of the regulation and regulation plan providing model for optimizing the loss until the output value of the loss function is converged to a preset threshold value, and then determining that the training is finished.

In particular, in order to minimize the loss function efficiently and complete the training as quickly as possible, some training methods that can automatically optimize the learning rate may be used. For example, Batch Gradient Descent (BGD), Stochastic Gradient Descent (SGD), Momentum method (Momentum), Adaptive grad (which is an Adaptive learning rate algorithm), Adaptive Moment Estimation (Adam) method, and the like can be used.

Optionally, the mediation plan providing model further outputs mediation success rates corresponding to the mediation plans. Optionally, the mediation success rate is obtained according to the prediction value of the obtained mediation plan, that is, the prediction value of each mediation plan content tag is classified according to the Sigmod function output by the output layer, for example.

In some embodiments of the present disclosure, a schematic diagram of a providing apparatus 120 for a mediation plan is shown in fig. 12. The specific implementation of the providing device 120 of the mediation plan may refer to the previous providing method of the mediation plan, and will not be described repeatedly here.

The providing means 120 of the mediation plan may comprise:

the data obtaining unit 121 is configured to obtain case information of a case to be mediated;

an information obtaining unit 122, configured to obtain mediation reference information of the case to be mediated according to the case information;

an input unit 123, configured to input the case information and the mediation reference information to a mediation plan providing model 124, so as to obtain one or more mediation plans of the case to be mediated.

extracting a first case from a preset referee document library;

Optionally, the referee library is obtained by:

Optionally, the method further includes:

and structuring the referee text library.

In some embodiments of the present disclosure, a schematic diagram of a training apparatus 130 suitable for a mediation plan providing model for mediation plan providing is shown in fig. 13. The specific implementation of the training apparatus 130 may provide a training method of the model according to the previous mediation plan, and details are not repeated here.

The training apparatus 130 includes:

a data obtaining unit 131, configured to obtain a sample data set; wherein the sample data set comprises a training data set comprising: case information about each historical case, mediation reference information obtained according to the case information, and a mediation result;

the training unit 132 inputs the sample data set to the mediation plan providing model 133, so that the mediation plan providing model is trained according to the loss between the mediation plan obtained by the mediation plan providing model and the mediation result.

extracting a first case from a preset referee document library;

Optionally, the referee library is obtained by:

Optionally, the method further includes:

and structuring the referee text library.

It should be noted that each unit in the embodiments of fig. 12 and 13 may be implemented by software, a combination of software and hardware, or a hardware circuit.

The present specification further provides an electronic device 140, such as the structural schematic diagram of the electronic device 140 shown in fig. 14, where the electronic device 140 may include a memory 141 and a processor 142, where the memory 141 stores computer instructions executable on the processor, and the processor executes the computer instructions to perform the steps of the mediation plan providing method according to any one of the foregoing embodiments or the training method suitable for the mediation plan providing model according to any one of the foregoing embodiments.

In particular implementations, the processor 142 may be a Central Processing Unit (CPU), an Application Specific Integrated Circuit (ASIC), or one or more Integrated circuits configured to implement embodiments of the present invention. The memory may comprise high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as at least one disk memory. The processor 142 and the memory 141 may communicate with each other via a bus or the like, and the chips may communicate with each other via corresponding communication interfaces.

In a specific implementation, the electronic device 140 may further include: a communicator 143; the communicator can comprise one or more of a wired network card, a wireless network card, a 2G/3G/4G/5G module and the like, and can interact information with the outside.

In some examples, the electronic device 140 may be loaded with a trained reconciliation plan model and run the reconciliation plan providing method to obtain one or more reconciliation plans corresponding to the case to be reconciled.

For example, the electronic device 140 may be applied in the scenarios of fig. 1A, 1B and 1C, and provide data of one or more mediation plans to the user terminal 10 of the mediator as a service terminal in communication connection with the user terminal. Wherein the service terminal may be a server/server group or other electronic device 140; in the scenario of fig. 1C, the electronic device 140 may be used to implement the "smart mediation platform". The way for the electronic device 140 to obtain the input information of the case to be mediated may include: read from local memory or other communicatively available storage media, transmitted by a user terminal of a moderator or other party, or entered manually.

For another example, the electronic device 140 may also be implemented as the user terminal 10 on the side of the moderator in fig. 1A, 1B, or 1C, for example, to load the trained moderator plan model in a local memory, obtain one or more moderator plans corresponding to the case to be moderator by locally running the moderator plan providing method, and display a moderator plan interface through a display screen of the user terminal 10.

The present specification further provides a computer readable storage medium, on which computer instructions are stored, which when executed perform the steps of the mediation plan providing method according to any one of the foregoing embodiments or the training method for a mediation plan providing model suitable for mediation plan provision according to any one of the foregoing embodiments.

In particular implementations, the computer-readable storage medium may include, for example, any suitable type of memory unit, memory device, memory article, memory medium, storage device, storage article, storage medium and/or storage unit, for example, memory, removable or non-removable media, erasable or non-erasable media, writeable or re-writeable media, digital or analog media, hard disk, floppy disk, compact disk read Only memory (CD-ROM), compact disk recordable (CD-R), compact disk Rewriteable (CD-RW), optical disk, magnetic media, magneto-optical media, removable memory cards or disks, various types of Digital Versatile Disk (DVD), a tape, a cassette, or the like.

The computer instructions may include any suitable type of code, such as source code, compiled code, interpreted code, executable code, static code, dynamic code, encrypted code, and the like, implemented using any suitable high-level, low-level, object-oriented, visual, compiled and/or interpreted programming language.

Specific implementation manners, operation principles, specific actions and effects of each device, system, equipment or system in the embodiments of the present invention may be referred to in the detailed descriptions of the corresponding method embodiments.

Although the embodiments of the present invention are disclosed above, the present invention is not limited thereto. Various changes and modifications may be effected by one skilled in the art without departing from the spirit and scope of the embodiments of the invention as defined in the appended claims.

Claims

1. A provision method of a mediation plan, comprising:

acquiring case information of a case to be mediated;

2. The provision method of the mediation plan according to claim 1, wherein the mediation reference information includes class plan information; the class information is obtained by the following method:

extracting a first case from a preset referee document library;

3. A provision method of a mediation plan according to claim 2, wherein the official document library is obtained by:

4. A method of providing a mediation plan as claimed in claim 3, wherein the method further comprises:

and structuring the referee text library.

5. The provision method of a mediation plan as set forth in claim 3, wherein the extracting of the first case in the preset official document library includes:

6. The provision method of the mediation plan as claimed in claim 5, wherein the calculating the similarity of the case to be mediated and the first case according to the case information includes:

7. The provision method of a mediation plan according to claim 2, wherein the class information includes a statistical result of litigation results of a class.

8. The provision method of the mediation plan as set forth in claim 1, wherein the mediation reference information further includes party characteristic information, the party characteristic information including: and presentation information of one or more combinations of the party's complaint records, executed records, and risk preferences.

9. The provision method of a mediation plan according to claim 1, wherein the mediation reference information includes: appeal result prediction information; the appeal result prediction information is obtained through the following method:

10. A training method of a mediation plan provision model adapted to mediation plan provision, comprising:

11. The training method of claim 10, wherein the mediation reference information includes class information; the class information is obtained by the following method:

extracting a first case from a preset referee document library;

12. Training method according to claim 11, wherein the official document library is obtained by:

13. The training method of claim 12, wherein the method further comprises:

and structuring the referee text library.

14. The training method of claim 12, wherein said extracting the first case in a predetermined official document library comprises:

15. The training method of claim 14, wherein said calculating a similarity of said historical case to said first case based on said case information comprises:

16. The training method of claim 11, wherein the class information comprises statistical results of litigation outcomes for a class.

17. The training method of claim 10, wherein the mediation reference information further includes party characteristic information, the party characteristic information including: and presentation information of one or more combinations of the party's complaint records, executed records, and risk preferences.

18. The training method of claim 10, wherein the mediation reference information further comprises: appeal result prediction information; the appeal result prediction information is obtained through the following method:

19. A provision apparatus of a mediation plan, comprising:

20. A training apparatus adapted to mediate provision of a plan providing model, comprising:

and the training unit is used for inputting the sample data set into the mediation plan providing model so as to enable the mediation plan providing model to obtain training according to the loss between the mediation plan obtained by the mediation plan providing model and the mediation result.

21. An electronic device comprising a memory and a processor, the memory having stored thereon computer instructions executable on the processor, wherein the processor, when executing the computer instructions, performs the method of providing a mediation plan of any of claims 1 to 9 or the method of training of any of claims 10 to 18.

22. A computer readable storage medium having stored thereon computer instructions, wherein the computer instructions when executed perform a method of providing a mediation plan as defined in any one of claims 1 to 9 or a method of training as defined in any one of claims 10 to 18.