A Method for Intelligent Summary Generation of Mediation Cases Based on Multi-Source Information and Semantic Aggregation

Abstract

This invention relates to the field of semantic aggregation, and more particularly to an intelligent method for generating summaries of mediation cases based on multi-source information and semantic aggregation. The method includes: using a semantic extraction model to determine candidate case keywords from a mediation application; using a legal entity-irrelevant word filtering algorithm to generate high-frequency mediation-related text from the mediation record text; matching the high-frequency mediation-related text and candidate case keywords using a large language model to determine concise case keywords; expanding the concise case keywords with synonyms to determine case semantic terms; and performing multi-path retrieval on multi-source mediation materials based on the concise case keywords and case semantic terms to generate a structured case summary. This invention improves the accuracy and practicality of generating mediation case summaries from multi-source mediation materials.

Description

Technical Field

[0001] This invention relates to the field of semantic aggregation, and in particular to a method for generating intelligent summaries of mediation cases based on multi-source information and semantic aggregation. Background Technology

[0002] With the continuous improvement of the social governance system, dispute mediation, as an important means of efficiently resolving civil disputes and reducing judicial pressure, is increasingly widely applied, covering civil scenarios such as labor disputes, neighborhood disputes, and marriage and family disputes, as well as commercial scenarios such as contract disputes, business activity disputes, commercial transaction disputes, and disputes between commercial entities. The mediation process generates diverse and heterogeneous case materials, mainly including mediation applications, mediation records, statements from the parties, and evidentiary materials. In the actual handling of mediation cases, the core need for mediators when receiving case materials is to quickly grasp the overall picture of the case, laying the foundation for subsequent mediation work.

[0003] However, in practice, the case materials submitted by the parties often have obvious problems of being lengthy and fragmented, containing a large number of emotional expressions and irrelevant details. The evidence materials are mostly spliced ​​together from original documents and lack systematic organization, making it difficult for mediators to quickly filter out the core information and thus make it difficult to quickly grasp the whole picture of the case.

[0004] Currently, the generation of case summaries for such cases still relies primarily on traditional manual methods. Mediators must review all lengthy and fragmented statements and evidence materials one by one, extracting key information, identifying points of contention, and summarizing the core demands of the parties involved, all manually compiling the case summary. This method is not only time-consuming and labor-intensive—actual surveys show that the average manual summary compilation for each mediation case takes up to 30 minutes—but also significantly consumes mediators' time given the surge in cases. More importantly, the quality of the summary depends entirely on the mediator's subjective experience, professional competence, and attention to detail. Different mediators have different standards for judging the core information of a case, resulting in inconsistent quality of summaries for the same case compiled by different mediators. Some summaries suffer from omissions of core information, lack of emphasis, and non-standard expression, failing to accurately reflect the core context of the case. This severely impacts the efficiency of subsequent mediation work and is ill-suited to the actual needs of efficient and standardized grassroots mediation work.

[0005] To address the pain points of manual summary generation, existing technologies have gradually developed case summary generation methods based on natural language processing, attempting to extract case information and generate summaries through automated means.

[0006] However, existing technologies still have many shortcomings. On the one hand, existing methods only target a single type of mediation material, such as processing only mediation applications or analyzing only mediation records, failing to fully integrate complementary information from multiple sources of mediation materials. This results in one-sided summary information that cannot fully cover the core elements of the case's claims, facts, and points of contention, and is out of touch with the actual scenarios in which the summary is used. On the other hand, existing methods lack targeted processing of legal entities and synonyms during the information extraction process, thus affecting the accuracy and practicality of the summary.

[0007] In summary, how to apply artificial intelligence natural language processing models to the semantic intelligent extraction of multi-source mediation materials, thereby improving the accuracy and practicality of semantic aggregation in generating mediation case summaries, is a technical problem that needs to be solved. Summary of the Invention

[0008] To address this issue, the present invention provides a method for generating intelligent summaries of mediation cases based on multi-source information and semantic aggregation. This method overcomes the problem in existing technologies where artificial intelligence natural language processing models cannot be applied to the semantic intelligent extraction of multi-source mediation materials, resulting in poor accuracy and practicality of semantic aggregation-based mediation case summaries.

[0009] To achieve the above objectives, this invention proposes a method for generating intelligent summaries of mediation cases based on multi-source information and semantic aggregation, comprising: The mediation application is processed through a semantic extraction model to determine candidate case keywords. The semantic extraction model includes a pre-trained Sentence-BERT sub-model, a pre-trained Legal-BERT-zh sub-model, an HDBSCAN clustering algorithm, and a lightweight large language model. The mediation record text is filtered through a legal entity-irrelevant word filtering algorithm to generate high-frequency mediation-related text. The high-frequency mediation-related text and candidate case keywords are then semantically matched using a large language model to determine concise case keywords. The simplified case keywords are expanded using a legal thesaurus to determine the case semantic terms. Based on the simplified case keywords and case semantic terms, multi-source mediation materials are retrieved in multiple ways to generate a structured case summary.

[0010] Furthermore, the process of obtaining keywords for candidate cases includes: The mediation application is processed through a pre-trained Sentence-BERT sub-model and a pre-trained Legal-BERT-zh sub-model to generate comprehensive candidate topic word vectors; The candidate topic words corresponding to the comprehensive candidate topic word vectors are classified using the HDBSCAN clustering algorithm, and divided into candidate outlier removal topic words and original candidate topic words. The original candidate keywords are filtered out to select initial candidate case keywords through semantic word frequency statistics; The candidate outlier keywords are used to determine whether they are related to the initial candidate case keywords using a lightweight large language model. If they are not related, the candidate outlier keywords are discarded. If they are related, the candidate outlier keywords are merged with the initial candidate case keywords to determine the candidate case keywords.

[0011] Furthermore, the process of selecting initial candidate case keywords includes: Calculate the mean of the original candidate keyword vectors corresponding to all the original candidate keywords of the mediation application to obtain the semantic vector of the application; Calculate the cosine similarity between the original candidate topic word vectors and the application semantic vectors to obtain a semantic relevance score; The frequency of the original candidate keywords in the mediation application is calculated, and the ratio of the frequency of occurrence to the maximum frequency of occurrence of the original candidate keywords in the mediation application is calculated to obtain the word frequency significance score. The comprehensive score of candidate topic words is obtained by weighted summation of the semantic relevance score and the word frequency significance score; The initial candidate case keywords are selected by ranking the comprehensive scores of the original candidate keywords. The integrated candidate topic word vectors include the original candidate topic word vectors.

[0012] Furthermore, the process of determining whether candidate outlier removal keywords are related to the initial candidate case keywords using a lightweight large language model includes: Based on the candidate outlier removal keywords, initial candidate case keywords, and mediation application, relevant prompt keywords are constructed. The relevant prompts are input into a lightweight large language model to determine whether the candidate outlier removal keywords are related to the initial candidate case keywords.

[0013] Furthermore, the process of identifying high-frequency mediation-related texts includes: The mediation record text is processed through a pre-filtering algorithm to determine the initial high-frequency mediation-related text; The initial high-frequency mediation-related text is processed using an entity extraction algorithm to determine word type weights; Calculate the word frequency inverse text weights of the initial high-frequency mediation-related text; Obtain the number of paragraph occurrences in the initial high-frequency mediation-related text; The product of the number of occurrences of the paragraph, the inverse text weight of word frequency, and the weight of word type is calculated to obtain the initial paragraph word score; Calculate the initial paragraph word score by dividing it by the number of words in the paragraph to obtain the normalized paragraph word score; Based on the ranking of normalized paragraph word scores in the initial high-frequency mediation-related texts, the high-frequency mediation-related texts are selected. The legal entity-irrelevant word filtering algorithm includes a pre-filtering algorithm and an entity extraction algorithm.

[0014] Furthermore, the process of determining concise case keywords includes: Based on the high-frequency mediation-related text and candidate case keywords, guiding prompts are generated; The guiding prompts are input into a large language model for matching to determine concise case keywords.

[0015] Furthermore, the process of generating structured case summaries includes: The simplified case keywords are used to search multi-source mediation materials using a first retrieval algorithm to determine the contextual text of the simplified case. The semantic terms of the cases are retrieved from multi-source mediation materials using a second retrieval algorithm to identify semantically similar case context texts; The simplified case context text and semantically similar case context text are sorted using a cross-encoder to determine the comprehensive retrieval results; The comprehensive search results are grouped and filled into the case summary template to generate a structured case summary; The multi-path retrieval includes a first retrieval algorithm, a second retrieval algorithm, and a cross encoder.

[0016] Furthermore, the process of determining the comprehensive search results includes: Use the simplified case keywords as query input, and use the simplified case context text and semantically similar case context text as candidate text input; The query input and candidate text input are passed through a cross encoder to generate text scores, and the overall retrieval results are determined based on the ranking of the text scores.

[0017] Furthermore, the cross-encoder is a BGE Reranker architecture, which includes a raw encoding layer, a plot matching output layer, a claim matching output layer, and an entity overlap output layer. The process of generating a text score includes: The concatenated sequence of the query input and candidate text is passed through the original encoding layer to generate a semantic encoding vector; The semantic encoding vector is input into the plot matching degree output layer, the appeal matching degree output layer, and the entity overlap degree output layer, respectively, to generate the text score.

[0018] Furthermore, the process of determining the comprehensive search results also includes: A comprehensive loss function is constructed based on the triplet loss function and the binary cross-entropy loss function, and the cross encoder is fine-tuned based on the comprehensive loss function.

[0019] Compared with existing technologies, the beneficial effects of this invention are that it generates comprehensive candidate topic word vectors through the Sentence-BERT and Legal-BERT-zh twin models, which simultaneously cover the general semantics of everyday case facts in mediation applications and the professional legal semantics of statutory claims and legal relationships. It completes semantic-level initial screening through HDBSCAN density clustering and completes secondary verification of outliers through a lightweight large language model. By anchoring the basic initial claims of the case to the mediation application, the accuracy of topic word extraction is improved.

[0020] In particular, this invention constructs a large-scale model matching and calibration mechanism guided by two sources: high-frequency mediation core text and candidate case keywords. It connects the two major steps of front-end precise noise reduction and candidate keyword extraction, and realizes semantic aggregation that achieves bidirectional precise alignment between the static initial claims in the mediation application and the dynamic real disputes in the mediation record. This completely solves the problems of traditional keyword extraction that focus on unilateral application claims, neglect the actual mediation core, and are disconnected from mediation practice. It reduces the invalid information recall rate in the retrieval process, realizes semantic intelligent extraction of multi-source mediation materials, and thus improves the accuracy and practicality of semantically aggregated mediation case summary generation. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of determining candidate case keywords in the intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation, according to an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the process of determining and simplifying case keywords in the intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation, according to an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the process of generating structured case summaries using the intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation, as described in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0025] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] like Figures 1 to 4 As shown, this invention provides a method for generating intelligent summaries of mediation cases based on multi-source information and semantic aggregation, which overcomes the problem in the prior art that artificial intelligence natural language processing models cannot be applied to the semantic intelligent extraction of multi-source mediation materials, resulting in poor accuracy and practicality of the semantic aggregation-based mediation case summaries.

[0027] like Figure 1 As shown, this embodiment proposes a method for generating intelligent summaries of mediation cases based on multi-source information and semantic aggregation, including: The mediation application is processed through a semantic extraction model to determine candidate case keywords. The semantic extraction model includes a pre-trained Sentence-BERT sub-model, a pre-trained Legal-BERT-zh sub-model, an HDBSCAN clustering algorithm, and a lightweight large language model. The mediation record text is filtered through a legal entity-irrelevant word filtering algorithm to generate high-frequency mediation-related text. The high-frequency mediation-related text and candidate case keywords are then semantically matched using a large language model to determine concise case keywords. The simplified case keywords are expanded using a legal thesaurus to determine the case semantic terms. Based on the simplified case keywords and case semantic terms, multi-source mediation materials are retrieved in multiple ways to generate a structured case summary.

[0028] It should be noted that the semantic extraction model and large language model described in this embodiment are both local offline versions. They are not connected to the internet, are not uploaded to the public internet, and do not transmit case data to any external servers, cloud platforms, or model service providers. They do not need to call cloud interfaces or connect to external computing power. Therefore, in this process, the judicial confidentiality sensitive materials of the mediation application, mediation record text, and multi-source mediation materials are not uploaded to the network or transmitted to the outside world. Under the premise of complying with the confidentiality norms of judicial mediation, the accurate generation of the mediation case summary is achieved.

[0029] Specifically, the case intelligent summary generation method described in this embodiment is preferably applied to mediation cases in commercial scenarios such as contract disputes, business activity disputes, commercial transaction disputes, and disputes between commercial entities.

[0030] Specifically, the mediation record text is a collection of multi-source materials automatically collected by the system, including the mediation application submitted by the parties, historical chat logs, evidence lists, and voice-to-text transcripts. These multi-source mediation materials include historical chat logs, evidence lists, and voice-to-text transcripts submitted by the parties.

[0031] It is understandable that the mediation application is a unilateral, pre-emptive, static document submitted by the parties to initiate mediation proceedings, clearly recording the applicant's claims and factual assertions. The mediation record, also known as the mediation transcript, is a multi-party, dynamic document produced by the mediation organization during the mediation process, clearly recording the statements of all parties, the points of contention, and the negotiation process throughout the entire mediation process. Therefore, the mediation application has a high information concentration and low levels of invalid or interfering information, while the mediation record has uneven information density, with core points of contention scattered and a high proportion of invalid or interfering information. Thus, the standardized, low-noise, and clearly defined claims of the mediation application are well-suited to the needs of candidate keyword extraction, avoiding the inefficiency and bias of extracting from messy texts, and quickly anchoring the fundamentals of the case. The full-process, authentic dispute, and multi-party consensus characteristics of the mediation record are well-suited to the needs of keyword calibration and simplification, avoiding a disconnect between unilateral claims and the actual points of mediation, ensuring that the simplified case keywords truly match the points of conflict in the mediation case, and providing a precise anchor for subsequent structured summary generation.

[0032] like Figure 2 As shown, the process of obtaining candidate case keywords further includes: The mediation application is processed through a pre-trained Sentence-BERT sub-model and a pre-trained Legal-BERT-zh sub-model to generate comprehensive candidate topic word vectors; The candidate topic words corresponding to the comprehensive candidate topic word vectors are classified using the HDBSCAN clustering algorithm, and divided into candidate outlier removal topic words and original candidate topic words. The original candidate keywords are filtered out to select initial candidate case keywords through semantic word frequency statistics; The candidate outlier keywords are used to determine whether they are related to the initial candidate case keywords using a lightweight large language model. If they are not related, the candidate outlier keywords are discarded. If they are related, the candidate outlier keywords are merged with the initial candidate case keywords to determine the candidate case keywords.

[0033] Specifically, the process of generating comprehensive candidate topic word vectors is as follows: the mediation application is cleaned by removing common format noise, including headers, footers, page numbers, tables, and special symbols, and unifying full-width characters and converting traditional Chinese to simplified Chinese. After the cleaned mediation application is segmented using the Chinese word segmentation tool jieba, the segmented words are converted into elements of the comprehensive candidate topic word vectors through the pre-trained Sentence-BERT sub-model and the pre-trained Legal-BERT-zh sub-model, respectively.

[0034] Understandably, the pre-trained Sentence-BERT sub-model is specifically optimized for sentence-level semantic embedding. Through Siamese network training, it achieves accurate semantic similarity representation of dense vectors, precisely capturing natural language content such as case facts, event timelines, and everyday demands stated by the parties in mediation applications. Even logically fragmented factual statements written by the parties can reliably extract their core semantic themes, avoiding the omission of non-legal factual content by single legal models. The pre-trained Legal-BERT-zh sub-model is pre-trained on a massive Chinese legal corpus, exhibiting strong adaptability to the semantic representation of Chinese legal terminology, legal relationships, statutory demands, and legal concepts. It can accurately identify professional legal expressions in mediation applications, avoiding semantic bias in legal terminology and preventing single models from confusing "non-compete agreements" with "confidentiality obligations," and "statutory inheritance" with "testamentary succession." Therefore, by using the outputs of both the pre-trained Sentence-BERT sub-model and the pre-trained Legal-BERT-zh sub-model as elements of the comprehensive candidate topic word vector, it is possible to ensure that the extracted candidate topic words are consistent with the parties' real demands and the actual circumstances of the case, while also complying with legal norms.

[0035] Specifically, the process of removing outliers using the HDBSCAN clustering algorithm is as follows: The elements of the comprehensive candidate topic word vectors are normalized to eliminate the interference of vector length differences on the distance calculation of the clustering algorithm; cosine distance is selected as the distance metric for the normalized preprocessed vectors, and the cosine distance between the vector elements corresponding to each pair of comprehensive candidate topic words is calculated one by one to form a complete semantic distance matrix, clearly depicting the relative positional relationship of all candidate topic words in the semantic space; for each candidate topic word element, the nearest clustering core parameter neighbor vectors are found, and the distance of the farthest one is defined as the core distance of that vector; if a candidate topic word element can cover at least the clustering core parameter neighbors within its own core distance range, it is marked as a core object, corresponding to candidate topic words with a clear semantic association group; otherwise, it is temporarily marked as a non-core object; and so on. Using core objects as nodes, the reachability distances between nodes are defined. A minimum spanning tree algorithm is used to connect all core objects with the shortest reachability distance, forming a semantic connection tree to initially distinguish different semantic clustering regions. The longest edge is gradually removed from the semantic connection tree, splitting the tree into multiple subtrees, each corresponding to a candidate semantic cluster. The size of each candidate semantic cluster is verified: if the number of core objects contained in the cluster is greater than or equal to the preset minimum cluster number, it is marked as a valid semantic cluster. All core objects in the cluster and non-core objects directly connected to them are classified as original candidate keywords. Original candidate keywords are semantically focused core keywords with group associations. If the number of core objects in the cluster is less than the preset minimum cluster number, or if there are non-core objects that are not included in any valid semantic clusters, they are classified as candidate outlier removal keywords, which are semantically isolated and potential noise words without clear associations.

[0036] The core clustering parameter is set to 3 because after a mediation application is extracted by a dual-pre-trained model, the number of candidate keywords is usually between 10 and 30. A threshold of 3 avoids misclassifying scattered noise words as core objects due to an excessively low threshold, and also avoids excluding valuable small semantic clusters from core objects due to an excessively high threshold. The preset minimum cluster size is 4 because there may be a small number of semantically unrelated words in the candidate keyword vectors that happen to be similar in vector, such as non-core time and place words mentioned by the parties. These accidentally clustered clusters are usually very small, mostly consisting of 2 to 3 words. A threshold of 4 can directly filter out these invalid small clusters, reducing the interference of keywords unrelated to the application's theme from the source.

[0037] In one implementation of this embodiment, a typical commercial mediation case concerning a business dispute, the facts and reasons stated in the mediation application are as follows: The respondent, Li Si, borrowed RMB 150,000 from the applicant, citing financial difficulties in his fruit shop. Both parties signed a written promissory note on the spot, stipulating a monthly interest rate of 1% (1% per month), a loan term of 6 months, and a lump-sum repayment of principal and interest upon maturity. The respondent, Wang Wu, signed and affixed his fingerprint to the promissory note as a joint guarantor, promising to assume full repayment responsibility if Li Si was unable to repay the loan on time. On the same day, the applicant transferred the full amount of RMB 150,000 to Li Si's bank account via bank transfer, thus fully fulfilling his lending obligations. After the loan matured, Li Si only repaid RMB 30,000 of the principal to the applicant, leaving the remaining RMB 120,000 of principal and corresponding RMB 14,400 of unpaid interest unpaid to date. After the loan became overdue, the applicant repeatedly demanded that Li Si fulfill his repayment obligations by making phone calls and visiting the fruit shop in person. Li Si made excuses and delayed repayments, citing business losses and financial difficulties. The applicant also demanded that Wang Wu assume guarantee responsibility, but Wang Wu refused to fulfill his repayment obligations for various reasons.

[0038] Using the aforementioned dual-pre-trained model and HDBSCAN clustering algorithm, the identified candidate outlier removal keywords are bank account, bank transfer, and business loss. It can be seen that the dual-pre-trained model and the HDBSCAN clustering algorithm with a core clustering parameter of 3 and a minimum cluster number of 4 will remove details that do not fit the core of the case and semantically isolated non-key content (such as non-important excuses for business loss) as candidate outlier removal keywords.

[0039] Furthermore, the process of selecting initial candidate case keywords includes: Calculate the mean of the original candidate keyword vectors corresponding to all the original candidate keywords of the mediation application to obtain the semantic vector of the application; Calculate the cosine similarity between the original candidate topic word vectors and the application semantic vectors to obtain a semantic relevance score; The frequency of the original candidate keywords in the mediation application is calculated, and the ratio of the frequency of occurrence to the maximum frequency of occurrence of the original candidate keywords in the mediation application is calculated to obtain the word frequency significance score. The comprehensive score of candidate topic words is obtained by weighted summation of the semantic relevance score and the word frequency significance score; The initial candidate case keywords are selected by ranking the comprehensive scores of the original candidate keywords. The integrated candidate topic word vectors include the original candidate topic word vectors.

[0040] Specifically, the process of obtaining the semantic relevance score is as follows: First, the first original candidate topic word vector belonging to the original candidate topic word from the comprehensive candidate topic word vector output by the pre-trained Sentence-BERT sub-model. The mean of all first original candidate topic word vectors in the mediation application is used as the first application semantic vector. The cosine similarity between the original candidate topic word and the first application semantic vector is calculated to obtain the first semantic relevance score. Second, the second original candidate topic word vector belonging to the original candidate topic word from the comprehensive candidate topic word vector output by the pre-trained Legal-BERT-zh sub-model. The mean of all first original candidate topic word vectors in the mediation application is used as the second application semantic vector. The cosine similarity between the original candidate topic word and the second application semantic vector is calculated to obtain the second semantic relevance score. Min-Max normalization is performed on the first and second semantic relevance scores respectively to eliminate the differences in the sub-model outputs, resulting in a first normalized semantic relevance score and a second normalized semantic relevance score. The mean of the first and second normalized semantic relevance scores is used as the semantic relevance score of the original candidate topic word.

[0041] Specifically, the core disputes in a petition are often expressed through a few keywords, but semantically related words may not appear frequently. Therefore, the semantic relevance score and the word frequency significance score are weighted and summed with weighting coefficients of 0.7 and 0.3, respectively, to obtain the comprehensive score of candidate topic words. It is understandable that the weighting coefficient of 0.7 indicates that measuring the degree of matching between candidate words and the entire petition in the semantic space relies on the deep semantics captured by the pre-trained models (Sentence-BERT, Legal-BERT-zh), which can identify those words that, even if they appear infrequently, are core concepts. This aligns with the characteristic that semantic relevance is more important than simple word frequency in mediation petitions. The word frequency significance score reflects the frequency of candidate words in the petition; high-frequency words are often what the parties repeatedly emphasize, directly reflecting their focus. The weighting coefficient of 0.3 can help the large language model further uncover implicit themes.

[0042] Therefore, the generated initial candidate case keywords eliminate low-frequency and semantically redundant expressions, retaining high-frequency keywords that anchor the core of the case. For example, in the above implementation process, the initial candidate case keywords are: Zhang San, Li Si, Wang Wu, lender, borrower, guarantor, private lending, loan, IOU, 150,000 yuan loan, monthly interest rate of 1%, loan term of 6 months, bank transfer, 30,000 yuan principal, 120,000 yuan remaining principal, 14,400 yuan overdue interest, loan overdue, multiple collections, joint and several liability, guarantor's signature, mediation application, repayment of loan principal and interest, assuming joint and several liability. It can be seen that by weighting the semantic relevance score and the word frequency significance score with weighting coefficients of 0.7 and 0.3 respectively, the comprehensive score of the candidate keywords is obtained, which ensures that deep semantics are extracted while considering the importance of high-frequency expressions.

[0043] Furthermore, the process of determining whether candidate outlier removal keywords are related to the initial candidate case keywords using a lightweight large language model includes: Based on the candidate outlier removal keywords, initial candidate case keywords, and mediation application, relevant prompt keywords are constructed. The relevant prompts are input into a lightweight large language model to determine whether the candidate outlier removal keywords are related to the initial candidate case keywords.

[0044] Specifically, the lightweight large language model is the Qwen-7B model, more specifically the Qwen-7B-Chat-Int4 quantized version, to enable the offline deployment of the lightweight large language model on the GPU.

[0045] Specifically, the relevance prompt is as follows: "You are an experienced mediator analyzing a mediation application. The full text of the application is as follows: Mediation Application, Initial Candidate Case Keywords List, There is a word / phrase marked as an outlier in the initial clustering: Candidate Outlier Removal Keywords to be Judged. Based on the content of the application and the existing candidate keywords, please determine whether this outlier word is relevant to the core disputed issue of this case. Judgment criteria: If the word appears in the application and its meaning is similar to any candidate keyword, or can supplement important details of the case, it is considered relevant. Please output: Yes or No."

[0046] Therefore, by automatically filling the candidate outlier removal keywords, the initial candidate case keywords, and the mediation application into the corresponding positions of the prompt word template, the relevant prompt words can be obtained. By automatically inputting the relevant prompt words into a lightweight large language model, the judgment result of whether the sampled candidate outlier removal keywords are related to the initial candidate case keywords can be obtained.

[0047] Specifically, experiments were conducted to demonstrate that the semantic extraction model, composed of the pre-trained Sentence-BERT sub-model, the pre-trained Legal-BERT-zh sub-model, the HDBSCAN clustering algorithm, and the lightweight large language model, achieved superior accuracy in extracting subject terms from candidate mediation cases. All experimental procedures used the officially released pre-trained weights without any fine-tuning specific to this task; only the task processing flow was modified to eliminate performance interference from fine-tuning. The experiments used identical hardware, inference frameworks, and sampling parameters. The experimental dataset consisted of real, anonymized grassroots people's mediation case files from a local judicial bureau in a prefecture-level city in China, collected in a specific year.

[0048] To ensure the authority and professionalism of the evaluation, the gold standard keywords for each sample were jointly labeled by three full-time mediators with more than five years of grassroots mediation experience and two civil and commercial law professionals. The labeling rules are that the keywords of candidate cases must simultaneously cover the core dimensions of the core legal relationship of the case, the core claims of the parties, the key facts of the case, and the core points of contention. The labeling results adopt a majority voting system, and only keywords that are agreed upon by at least four labelers can be included in the gold standard keywords.

[0049] After conducting single-stage elimination ablation experiments, the F1-Score of the complete experimental group using the pre-trained Sentence-BERT sub-model, the pre-trained Legal-BERT-zh sub-model, the HDBSCAN clustering algorithm, and the lightweight large language model was 90.64%. The F1-Score of the complete experimental group using only Sentence-BERT after eliminating Legal-BERT-zh was 84.37%. The F1-Score of the complete experimental group after eliminating the HDBSCAN clustering stage was 81.25%. The F1-Score of the complete experimental group after eliminating the semantic word frequency screening stage was 86.59%. The F1-Score of the complete experimental group after eliminating the Qwen-7B omission filling stage was 87.12%.

[0050] It is evident that each step in the semantic extraction model process contributes positively to the final accuracy, improving the accuracy of extracting keywords from candidate mediation cases. This enables the determination of preliminary keywords for cases by integrating the outputs of multiple models, even with limited computing power on the offline local end.

[0051] like Figure 3 As shown, the process of identifying high-frequency mediation-related text further includes: The mediation record text is processed through a pre-filtering algorithm to determine the initial high-frequency mediation-related text; The initial high-frequency mediation-related text is processed using an entity extraction algorithm to determine word type weights; Calculate the word frequency inverse text weights of the initial high-frequency mediation-related text; Obtain the number of paragraph occurrences in the initial high-frequency mediation-related text; The product of the number of occurrences of the paragraph, the inverse text weight of word frequency, and the weight of word type is calculated to obtain the initial paragraph word score; Calculate the initial paragraph word score by dividing it by the number of words in the paragraph to obtain the normalized paragraph word score; Based on the ranking of normalized paragraph word scores in the initial high-frequency mediation-related texts, the high-frequency mediation-related texts are selected. The legal entity-irrelevant word filtering algorithm includes a pre-filtering algorithm and an entity extraction algorithm.

[0052] Specifically, the pre-filtering algorithm removes meaningless function words and modal particles, removes common high-frequency words in non-legal fields such as "today," "okay," and "thank you," removes low-value modifiers, and removes obvious typos or garbled characters.

[0053] Specifically, the entity extraction algorithm is a general NER (Named Entity Recognition) algorithm, which identifies the following entity types and their corresponding word type weights: word type weight 2.0 for the party's name, word type weight 1.8 for monetary amounts, word type weight 1.5 for time nodes, word type weight 1.5 for evidence names, and word type weight 1.0 for common keywords. Therefore, the word type weights ensure that dialogues containing key party statements and mentioning specific amounts are prioritized, avoiding the omission of important information due to low word frequency.

[0054] Therefore, the larger the value of the inverse text weight of word frequency, the more distinctive the word is, which can effectively reduce the score of generic words that frequently appear in legal texts such as questions, situations, and disputes but have no specific target.

[0055] Specifically, based on the ascending order of the normalized paragraph word scores in the initial high-frequency mediation-related texts, the initial high-frequency mediation-related texts that are less than 80% of the maximum input length of the large language model are selected as high-frequency mediation-related texts. That is, within the maximum length allowed by the large language model, the initial high-frequency mediation-related texts with the highest normalized paragraph word scores are input into it.

[0056] Therefore, the initial paragraph word score is the product of the count, the inverse text weight of the word frequency, and the word type weight. It takes into account the word occurrence frequency, global discriminative power, and domain specificity, which can balance the statistical significance and semantic importance of words. This allows the text paragraphs input into the large language model to retain key information, including the focus of the controversy, the expression of demands, and the statement of facts, to the maximum extent possible without exceeding the length limit.

[0057] Furthermore, the process of determining concise case keywords includes: Based on the high-frequency mediation-related text and candidate case keywords, guiding prompts are generated; The guiding prompts are input into a large language model to determine concise case keywords.

[0058] Therefore, the generated high-frequency mediation-related texts and simplified case keywords retain only the core rights and obligations asserted by both parties, frequently occurring core elements of the case, key texts anchoring the focus of the dispute, and core agreements of the mediation plan. For example, in the above implementation process, the high-frequency mediation-related text is: Private lending dispute, both parties agree on the loan fact, the IOU is authentic, the loan of 150,000 yuan has been fully delivered; 30,000 yuan of principal has been repaid, the remaining principal is 120,000 yuan, and interest of 14,400 yuan is owed, with an agreed monthly interest rate of 1% (1% per month), and both parties agree on this. Candidate case keywords include: private lending dispute, lender, borrower, joint guarantor, no dispute on the loan fact, authentic IOU, 150,000 yuan loan delivered, monthly interest rate of 1%, etc.

[0059] As can be seen, the initial paragraph word score is the product of the count, the inverse text weight of the word frequency, and the word type weight. This achieves a good balance between the statistical significance and semantic importance of words by simultaneously considering the frequency of word occurrence, global distinguishability, and domain specificity.

[0060] Specifically, the large language model is the offline-deployed Qwen3-8B. Qwen3-8B has dual-mode reasoning with both thinking and dialogue modes, and good long context processing capabilities, making it suitable for summarizing long mediation records. Therefore, it achieves a solution for processing non-public long mediation record texts, using a model with superior instruction following and context understanding capabilities under the limited performance and resource consumption of offline deployment.

[0061] Specifically, the guiding prompt is: "You are an experienced mediator analyzing a mediation application. The application is as follows: High-frequency mediation-related text. Candidate keywords: Candidate case keywords. Based on the application content and candidate keywords, please determine whether the candidate keywords are related to the core disputed issues of this case. Judgment criteria: If the keyword appears in the application and its meaning is similar, or it can supplement important details of the case, it is considered relevant. Please output yes or no." Therefore, the prompt simultaneously includes high-frequency mediation-related text (reflecting the actual disputed points in the dialogue) and candidate case keywords (derived from the initial refinement of the application), enabling the model to integrate information from both "party dialogue" and "written statements," avoiding bias from a single source. This allows the model to select candidate case keywords deemed relevant by the large language model as concise case keywords, and then perform retrieval based on these concise case keywords to generate a case summary, achieving conciseness, succinctness, and accuracy in the case summary.

[0062] like Figure 4 As shown, the process of determining the semantic terms of a case further includes: The simplified case keywords were expanded using a legal thesaurus to determine the semantic terms for the cases.

[0063] Specifically, the legal thesaurus is the Tsinghua University Open Chinese Thesaurus (THUOCL), which can expand authoritative legal thesaurus and avoid omissions caused by legal thesaurus not mentioned in the mediation application and mediation record text appearing in multi-source mediation materials.

[0064] like Figure 4 As shown, the process of generating a structured case summary further includes: The simplified case keywords are used to search multi-source mediation materials using a first retrieval algorithm to determine the contextual text of the simplified case. The semantic terms of the cases are retrieved from multi-source mediation materials using a second retrieval algorithm to identify semantically similar case context texts; The simplified case keywords, simplified case context text, and semantically similar case context text are sorted using a cross-encoder to determine the comprehensive search results; The comprehensive search results are grouped and filled into the case summary template to generate a structured case summary; The multi-path retrieval includes a first retrieval algorithm, a second retrieval algorithm, and a cross encoder.

[0065] Specifically, the first retrieval algorithm is Elasticsearch BM25, which precisely matches entity names and legal terms, and the second retrieval algorithm is Faiss, which can recall semantically similar but literally different expressions.

[0066] Furthermore, the process of determining the comprehensive search results includes: Use the simplified case keywords as query input, and use the simplified case context text and semantically similar case context text as candidate text input; The query input and candidate text input are passed through a cross encoder to generate text scores, and the overall retrieval results are determined based on the ranking of the text scores.

[0067] Furthermore, the cross-encoder is a BGE Reranker architecture, which includes a raw encoding layer, a plot matching output layer, a claim matching output layer, and an entity overlap output layer. The process of generating a text score includes: The concatenated sequence of the query input and candidate text is passed through the original encoding layer to generate a semantic encoding vector; The semantic encoding vector is input into the plot matching degree output layer, the appeal matching degree output layer, and the entity overlap degree output layer, respectively, to generate the text score.

[0068] Specifically, the cross encoder is a BGE Reranker architecture that performs deep semantic interaction calculations on the input query input and candidate text pairs to output a relevance text score between the query and candidate texts.

[0069] More specifically, the original encoding layer is the core encoding layer of the BGE Reranker architecture, with no structural modifications. It includes an embedding layer, multiple Transformers containing multi-head attention and feedforward networks, layer normalization, and residual connections. Multi-head attention allows the model to capture information simultaneously from different perspectives, providing a foundation for improvements to the output layer. Layer normalization and residual connections enable semantic information to flow smoothly within the deep network, ultimately resulting in a semantic encoding vector containing rich interactive information.

[0070] Specifically, the plot matching degree output layer, the demand matching degree output layer, and the entity overlap degree output layer are all simple fully connected neural network layers and sigmoid function structures. The three output layers are used to achieve three-dimensional output based on factual plot matching degree, demand matching degree, and entity overlap degree, so as to sort and select the corresponding comprehensive search results based on the text scores of the three dimensions, and to achieve the classification and retrieval of three-dimensional information of case summary.

[0071] Furthermore, the process of determining the comprehensive search results also includes: A comprehensive loss function is constructed based on the triplet loss function and the binary cross-entropy loss function, and the cross encoder is fine-tuned based on the comprehensive loss function.

[0072] Specifically, the binary cross-entropy loss function measures the difference between the model's predicted relevance score and the true label, and is suitable for determining whether a query and document pair are relevant. The triplet loss function is used to bring similar samples closer together and push dissimilar samples away. The combined loss function is constructed by weighting and summing the triplet loss function and the binary cross-entropy loss function, and its weight coefficients are fine-tuned on the validation set.

[0073] Specifically, to verify the recall completeness, matching accuracy, and ranking rationality of multi-path retrieval using the first retrieval algorithm, the second retrieval algorithm, and the cross-encoder in retrieving multi-source materials for mediation cases, an experimental dataset of 300 anonymous commercial mediation cases was constructed. Three benchmark labels—core facts, core claims, and key entities—were manually labeled. The experimental group used multi-path retrieval, control group 1 used only the first retrieval algorithm (Elasticsearch BM25), and control group 2 used only the second retrieval algorithm (Faiss). For the F1@10 experimental metric: the experimental group achieved 91.1%, control group 1 achieved 82.0%, and control group 2 achieved 78.4%. This demonstrates that multi-path retrieval balances precise matching and semantic similarity, resulting in more complete recall, more accurate matching, and more reasonable ranking.

[0074] Specifically, for example, in the above implementation process, the final structured summary includes: Loan Facts: On March 15, 2024, borrower Li Si borrowed RMB 150,000 from lender Zhang San for working capital needs in operating a fruit shop. Both parties signed a legally valid promissory note, agreeing to a monthly interest rate of 1% (1% monthly interest), a loan term of 6 months, and a maturity date of September 14, 2024. Guarantor Wang Wu signed the promissory note, voluntarily assuming joint and several liability. Payment of Funds: On the same day, lender Zhang San transferred the full amount of RMB 150,000 to borrower Li Si's bank account via bank transfer, fully fulfilling his lending obligations. Both parties have no objection to the fact of loan payment.

[0075] It is evident that in multi-way retrieval with different keywords, entity overlap ensures that only relevant texts of entities such as Zhang San, Li Si, Wang Wu, lender, borrower, guarantor, private lending, loan, and IOU are extracted; factual and plot matching ensures the consistency of mediation materials and related texts; and demand matching ensures the complete extraction of core demand information.

[0076] In this embodiment, a comprehensive candidate topic word vector is generated using a dual model of Sentence-BERT and Legal-BERT-zh. This vector covers both the general semantics of everyday case facts in the mediation application and the professional legal semantics of legal claims and legal relationships. HDBSCAN density clustering is used for initial semantic screening, and a lightweight large language model is used for secondary verification of outliers. By anchoring the initial claims in the mediation application to the fundamentals of the case, the accuracy of topic word extraction is improved. The constructed high-frequency core mediation text and candidate case topic words, guided by a dual-source large model matching and calibration mechanism, connect the two major stages of front-end precise noise reduction and candidate topic word extraction. This achieves bidirectional and precise semantic aggregation of the static initial claims in the mediation application and the dynamic real disputes in the mediation record. It completely solves the problems of traditional topic word extraction that emphasize unilateral claims, neglect the actual core of mediation, and are disconnected from mediation practice. This reduces the invalid information recall rate in the retrieval process, enables intelligent semantic extraction of multi-source mediation materials, and thus improves the accuracy and practicality of semantically aggregated mediation case summaries.

[0077] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0078] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating intelligent summaries of mediation cases based on multi-source information and semantic aggregation, characterized in that, include: The mediation application is processed through a semantic extraction model to determine candidate case keywords. The semantic extraction model includes a pre-trained Sentence-BERT sub-model, a pre-trained Legal-BERT-zh sub-model, an HDBSCAN clustering algorithm, and a lightweight large language model. The mediation record text is filtered through a legal entity-irrelevant word filtering algorithm to generate high-frequency mediation-related text. The high-frequency mediation-related text and candidate case keywords are then semantically matched using a large language model to determine concise case keywords. The simplified case keywords are expanded using a legal thesaurus to determine the case semantic terms. Based on the simplified case keywords and case semantic terms, multi-source mediation materials are retrieved in multiple ways to generate a structured case summary.

2. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to claim 1, characterized in that, The process of obtaining keywords for candidate cases includes: The mediation application is processed through a pre-trained Sentence-BERT sub-model and a pre-trained Legal-BERT-zh sub-model to generate comprehensive candidate topic word vectors; The candidate topic words corresponding to the comprehensive candidate topic word vectors are classified using the HDBSCAN clustering algorithm, and divided into candidate outlier removal topic words and original candidate topic words. The original candidate keywords are filtered out to select initial candidate case keywords through semantic word frequency statistics; The candidate outlier keywords are used to determine whether they are related to the initial candidate case keywords using a lightweight large language model. If they are not related, the candidate outlier keywords are discarded. If they are related, the candidate outlier keywords are merged with the initial candidate case keywords to determine the candidate case keywords.

3. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to claim 2, characterized in that, The process of selecting initial candidate case keywords includes: Calculate the mean of the original candidate keyword vectors corresponding to all the original candidate keyword vectors of the mediation application to obtain the semantic vector of the application; Calculate the cosine similarity between the original candidate topic word vectors and the application semantic vectors to obtain a semantic relevance score; The frequency of the original candidate keywords in the mediation application is calculated, and the ratio of the frequency of occurrence to the maximum frequency of occurrence of the original candidate keywords in the mediation application is calculated to obtain the word frequency significance score. The comprehensive score of candidate topic words is obtained by weighted summation of the semantic relevance score and the word frequency significance score; The initial candidate case keywords are selected by ranking the comprehensive scores of the original candidate keywords. The integrated candidate topic word vectors include the original candidate topic word vectors.

4. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to claim 2, characterized in that, The process of determining whether candidate outlier removal keywords are related to the initial candidate case keywords using a lightweight large language model includes: Based on the candidate outlier removal keywords, initial candidate case keywords, and mediation application, relevant prompt keywords are constructed. The relevant prompts are input into a lightweight large language model to determine whether the candidate outlier removal keywords are related to the initial candidate case keywords.

5. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to claim 1, characterized in that, The process of identifying high-frequency mediation-related texts includes: The mediation record text is processed through a pre-filtering algorithm to determine the initial high-frequency mediation-related text; The initial high-frequency mediation-related text is processed using an entity extraction algorithm to determine word type weights; Calculate the word frequency inverse text weights of the initial high-frequency mediation-related text; Obtain the number of paragraph occurrences in the initial high-frequency mediation-related text; The product of the number of occurrences of the paragraph, the inverse text weight of word frequency, and the weight of word type is calculated to obtain the initial paragraph word score; Calculate the initial paragraph word score and divide it by the number of words in the paragraph to obtain the normalized paragraph word score; Based on the ranking of normalized paragraph word scores in the initial high-frequency mediation-related texts, the high-frequency mediation-related texts are selected. The legal entity-irrelevant word filtering algorithm includes a pre-filtering algorithm and an entity extraction algorithm.

6. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to any one of claims 1 to 5, characterized in that, The process of determining concise case keywords includes: Guiding prompts are generated based on the high-frequency mediation-related texts and candidate case keywords; The guiding prompts are input into a large language model for matching to determine concise case keywords.

7. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to any one of claims 1 to 5, characterized in that, The process of generating a structured case summary includes: The simplified case keywords are used to search multi-source mediation materials using a first retrieval algorithm to determine the contextual text of the simplified case. The semantic terms of the cases are retrieved from multi-source mediation materials using a second retrieval algorithm to identify semantically similar case context texts; The simplified case context text and semantically similar case context text are sorted using a cross-encoder to determine the comprehensive retrieval results; The comprehensive search results are grouped and filled into the case summary template to generate a structured case summary; The multi-path retrieval includes a first retrieval algorithm, a second retrieval algorithm, and a cross encoder.

8. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to claim 7, characterized in that, The process of determining the comprehensive search results includes: Use the simplified case keywords as query input, and use the simplified case context text and semantically similar case context text as candidate text input; The query input and candidate text input are passed through a cross-encoder to generate text scores, and the overall retrieval results are determined based on the ranking of the text scores.

9. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to claim 8, characterized in that, The cross-encoder is a BGE Reranker architecture, which includes a raw encoding layer, a plot matching output layer, a claim matching output layer, and an entity overlap output layer. The process of generating text scores includes: The concatenated sequence of the query input and candidate text is passed through the original encoding layer to generate a semantic encoding vector; The semantic encoding vector is input into the plot matching degree output layer, the appeal matching degree output layer, and the entity overlap degree output layer, respectively, to generate the text score.

10. The intelligent summary generation method for mediation cases based on multi-source information and semantic aggregation according to claim 7, characterized in that, The process of determining the comprehensive search results also includes: A comprehensive loss function is constructed based on the triplet loss function and the binary cross-entropy loss function, and the cross encoder is fine-tuned based on the comprehensive loss function.

Images