A problem recommendation method, device and equipment

Abstract

This application discloses a method, apparatus, and device for question recommendation. First, a keyword set and the word frequency of each keyword in the target court hearing transcript are obtained based on the target transcript. Then, multiple feature vectors are retrieved from a pre-constructed co-occurrence matrix based on the keyword set. In the co-occurrence matrix, the two dimensions represent the question and the keyword, respectively. The feature vectors corresponding to the relevant keywords are then subjected to feature enhancement processing using the aforementioned word frequencies to obtain multiple enhanced feature vectors. These enhanced feature vectors are then summed to obtain a result vector. Finally, a question recommendation list for the target court hearing transcript is generated based on the values ​​in the result vector. This application achieves question recommendation through automated technical means, helping to improve the efficiency and accuracy of question extraction from textual materials in application scenarios such as court hearing investigations, and reducing reliance on the experience and ability of individual questioners.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus and device for solving a problem. Background Technology

[0002] An arbitration tribunal is a temporary organization that undertakes the task of arbitrating disputes between the parties involved. It is capable of handling various types of disputes, including civil and commercial arbitration, labor dispute arbitration (also known as labor arbitration), agricultural contract dispute arbitration, and maritime arbitration. During the arbitration hearing process, the tribunal conducts questioning based on written materials, i.e., the hearing investigation. This stage is crucial for the arbitrators to proactively investigate the facts and legal issues of the case. Its significance goes far beyond simple questioning; it is a core procedure that directly affects the direction of the case and the award outcome. It embodies the combination of the "flexibility" and "inquisitorial" characteristics of arbitration and is an important guarantee for ensuring the fairness and efficiency of the award.

[0003] The following section uses labor arbitration as an example. Labor arbitration specifically handles disputes between employers and employees arising from labor rights and obligations, such as the confirmation of employment relationships, payment of wages, and work-related injury benefits. This type of arbitration is a prerequisite for labor dispute litigation; generally, one cannot directly file a lawsuit in court without arbitration. Labor arbitration is an important component of the human resources and social security system and the preferred method for resolving labor disputes. A sound labor arbitration system can quickly and effectively protect vulnerable groups at critical moments, effectively safeguard the legitimate rights and interests of both workers and employers, reduce unnecessary civil disputes arising from labor disputes, and alleviate some of the pressure on the people's courts. The trial process of labor arbitration includes statements and responses, examination of evidence, questioning by the arbitration tribunal (trial investigation), and arguments from both parties. However, in practice, the progress of the trial investigation stage faces numerous difficulties.

[0004] First, the trial investigation phase places high demands on the arbitrator's ability to extract key questions. Arbitrators must comprehensively study and accurately grasp the arbitration applications, responses, and evidence submitted by both parties, meticulously analyzing them to extract the core issues that will substantially impact the award. This process not only requires arbitrators to possess extremely high professional competence but also demands a significant investment of time and energy; therefore, it is a profound test of their cognitive abilities.

[0005] Furthermore, human questioning has certain limitations. Current arbitration procedures rely entirely on human questioning by arbitrators, and its effectiveness is closely tied to the individual experience of each arbitrator. Because different arbitrators have varying levels of experience, oversights on key issues are inevitable, and may even lead to the omission of important facts, thereby affecting the fairness and accuracy of the award.

[0006] Furthermore, there is a severe shortage of experienced arbitrators. In recent years, with the increasing legal awareness of the working public, the number of labor arbitration cases has been on the rise, which has put increasing pressure on arbitrators. Therefore, there is an urgent need for auxiliary measures to alleviate the pressure and improve efficiency. Summary of the Invention

[0007] To address the aforementioned issues, this application provides a question recommendation method, apparatus, and device. The aim is to achieve question recommendation through automated technical means, thereby improving the efficiency and accuracy of extracting questions from textual materials in application scenarios such as court investigations, and reducing reliance on the experience and abilities of individual human questioners.

[0008] The embodiments of this application disclose the following technical solutions:

[0009] The first aspect of this application provides a problem recommendation method, which includes:

[0010] Based on the target court hearing transcript, obtain a set of keywords and the word frequency of each keyword in the target court hearing transcript;

[0011] Multiple feature vectors are retrieved from a pre-constructed co-occurrence matrix based on the keyword set; each feature vector corresponds one-to-one with a keyword in the keyword set; the co-occurrence matrix is ​​a two-dimensional matrix, where the first dimension represents the question and the second dimension represents the keyword; wherein the first dimension of the co-occurrence matrix represents the question statement extracted from the historical court transcript set, and the second dimension of the co-occurrence matrix represents the text keywords extracted from the historical court transcript set;

[0012] The word frequency is used to perform feature enhancement processing on the feature vectors corresponding to the respective keywords to obtain multiple enhanced feature vectors;

[0013] The summation operation is performed on the multiple enhanced feature vectors to obtain the result vector;

[0014] A list of recommended questions for the target court hearing transcript is generated based on the values ​​in the result vector.

[0015] A second aspect of this application provides a problem recommendation apparatus, the apparatus comprising:

[0016] The first acquisition module is used to acquire a set of keywords and the word frequency of each keyword in the target court hearing transcript based on the target court hearing transcript.

[0017] The retrieval module is used to retrieve multiple feature vectors from a pre-constructed co-occurrence matrix based on the keyword set; the multiple feature vectors correspond one-to-one with each keyword in the keyword set; the co-occurrence matrix is ​​a two-dimensional matrix, where the first dimension represents the question and the second dimension represents the keyword; wherein the first dimension of the co-occurrence matrix represents the question statement extracted from the historical court transcript set, and the second dimension of the co-occurrence matrix represents the text keywords extracted from the historical court transcript set;

[0018] The enhancement module is used to perform feature enhancement processing on the feature vectors corresponding to the keywords using the word frequency, so as to obtain multiple enhanced feature vectors;

[0019] The summation module is used to sum the multiple enhanced feature vectors to obtain a result vector;

[0020] The recommendation module is used to generate a list of recommended questions for the target court transcript based on the values ​​in the result vector.

[0021] A third aspect of this application provides a problem recommendation device, the device comprising: a memory and a processor;

[0022] The memory is used to store computer programs;

[0023] The processor is configured to run the computer program, which, when running, performs the steps of the problem recommendation method as described in the first aspect.

[0024] Compared with the prior art, this application has the following advantages:

[0025] The question recommendation method proposed in this application utilizes a pre-constructed co-occurrence matrix reflecting the association between keywords and questions. By extracting only the keywords from the target court transcript, feature vector retrieval can be achieved within the co-occurrence matrix. The retrieved feature vectors numerically reflect the historical co-occurrence relationship between each question and the keyword. Based on the retrieved feature vectors, word frequency enhancement processing is applied to strengthen the significance of the word frequency in the target court transcript, enhancing the specific role of word frequency in question recommendation. Furthermore, a result vector is obtained by summing multiple enhanced feature vectors. This result vector systematically reflects the closeness of the connection between each question and the target court transcript through numerical values, thus logically completing question recommendations specifically tailored to the target court transcript from multiple questions. This application achieves question recommendation through automated technical means, assisting in improving the efficiency and accuracy of question extraction from textual materials in application scenarios such as court investigations, and reducing reliance on the experience and ability of individual questioners. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart illustrating a problem recommendation method provided in this application embodiment;

[0028] Figure 2 An architecture diagram of a problem recommendation method provided in an embodiment of this application;

[0029] Figure 3 This is a flowchart illustrating the data preparation stage prior to co-occurrence matrix construction in an embodiment of this application.

[0030] Figure 4 This is a schematic diagram illustrating the construction of a co-occurrence matrix provided in an embodiment of this application;

[0031] Figure 5 An architecture diagram of another problem recommendation method provided in the embodiments of this application;

[0032] Figure 6 A flowchart illustrating the optimization of the co-occurrence matrix using the PageRank enhancement mechanism is provided in this application embodiment.

[0033] Figure 7 This is a schematic diagram of a problem recommendation device provided in an embodiment of this application. Detailed Implementation

[0034] In scenarios such as project management, market research and analysis, policy formulation and evaluation, community building and management, and arbitration, it is often necessary to carefully study the materials and extract the questions to be asked. Taking arbitration as an example, arbitrators need to extract the core issues that have a substantial impact on the award by combining several written materials, which places high demands on the arbitrators' personal abilities and work experience. However, the current number of senior arbitrators is insufficient to meet the growing demand for arbitration, and manual questioning has certain limitations. Therefore, manually completing the investigation phase of arbitration hearings generally presents challenges in terms of efficiency, accuracy, and cost.

[0035] To address the above problems, the inventors have proposed a fully automated question recommendation method. By leveraging a co-occurrence matrix constructed from historical court transcripts, which displays the co-occurrence of questions and keywords, when faced with a new court transcript requiring questions, only keywords and their frequencies within that transcript need to be extracted. This allows for the retrieval of multiple feature vectors from the co-occurrence matrix. Subsequently, the feature vectors are enhanced using word frequencies, and the enhanced feature vectors are summed. Finally, a new list of recommended questions for the new court transcript is generated based on the values ​​reflected in the resulting vectors. This method does not rely on human experience or ability and can be executed automatically. Using historical court transcript data as a knowledge base, it efficiently and accurately recommends questions for court materials using the co-occurrence matrix. This effectively assists the arbitration tribunal in asking questions.

[0036] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0037] See Figure 1 and Figure 2 , Figure 1 A flowchart illustrating a problem recommendation method provided in this application embodiment. Figure 2 This is an architecture diagram of a problem recommendation method provided in an embodiment of this application. Wherein, Figure 1 The recommendation process is presented from top to bottom, showing a sequential relationship. Figure 2 The implementation architecture of the new court transcript generation question recommendation list is then displayed from left to right.

[0038] S101. Obtain the keyword set and the word frequency of each keyword in the target court hearing transcript based on the target court hearing transcript.

[0039] The target court hearing transcript can be understood as a newly received court hearing transcript. In the application scenario, it is necessary to recommend questions for the target court hearing transcript. In this step, the key features of the target court hearing transcript are extracted, including keywords and their frequencies. Since there are often more than one extracted keyword, a keyword set is defined to specifically contain the keywords extracted from the target court hearing transcript.

[0040] In the specific implementation of this step, as the first optional approach, a pre-built keyword library can be used to extract words matching the keyword library from the target court transcript and add them to the keyword set of the target court transcript. In the second optional approach, considering the potential semantic similarity between words with different expressions, a mapping relationship between similar words and their representative words can be further constructed when pre-building the keyword library. Based on the keyword library and the above mapping relationship, a corresponding set of representative words can be derived. When executing this step, combining the mapping relationship between similar words and representative words, and the specific words in the representative word library, matching representative words are queried for the target court transcript, and the matching representative words are added to the keyword set of the target court transcript.

[0041] The frequency of keywords can be obtained through statistical data analysis. If the text material has fewer words, the keyword frequency will be significantly higher compared to text materials with more words, resulting in that particular court transcript contributing more to the co-occurrence matrix than other transcripts. Therefore, this application proposes that, in the stages of constructing and applying the co-occurrence matrix, the frequency of a keyword is not based on the number of times it appears, but rather on the ratio of the number of times a keyword appears in a particular court transcript to the sum of the number of times all keywords appear in the same court transcript. This ratio is then used to obtain the frequency of each keyword in the target court transcript's keyword set. This standardized approach reduces the impact of the number of words in the text material on the application of keyword frequency.

[0042] Similarly, considering the potential semantic similarities between words in different expressions, for the second implementation method of obtaining the keyword set mentioned above, when calculating keyword frequencies, the sum of the occurrences of similar words in the target court transcript that have a mapping relationship with a representative word, along with the number of times that representative word appears, can be used as the numerator for standardizing the frequency of that representative word. Then, the frequency of the keyword is obtained through division. Thus, through this implementation method, and by leveraging the mapping relationship between similar words and representative words, the frequency of each representative word in the keyword set of the target court transcript can be obtained. For easy indexing, these calculated word frequencies can be displayed in the form of a word frequency list.

[0043] The purpose of obtaining keywords in this step is to retrieve feature vectors corresponding to each word in the keyword set from the pre-constructed co-occurrence matrix, so as to reflect the close relationship between each question and the word through the feature vectors. The purpose of obtaining the word frequency of each word in the keyword set in this step is to enhance the specific influence of these words in the target trial materials, so that they can help to more accurately recommend questions that are meaningful to the scenario. The process of retrieving feature vectors is described below in conjunction with step S102; the application of word frequency is introduced in step S103.

[0044] S102. Based on the keyword set, retrieve multiple feature vectors from the pre-constructed co-occurrence matrix.

[0045] In this embodiment, the co-occurrence matrix is ​​a two-dimensional matrix, where the first dimension represents the question and the second dimension represents the keyword. Here, the first and second dimensions are used to distinguish the two different dimensions of the co-occurrence matrix: rows and columns. For example, if the first dimension is the rows of the matrix and the second dimension is the columns, then each row of the co-occurrence matrix corresponds to a question, and each column corresponds to a keyword. In another example, if the first dimension is the columns of the matrix and the second dimension is the rows, then each column of the co-occurrence matrix corresponds to a question, and each row corresponds to a keyword. In the embodiments described below, for ease of understanding and description, an example where the first dimension is the rows and the second dimension is the columns will be used. It is understood that even when the first dimension is the columns and the second dimension is the rows, the technical concept described in this application can still be adaptively applied.

[0046] The co-occurrence matrix is ​​built upon a large amount of historical court transcripts by analyzing the co-occurrence of keywords and questions. It reflects the close relationship between keywords and questions in the historical transcripts in matrix form. The first dimension of the co-occurrence matrix represents the questions extracted from the historical transcript dataset; the second dimension represents the keywords extracted from the historical transcript dataset. The larger the value of an element at a certain position in the co-occurrence matrix, the closer the relationship between the corresponding keyword and the question. In this application, the co-occurrence matrix, as a dataset reflecting the close relationship between keywords and questions based on historical data, can be used to recommend suitable questions for new court transcripts (i.e., target court transcripts), assisting the arbitration tribunal in the trial investigation phase. In summary, the established co-occurrence matrix is ​​an efficient and convenient query tool for this application, and also a data treasure trove.

[0047] To facilitate understanding, the construction process of the co-occurrence matrix will be introduced below.

[0048] First, a dataset of historical court hearing transcripts was obtained (containing multiple historical court hearing transcripts). Next, questions and keywords were extracted from each historical court hearing transcript, constructing a question set and keyword set for each transcript. The word frequency of each keyword in the keyword set within its corresponding historical court hearing transcript was calculated, constructing keyword-word frequency pairs for each transcript. Furthermore, based on the constructed keyword sets and the semantic similarity between different keywords, a representative word set corresponding to the historical court hearing transcript dataset was constructed; and based on the constructed question sets and the semantic similarity between different questions, a representative question set corresponding to the historical court hearing transcript dataset was constructed. Based on the semantic similarity between different keywords and the representative word set, the keyword-word frequency pairs for each historical court hearing transcript were corrected, resulting in corrected keyword-word frequency pairs for each transcript. A co-occurrence matrix was constructed based on the representative word set, the representative question set, and the corrected keyword-word frequency pairs for each historical court hearing transcript. The above process is explained in detail below.

[0049] Figure 3 This is a flowchart illustrating the data preparation stage prior to co-occurrence matrix construction in an embodiment of this application. From... Figure 3 As can be seen, keywords can be extracted from the applicant's application, the respondent's response, and the arbitral tribunal's cross-examination in the court hearing transcript to form a keyword database; questions can be extracted from the arbitrators' questions during the court hearing investigation to form a question database.

[0050] As an example, 4919 historical court hearing transcripts from a certain region were retrieved, and questions and keywords were extracted from each. A large language model can be applied to each historical court hearing transcript to extract keywords crucial to the questions. For example, the prompt might be: "[Instruction] The materials contain the applicant's application, the respondent's response, and the examination of evidence in a court hearing transcript. Please find keywords that are important for subsequent investigation questions. For example, if the question involves 'Should the applicant's nursing care expenses be paid?' and the materials contain related terms like 'nursing care expenses,' then it is considered a keyword. [Materials]... [Investigation Questions]...". This constructs the keyword set corresponding to that historical court hearing transcript. After deduplicating and merging the keyword sets of each historical court hearing transcript, a keyword database is formed.

[0051] From the investigation section of each historical court hearing transcript, all questions asked by the arbitrator to the applicant and respondent are extracted. These questions are then segmented, stop words are removed, and only the core semantics are retained. This step can also be achieved by designing prompts and applying a large language model to each historical court hearing transcript. The prompt is: "[Instruction] I want to convert the question list into a vector format, but I need you to preprocess the questions, including removing stop words and irrelevant information such as time, location, names, and amounts. If the text is empty, please output 'None'."

Requirements

[0052] Given that some materials have very few words, resulting in high keyword frequencies, these documents contribute more to the co-occurrence matrix than others. Therefore, this scheme adopts a standardized term frequency formula: Keyword Frequency = Number of times the keyword appears / Sum of the total number of times all keywords appear, ensuring that each document contributes 1 to the matrix. Assume a document contains the following keywords: "labor": 3 times; "dispute": 2 times; "arbitration": 1 time. The standardized term frequencies are calculated as follows:

[0053]

[0054]

[0055]

[0056] When constructing the co-occurrence matrix, it's also necessary to consider the semantic similarity between keywords and between questions. Analyzing semantic similarity allows us to establish semantic connections between similar questions and representative questions, and between similar words and representative words. This prevents the diversification of questions and keywords with similar semantics, avoiding interference with the efficient construction of the co-occurrence matrix or increasing the complexity of its application. Regarding keywords, the keyword library can be vectorized, such as... Figure 3 As shown, the K-means clustering method is used: all similar words are clustered together, and the optimal number of clusters K is determined using the elbow rule. Finally, each cluster and its center are obtained, thus constructing a mapping relationship between similar words and representative words. Regarding the problem, compared to words, sentences have more complex semantics and blurred category boundaries, making forced clustering unsuitable. Therefore, for this problem, a method such as... Figure 3 The cosine similarity threshold classification method shown ultimately obtains the mapping relationship between each representative question and its similar questions through classification. Based on the construction of the above mapping relationship, the keyword library mentioned above can be optimized into a set of representative words corresponding to the historical court transcript dataset, and the question library mentioned above can be optimized into a set of representative questions corresponding to the historical court transcript dataset.

[0057] Of course, before optimizing the keyword and question database based on semantic similarity in the previous stage, it was already possible to traverse and obtain the keyword and word frequency combinations (i.e., keyword-word frequency pairs) for each historical court transcript. After optimization, the above keyword-word frequency pairs can also be corrected to avoid the interference of diversified presentation of similar semantic questions and keywords, which would affect the accuracy of the keyword-word frequency pairs and improve the accuracy of the co-occurrence matrix to be constructed. The logic for correcting keyword-word frequency pairs is still to combine the semantic similarity between different keywords. For example, a representative word A has two similar words, namely similar word B and similar word C. Previously, the keyword-word frequency pair of the representative word A in a certain historical court transcript was calculated as: representative word A: 0.12, the keyword-word frequency pair of similar word B: similar word B: 0.1, and the keyword-word frequency pair of similar word C: similar word C: 0.15. Among them, 0.12, 0.1, and 0.15 all represent word frequencies. Through correction, the keyword-word frequency pair contained in this historical court transcript can be obtained as: representative word A: 0.37. In other words, the word frequencies of the representative word and its similar words are combined to obtain the accurate word frequency of the representative word in the historical court transcript, thereby correcting the keyword-word frequency pair of the representative word and its similar words.

[0058] Once the aforementioned preliminary work is completed, a co-occurrence matrix can be constructed based on the representative word set, the representative question set, and the revised keyword-word frequency pairs corresponding to each historical court transcript. Figure 4This is a schematic diagram of constructing a co-occurrence matrix provided in an embodiment of this application. The following is in conjunction with... Figure 4 The form of the constructed co-occurrence matrix is ​​introduced. Figure 4 This provides a simplified illustration of the local effects during an intermediate stage of co-occurrence matrix construction, showing only the values ​​in the four rows and four columns. Specifically, Figure 4 The right side displays four different representative questions: Question 1, Question 2, Question 3, and Question 4; and four different representative words: Word 1, Word 2, Word 3, and Word 4. Figure 4 The left side shows the representative questions mapped to the question set corresponding to a certain historical court transcript, including: Question 2 and Question 4. In addition, Figure 4 The left side also displays the revised keyword-frequency pairs of the historical court transcript, including word 1: 0.4 and word 3: 0.6.

[0059] In the optional implementation, a co-occurrence matrix is ​​constructed based on the representative word set, the representative question set, and the corrected keyword-word frequency pairs corresponding to each historical court transcript, including:

[0060] Construct an initial matrix with each representative question from the representative question set as the first dimension index and each representative word from the representative word set as the second dimension index, with all elements set to 0. In this example, the first dimension of the co-occurrence matrix is ​​the row, and the second dimension is the column. Therefore, the first dimension index could refer to the row index, and the second dimension index could refer to the column index. (See reference...) Figure 4 Example, Figure 4 On the right side, what is shown is only the matrix effect after traversing the first historical court transcript. Before that, the matrix effect of questions 1 to 4 and words 1 to 4 in the initial matrix should be that all elements in each of the four rows and four columns are set to 0.

[0061] The process iterates through each historical court transcript in the dataset. Based on the semantic similarity between different questions, the set of questions corresponding to the currently visited historical court transcript is transformed into a first representative question set. Similarly, based on the semantic similarity between different keywords, the set of keywords corresponding to the currently visited historical court transcript is transformed into a first representative word set. The purpose of this transformation process is to determine which representative questions and words are involved in each historical court transcript. Through these transformation operations, the keywords and questions of the visited historical court transcripts are aligned with the row and column references of the established initial matrix. Figure 4 Words 1 and 3 shown belong to the first representative set of words in this historical court transcript. Figure 4 Questions 2 and 4 shown belong to the first representative set of questions in this historical court transcript.

[0062] The first representative question set and the first representative word set are used to query the element at the corresponding position in the current initial matrix. The frequency of the corrected keyword-word frequency pair corresponding to the currently traversed historical court transcript is then added to the value of the element at the corresponding position in the current initial matrix to complete one update of the initial matrix. See also... Figure 4 Example, Figure 4 The historical court transcript shown displays the corrected keyword-frequency pairs, including word 1: 0.4 and word 3: 0.6. Furthermore, as learned through the aforementioned steps, the first representative question set only includes questions 2 and 4. Therefore, we can locate the positions in the second row, first column (2,1) = 0, the second row, third column (2,3) = 0, the fourth row, first column (4,1) = 0, and the fourth row, third column (4,3) = 0 in the current initial matrix. For the two positions (2,1) and (4,1) in the first column, since the first column corresponds to word 1 and word 1 has a frequency of 0.4, the frequency 0.4 is added to the two positions (2,1) and (4,1) in the current initial matrix. For the two positions (2,3) and (4,3) in the third column, since the third column corresponds to word 3 and word 3 has a frequency of 0.6, the frequency 0.6 is added to the two positions (2,3) and (4,3) in the current initial matrix. This demonstrates... Figure 4 The matrix effect shown on the right.

[0063] In the optional implementation, the first representative question set corresponding to the traversed historical trial records includes m representative questions, and the first representative word set includes n representative words; both m and n are positive integers. The following example illustrates the process of updating the matrix once. Traverse the first representative question set and the first representative word set. For the i-th representative question in the first representative question set and the j-th representative word in the first representative word set, query the target position in the current initial matrix, i.e., the position (i,j) in the i-th row and j-th column, and obtain the current value of the element at the target position; i takes values ​​from 1 to m, and j takes values ​​from 1 to n, where i and j are both integers. In the corrected keyword-frequency pairs corresponding to the currently traversed historical trial records, determine the frequency of the j-th representative word in the first representative word set; add the frequency of the j-th representative word to the current value of the element at the target position to complete the value update of the element at the target position. After traversing both the first representative question set and the first representative word set, complete one update of the initial matrix.

[0064] It should be noted that, Figure 4The matrix shown only represents the result after iterating through the first historical court transcript; it's an intermediate stage in the co-occurrence matrix construction process and does not represent the final result. After all historical court transcripts in the dataset have been traversed, the initial matrix with the last update will be used as the co-occurrence matrix. The implementation of adding word frequencies to the current matrix during each iteration can be found in the example process described earlier, and will not be repeated here.

[0065] Understandably, the matrix was updated multiple times by traversing the historical trial transcript dataset. Therefore, the resulting co-occurrence matrix is ​​based on all historical trial transcripts in the dataset. By presenting this data, it reflects the overall closeness of the relationship between representative issues and representative terms. Within the same row of the co-occurrence matrix, higher values ​​indicate a closer relationship between the corresponding term and the representative issue of that row; similarly, within the same column, higher values ​​indicate a closer relationship between the corresponding representative issue and the representative term of that column.

[0066] Based on the co-occurrence matrix that has already been constructed, data indexing can be completed using the words in the keyword set obtained in S101. For each word in the keyword set, a vector is indexed from the co-occurrence matrix. Since each column in the co-occurrence matrix corresponds to a representative word, the data in the column corresponding to that word can be used as an element in a feature vector. For example, if there are five representative words in the keyword set, five feature vectors can be extracted, and each extracted feature vector corresponds one-to-one with a representative word in the keyword set.

[0067] S103. Use word frequency to perform feature enhancement processing on the feature vectors corresponding to the keywords to obtain multiple enhanced feature vectors.

[0068] In this step, feature enhancement processing is performed on the feature vector using word frequency, strengthening the crucial role of word frequency in the target court transcript. This enhancement process creates a strong correlation between the enhanced feature vector and the inherent features of the target court transcript. Therefore, the enhanced feature vector not only reflects the co-occurrence characteristics of questions and keywords in the historical court transcript dataset but also reflects the significant contribution of keywords in the target court transcript to question extraction and recommendation.

[0069] In an optional implementation of this step, feature enhancement can be achieved by multiplying the feature vector corresponding to the keyword by the word frequency of that keyword in the target court transcript, thereby obtaining the enhanced feature vector of the keyword. For example, if the target court transcript contains the keyword "nursing expenses" with a word frequency of 0.3 and the keyword "work injury insurance" with a word frequency of 0.7, then the columns corresponding to "nursing expenses" and "work injury insurance" can be extracted from the co-occurrence matrix to construct two feature vectors, and then multiplied by 0.3 and 0.7 respectively.

[0070] In practical applications, this step is not limited to the feature enhancement method of directly multiplying the feature vector by the word frequency of the keyword. For example, summation or other operations involving summation, multiplication, or exponentiation can also be used to achieve feature enhancement.

[0071] S104. Summate multiple enhanced feature vectors to obtain the result vector.

[0072] Summing multiple enhanced feature vectors essentially involves accumulating the values ​​of elements at the same positions in these vectors. The purpose is to reflect the matching degree between each issue and the target trial transcript from the perspective of the overall target trial transcript by using the value of elements at each position in the result vector.

[0073] S105. Generate a list of recommended questions for the target court hearing transcript based on the values ​​in the result vector.

[0074] In an optional implementation, this step can be specifically as follows: based on the one-to-one correspondence between the elements at each position of the result vector and the questions, determine the questions corresponding to the elements with positive values ​​in the result vector; arrange the determined questions in descending order according to the numerical values ​​of the corresponding elements to generate a recommended list of questions for the target court hearing transcript.

[0075] For example, the summation result vector is represented as [11 13 25 0 0 10], with each position corresponding to question 1, question 2, question 3, question 4, question 5, and question 6, respectively. From this result vector, we can see that question 3 has the highest matching degree with the target court transcript and is more suitable as a question recommended for the court investigation stage. The resulting order is: ① Question 3; ② Question 2; ③ Question 1; ④ Question 6.

[0076] In the question recommendation method described in the above embodiments, by leveraging a pre-constructed co-occurrence matrix reflecting the association between keywords and questions, feature vector retrieval can be achieved by extracting only the keywords of the target court hearing transcript from the co-occurrence matrix. In the retrieved feature vectors, numerical values ​​reflect the historical co-occurrence relationship between each question and the keyword. Based on the retrieved feature vectors, word frequency feature enhancement processing is applied to strengthen the significance of the word frequency in the target court hearing transcript, enhancing the specific role of word frequency in question recommendation. Furthermore, a result vector is obtained by summing multiple enhanced feature vectors. In the result vector, the closeness of the connection between each question and the target court hearing transcript can be systematically reflected numerically, thus logically completing question recommendations specifically for the target court hearing transcript among multiple questions. This application achieves question recommendation through automated technical means, assisting in improving the efficiency and accuracy of question extraction based on textual materials in application scenarios such as court hearing investigations, and reducing reliance on the individual experience and ability of human questioners.

[0077] In step S101 of the previous embodiment, it was mentioned that the frequency of keywords in the target court transcript was statistically analyzed. The following describes a specific implementation of this process: the standardization process of keyword frequency analysis.

[0078] In the optional implementation, taking the first target keyword in the keyword set corresponding to the target court hearing transcript as an example: obtain the number of times the first target keyword in the keyword set appears in the target court hearing transcript, and obtain the number of times other keywords in the keyword set appear in the target court hearing transcript; use the sum of the number of times the first target keyword appears in the target court hearing transcript and the number of times other keywords appear in the target court hearing transcript as the denominator, and use the number of times the first target keyword appears in the target court hearing transcript as the numerator, and use the ratio obtained as the word frequency of the first target keyword in the target court hearing transcript.

[0079] The above process describes the standardized calculation process for the word frequency of keywords in the target court transcript.

[0080] Similarly, the following describes the technical implementation of extracting keyword frequencies from historical court transcripts in a dataset, namely, the standardization process of word frequencies. The rationale for using the standardization calculation method for word frequencies has already been explained above and will not be repeated here.

[0081] In the optional implementation, taking the target historical court hearing transcript and its keyword set in the historical court hearing transcript dataset as an example: obtain the number of times the second target keyword appears in the target historical court hearing transcript and the number of times other keywords appear in the target court hearing transcript in the keyword set corresponding to the target historical court hearing transcript in the historical court hearing transcript dataset; use the sum of the number of times the second target keyword appears in the target historical court hearing transcript and the number of times other keywords appear in the target historical court hearing transcript as the denominator, and use the number of times the second target keyword appears in the target historical court hearing transcript as the numerator, and use the obtained ratio as the word frequency of the second target keyword in the target historical court hearing transcript, and construct the keyword-word frequency pair of the second target keyword.

[0082] The above process describes the standardized calculation process for keyword frequencies in historical court transcripts.

[0083] The following section continues with a discussion of word frequency, introducing methods for correcting keyword-frequency pairs corresponding to historical court transcripts. The correction of keyword-frequency pairs was already described in section S102, primarily involving the mapping relationship between representative words and their similar words. The real significance of this correction lies in integrating the word frequencies of semantically similar keywords. In optional implementations, based on the semantic similarity between different keywords and the set of representative words, the keyword-frequency pairs corresponding to each historical court transcript are corrected, including:

[0084] If the second target keyword is one of several representative words in the representative word set, then identify the similar keywords in the keyword set corresponding to the target historical court transcript that have a similar word mapping relationship with the second target keyword. Add the word frequency of the similar keywords in the target historical court transcript to the word frequency of the second target keyword in the target historical court transcript to obtain the first word frequency sum. Use this first word frequency sum to correct the keyword-word frequency pair of the second target keyword. This means that the word frequencies of similar words are superimposed to correct the word frequency of the second target keyword.

[0085] If the second target keyword is not one of the multiple representative words in the representative word set, then the target representative word corresponding to the second target keyword in the representative word set is determined based on the similarity word mapping relationship. The word frequency of the second target keyword in the target historical court transcript, the word frequency of the target representative word in the target historical court transcript, and the word frequencies of other keywords with similarity word mapping relationships with the target representative word in the target historical court transcript are added together to obtain the second word frequency sum. The keyword-word frequency pair of the target representative word corresponding to the target historical court transcript is corrected using the second word frequency sum. This means that the word frequencies of similar words are superimposed on the word frequencies of their representative words, and the final superposition result is used as the corrected word frequency of the representative word.

[0086] The above analysis, considering both representative and non-representative keywords for the second target keyword, outlines the operations performed during the correction process for these two different scenarios. Regardless of whether the second target keyword is a representative keyword, its originally calculated word frequency can be corrected by overlaying the frequency of the representative keyword. This ensures the accuracy of the values ​​in the constructed co-occurrence matrix and improves its effectiveness in question recommendation scenarios.

[0087] In practical applications, certain questions often have a fixed questioning order. For example, "When did the applicant start working at the applied-for unit?" and "What job do they do?" are asked in a sequential order. In the question recommendation method embodiment described above, the questioning order factor was not considered during the generation of the recommendation list. In this application, to make the question order presented in the question recommendation list more in line with human questioning habits and language logic, a sequence-aware weighted algorithm is proposed. This algorithm uses subsequence mining to weight the sequential elements in the score of the questions in the question recommendation list. Specifically, a set of questions from several historical court transcripts can be mapped to a representative question set based on the semantic similarity of the questions. Optionally, these representative question sets can be mined using the PrefixSpan library in Python to obtain subsequences of length 2 with a support greater than 500, thus obtaining high-frequency binary question sequences. The first question in each binary question sequence is called position 0, and the second is called position 1. For each binary question sequence, if both questions are in the question recommendation list generated by this scheme, the element at position 1 is moved after position 0. Furthermore, when there is a competing order between two or more sets of binary problem sequences, the sequence with higher support is preferred.

[0088] The following describes the possible implementations of the problem recommendation method based on the above concept. Among the possible implementations, the problem recommendation method also includes:

[0089] Obtain a dataset of historical court trial transcripts; this dataset includes multiple historical court trial transcripts. Extract questions from each historical court trial transcript to construct a question set corresponding to each transcript. Based on the constructed question sets and the semantic similarity between different questions, construct a representative question set corresponding to the historical court trial transcript dataset. Note that if the historical court trial transcript dataset has already been obtained and a representative question set has been constructed in this method, this step can be omitted. Mine subsequences with a length of 2 that meet the preset support condition from the representative question set. Add the mined subsequences as high-frequency binary question sequences to the mined sequence set; the question priority of the first question in the high-frequency binary question sequence is higher than that of the second question. The preset condition can be set according to actual needs, for example, a support greater than 500 can be set. There is no limit to the support threshold for the preset condition here.

[0090] After generating the recommended list of questions for the target court hearing transcript, the method also includes:

[0091] If two questions in the question recommendation list both belong to the target high-frequency binary question sequence, and the question recommended first is the first question in the target high-frequency binary question sequence, and the question recommended second is the second question in the target high-frequency binary question sequence, then the order of the two questions in the question recommendation list belonging to the target high-frequency binary question sequence will not be adjusted. The target high-frequency binary question sequence can be any high-frequency binary question sequence in the mining sequence set. In the above case, it means that the order of these two questions in the current question recommendation list matches the order in the target high-frequency binary sequence, which conforms to human questioning habits and language logic, therefore no further adjustment is needed.

[0092] If two questions in the question recommendation list both belong to the target high-frequency binary question sequence, and the question recommended first is the second question in the target high-frequency binary question sequence, while the question recommended later is the first question in the target high-frequency binary question sequence, then in the question recommendation list, the question recommended first will be moved after the question recommended later. This situation means that the order of these two questions in the current question recommendation list contradicts their order in the target high-frequency binary sequence, which does not conform to human questioning habits and language logic. Therefore, the order of these two questions in the question recommendation list needs to be adjusted again.

[0093] Based on the above embodiments, it is easy to see that the question recommendation method proposed in this application not only facilitates the work of arbitrators by automatically extracting high-value questions to be asked from the target court transcript, but also adds sequential elements to the recommended question list. This allows for consideration of questioning habits and language logic when asking questions in the order of the question recommendation list, resulting in better application effects in the court investigation stage.

[0094] Building upon this, the inventors proposed three additional optimization ideas for the co-occurrence matrix. Optimizing the co-occurrence matrix can further improve the effectiveness of question recommendations. Figure 5 An architecture diagram of another problem recommendation method provided in the embodiments of this application.

[0095] Figure 5 The document outlines four stages: data preparation, co-occurrence matrix construction, question recommendation, and matrix optimization. Data preparation involves constructing representative question sets and representative word sets based on a dataset of historical court transcripts and the semantic similarity between questions and keywords. Furthermore, the data preparation stage requires calculating the word frequency of each representative word in each historical court transcript to aid in constructing the co-occurrence matrix in the next stage. The co-occurrence matrix construction stage combines the data obtained in the data preparation stage, continuously updating the values ​​of elements at each position in the co-occurrence matrix by traversing each historical court transcript. After traversal, the co-occurrence matrix is ​​obtained as a vector retrieval tool for the next stage. In the question recommendation stage, keywords are extracted from new court transcripts and converted into representative words. Word frequencies are calculated, and feature vectors for each extracted representative word are retrieved from the co-occurrence matrix. These feature vectors are then enhanced using the corresponding word frequencies. After this step, the enhanced feature vectors are summed to obtain a result vector. The questions are then sorted based on the numerical values ​​in the result vector, ultimately forming a question recommendation list for the new court transcript. This completes the recommendation of questions for the new court transcript. Figure 5 The matrix optimization shown can improve the matrix in various ways. For example, it can be done by one or more of the following methods: sparsification, semantic similarity weighting, or PageRank enhancement.

[0096] The optimized implementation of the co-occurrence matrix is ​​introduced below. The sparsity processing, semantic similarity weighting, or PageRank enhancement mechanisms mentioned above mainly target three optimization directions, and the reasons for proposing optimization are as follows: ①, ②, and ③. ① In the constructed co-occurrence matrix, there may be low-frequency noise interference. Specifically, in the co-occurrence matrix, there may be a large number of low co-occurrence values ​​(accidental co-occurrence of questions and keywords), diluting the correlation between keywords and questions. ② The semantic granularity of the constructed co-occurrence matrix is ​​coarse. Specifically, the co-occurrence matrix is ​​mainly constructed based on word frequency statistics, which may be limited in its ability to capture the deep semantic relationship between questions and materials. ③ Importance transfer is missing. Specifically, the global influence of question vertices in the network is not considered in the constructed co-occurrence matrix.

[0097] Regarding point ②, among the optional implementations, after constructing the co-occurrence matrix, the recommended approach for this problem also includes:

[0098] For each representative word in the representative word set, extract its semantic vector; for each representative question in the representative question set, extract its semantic vector. This step represents the semantics of the representative words and questions in vector form, facilitating similarity calculation. This step can be implemented using a pre-trained BERT model. For example, a BERT model can be used to extract a 768-dimensional question semantic vector for each representative question and a 768-dimensional word semantic vector for each representative word.

[0099] The cosine similarity between the extracted word semantic vector and the question semantic vector is calculated, and a cosine similarity matrix is ​​constructed based on the calculated cosine similarities. In the cosine similarity matrix, the first dimension represents the representative question, and the second dimension represents the representative word. In one example, if the first dimension of the matrix is ​​the row and the second dimension is the column, then each row in the cosine similarity matrix corresponds to a different representative question, and each column corresponds to a different representative word. In another example, if the first dimension of the matrix is ​​the column and the second dimension is the row, then each column in the cosine similarity matrix corresponds to a different representative question, and each row corresponds to a different representative word. The definitions of the first and second dimensions should be aligned in the cosine similarity matrix and the co-occurrence matrix to ensure the feasibility of subsequent matrix fusion operations.

[0100] The following section will introduce the relevant information using a previous example. In this step, the semantic vectors of the words and the semantic vectors of the questions are compared. The cosine similarity value reflects the semantic difference between the representative words and the representative vectors. Furthermore, the deep semantic relationship between the questions and the material can be easily captured from the cosine similarity matrix. The element s at the i-th row and j-th column of the cosine similarity matrix S is... ijIt can be calculated using the following formula:

[0101]

[0102] In the formula, q i Let k represent the semantic vector of the question representing the i-th row. j This represents the semantic vector of the representative word in column j.

[0103] The co-occurrence matrix and the cosine similarity matrix are fused to obtain the first optimized co-occurrence matrix. Assuming that before optimization, the co-occurrence matrix is ​​represented as M and the cosine similarity matrix as S, the process of fusing them to obtain the first optimized co-occurrence matrix M' can be expressed as:

[0104]

[0105] In the formula, This represents the fusion weight coefficient of S, as an example. =0.2. By calculating the cosine similarity between the question semantic vector and the word semantic vector, and constructing a cosine similarity matrix, and then fusing the cosine similarity matrix with the original co-occurrence matrix, we can capture the deep semantic connections between the question and the material, making the semantic granularity of the first optimized co-occurrence matrix more refined. Testing revealed that this optimization strategy significantly improved the recall rate for long-tail questions.

[0106] Regarding point ①, among the optional implementations, after obtaining the first optimized co-occurrence matrix, the recommended approach for this problem also includes:

[0107] The values ​​of elements at each position in the first optimized co-occurrence matrix are filtered using a preset threshold. Elements with values ​​below the preset threshold are reset to 0, while elements with values ​​above or equal to the preset threshold are retained, resulting in a second optimized co-occurrence matrix. Essentially, this optimization strategy applies threshold filtering to the elements at each position in the first optimized co-occurrence matrix. The preset threshold can be set according to actual needs; for example, a lower threshold can be set if stronger resistance to low-frequency noise interference is required. As an example, a preset threshold of 1 means that elements greater than or equal to 1 are retained, while elements less than 1 are reset to 0. This sparsification of the co-occurrence matrix before optimization effectively eliminates low-frequency noise interference, improving the accuracy of the recommended questions.

[0108] After obtaining the second optimized co-occurrence matrix, it is clear that issues ① and ② have been overcome. Therefore, when applying the co-occurrence matrix for feature vector retrieval in the face of new court transcripts, we can specifically retrieve multiple feature vectors from the second optimized co-occurrence matrix based on the keyword set of the target court transcript (i.e., the new court transcript). This optimizes the subsequent vector retrieval effect and question recommendation performance.

[0109] Regarding point ③, through Figure 6 Demonstrate the matrix optimization process. Figure 6 This application provides a flowchart illustrating the optimization of the co-occurrence matrix using the PageRank enhancement mechanism. The PageRank algorithm is used to evaluate the importance of web pages on the internet. It determines the importance of a page by analyzing the link structure between pages and is commonly used for ranking search engine results, making it well-suited for question recommendation scenarios. However, the traditional PageRank mechanism has two significant limitations in question recommendation: first, the assumption of uniformly distributing edge weights fails to capture the semantic correlation strength between questions; second, it ignores the domain statistical characteristics of the nodes themselves. See also... Figure 6 This reveals the inventor's improvement strategy for the above two problems in this application.

[0110] like Figure 6 As shown, among the optional implementation methods, the recommended approach also includes:

[0111] S601. Based on the co-occurrence of pairs of questions in each historical trial transcript, construct the question transition matrix of the historical trial transcript dataset.

[0112] In the problem transition matrix, both the first and second dimensions represent representative problems; that is, each row corresponds to a different representative problem, and each column also corresponds to a different representative problem. The representative problem corresponding to the row containing the element on the diagonal of the problem transition matrix is ​​the same as the representative problem corresponding to the column containing that element. The value of the element in the i-th row and j-th column of the problem transition matrix represents the transition probability from the i-th representative problem to the j-th representative problem.

[0113] In an optional implementation, this step may include:

[0114] The process iterates through the historical court transcript dataset, counting the co-occurring issues in each transcript and assigning these counts to the two symmetrical positions in the first transition matrix associated with the co-occurring issues. This process continues until the dataset is fully iterated, yielding the second transition matrix. The first transition matrix is ​​an initialized problem transition matrix, with all elements set to 0 during initialization. Both the first and second dimensions of the first transition matrix represent different representative issues, with each row and column corresponding to a different representative issue. The elements of each row in the second transition matrix are then normalized to obtain the final problem transition matrix.

[0115] For example, regarding the problem of co-occurrence of each pair (q) i , q j In the initialized question transition matrix, the positions (i, j) in the i-th row and j-th column, and (j, i) in the j-th row and i-th column, are incremented by 1. This process searches for co-occurrence relationships between pairs of questions within the historical court transcript dataset, using numerical counts to reflect the frequency of co-occurrence. After normalizing the second transition matrix, it is necessary to ensure that the sum of the values ​​in each row is 1. That is, the sum of the transition probabilities from each question to other questions is 1. Normalizing a row of data is a mature technique in this field, therefore the specific normalization process will not be elaborated here.

[0116] S602. Based on the representative problem set and the problem transition matrix, construct a problem co-occurrence graph with representative problems as vertices and transition probabilities as edge weights.

[0117] In this application, a problem co-occurrence graph G can be constructed based on a representative problem set from a historical court transcript dataset. The vertex set in this graph can be represented by V, where V is essentially the representative problem set, and each vertex in V represents a representative problem from that set. As mentioned earlier, a problem transition matrix is ​​constructed in S601, where each position represents the transition probability from one problem to another. For example, the value at the i-th row and j-th column represents problem q. i To question q j The transition probability, where the position in row j and column i represents problem q. j To question q i The transition probability is calculated as follows: The problem co-occurrence graph G constructed in this application is a directed graph, and the edges between two vertices have a direction. The transition probability is then used as the edge weight of the directed edge between vertices in the problem co-occurrence graph G. For example, representing problem q... i The vertex pointing to represents the problem q j An edge to a vertex, with edge weight w ij It equals the value of the element at the i-th row and j-th column of the problem transition matrix.

[0118] Traditional PageRank defines edge weights as uniformly distributed (all outgoing edge weights = 1 / out-degree), which ignores domain knowledge that "certain question combinations require joint questioning." This approach uses the co-occurrence value between questions as edge weights, automatically strengthening the transition probability of frequently co-occurring questions and ensuring that the recommended question chain conforms to actual dispute patterns. This solves the problem that the assumption of uniformly distributed edge weights in traditional approaches fails to capture the semantic correlation strength between questions.

[0119] In traditional PageRank, all nodes have the same initial weight (1 / N). However, this application uses the newly generated court transcripts to generate a list of recommended questions using a previously optimized co-occurrence matrix, and uses the score of each question as the initial weight of the question vertex. This helps high-frequency questions (such as points of contention) receive higher initial weights, continuously influencing the results during iteration. The implementation strategy of this optimization process is described below through sections S603-S606.

[0120] S603. Obtain the semantic weighted score of each question in the question recommendation list based on the result vector, and perform normalization processing based on the semantic weighted score of each question in the question recommendation list to obtain the initial weight of each question and the initial weight vector.

[0121] The resulting vector here can be the vector obtained after vector retrieval, feature enhancement processing, and vector summation using the second optimized co-occurrence matrix. Of course, if the PageRank improvement is independent of the improvements in points ① and ② above, then the resulting vector here can also be the initially created co-occurrence matrix without optimization. This section uses the example of obtaining the final resulting vector using the second optimized co-occurrence matrix and recommending a list of recommended questions to illustrate this step and subsequent steps.

[0122] It is understandable that the numerical values ​​of elements at different positions in the result vector represent the semantic weighted score of each representative question. These scores can be normalized, and the normalized values ​​serve as the initial weights for the representative questions at their respective positions. The vector containing the initial weights of each question—that is, the normalized vector of the result vector—can be called the initial weight vector. The initial weight vector includes the initial weights of each vertex in the problem co-occurrence graph. The normalization formula is as follows:

[0123]

[0124] In the formula, q i Let s represent the i-th question. i This represents the problem q in the result vector. i The semantic weighted score, PR0(q) i ) represents the problem q obtained after normalization. iThe initial weights are calculated using the above formula. The initial weights for other problems can also be calculated using the above formula. This application employs an iterative optimization method, asynchronously updating the weights of each vertex in the co-occurrence graph of the problem, ensuring continuous updates based on the initial weights. See the description in S604.

[0125] S604. The initial weight vector is iteratively optimized using the problem transition matrix, the initial weight vector, and the total number of vertices in the co-occurring problem graph to obtain the weight vector after multiple iterations.

[0126] One possible implementation of this step is shown in the formula below:

[0127]

[0128] In the formula, represents the damping coefficient, which can be 0.8; T represents the problem transition matrix, obtained after step S601; N represents the total number of vertices in the co-occurring problem graph, which can also be understood as representing the total number of problems in the problem set; t represents the iteration number, when t=0, w (0) The initial weight vector is obtained after step S603. The '1' in the above formula represents a vector with all elements equal to 1. Thus, the sum of vectors can still be represented as a vector. Using the above formula, a new weight vector containing the weights of each vertex can be calculated after one iteration. As an example, 100 iterations can be performed to obtain the final vector w. (100) .

[0129] S605. For each question in the question recommendation list, the final score of the question is obtained by using the semantic weighted score of the question and the weight of the question in the weight vector after multiple iterations.

[0130] The following is an example implementation of this step:

[0131]

[0132] For each question, the final score can be calculated using the above formula. Taking a specific question as an example, PR... x Score represents the weight of the problem after x iterations, that is, the weight of the vertices representing the problem after x iterations of the problem co-occurrence graph. semantic This represents the semantic weighted score for the problem, and its value is equivalent to the element value at the corresponding position in the final result vector of the second optimization co-occurrence matrix. β represents the weight of the semantic weighted score, and correspondingly, 1-β represents the weight of the vertex in the graph. The calculated Score... final This represents the final score for the question. For example, β = 0.9. This means that the semantically weighted score has a greater weight, but the score also considers the weight of the vertex after multiple iterations.

[0133] S606. Based on the final scores of each question in the question recommendation list, rearrange the order of the questions in the question recommendation list.

[0134] As an example, the questions in the recommended list are sorted from highest to lowest based on their final scores obtained from S605. This sorting method takes into account the domain statistical characteristics of the vertices themselves, which helps high-frequency questions (such as points of contention) to obtain higher initial weights and continuously influence the results during iterations.

[0135] The inventors in this application propose a three-level optimization framework for the co-occurrence matrix: noise is eliminated through sparsification (threshold filtering) while retaining significant co-occurrence relationships; similarity weighting is introduced to calculate the semantic similarity between the recommended questions and material keywords, dynamically enhancing the weight of questions that conform to the core semantics of the case; PageRank weighting is applied to iteratively update the importance of questions through the topological structure of the co-occurrence network, ensuring that the recommendation results reflect both local co-occurrence features and conform to the global importance distribution, taking into account the global influence of vertices in the network.

[0136] This application achieves a deep integration of statistical learning and domain knowledge by combining traditional statistical methods such as co-occurrence matrices with cutting-edge large language modeling technology. This innovation not only significantly improves the accuracy of semantic understanding but also provides key clues for the intelligent generation of investigation questions. Compared to purely manual questioning, this solution represents a qualitative leap in efficiency while effectively avoiding the problem of missing key information due to human oversight. Traditional manual questioning requires arbitrators to spend an average of 4-6 hours studying materials and designing questions, while this application proposes a complete framework for automatically generating trial investigation questions. This framework, based on co-occurrence matrix technology, utilizes keywords from the materials submitted by both parties before the trial to generate a list of compliance questions in seconds. Arbitrators only need to roughly understand the needs of both parties during the trial and can then ask questions based on the intelligently recommended question list in this solution, listen to the statements of both parties to reveal the truth of the case, and thus provide a basis for a fair ruling, greatly reducing the workload of traditionally entirely manual questioning. In addition, the automation process reduces repetitive manual labor, and is expected to reduce the overall cost of arbitration services by more than 30%, especially suitable for high-concurrency case processing scenarios.

[0137] Furthermore, this application fills a gap in intelligent questioning in labor arbitration. Currently, there is no fully automated intelligent recommendation scheme for generating investigation questions in labor arbitration hearings. The technology proposed in this application fills this gap, aiming to help arbitrators reduce workload, improve work efficiency, and prevent the omission of important information due to human error.

[0138] Co-occurrence matrices are a commonly used technique in natural language processing and text generation tasks to capture relationships between words. They construct a matrix representing the associations between words by statistically analyzing the frequency of their co-occurrence in a specific context, quantifying co-occurrence relationships and revealing their semantic connections. However, given the limitations of co-occurrence matrix applications, it is clearly impossible to directly use co-occurrence matrices to statistically analyze the connections between the materials submitted by the applicant and respondent and the questions raised by the arbitrator in the intelligent recommendation scenario of arbitration hearing investigation questions. Therefore, for the intelligent recommendation scenario of arbitration hearing investigation questions, this invention, utilizing co-occurrence matrix-related technologies and large-scale models, integrates labor arbitration business knowledge, analyzes the process of raising investigation questions in the hearing transcript, and learns the associations between the materials submitted by both parties and the questions, as well as between the questions themselves, through various algorithms such as PageRank, ultimately achieving the recommendation of an optimal question list based on the materials.

[0139] This application marks the first time that PageRank has been integrated and improved in the judicial field: it organically combines co-occurrence statistical features with graph neural representation; and it also achieves dynamic correction of transition probabilities based on court trial scenarios. This application improves upon traditional PageRank by abandoning the average edge weight definition and instead using per-present value to automatically strengthen the transition probability of high-frequency co-occurrence issues, ensuring that the recommended issue chain conforms to the actual dispute pattern; it also abandons the average node weight definition and instead uses the existing score of each node as the initial node weight, which helps high-frequency issues obtain higher initial weights, continuously influencing the results during iteration, making the issue recommendations more accurate and meeting the needs of the scenario.

[0140] Based on the problem recommendation method described in the foregoing embodiments, this application also provides a problem recommendation device. Figure 7 This is a schematic diagram of the device. Figure 7 The problem recommendation device shown includes:

[0141] The first acquisition module 71 is used to acquire a keyword set and the word frequency of each keyword in the keyword set in the target court hearing transcript based on the target court hearing transcript.

[0142] The retrieval module 72 is used to retrieve multiple feature vectors from a pre-constructed co-occurrence matrix based on the keyword set; the multiple feature vectors correspond one-to-one with each keyword in the keyword set; the co-occurrence matrix is ​​a two-dimensional matrix, wherein the first dimension represents the question and the second dimension represents the keyword; wherein the first dimension of the co-occurrence matrix represents the question statement extracted from the historical court transcript set, and the second dimension of the co-occurrence matrix represents the text keywords extracted from the historical court transcript set;

[0143] Enhancement module 73 is used to perform feature enhancement processing on the feature vector corresponding to the keyword using the word frequency to obtain multiple enhanced feature vectors;

[0144] Summation module 74 is used to sum the multiple enhanced feature vectors to obtain a result vector;

[0145] Recommendation module 75 is used to generate a list of recommended questions for the target court transcript based on the values ​​in the result vector.

[0146] In optional implementations, the problem recommendation device also includes:

[0147] The second acquisition module is used to acquire the historical court hearing transcript dataset; the historical court hearing transcript dataset includes multiple historical court hearing transcripts.

[0148] The first construction module is used to extract questions and keywords from each of the historical court hearing transcripts and construct a question set and keyword set corresponding to each of the historical court hearing transcripts.

[0149] The statistics module is used to count the word frequency of each keyword in the keyword set in the corresponding historical court transcript, and to construct a keyword-word frequency pair for each historical court transcript.

[0150] The second construction module is used to construct a set of representative words corresponding to the historical court transcript dataset based on the constructed sets of keywords and the semantic similarity between different keywords; and to construct a set of representative questions corresponding to the historical court transcript dataset based on the constructed sets of questions and the semantic similarity between different questions.

[0151] The correction module is used to correct the keyword-word frequency pair corresponding to each of the historical court hearing transcripts based on the semantic similarity between different keywords and the representative word set, so as to obtain the corrected keyword-word frequency pair corresponding to each of the historical court hearing transcripts.

[0152] The matrix construction module is used to construct the co-occurrence matrix based on the representative word set, the representative question set, and the corrected keyword-word frequency pairs corresponding to each of the historical court transcripts.

[0153] In an optional implementation, the matrix construction module includes:

[0154] An initial matrix construction unit is used to construct an initial matrix with each representative question in the representative question set as the first dimension index, each representative word in the representative word set as the second dimension index, and each element at each position set to 0.

[0155] The update unit is used to traverse each historical trial transcript in the historical trial transcript dataset, convert the set of questions corresponding to the currently traversed historical trial transcript into a first representative question set based on the semantic similarity between different questions, and convert the set of keywords corresponding to the currently traversed historical trial transcript into a first representative word set based on the semantic similarity between different keywords; use the first representative question set and the first representative word set to query the elements at the corresponding positions of the current initial matrix, and add the word frequency of the corrected keyword-word frequency pair corresponding to the currently traversed historical trial transcript to the value of the element at the corresponding position of the current initial matrix to complete one update of the initial matrix;

[0156] The co-occurrence matrix determination unit is used to take the initial matrix that has been last updated as the co-occurrence matrix after the historical court transcript dataset has been traversed.

[0157] In an optional implementation, the first representative question set includes m representative questions, and the first representative word set includes n representative words; m and n are both positive integers.

[0158] The update unit is specifically used for:

[0159] Traverse the first representative question set and the first representative word set. For the i-th representative question in the first representative question set and the j-th representative word in the first representative word set, query the target position in the current initial matrix and obtain the current value of the element at the target position.

[0160] i takes values ​​from 1 to m, and j takes values ​​from 1 to n, where i and j are both integers;

[0161] In the corrected keyword-frequency pairs corresponding to the currently traversed historical court transcripts, determine the frequency of the j-th representative word of the first representative word set;

[0162] The frequency of the j-th representative word is added to the current value of the element at the target position to complete the value update of the element at the target position;

[0163] After both the first representative question set and the first representative word set have been traversed, the initial matrix is ​​updated once.

[0164] In an optional implementation, the first acquisition module 71 includes:

[0165] The first acquisition unit is used to acquire the number of times the first target keyword in the keyword set appears in the target court transcript;

[0166] The second acquisition unit is used to acquire the number of times other keywords in the keyword set appear in the target court transcript;

[0167] The first word frequency calculation unit is used to take the sum of the number of times the first target keyword appears in the target court transcript and the number of times the other keywords appear in the target court transcript as the denominator, and the number of times the first target keyword appears in the target court transcript as the numerator, and the obtained ratio is taken as the word frequency of the first target keyword in the target court transcript.

[0168] The statistics module includes:

[0169] The third acquisition unit is used to acquire the number of times the second target keyword appears in the target historical court hearing transcript and the number of times other keywords appear in the target court hearing transcript in the keyword set corresponding to the target historical court hearing transcript in the historical court hearing transcript dataset.

[0170] The second word frequency calculation unit is used to take the sum of the number of times the second target keyword appears in the target historical court transcript and the number of times other keywords appear in the target historical court transcript as the denominator, and the number of times the second target keyword appears in the target historical court transcript as the numerator, and the obtained ratio is taken as the word frequency of the second target keyword in the target historical court transcript, and constructs the keyword-word frequency pair of the second target keyword.

[0171] In an optional implementation, the correction module is specifically used for:

[0172] If the second target keyword is one of the multiple representative words in the representative word set, then determine the similar keywords in the keyword set corresponding to the target historical court transcript that have a similar word mapping relationship with the second target keyword;

[0173] The word frequency of the similar keywords in the target historical court transcript is added to the word frequency of the second target keyword in the target historical court transcript to obtain the first word frequency sum. The keyword-word frequency pair of the second target keyword is then corrected using the first word frequency sum.

[0174] If the second target keyword is not one of the multiple representative words in the representative word set, then the target representative word corresponding to the second target keyword in the representative word set is determined based on the similar word mapping relationship;

[0175] The word frequency of the second target keyword in the target historical court transcript, the word frequency of the target representative word in the target historical court transcript, and the word frequencies of other keywords with similar word mapping relationships to the target representative word in the target historical court transcript are added together to obtain a second word frequency sum. The keyword-word frequency pair of the target representative word corresponding to the target historical court transcript is corrected using the second word frequency sum.

[0176] In an optional implementation, the recommendation module 75 is specifically used for:

[0177] Based on the one-to-one correspondence between the elements at each position of the result vector and the questions, determine the questions corresponding to the elements with positive values ​​in the result vector;

[0178] The identified questions are sorted in descending order according to the numerical value of their corresponding elements to generate a recommended list of questions for the target court hearing transcript.

[0179] In optional implementations, the problem recommendation device also includes:

[0180] The second acquisition module is used to acquire a dataset of historical court hearing transcripts; the dataset of historical court hearing transcripts includes multiple historical court hearing transcripts.

[0181] The third construction module is used to extract questions from each of the historical court hearing transcripts and construct a question set corresponding to each of the historical court hearing transcripts.

[0182] The fourth construction module is used to construct a representative question set corresponding to the historical court transcript dataset based on the constructed question sets and the semantic similarity between different questions;

[0183] The sequence mining module is used to mine subsequences with a length of 2 that meet preset support conditions from the representative question set, and add the mined subsequences as high-frequency binary question sequences to the mining sequence set; in the high-frequency binary question sequence, the first question in the sequence has a higher question priority than the second question in the sequence.

[0184] The problem recommendation device also includes a problem ranking adjustment module, used for:

[0185] If there are two questions in the question recommendation list that both belong to the target high-frequency binary question sequence, and the question recommended first is the first question in the target high-frequency binary question sequence, and the question recommended second is the second question in the target high-frequency binary question sequence, then the order of the two questions in the question recommendation list that both belong to the target high-frequency binary question sequence will not be adjusted.

[0186] If there are two questions in the question recommendation list that both belong to the target high-frequency binary question sequence, and the question recommended first is the second question in the target high-frequency binary question sequence, and the question recommended second is the first question in the target high-frequency binary question sequence, then in the question recommendation list, the question recommended first is moved after the question recommended second.

[0187] In an optional implementation, the enhancement module 73 is specifically used for:

[0188] The enhanced feature vector of the keyword is obtained by multiplying the feature vector corresponding to the keyword with the word frequency of the keyword in the target court transcript.

[0189] In optional implementations, the problem recommendation device also includes:

[0190] The semantic vector extraction module is used to extract word semantic vectors for each representative word in the representative word set and to extract question semantic vectors for each representative question in the representative question set.

[0191] The similarity calculation module is used to calculate the cosine similarity between each extracted word semantic vector and each question semantic vector, and to construct a cosine similarity matrix based on the calculated cosine similarities; in the cosine similarity matrix, the first dimension represents the representative question and the second dimension represents the representative word;

[0192] The matrix fusion module is used to fuse the co-occurrence matrix with the cosine similarity matrix to obtain a first optimized co-occurrence matrix.

[0193] In optional implementations, the problem recommendation device also includes:

[0194] The numerical filtering module is used to filter the numerical values ​​of the elements at each position in the first optimized co-occurrence matrix using a preset threshold. Elements with values ​​lower than the preset threshold are reset to 0, while elements with values ​​higher than or equal to the preset threshold are retained, thereby obtaining the second optimized co-occurrence matrix.

[0195] The retrieval module 72 is specifically used for:

[0196] Multiple feature vectors are retrieved from the second optimized co-occurrence matrix based on the keyword set.

[0197] In optional implementations, the problem recommendation device also includes:

[0198] The fifth construction module is used to construct a problem transition matrix for the historical trial transcript dataset based on the co-occurrence of pairwise questions in each of the historical trial transcripts. In the problem transition matrix, both the first and second dimensions represent representative questions, and the representative questions corresponding to the elements on the diagonal of the problem transition matrix are the same as the representative questions corresponding to the elements in the columns. The value of the element in the i-th row and j-th column of the problem transition matrix represents the transition probability from the i-th representative question to the j-th representative question.

[0199] The graph construction module is used to construct a problem co-occurrence graph with representative problems as vertices and transition probabilities as edge weights based on the representative problem set and the problem transition matrix;

[0200] The normalization processing module is used to obtain the semantic weighted score of each question in the question recommendation list based on the result vector, and to perform normalization processing on each question in the question recommendation list to obtain the initial weight of each question and the initial weight vector; the initial weight vector includes the initial weight of each vertex in the question co-occurrence graph;

[0201] The weight vector optimization module is used to optimize the initial weight vector multiple times using the problem transition matrix, the initial weight vector, and the total number of vertices in the co-occurring problem graph to obtain the weight vector after multiple iterations.

[0202] The score calculation module is used to calculate the final score of each question in the question recommendation list by using the semantic weighted score of the question and the weight of the question in the weight vector after multiple iterations.

[0203] The question ranking readjustment module is used to readjust the ranking of questions in the question recommendation list based on the final scores of each question in the question recommendation list.

[0204] In an optional implementation, the fifth building module is specifically used for:

[0205] The historical court transcript dataset is traversed, and the co-occurring issues in each historical court transcript are counted. The counts are then entered into two symmetrical positions in the first transition matrix related to the two co-occurring issues, until the traversal is completed and a second transition matrix is ​​obtained. The first transition matrix is ​​an initialized issue transition matrix, in which the elements at each position are set to 0 during initialization. In the first transition matrix, both the first dimension and the second dimension represent representative issues.

[0206] The problem transition matrix is ​​obtained by performing numerical normalization on each row of the second transition matrix.

[0207] Based on the problem recommendation method and problem recommendation apparatus described in the foregoing embodiments, this application also provides a problem recommendation device, which includes: a memory and a processor;

[0208] The memory is used to store computer programs;

[0209] The processor is configured to run the computer program, which, when running, executes the steps of the problem recommendation method as described in any of the implementations in the method embodiments.

[0210] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and equipment embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0211] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A problem recommendation method, characterized in that, include: Based on the target court hearing transcript, obtain a set of keywords and the word frequency of each keyword in the target court hearing transcript; Multiple feature vectors are retrieved from a pre-constructed co-occurrence matrix based on the keyword set; each feature vector corresponds one-to-one with a keyword in the keyword set; the co-occurrence matrix is ​​a two-dimensional matrix, where the first dimension represents the question and the second dimension represents the keyword; wherein the first dimension of the co-occurrence matrix represents the question statement extracted from the historical court transcript set, and the second dimension of the co-occurrence matrix represents the text keywords extracted from the historical court transcript set; The enhanced feature vector of the keyword is obtained by multiplying the feature vector corresponding to the keyword with the word frequency of the keyword in the target court transcript. The multiple enhanced feature vectors are summed to obtain a result vector; different positions in the result vector correspond to different questions, and the values ​​at different positions are used to reflect the degree of matching between the corresponding questions and the target court transcript; A list of recommended questions for the target court hearing transcript is generated based on the values ​​in the result vector.

2. The method according to claim 1, characterized in that, Also includes: Obtain the historical court hearing transcript dataset; the historical court hearing transcript dataset includes multiple historical court hearing transcripts; Extract questions and keywords from each of the aforementioned historical court hearing transcripts, and construct a question set and keyword set corresponding to each of the aforementioned historical court hearing transcripts; The frequency of each keyword in the keyword set is counted in the corresponding historical court transcripts, and keyword-frequency pairs are constructed for each historical court transcript. Based on the constructed keyword sets and the semantic similarity between different keywords, a representative word set corresponding to the historical court transcript dataset is constructed. Furthermore, based on the constructed question sets and the semantic similarity between different questions, a representative question set corresponding to the historical court transcript dataset is constructed; Based on the semantic similarity between different keywords and the representative word set, the keyword-word frequency pair corresponding to each of the historical court hearing transcripts is corrected to obtain the corrected keyword-word frequency pair corresponding to each of the historical court hearing transcripts; The co-occurrence matrix is ​​constructed based on the representative word set, the representative question set, and the corrected keyword-word frequency pairs corresponding to each of the historical court transcripts.

3. The method according to claim 2, characterized in that, The co-occurrence matrix is ​​constructed based on the representative word set, the representative question set, and the corrected keyword-word frequency pairs corresponding to each of the historical court transcripts, including: Construct an initial matrix with each representative question in the representative question set as the first dimension index and each representative word in the representative word set as the second dimension index, and set the elements at each position to 0; The process iterates through each historical court trial transcript in the dataset. Based on the semantic similarity between different questions, the set of questions corresponding to the currently traversed historical court trial transcript is converted into a first representative question set. Based on the semantic similarity between different keywords, the set of keywords corresponding to the currently traversed historical court trial transcript is converted into a first representative word set. The first representative question set and the first representative word set are used to query the elements at the corresponding positions of the current initial matrix. The word frequency of the corrected keyword-word frequency pair corresponding to the currently traversed historical court trial transcript is superimposed on the value of the element at the corresponding position of the current initial matrix to complete one update of the initial matrix. After the historical court transcript dataset has been traversed, the initial matrix that has been updated for the last time will be used as the co-occurrence matrix.

4. The method according to claim 3, characterized in that, The first representative question set includes m representative questions, and the first representative word set includes n representative words; m and n are both positive integers. The step of querying the corresponding element of the current initial matrix using the first representative question set and the first representative word set, and then updating the initial matrix by adding the word frequency of the corrected keyword-word frequency pair corresponding to the currently traversed historical court transcript to the value of the corresponding element of the current initial matrix, includes: Traverse the first representative question set and the first representative word set. For the i-th representative question in the first representative question set and the j-th representative word in the first representative word set, query the target position in the current initial matrix and obtain the current value of the element at the target position. i takes values ​​from 1 to m, and j takes values ​​from 1 to n, where i and j are both integers; In the corrected keyword-frequency pairs corresponding to the currently traversed historical court transcripts, determine the frequency of the j-th representative word of the first representative word set; The frequency of the j-th representative word is added to the current value of the element at the target position to complete the value update of the element at the target position; After both the first representative question set and the first representative word set have been traversed, the initial matrix is ​​updated once.

5. The method according to claim 2, characterized in that, Based on the target court hearing transcript, the word frequency of each keyword in the keyword set in the target court hearing transcript is obtained, including: Obtain the number of times the first target keyword in the keyword set appears in the target court hearing transcript, and obtain the number of times other keywords in the keyword set appear in the target court hearing transcript; The sum of the number of times the first target keyword appeared in the target court transcript and the number of times the other keywords appeared in the target court transcript was used as the denominator, and the number of times the first target keyword appeared in the target court transcript was used as the numerator. The ratio obtained was taken as the word frequency of the first target keyword in the target court transcript. The method involves statistically analyzing the word frequency of each keyword in the keyword set within its corresponding historical court transcript, constructing keyword-word frequency pairs for each historical court transcript, including: Obtain the number of times the second target keyword appears in the target historical court hearing transcript and the number of times other keywords appear in the target court hearing transcript from the keyword set corresponding to the target historical court hearing transcript in the target historical court hearing transcript; The sum of the number of times the second target keyword appears in the target historical court hearing transcript and the number of times other keywords appear in the target historical court hearing transcript is used as the denominator, and the number of times the second target keyword appears in the target historical court hearing transcript is used as the numerator. The ratio obtained is used as the word frequency of the second target keyword in the target historical court hearing transcript, and a keyword-word frequency pair of the second target keyword is constructed.

6. The method according to claim 5, characterized in that, The process of correcting the keyword-frequency pairs corresponding to each of the historical court transcripts based on the semantic similarity between different keywords and the representative word set includes: If the second target keyword is one of the multiple representative words in the representative word set, then determine the similar keywords in the keyword set corresponding to the target historical court transcript that have a similar word mapping relationship with the second target keyword; The word frequency of the similar keywords in the target historical court transcript is added to the word frequency of the second target keyword in the target historical court transcript to obtain the first word frequency sum. The keyword-word frequency pair of the second target keyword is then corrected using the first word frequency sum. If the second target keyword is not one of the multiple representative words in the representative word set, then the target representative word corresponding to the second target keyword in the representative word set is determined based on the similar word mapping relationship; The word frequency of the second target keyword in the target historical court transcript, the word frequency of the target representative word in the target historical court transcript, and the word frequencies of other keywords with similar word mapping relationships to the target representative word in the target historical court transcript are added together to obtain a second word frequency sum. The keyword-word frequency pair of the target representative word corresponding to the target historical court transcript is corrected using the second word frequency sum.

7. The method according to claim 1, characterized in that, The process of generating the question recommendation list for the target court transcript based on the values ​​in the result vector includes: Based on the one-to-one correspondence between the elements at each position of the result vector and the questions, determine the questions corresponding to the elements with positive values ​​in the result vector; The identified questions are sorted in descending order according to the numerical value of their corresponding elements to generate a recommended list of questions for the target court hearing transcript.

8. The method according to claim 7, characterized in that, Also includes: Obtain a dataset of historical court hearing transcripts; the dataset of historical court hearing transcripts includes multiple historical court hearing transcripts. Extract questions from each of the aforementioned historical court hearing transcripts and construct a question set corresponding to each of the aforementioned historical court hearing transcripts; Based on the constructed question sets and the semantic similarity between different questions, a representative question set corresponding to the historical court transcript dataset is constructed. From the representative problem set, subsequences with support satisfying preset conditions and length of 2 are mined, and the mined subsequences are added to the mined sequence set as high-frequency binary problem sequences; In the high-frequency binary question sequence, the first question in the sequence has a higher question priority than the second question in the sequence. After generating the question recommendation list for the target court transcript, the method further includes: If there are two questions in the question recommendation list that both belong to the target high-frequency binary question sequence, and the question recommended first is the first question in the target high-frequency binary question sequence, and the question recommended second is the second question in the target high-frequency binary question sequence, then the order of the two questions in the question recommendation list that both belong to the target high-frequency binary question sequence will not be adjusted. If there are two questions in the question recommendation list that both belong to the target high-frequency binary question sequence, and the question recommended first is the second question in the target high-frequency binary question sequence, and the question recommended second is the first question in the target high-frequency binary question sequence, then in the question recommendation list, the question recommended first is moved after the question recommended second.

9. The method according to claim 2, characterized in that, After constructing the co-occurrence matrix, the method further includes: Extract word semantic vectors for each representative word in the representative word set, and extract question semantic vectors for each representative question in the representative question set; The cosine similarity between the semantic vector of each extracted word and the semantic vector of each question is calculated, and a cosine similarity matrix is ​​constructed based on the calculated cosine similarity. In the cosine similarity matrix, the first dimension represents the representative question, and the second dimension represents the representative word. The co-occurrence matrix is ​​fused with the cosine similarity matrix to obtain the first optimized co-occurrence matrix.

10. The method according to claim 9, characterized in that, After obtaining the first optimized co-occurrence matrix, the method further includes: The values ​​of the elements at each position in the first optimized co-occurrence matrix are filtered by a preset threshold. Elements with values ​​lower than the preset threshold are reset to 0, and elements with values ​​higher than or equal to the preset threshold are retained to obtain the second optimized co-occurrence matrix. The process of retrieving multiple feature vectors from a pre-constructed co-occurrence matrix based on the keyword set includes: Multiple feature vectors are retrieved from the second optimized co-occurrence matrix based on the keyword set.

11. The method according to claim 10, characterized in that, Also includes: Based on the co-occurrence of pairs of questions in each of the historical trial transcripts, a question transition matrix is ​​constructed for the historical trial transcript dataset. In the problem transition matrix, both the first and second dimensions represent representative problems. The representative problems corresponding to the rows where the elements on the diagonal of the problem transition matrix are located are the same as the representative problems corresponding to the columns where the elements are located. The value of the element in the i-th row and j-th column of the problem transition matrix represents the transition probability from the i-th representative problem to the j-th representative problem. Based on the representative problem set and the problem transition matrix, a problem co-occurrence graph is constructed with representative problems as vertices and transition probabilities as edge weights. The semantic weighted score of each question in the question recommendation list is obtained based on the result vector, and the initial weight of each question is obtained by normalization based on the semantic weighted score of each question in the question recommendation list, as well as the initial weight vector; the initial weight vector includes the initial weight of each vertex in the question co-occurrence graph; The initial weight vector is iteratively optimized using the problem transition matrix, the initial weight vector, and the total number of vertices in the problem co-occurrence graph to obtain a weight vector after multiple iterations. For each question in the question recommendation list, the final score of the question is obtained by using the semantic weighted score of the question and the weight of the question in the weight vector after multiple iterations; Based on the final scores of each question in the question recommendation list, the question ranking in the question recommendation list is readjusted.

12. The method according to claim 11, characterized in that, The process of constructing a question transition matrix for the historical trial transcript dataset based on the co-occurrence of pairwise questions in each of the historical trial transcripts includes: The historical court transcript dataset is traversed, and the co-occurring issues in each historical court transcript are counted. The counts are then entered into two symmetrical positions in the first transition matrix related to the two co-occurring issues, until the traversal is completed and a second transition matrix is ​​obtained. The first transition matrix is ​​an initialized issue transition matrix, in which the elements at each position are set to 0 during initialization. In the first transition matrix, both the first dimension and the second dimension represent representative issues. The problem transition matrix is ​​obtained by performing numerical normalization on each row of the second transition matrix.

13. A problem recommendation device, characterized in that, include: The first acquisition module is used to acquire a set of keywords and the word frequency of each keyword in the target court hearing transcript based on the target court hearing transcript. The retrieval module is used to retrieve multiple feature vectors from a pre-constructed co-occurrence matrix based on the keyword set; the multiple feature vectors correspond one-to-one with each keyword in the keyword set; the co-occurrence matrix is ​​a two-dimensional matrix, where the first dimension represents the question and the second dimension represents the keyword; wherein the first dimension of the co-occurrence matrix represents the question statement extracted from the historical court transcript set, and the second dimension of the co-occurrence matrix represents the text keywords extracted from the historical court transcript set; The enhancement module is used to multiply the feature vector corresponding to the keyword with the word frequency of the keyword in the target court transcript to obtain the enhanced feature vector of the keyword. The summation module is used to sum the multiple enhanced feature vectors to obtain a result vector; different positions in the result vector correspond to different questions, and the values ​​at different positions are used to reflect the degree of matching between the corresponding questions and the target court transcript; The recommendation module is used to generate a list of recommended questions for the target court transcript based on the values ​​in the result vector.

14. A problem recommendation device, characterized in that, include: Memory and processor; The memory is used to store computer programs; The processor is configured to run the computer program, which, when running, performs the steps of the problem recommendation method as described in any one of claims 1-12.

Images