Intelligent government affair hotline group appeal analysis and detection method

Abstract

The invention provides an intelligent government affair hotline group appeal analysis and detection method. The method comprises the following steps: regularly pulling daily work order data; performing vectorization coding on work order content in the work order data; aiming at the vector chemical order data, using a DBSCAN clustering algorithm to obtain a classification cluster; traversing the classification clusters, and expanding the classification clusters by using the DBSCAN clustering algorithm again by using address features and complaint features of the classification clusters and data which are not included in any cluster, so as to obtain more accurate complaint amount and complaint number information; obtaining a complaint summary of each classification cluster through a large model; and writing the data into the data. According to the method, group appeal detection considering semantic and spatial information can be realized, redundant and repeated work of original manual statistics is replaced by an intelligent mode, and similar appeals can be effectively identified and integrated, so that the processing efficiency is improved, the service quality is optimized, and a daily group event list is quickly and efficiently provided.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis, and specifically to an intelligent analysis and detection method for group demands in government hotlines. Background Art

[0002] According to the prior art, such as a method, device and storage medium for public opinion early warning analysis of citizens provided in patent document CN116805207A, it analyzes the public opinion events corresponding to each work order, and constructs a public opinion early warning model corresponding to different event types; when it monitors that the number of reports and complaints corresponding to each event type reaches a preset trigger threshold based on the public opinion early warning model, it triggers a public opinion early warning for the public opinion events corresponding to each event type.

[0003] However, in the above prior art, the public opinion early warning is based on different event types, that is, under the same event type, there will only be a corresponding public opinion level early warning. For example, in the case of property disputes, although the complaints are scattered in various communities in the city, if group demands are to be detected, there should be certain spatial constraints.

[0004] On the other hand, in the existing solutions, the processing method of the hotline mainly relies on the operator to manually summarize and count the group demands every day. However, this traditional manual counting method is not only time-consuming and laborious, but also prone to delays in information processing, affecting the timely response to urgent problems. Therefore, the existing methods either rely on a preset complaint database or do not consider spatial information, being too broad. And the above existing solutions are not conducive to law enforcement and problem-solving. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide an intelligent analysis and detection method for group demands in government hotlines to solve the problems raised in the above background art. The present invention can achieve group demand detection that takes into account both semantics and spatial information, replaces the redundant and repetitive work of the original manual statistics through an intelligent method, can effectively help identify and integrate similar demands, thereby improving the processing efficiency, optimizing the service quality, and achieving the rapid and efficient provision of a daily list of group events.

[0006] To achieve the above purpose, the present invention is implemented through the following technical solutions: an intelligent analysis and detection method for group demands in government hotlines, including two parts: key word extraction of work order information and detection of mass demands. The key word extraction of work order information includes the following steps:

[0007] S1.1. Regularly pull unanalyzed work order data;

[0008] S1.2. Batch analyze work order data;

[0009] S1.3. Analyze the model results;

[0010] S1.4, Result storage;

[0011] The mass appeal detection includes the following steps:

[0012] S2.1, Regularly pull the work order data of the current day, pull the work order data from the business database and the analysis data associated with the work order from ES respectively;

[0013] S2.2, Vectorize and encode the work order content in the work order data;

[0014] S2.3, Cluster the work order data;

[0015] S2.4, Summarize the group complaint events, and count the number of complaints and the number of complainants;

[0016] S2.5, Store the group complaint events in the database.

[0017] Furthermore, in the process of extracting keywords from the work order information, the work order content of the unanalyzed work order is obtained by pulling the unanalyzed work order data, and the detailed work order address is detailed; the process of batch analyzing the work order data includes: inputting the work order content, the detailed work order address; constructing the input of the large language model; batch calling the interface deployed by the glm4-9b-chat model to obtain the result.

[0018] Furthermore, the model result parsing process includes parsing the keywords in the output json format text string and storing the parsing result in Elasticsearch.

[0019] Furthermore, in step S2.2, the open-source text2vec-base-chinese model is used for vectorization processing. In this process, the content and address of each work order are first input into the model, and the corresponding vector representation is obtained through the output layer of the BERT model and used as the feature vector of the text for subsequent similarity calculation, classification or other tasks; through vectorization, the semantic information of the work order content and address is retained and converted into a numerical format that can be processed by the computer.

[0020] Furthermore, the work order data clustering process in step S2.3 specifically includes the following content:

[0021] S2.3.1, Calculate the cosine similarity between the work order content vectors;

[0022] S2.3.2, Calculate the cosine similarity between the work order address vectors;

[0023] S2.3.3. Calculate the similarity between work orders with weighting, where the semantic similarity accounts for 0.5 and the spatial similarity accounts for 0.5, to obtain the work order similarity, and the value range is 0 - 1. The larger the value, the more similar;

[0024] S2.3.4. Use 1 - work order similarity to convert and obtain the work order similarity distance. The similarity distance range is 0 - 1, and the smaller the value, the closer the distance;

[0025] S2.3.5. Run the dbscan algorithm, set eps to 0.05, and num_samples to 5;

[0026] S2.3.6. Obtain the classification cluster results of the vectorized clustering calculation;

[0027] S2.3.7. Clean the unlabeled data and filter out specific classification data;

[0028] S2.3.8. Keyword clustering. Keywords are divided into object - type keywords and complaint - reason - type keywords;

[0029] S2.3.9. Fusion of clustering results. Fusion of the clustering results in step 2.3.6 and the result list in 2.3.8.3 to obtain the basic list of group complaint events;

[0030] S2.3.10. Expand the data in the group complaint list.

[0031] Furthermore, the similarity between two vectors is judged by calculating the cosine value of the included angle between them. The value range is from - 1 to 1, where 1 means completely similar, 0 means irrelevant, and - 1 means completely opposite. The calculation formula is:

[0032]

[0033] Furthermore, the steps of the dbscan algorithm include: select an unvisited point from the dataset and judge whether it is a core point; if it is a core point, group all the density - reachable points in its neighborhood into the same cluster; for non - core points, if it is adjacent to a core point, it becomes part of the cluster, otherwise it is marked as noise.

[0034] Furthermore, in step S2.3.8, by dividing keywords, vectorizing keywords, then clustering object - type keywords based on a large model, traversing object - type keywords whose quantity meets the threshold, and finally using the large model to judge whether the several words most similar to it point to the same entity.

[0035] Further, traverse the obtained object class keyword classification clusters, cluster the reason class keywords corresponding to the data within the cluster, use the DBSCAN clustering method, set the eps parameter to 0.3 and the min_samples parameter to 5, and then traverse the clustering results to obtain the results of the keyword clustering step.

[0036] Further, in the expansion stage of the group complaint list data, first use the cluster feature to screen the unclassified data, which requires its semantic similarity with the cluster feature to be greater than 0.8, and then screen the cluster space feature to obtain the data to be screened that simultaneously meets the semantic and spatial similarity thresholds; then mix the data to be screened with the classified data, use DBSCAN clustering, and gradually increase the eps parameter and the min_samples parameter until the classified data is still in the same category after reclustering. At this time, other data in the same category as the classified data is regarded as the expanded data.

[0037] Advantages of the present invention:

[0038] 1. This intelligent analysis and detection method for group appeals in government affairs hotlines no longer needs to wait until a fixed time every day to give the list of group events on the same day. There is only a half-hour delay between the group detection data and the actual data, with high near-real-time performance.

[0039] 2. This intelligent analysis and detection method for group appeals in government affairs hotlines can replace the traditional manual statistics method and adopt a more intelligent and automatic statistics method. At the same time, this method can not only count the information of group complaint events on the same day, but can actually also be used to count the information of group events in the same week, month, and quarter, providing analysis results of group events in different dimensions.

[0040] 3. This intelligent analysis and detection method for group appeals in government affairs hotlines is more comprehensive, can more effectively and accurately judge group complaint events, display them in the group event list, and can obtain data from multiple channels. Description of the Drawings

[0041] Figure 1 Schematic diagram of key word extraction for work order information provided by the present invention;

[0042] Figure 2 Schematic diagram of a case of key word extraction for work orders of the present invention;

[0043] Figure 3 Schematic diagram of the total process of group complaint detection of the present invention;

[0044] Figure 4 Clustering of work order data of the present invention;

[0045] Figure 5 Schematic diagram of key word clustering of the present invention. Detailed implementation manners

[0046] To make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below in conjunction with specific implementation manners.

[0047] Please refer to Figures 1 to 5 , the present invention provides the following technical solutions: An intelligent analysis and detection method for group demands of government affairs hotlines, which mainly includes the following contents:

[0048] 1. Regularly pull daily work order data

[0049] 2. Vectorize and encode the work order content in the work order data

[0050] 3. Use the DBSCAN clustering algorithm for the vectorized work order data to obtain classification clusters

[0051] 4. Traverse the classification clusters, utilize the address features and complaint features of the classification clusters, and the data not included in any cluster, and use the DBSCAN clustering algorithm again to expand the classification clusters (the purpose is to obtain more accurate information on the number of complaints and the number of complainants)

[0052] 5. Obtain the complaint summary of each classification cluster through a large model

[0053] 6. Write the data into the data.

[0054] Through the above method, it is no longer necessary to wait until a fixed time every day to give the list of group events on the same day. There is only a half-hour delay between the group detection data and the actual data, with high near-real-time performance, and it can more effectively and accurately judge group complaint events, display them in the list of group events, and can obtain data from multiple channels.

[0055] In this embodiment, the following technical processes are also provided to explain the above method. This embodiment includes two parts: extraction of work order information keywords and detection of mass demands, where:

[0056] The schematic diagram of the work order keyword extraction process is as Figure 1 shown, and the schematic diagram of the work order keyword extraction case is as Figure 2 shown. The specific process of extracting work order information keywords is as follows:

[0057] Regularly pull the work order data that has not been analyzed

[0058] 1.1 Obtain the work order content of the unanalyzed work order, and the detailed address of the work order

[0059] 1.2. Batch analyze the work order data

[0060] 1.2.1. Input the work order content and the detailed address of the work order.

[0061] 1.2.2. Construct the input of the large language model

[0062] 1.2.3. Batch call the interface deployed by the glm4-9b-chat model (an open-source large language model) to obtain the results

[0063] 1.3. Model result parsing

[0064] Parse the keywords in the output JSON format text string

[0065] 1.4. Result storage

[0066] Store the parsed results in Elasticsearch (similar to a database, a way to store and retrieve data)

[0067] This embodiment also provides a method for detecting public demands. The schematic diagram of the public demand detection process is as Figure 3 shown. The schematic diagram of the work order data clustering process is as Figure 4 shown. The schematic diagram of the keyword clustering process is as Figure 5 shown. The specific detection process includes the following steps:

[0068] 2.1 Regularly pull the work order data of the current day

[0069] Pull the work order data from the business database and the analysis data (keyword data) associated with the work order from ES respectively

[0070] 2.2 Use the open-source text2vec-base-chinese model to vectorize the work order content and the work order address respectively

[0071] Vectorization is the process of converting text into a fixed-length vector, aiming to convert text data into a numerical format that can be processed by the model. In natural language processing (NLP), the goal of vectorization is to retain the semantic information in the text in the numerical space for subsequent calculation and analysis.

[0072] The BERT encoding model (Bidirectional Encoder Representations from Transformers) is based on the Transformer architecture and learns the context information of text through bidirectional training. BERT uses two pre-training tasks, Masked Language Model (MLM) and Next Sentence Prediction (NSP), enabling the model to understand the meaning of words in context. Different from traditional unidirectional language models (such as GPT), BERT captures richer semantic information by considering both the left and right contexts in the text.

[0073] text2vec-base-chinese is a BERT model for Chinese text. During pre-training, it uses a large amount of Chinese corpus and can handle the grammar and semantics of Chinese text well. Using this model, the work order content and work order address can be converted into vectors respectively. Specifically, first, the content and address of each work order are input into the model respectively, and the corresponding vector representations are obtained through the output layer of the BERT model. These vectors can be used as feature vectors of the text for subsequent similarity calculation, classification, or other tasks.

[0074] Through vectorization, the semantic information of the work order content and address is retained and converted into a numerical format that can be processed by a computer, providing a basis for further analysis and prediction. For example, the text: "Hello"

[0075] The vectorization is:

[0076] [-0.2139309197664261,0.3746878206729889,0.6576337814331055,.....,0.2119288593530655,-0.9286743402481079,-0.12038318812847137] (length 648)

[0077] 2.3 Work order data clustering

[0078] 2.3.1 Calculate the cosine similarity (semantic similarity) between work order content vectors

[0079] Cosine similarity is a method for measuring the similarity between two vectors and is commonly used in text analysis. It determines their similarity by calculating the cosine value of the angle between the two vectors. The value ranges from -1 to 1, where 1 indicates complete similarity, 0 indicates irrelevance, and -1 indicates complete opposition.

[0080] The calculation formula is:

[0081]

[0082] For example, vector A [1, 2, 3] and vector B [4, 5, 6]

[0083] Calculate the similarity:

[0084] A·B (dot product of vector A and B) = 1 * 4 + 2 * 5 + 3 * 6 = 32

[0085]

[0086] Similarity: 32 / 32.8 ≈ 0.975

[0087] The growth trends of vector A and b are similar, and they are very similar after calculating the overall similarity.

[0088] 2.3.2 Calculate the cosine similarity (spatial similarity) between work order address vectors

[0089] 2.3.3 Calculate the similarity between work orders with weights. The semantic similarity accounts for 0.5, and the spatial similarity accounts for 0.5, to obtain the work order similarity (value range 0 - 1, the larger the value, the more similar)

[0090] For example, if the semantic similarity is 0.9 and the spatial similarity is 0.8, then the final similarity between work orders is:

[0091] 0.9 * 0.5 + 0.8 * 0.5 = 0.85

[0092] 2.3.4 Use 1 - work order similarity to transform and obtain the work order similarity distance (0 - 1, the smaller the value, the closer the distance)

[0093] The following clustering algorithm is based on distance (the more similar, the closer the distance). The current metric is similarity (the more similar, the closer to 1). Therefore, a transformation is needed.

[0094] 2.3.5 Run the DBSCAN algorithm, set eps to 0.05 (within what distance is considered similar and belongs to the same cluster), and num_samples to 5 (how many similarities together are considered to form a cluster).

[0095] DBSCAN (Density - Based Spatial Clustering of Applications with Noise) is a density - based clustering algorithm. It forms clusters by density - connected points, does not require specifying the number of clusters, and is suitable for processing data with noise and irregular shapes.

[0096] The main parameters of DBSCAN are:

[0097] - eps (maximum distance of the neighborhood): Specify the neighborhood radius of a point. If the distance between two points is less than `eps`, they belong to the same cluster.

[0098] - min_samples (minimum number of neighborhood points for a core point): A point is a core point if the number of points in its neighborhood is greater than or equal to `min_samples`.

[0099] Algorithm steps:

[0100] 1. Select an unvisited point from the dataset and determine whether it is a core point.

[0101] 2. If it is a core point, group all the density-reachable points in its neighborhood into the same cluster.

[0102] 3. For non-core points, if it is adjacent to a core point, it becomes part of that cluster; otherwise, it is marked as noise.

[0103] By setting eps = 0.05 and min_samples = 5, DBSCAN will cluster points that are close to each other into a cluster, and only sufficiently dense points can form valid clusters, while sparse points are regarded as noise.

[0104] For example: There are 1 - 10 work orders. After using the dbscan clustering algorithm with parameters eps 0.15 and min_samples 3, the result is: [0, 0, 0, -1, -1, -1, 1, 1, 1, 1]

[0105] This means that work orders 1 - 3 are one cluster, 4 - 6 are unlabeled work orders, and 7 - 10 are one cluster. Here, the classified clusters can be regarded as the group complaint events we expect to obtain. Because the group complaint events we expect to find are those with the same complaint reason and the same complaint object, and we expect to aggregate and display them to represent the intensity of the public's demands. Therefore, when calculating the similarity between them pairwise, there will be a relatively high similarity value, while for other work orders with inconsistent complaint reasons or objects, the similarity value will be low. Thus, the dbscan clustering algorithm can be used to efficiently discover group complaint events. Additionally, the min_samples parameter can be considered as the minimum number of complaints represented by a group complaint event. The number of work orders within the cluster obtained by clustering must be greater than or equal to the min_samples value, and it acts as a threshold.

[0106] 2.3.6 Obtain the results of the classified clusters (basic list of group complaint events) from the vectorized clustering calculation;

[0107] 2.3.7 Clean the unlabeled data and filter out specific classified data.

[0108] Such as: education and training classification, consumer rights classification and other data

[0109] Because these specially classified data may have the same complaint objects, but different reasons for the complaints.

[0110] The final demands may be the same, which means that such incidents cannot be detected well based on semantics alone. In addition, the names of some companies are very small, with only one letter difference. It is not appropriate to use vectorization to calculate the distance, which will cause some companies to be mixed together, which is not conducive to law enforcement agencies to check and handle.

[0111] In the education and training category, such situations often occur, such as the same company is complained, but for different reasons, for example, one was defrauded because of applying for a social worker certificate, and the other was defrauded because of reporting academic upgrading. The two have different reasons, and the core demand is to refund. In such a complex situation, the similarity clustering method of the appealed work order is not applicable.

[0112] 2.3.8 Keyword Clustering

[0113] Keywords are divided into two categories: one is object-related keywords, such as community, company name, organization name, etc., and the other is complaint reason-related keywords.

[0114] 2.3.8.1 Keyword Division and Keyword Vectorization

[0115] 2.3.8.2 Object Class Keyword Clustering Based on Large Model

[0116] 2.3.8.2.1 Object class keywords whose traversal quantity meets the threshold (e.g., object class keywords that appear more than 3 times)

[0117] 2.3.8.2.2 Use the large model to determine whether the top 5 most similar words refer to the same entity.

[0118] Example:

[0119] Construct the following large model input:

[0120] Determine whether the company complained about in the following two incidents is the same:

[0121] Complaint A: In 2023, I signed up for a social worker certificate course at Juxue Education and paid

[0122] 2,700 yuan, and they promised that I would pass the exam, but I failed in the end and asked for a refund.

[0123] Complaint B: I previously signed up for an adult college course at Wuhan Juxue Education Technology Co., Ltd. and paid 4,500 yuan. They promised that there would be tutorials and a guaranteed pass, but no tutorials were sent. I demand an immediate refund.

[0124] Reason: xxx

[0125] Conclusion: 1 / 0"

[0126] The red part is dynamically changed, and the large model can better determine the object class.

[0127] Whether the keywords belong to the same entity can obtain more accurate clustering results.

[0128] Through the big model, it can be judged that Juxue Education and Wuhan Juxue Education Technology Co., Ltd. are actually the same entity, Juxue Education is the abbreviation.

[0129] The reason why vector similarity is not used here is that the effect is poor. For example, although the abbreviation and the full name have the same four words "Juxue Education", the full name of Wuhan, Technology Co., Ltd., etc., will tend to be more similar to the full names of other similar companies, resulting in multiple different companies being considered as the same one, resulting in incorrect statistics.

[0130] 2.3.8.3 Traverse the object class keyword classification clusters obtained in the previous step

[0131] Cluster the cause-type keywords corresponding to the data in the cluster, use the dbscan clustering method, set the eps parameter to 0.3, the min_samples parameter to 5, and then traverse the clustering results to get the results of the keyword clustering step. In the case of the same complaint object, the reasons can be appropriately relaxed if they are similar.

[0132] 2.3.9 Fusion of clustering results

[0133] Combine the clustering results of step 2.3.6 with the result list of step 2.3.8.3 to obtain a basic group complaint event list.

[0134] 2.3.10 Group Complaint List Data Expansion

[0135] After obtaining the group lawsuit list, in order to obtain statistics on the number of complaint cases and the number of complainants that are closer to the actual data, it is necessary to target the data characteristics of each classification cluster and search the unclassified data for data that meets the characteristics of this classification cluster.

[0136] This embodiment first uses the classification cluster features (the data within the cluster is vectorized and the matrix is used to calculate the mean to obtain the cluster vector) to filter the unclassified data, which requires that the semantic similarity between the unclassified data and the classification cluster features is greater than 0.8. The classification cluster spatial features are filtered in the same way to obtain the data to be filtered that meets both the semantic and spatial similarity thresholds.

[0137] Mix the data to be screened with the classified data and use DBSCAN clustering. Here, gradually increase the eps parameter and the min_samples parameter. Until after the classified data is reclustered and remains in the same category, at this time, other data in the same category as the classified data is regarded as extended data.

[0138] In this embodiment, when clustering at one time, since the threshold is only a numerical value and cannot quantify whether the two complaints are indeed about the same thing, a relatively high similarity threshold is usually used for the initial vectorization clustering so that the obtained results are reliable.

[0139] For example, the complaint about high property fees in Community A and the complaint about high property fees in Community B actually have a high semantic similarity, resulting in a comprehensive similarity that may reach above 0.85.

[0140] The reclustering after obtaining the basic clustering result is to follow the clues. It is to find data that is consistent with the existing classification clusters from the existing data, and the restrictions can be appropriately relaxed to efficiently expand the data in the current classification cluster.

[0141] 2.4 Summary of group complaint events, statistics of the number of complaints and the number of complainants

[0142] Obtain the summary through two keywords of each group complaint event work order. Select the one with the most occurrences of the object class keyword,

[0143] and the one with the most occurrences of the complaint reason class keyword, and combine them to obtain the summary of the group complaint event.

[0144] For example, the one with the most occurrences in the object class is xx community, and the one with the most occurrences in the reason is high property fees

[0145] Then the final summary of this group complaint event is that the property fees in xx community are too high

[0146] 2.5 Store the group complaint events in the database.

[0147] The above shows and describes the basic principles, main features and advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic features of the present invention.

[0148] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An intelligent analysis and detection method for group demands in government service hotlines, characterized in that, It includes two parts: key word extraction of work order information and detection of public demands. The key word extraction of work order information includes the following steps: S1.

1. Regularly pull unanalyzed work order data; S1.

2. Batch analyze work order data; S1.

3. Analyze model results; S1.

4. Store results; The detection of public demands includes the following steps: S2.

1. Regularly pull the work order data of the current day, pull the work order data from the business database and the analysis data associated with the work order from ES respectively; S2.

2. Perform vectorization encoding on the work order content in the work order data; S2.

3. Cluster work order data; S2.

4. Summarize group complaint events, and count the number of complaints and the number of complainants; S2.

5. Store group complaint events in the database.

2. The intelligent analysis and detection method for mass demands of government affairs hotlines according to claim 1, characterized in that: During the process of key word extraction of work order information, the work order content of unanalyzed work orders is obtained by pulling unanalyzed work order data, and the detailed address of the work order is detailed; The process of batch analyzing work order data includes: inputting work order content and detailed work order address; constructing the input of the large language model; batch calling the interface deployed by the glm4-9b-chat model to obtain results.

3. The intelligent analysis and detection method for group demands of government affairs hotline according to claim 2, characterized in that: The process of analyzing model results includes parsing the key words in the output json format text string according to the output, and storing the parsing results in Elasticsearch.

4. The intelligent analysis and detection method for group demands of government affairs hotline according to claim 1, characterized in that: In step S2.2, an open-source text2vec-base-chinese model is used for vectorization processing. In this process, the content and address of each work order are first input into the model, and the corresponding vector representation is obtained through the output layer of the BERT model and used as the feature vector of the text for subsequent similarity calculation, classification or other tasks; through vectorization, the semantic information of the work order content and address is retained and converted into a numerical format that can be processed by a computer.

5. The intelligent analysis and detection method for mass demands of government affairs hotlines according to claim 1, characterized in that, The specific content of the work order data clustering process in step S2.3 is as follows: S2.3.

1. Calculate the cosine similarity between work order content vectors; S2.3.

2. Calculate the cosine similarity between work order address vectors; S2.3.

3. Calculate the similarity between work orders by weighted calculation, with the semantic similarity accounting for 0.5 and the spatial similarity accounting for 0.5, to obtain the work order similarity, and the value range is 0-1, and the larger the value, the more similar; S2.3.

4. Use 1 - work order similarity to convert to obtain the work order similarity distance, and the similarity distance range is 0-1, and the smaller the value, the closer the distance; S2.3.

5. Run the dbscan algorithm, set eps to 0.05, and num_samples to 5; S2.3.

6. Obtain the classification cluster results of vectorized clustering calculation; S2.3.

7. Clean the unlabeled data and filter out specific classification data; S2.3.

8. Cluster key words, and the key words are divided into object class key words and complaint reason class key words; S2.3.

9. Fusion of clustering results, fuse the clustering results in step 2.3.6 with the result list in 2.3.8.3 to obtain a basic list of group complaint events; S2.3.

10. Expand the data in the group complaint list.

6. The intelligent analysis and detection method for group demands of government affairs hotlines according to claim 5, characterized in that: The similarity between two vectors is judged by calculating the cosine value of the angle between them. The value ranges from -1 to 1, where 1 indicates complete similarity, 0 indicates irrelevance, and -1 indicates complete opposition. The calculation formula is as follows:

7. An intelligent analysis and detection method for mass demands of government affairs hotlines according to claim 5, characterized in that, The steps of the DBSCAN algorithm include: selecting an unvisited point from the dataset and judging whether it is a core point; if it is a core point, classifying all the density-reachable points in its neighborhood into the same cluster; for non-core points, if it is adjacent to a certain core point, it becomes part of the cluster, otherwise it is marked as noise.

8. An intelligent analysis and detection method for group demands of government affairs hotlines according to claim 5, characterized in that: In step S2.3.8, by dividing keywords, vectorizing the keywords, then clustering the object class keywords based on the large model, traversing the object class keywords whose quantity meets the threshold, and finally using the large model to judge whether several words most similar to it point to the same entity.

9. The intelligent analysis and detection method for group demands of government affairs hotlines according to claim 8, characterized in that: Traverse the obtained object class keyword classification clusters and the reason class keywords corresponding to the data within the clustering clusters. Use the DBSCAN clustering method, set the eps parameter to 0.3 and the min_samples parameter to 5, and then traverse the clustering results to obtain the results of the keyword clustering step.

10. The intelligent analysis and detection method for group demands of government affairs hotline according to claim 5, characterized in that: In the stage of expanding the group complaint list data, first use the classification cluster features to screen the unclassified data, requiring that its semantic similarity with the classification cluster features is greater than 0.8, and then screen the classification cluster space features to obtain the data to be screened that simultaneously meets the semantic and spatial similarity thresholds; then mix the data to be screened with the classified data and use DBSCAN clustering, gradually increasing the eps parameter and the min_samples parameter until after the classified data is reclustered, it is still the same category. At this time, other data of the same category as the classified data is regarded as the expanded data.

Images