Hot line brief report generation method and device, electronic equipment and storage medium

Abstract

The invention provides a hotline brief report generation method and device, electronic equipment and a storage medium, and the method comprises the steps: converting the image content contained in each hotline work order into a descriptive text, fusing the descriptive text with the text content contained in the corresponding hotline work order, and obtaining a complete work order text of each hotline work order; based on the semantic information of each complete work order text, clustering all hotline work orders to obtain a plurality of work order clusters; performing intra-cluster correlation analysis and inter-cluster correlation analysis on the hotline work orders in each work order cluster to obtain a correlation analysis result of each hotline work order; and based on the correlation analysis result of each hotline work order, generating an intelligent brief report of each hotline work order. According to the method, an end-to-end automatic process from an original work order to a decision support brief report is constructed, the defects of incomplete information utilization, single analysis dimension, rigid report form and the like in a traditional method are overcome, and the processing efficiency and insight ability of a government affair hotline or a customer service center are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method, apparatus, electronic device, and storage medium for generating hotline briefings. Background Technology

[0002] The government hotline platform generates massive amounts of unstructured business transaction data every day, but traditional processing methods are inefficient and make it difficult to conduct effective data mining and analysis.

[0003] Existing technologies include a briefing generation method based on a large language model, which directly processes the text data in hotline work orders and generates corresponding briefings. However, because this method only uses text information as input during analysis, the comprehensiveness and accuracy of the generated briefings are limited. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for generating hotline briefings, in order to overcome the deficiencies existing in the prior art.

[0005] This invention provides a method for generating hotline briefings, comprising the following steps: The image content contained in each hotline work order is converted into descriptive text, and the descriptive text is merged with the text content contained in the corresponding hotline work order to obtain the complete work order text of each hotline work order. Based on the semantic information of each complete work order text, all hotline work orders are clustered to obtain multiple work order clusters; Perform intra-cluster and inter-cluster correlation analysis on the hotline work orders in each work order cluster to obtain the correlation analysis results of each hotline work order. Based on the correlation analysis results of each hotline work order, an intelligent briefing is generated for each hotline work order.

[0006] According to a hotline briefing generation method provided by the present invention, the method involves clustering all hotline work orders based on the semantic information of each complete work order text to obtain multiple work order clusters, including: Based on the semantic information of each complete work order text, the structured information of the corresponding hotline work orders is extracted to obtain the structured information of each hotline work order; By integrating the complete work order texts and structured information of each hotline work order, the fusion characteristics of each hotline work order are obtained. Based on the fusion characteristics of each hotline work order, all hotline work orders are clustered to obtain multiple work order clusters.

[0007] According to a hotline briefing generation method provided by the present invention, the step of extracting structured information from the corresponding hotline work orders based on the semantic information of each complete work order text to obtain the structured information of each hotline work order includes: Using a large language model, structured information is extracted from the corresponding hotline work orders based on the semantic information of each complete work order text. The extracted structured information is then classified and labeled.

[0008] According to a hotline briefing generation method provided by the present invention, the method of fusing the complete work order text and structured information of each hotline work order to obtain the fused features of each hotline work order includes: Semantic encoding is performed on the complete work order text of each hotline work order to obtain the textual semantic features of each hotline work order; The structured information of each hotline work order is uniquely encoded to obtain the structured features of each hotline work order; We perform weighted fusion of the textual semantic features and structured features of each hotline work order to obtain the fused features of each hotline work order.

[0009] According to a hotline briefing generation method provided by the present invention, the method involves clustering all hotline work orders based on the fusion characteristics of each hotline work order to obtain multiple work order clusters, including: Treat each hotline ticket as an independent initial cluster; Based on the fusion features of each hotline work order, the similarity between each initial cluster is calculated. The two initial clusters with the highest similarity are iteratively merged into a new cluster until the final number of clusters meets the preset conditions, resulting in multiple work order clusters.

[0010] According to a hotline briefing generation method provided by the present invention, the step of performing intra-cluster correlation analysis and inter-cluster correlation analysis on hotline work orders in each work order cluster to obtain the correlation analysis results of each hotline work order includes: Using a large language model, the complete work order text of each hotline work order in each work order cluster is parsed to generate cluster class tags and cluster description text for each work order cluster; Using the large language model, intra-cluster association analysis is performed on the complete work order texts of all hotline work orders within the same work order cluster to extract the distribution of business elements and hotspots within each work order cluster. Using the large language model, cluster association analysis is performed based on the cluster category tags and cluster description text of each work order cluster to identify potential association relationships between different work order clusters; Based on the potential relationships between different work order clusters, the distribution of business elements and hotspots within each work order cluster, the correlation analysis results for each hotline work order are determined.

[0011] According to a hotline briefing generation method provided by the present invention, the step of generating intelligent briefings for each hotline work order based on the correlation analysis results of each hotline work order includes: Based on the correlation analysis results of each hotline work order, a briefing outline for each hotline work order and a description of the content to be written for each chapter in the briefing outline are generated. Based on the descriptions of the content to be written in each chapter of the briefing outline of each hotline work order, the briefing materials required for each chapter are retrieved and generated from the corresponding hotline work orders and the results of the correlation analysis. Based on the briefing materials of each hotline work order, the content of each chapter is written and integrated using a large language model to generate intelligent briefings for each hotline work order.

[0012] According to a hotline briefing generation method provided by the present invention, the step of retrieving and generating the briefing materials required for each chapter from the corresponding hotline work order and the correlation analysis results based on the description of the content to be written in each chapter of the briefing outline of each hotline work order includes: A hybrid retrieval strategy combining embedded vector similarity calculation and text semantic matching is adopted. Based on the description of the content to be written in each chapter of the hotline work order's briefing outline, targeted material retrieval is performed in the hotline work order database to generate the briefing materials required for each chapter. If any chapter type is a brief overview, a single-round search is performed; if any chapter type is an in-depth analysis, a large language model is used to autonomously initiate multiple rounds of iterative search until the semantic similarity between the new briefing material obtained in the current iteration and all the briefing materials obtained is higher than a threshold, or the number of iterations reaches a preset number, then all the briefing materials obtained are used as the briefing materials required for the corresponding chapter; each chapter type is determined based on the description of the content to be written in the corresponding chapter.

[0013] The present invention also provides a hotline briefing generation device, comprising the following modules: The fusion unit is used to convert the image content contained in each hotline work order into descriptive text, and to fuse the descriptive text with the text content contained in the corresponding hotline work order to obtain the complete work order text of each hotline work order. Clustering units are used to cluster all hotline work orders based on the semantic information of each complete work order text, resulting in multiple work order clusters; The analysis unit is used to perform intra-cluster correlation analysis and inter-cluster correlation analysis on the hotline work orders in each work order cluster, and obtain the correlation analysis results of each hotline work order. The generation unit is used to generate intelligent briefings for each hotline work order based on the correlation analysis results.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the hotline briefing generation method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the hotline briefing generation method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the hotline briefing generation method as described above.

[0017] The hotline briefing generation method, apparatus, electronic device, and storage medium provided by this invention first achieve comprehensive information extraction from multimodal work order data through image textification and intelligent fusion; then, based on text semantics, the work orders are accurately clustered, achieving automated classification of massive requests; next, through deep correlation analysis within and between clusters of the clustering results, potential problem patterns and root causes are uncovered; finally, based on the analysis results, multi-dimensional intelligent briefings containing images and text are automatically generated. This invention constructs an end-to-end automated process from original work orders to decision support briefings, not only overcoming the shortcomings of traditional methods such as incomplete information utilization, single analysis dimensions, and rigid report formats, but also greatly improving the processing efficiency and insight capabilities of government hotlines or customer service centers. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the hotline briefing generation method provided by the present invention.

[0020] Figure 2 This is a flowchart illustrating the process of extracting structured information from hotline work orders provided by the present invention.

[0021] Figure 3 This is a statistical diagram illustrating the distribution of business elements provided by the present invention.

[0022] Figure 4 This is a schematic diagram of hotspot distribution statistics provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the hotline briefing generation device provided by the present invention.

[0024] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] With the acceleration of the digitalization of government services, hotline platforms have become an important bridge for communication between the government and the public. These platforms receive and generate massive amounts of work order data daily. This data contains extremely high value, but it also exhibits multimodal and unstructured characteristics. Traditional processing methods mainly rely on manual recording and classification, which is not only inefficient and costly, but also prone to bias due to subjective factors.

[0027] In recent years, the rise of Large Language Models (LLM) has provided a new technical path for efficiently processing unstructured text. However, most existing applications are limited to single tasks such as work order classification and sentiment judgment, and lack a full-process intelligent solution that covers multi-dimensional parsing of work orders to intelligent briefing generation, can integrate multi-source data, and meet the needs of management decision-making.

[0028] To address the above issues, the industry has proposed several technical solutions, as follows: One approach is based on keyword matching and rule templates. This approach uses a pre-set keyword library to perform string matching on work order texts, achieving coarse-grained classification, and then uses fixed text templates to generate briefings. However, the classification accuracy of this approach heavily relies on the completeness of the keyword library and cannot understand complex contexts and synonyms. Furthermore, the briefings generated based on fixed templates are rigid and monotonous, failing to extract deep semantic information and thus failing to meet the needs of high-quality decision support.

[0029] The second approach is based on traditional natural language processing and statistical models. This approach uses feature engineering methods such as TF-IDF combined with machine learning classifiers to classify work orders and employs extractive summarization techniques such as TextRank to extract important sentences from the original text and combine them into a briefing. However, this approach requires a large amount of labeled data for model training, involves complex feature engineering, and has limited ability to understand deep semantics. Furthermore, its extractive summarization technique is merely a mechanical splicing of sentences, resulting in a briefing that lacks coherence, summarization, and comprehensive insight.

[0030] Thirdly, there's the large language model-based approach, which directly leverages the text generation capabilities of large language models to process collected hotline work order texts and generate briefings. However, this approach typically only processes text information, underutilizing multimodal data such as images in the work orders, leading to information gaps. During the briefing generation stage, the content is mostly single text paragraphs, lacking professional and readable presentation formats such as statistical charts and trend analyses, and exhibiting severely insufficient in-depth analytical capabilities.

[0031] In response, this invention provides a method for generating hotline briefings, aiming to solve the problems of insufficient utilization of hotline work order data, single analysis dimensions, rigid briefing formats, and insufficient insight in related technologies. It enables fully automated in-depth analysis of hotline work orders containing multimodal information such as text and images, and generates diversified intelligent briefings that integrate textual insights, statistical charts, and trend analysis, thereby greatly improving the efficiency of government service analysis and the quality of decision support.

[0032] in, Figure 1 This is a flowchart illustrating the hotline briefing generation method provided by the present invention, as shown below. Figure 1 As shown, the method includes steps 110, 120, 130 and 140.

[0033] Step 110: Convert the image content contained in each hotline work order into descriptive text, and merge the descriptive text with the text content contained in the corresponding hotline work order to obtain the complete work order text of each hotline work order.

[0034] Understandably, hotline work orders serve as a medium for recording user requests. They not only include text transcribed from the user's voice or directly entered text, such as the user's description of the problem and customer service records, but may also include images uploaded by the user. These images can be, but are not limited to, photos taken by the user at the scene, screenshots from their mobile phone or computer, etc., to more intuitively reflect the problem.

[0035] Specifically, to enable computers to uniformly process and understand this multimodal information, this step first requires textualizing the image content. One possible implementation involves using a pre-trained image understanding model to analyze each image in the hotline work order and output descriptive text about the image. For example, a large image understanding model can be used to perform deep image analysis; if a user uploads a photo, the model can generate descriptive text such as "The image shows traces of liquid leakage near the compressor of a refrigerator."

[0036] After obtaining the descriptive text of the image, it needs to be merged with the original text content of the corresponding hotline work order. This merging is not a simple text splicing, but rather aims to form a complete work order text that is semantically coherent and logically sound. In one example, a Large Language Model (LLM) can be used to perform this intelligent merging task. The LLM understands the context and intelligently inserts the image description text into the appropriate position in the original text of the hotline work order, or paraphrases and integrates it, thereby generating a high-quality, comprehensive, and complete work order text. In other words, this complete work order text not only integrates information from both image and text modalities, but also gains a structured understanding through the processing of the LLM, thus providing a unified, high-quality, and information-rich data foundation for subsequent automatic work order classification, semantic analysis, and briefing generation.

[0037] Optionally, before converting the image content contained in each hotline work order into descriptive text, the text content contained in each hotline work order can be cleaned, such as removing useless symbols, correcting typos, and standardizing entity information such as dates, locations, and names.

[0038] Step 120: Based on the semantic information of each complete work order text, cluster all hotline work orders to obtain multiple work order clusters.

[0039] Specifically, semantic information refers to the deeper meaning, intent, and theme contained in the complete work order text, rather than just the surface keywords. In one implementation, before clustering all hotline work orders, text embedding technology can be used, employing a finely tuned embedding model as an encoder to map each complete work order text into a high-dimensional text semantic vector that accurately captures the core semantics of the text.

[0040] After obtaining the text semantic vector of each complete work order text, clustering is performed based on these text semantic vectors to obtain multiple work order clusters. In a preferred implementation, a hierarchical clustering algorithm can be used. The advantage of this algorithm is that it does not require pre-specifying the number of clusters (K value). It iteratively calculates and merges the hotline work orders or work order clusters with the highest similarity until a certain stopping condition is met, such as the distance between clusters being greater than a threshold or reaching a preset range of the number of clusters, thereby adaptively discovering naturally formed cluster structures in the data.

[0041] After clustering, all hotline work orders were divided into multiple work order clusters. Each work order cluster represents a set of semantically highly related requests. For example, one cluster may consist entirely of complaints about "chaotic management of parking spaces in the community," while another cluster may focus on issues such as "traffic light malfunctions on a certain road section."

[0042] Step 130: Perform intra-cluster correlation analysis and inter-cluster correlation analysis on the hotline work orders in each work order cluster to obtain the correlation analysis results of each hotline work order.

[0043] Specifically, intra-cluster correlation analysis refers to analysis conducted within the same work order cluster. It is used to discover the distribution patterns, hotspot composition, development trends, and key characteristics within a specific category of issues. Intra-cluster correlation analysis can include: frequency and hotspot analysis, which involves counting the number of work orders within each work order cluster and their percentage of the total number of work orders, thereby identifying the most frequent and popular public demands—the TOP hotspot issues; and time-series trend analysis, which involves counting the number of work orders in each work order cluster over time dimensions such as day, week, or month, generating time trend charts to reveal the development trend of a certain type of issue, such as "sharp increase," "stable," or "seasonal fluctuations."

[0044] Inter-cluster correlation analysis refers to finding potential connections between different work order clusters. It is used to discover complex problem patterns that are not easily noticed, such as causal relationships, concurrent relationships, or evolutionary relationships between different request categories. This can be achieved by leveraging the powerful reasoning capabilities of large language models to analyze the topics of different clusters. For example, the model might find that a work order cluster about "home broadband connection failure" is highly correlated with another work order cluster about "firmware issues with a specific router model (e.g., router model XYZ123)" in terms of time and user groups, thus revealing that the root cause of the problem may lie in a defect of that router model.

[0045] Ultimately, the outputs of intra-cluster and inter-cluster correlation analyses constitute the correlation analysis results for each hotline work order. These results can take the form of structured data, trend chart code, or analytical descriptive text generated by a large model. For example, the correlation analysis result could be a JSON object containing an analysis of the "road flooding" work order cluster, along with code to generate a chart comparing the number of issues with rainfall.

[0046] Step 140: Based on the correlation analysis results of each hotline work order, generate an intelligent briefing for each hotline work order.

[0047] Here, a smart briefing is a multi-element, structured report that goes beyond traditional plain text summaries. It can include concise textual analysis, key data statistics, and visual statistical charts (such as pie charts and trend charts).

[0048] In a specific implementation, the intelligent briefing generation process may include: First, based on the correlation analysis results output in step 130 (such as total number of work orders, list of hot issues, time-series trend summary, root cause analysis conclusions, etc.), a structured briefing outline is automatically generated using a large language model. This outline can be in JSON format, with predefined chapter titles, chapter types, and specific requirements for the content to be written in each chapter. Then, based on the specific requirements for the content to be written in each chapter of the briefing outline, relevant materials are retrieved from the original work order database and correlation analysis results, and the large language model is invoked to complete the text writing for that chapter. Simultaneously, the chart code generated in step 130 is compiled into images and embedded into the corresponding positions in the report. Finally, all chapter content is integrated to form a complete intelligent briefing. The original work order database can be understood as a central information repository storing all raw, unprocessed work order data accessed from the hotline service platform. It contains all the initial information for each work order and is the starting point and basic data source for the entire analysis process.

[0049] The hotline briefing generation method provided in this embodiment first achieves comprehensive information extraction from multimodal work order data through image textification and intelligent fusion; then, it performs precise clustering of work orders based on text semantics, realizing automated classification of massive requests; next, it conducts deep correlation analysis within and between clusters on the clustering results to uncover potential problem patterns and root causes; finally, it automatically generates multi-dimensional intelligent briefings containing images and text based on the analysis results. This embodiment's method constructs an end-to-end automated process from original work orders to decision support briefings, overcoming the shortcomings of traditional methods such as incomplete information utilization, single analysis dimensions, and rigid report formats, and greatly improving the processing efficiency and insight capabilities of government hotlines or customer service centers.

[0050] Based on the above embodiments, and based on the semantic information of each complete work order text, all hotline work orders are clustered to obtain multiple work order clusters, including: Based on the semantic information of each complete work order text, the structured information of the corresponding hotline work orders is extracted to obtain the structured information of each hotline work order; By integrating the complete work order texts and structured information of each hotline work order, the fusion characteristics of each hotline work order are obtained. Based on the fusion characteristics of each hotline work order, all hotline work orders are clustered to obtain multiple work order clusters.

[0051] Specifically, structured information refers to the core elements of a work order that can be represented by predefined labels or a fixed format. As an optional implementation, a large language model can be used to receive the "complete work order text" as input and automatically identify and extract various key information elements from the text based on preset prompt text. The preset prompt text can be understood as natural language text describing the extraction task. It clearly defines the task to be completed, the key information fields to be extracted, and the expected output format for the model, thereby constraining the model's generation process and ensuring that the model extracts information according to a unified and standardized standard, rather than engaging in aimless free-form extraction.

[0052] For example, the preset prompt text could be: "You are a professional work order analysis assistant. Please extract three categories of information from the following 'complete work order text': 'core request,' 'departments involved,' and 'urgency level.' Please categorize the 'urgency level' as 'urgent,' 'moderate,' or 'low.' Please return the results in JSON format. Work order text: {Insert complete work order text here}." With such a preset prompt text, the model can be guided to generate structured information with a consistent format and accurate content for any work order text.

[0053] For example, given a hotline call about "noise pollution from construction work in the neighborhood at night," the large language model can extract the following structured information and classify and label it: "Background Information": For example, "The resident lives in XX community"; "Core Demand": For example, "Complain about nighttime construction noise and demand that it be stopped"; "Secondary Demand": For example, "Inquire whether the relevant construction permits are compliant"; "Relevant Departments": For example, "Environmental Protection Bureau", "Urban Management Bureau"; "Urgency Level": For example, "High", "Moderate", "Low".

[0054] Of course, the field types of the structured information mentioned above are only examples. In other implementations, they may include, but are not limited to, question types, key entities (such as names of people, places, and organizations), and sentiment polarity (such as "positive", "negative", and "neutral"). This embodiment does not make specific limitations on these.

[0055] in, Figure 2 This is a flowchart illustrating the process of extracting structured information from hotline work orders provided by the present invention, as shown below. Figure 2 As shown, firstly, multimodal separation is performed on hotline work orders containing both image and text content. This involves separating the image and text content from the hotline work order and then using a large language model to convert the image content into descriptive text. The separated text and image content retain their correspondence with the corresponding hotline work orders, ensuring that the text and image content of the same hotline work order can be merged after processing. The large language model used here can be the Spark Large Model. Figure 2 As shown, the text content in the hotline work order is "Request: There is a typo in the notice board under Building No. 2 of Xihua Community. Hope it can be resolved as soon as possible", and the descriptive text converted from the image content in the hotline work order is "The user complains that the word '扔' (should be '扔' instead of '仍') in the text 'Please don't throw cigarette butts into the pool and please throw garbage into the waste basket' mentioned in the picture is a typo, and it is circled with a red frame."

[0056] Then, the text content in the hotline work order is corrected. The "Building No. 2喽" is corrected to "Building No. 2", and the corrected text content is "Request: There is a typo in the notice board under Building No. 2 of Xihua Community. Hope it can be resolved as soon as possible". Then, the descriptive text generated in the previous step is integrated with the corrected text content to obtain a complete work order text containing multi-modal information: "The user's request is that there is a typo in the notice board under Building No. 2 of Xihua Community. The word '仍' should be changed to '扔'. Hope it can be resolved as soon as possible."

[0057] Next, based on this complete work order text, a large language model is used to extract structured information and classify and mark the extracted information. The obtained structured information is: "Core request: Correct the typo in the notice board of Building No. 2 in Xihua Community. Department involved: Xihua Community Property. Urgency level: Urgent."

[0058] Then, the complete work order text and structured information of each hotline work order are integrated to obtain the integrated features of each hotline work order. Specifically, in order to achieve accurate similarity calculation and classification of hotline work orders, each work order needs to be converted into a numerical vector that can comprehensively reflect its content, that is, the integrated feature. This integrated feature is not single-dimensional information, but a comprehensive representation that simultaneously contains the deep semantic information and key structured attributes of the hotline work order.

[0059] In a specific implementation, the process of obtaining this integrated feature may include: Semantic encoding is performed on the complete work order text of each hotline work order to obtain the text semantic features of each hotline work order. Specifically, a fine-tuned Embedding model can be used as the encoder. This encoder takes the "complete work order text" as the input and outputs a high-dimensional text semantic feature vector. This vector can deeply capture the overall semantic connotation of the work order text. Compared with the traditional keyword method, it can better understand synonyms, complex contexts, and deep appeal intentions.

[0060] In addition, one-hot encoding is performed on the structured information of each hotline work order to obtain the structured features of each hotline work order. For the discrete structured information extracted in the previous step, such as "problem type", "urgency level", etc., one-hot encoding (One-hot Encoding) and other methods can be used to convert it into a numerical structured feature vector. This encoding method can convert non-numerical categorical attributes into a format that can be processed by machine learning models.

[0061] The textual semantic features and structured features of each hotline work order are weighted and fused to obtain the fused features of each hotline work order. Specifically, the obtained textual semantic feature vector and structured feature vector can be concatenated or weighted and summed to form the final fused feature vector. The advantage of using a weighted approach is that the importance of semantic and structured information in the final features can be adjusted according to business needs. For example, in scenarios that focus on discovering new problems, the weight of textual semantic features can be increased, while in scenarios that require strict processing according to established classification rules, the weight of structured features can be increased.

[0062] Finally, based on the fusion characteristics of each hotline ticket, all hotline tickets are clustered to automatically aggregate hotline tickets with similar content and attributes, forming different ticket clusters. Each ticket cluster can be understood as a specific set of hot issues or a category of hot issues.

[0063] Specifically, the above clustering can be implemented using a hierarchical clustering algorithm. For example, each hotline work order is first treated as an independent initial cluster. Then, based on the fusion features of the hotline work orders, the similarity between each work order cluster is calculated. Iteratively, the two work order clusters with the highest similarity are merged into a new work order cluster until the final number of clusters meets a preset condition, at which point the iteration stops. The preset condition can be that the number of work order clusters reaches a certain preset number, or the similarity between clusters is lower than a certain threshold.

[0064] A significant advantage of using hierarchical clustering algorithms is that there is no need to pre-set the number of clusters (K value). It can autonomously discover potential cluster structures based on the distribution characteristics of the data itself, and can effectively identify noise points, such as effectively identifying isolated hotline work orders that are not similar to any other work orders.

[0065] Optionally, to improve the interpretability of the clustering results, after obtaining multiple work order clusters, the semantic generalization capability of a large language model can be further introduced to generate category labels and corresponding descriptions with clear semantic meanings for each automatically generated cluster. For example, the model can analyze the content of all work orders within a certain work order cluster and summarize it as "suggestions on installing elevators in old residential areas" or "complaints about improperly parked shared bicycles," etc.

[0066] By employing the aforementioned clustering technique based on fusion features, this embodiment transforms raw, scattered work order data into structured, semantic work order clusters. This approach not only considers the deep semantics of the work order text but also integrates key structured information, resulting in clustering results that are both accurate and highly interpretable for business applications, significantly improving the automation and intelligence of work order classification.

[0067] Based on any of the above embodiments, and based on the semantic information of each complete work order text, structured information is extracted from the corresponding hotline work orders to obtain the structured information of each hotline work order, including: Using a large language model, structured information is extracted from the corresponding hotline work orders based on the semantic information of each complete work order text. The extracted structured information is then classified and labeled.

[0068] Here, "large language model" can refer to a deep learning model trained on massive amounts of text data, possessing powerful natural language understanding and generation capabilities, such as a model based on the Transformer architecture. In the scenario of this embodiment, this model can deeply understand the complex semantics, contextual relationships, and underlying intentions in citizens' appeal texts, rather than simply relying on surface keyword matching.

[0069] Structured information refers to key information elements extracted from unstructured, complete work order texts according to pre-defined fields or dimensions with clear business meaning. Classification tags refer to assigning predefined category labels to these extracted information elements. These labels can be customized according to actual government management needs. For example, structured information can be classified and labeled as "background information," "core demands," "secondary demands," "involved departments," and "urgency level." Of course, this embodiment is not limited to this. In other embodiments, broader classifications can be set, such as only including "demand content" and "involved entities"; or more detailed classifications can be set, such as further subdividing "core demands" into subcategories such as "complaints," "suggestions," and "consultations."

[0070] To better understand the execution process of this embodiment, a specific hotline work order will be used as an example for explanation below: Suppose the complete work order text is: "I am a resident of Building 3, Area A. The convenience store downstairs uses a loudspeaker to promote sales every day from 8 am to 10 pm. It is so noisy that the elderly and children in my family cannot rest at all. I hope the relevant departments can do something about it. In addition, they are also occupying the road and piling up fruit boxes on the sidewalk, which makes it very inconvenient to walk." In this embodiment, after the complete work order text is input into the large language model, the model will use its powerful semantic understanding capabilities to analyze it and output the following formatted structured information: Background Information: Building 3, Area A, convenience store, business hours (8:00 AM to 10:00 PM). Core Complaint: Noise pollution (loudspeaker promotions). Secondary Complaint: Encroachment on sidewalk (fruit boxes piled up on the sidewalk). Relevant Departments: Market Supervision Administration, Urban Management Enforcement Bureau. Urgency Level: Relatively urgent.

[0071] This embodiment utilizes a large language model to perform deep semantic analysis on the complete work order text that integrates multimodal information. It can accurately and comprehensively extract the key structured information contained in the work order and automatically complete the classification and labeling. This overcomes the shortcomings of traditional keyword matching methods, such as the inability to understand the context and the tendency to miss information. It also avoids the drawbacks of low efficiency and strong subjectivity when manually organizing information.

[0072] Based on any of the above embodiments, by fusing the complete work order text and structured information of each hotline work order, the fused features of each hotline work order are obtained, including: Semantic encoding is performed on the complete work order text of each hotline work order to obtain the textual semantic features of each hotline work order; The structured information of each hotline work order is uniquely encoded to obtain the structured features of each hotline work order; We perform weighted fusion of the textual semantic features and structured features of each hotline work order to obtain the fused features of each hotline work order.

[0073] Specifically, semantic encoding refers to the process of converting the aforementioned text string into a numerical vector that a computer can understand and process using natural language processing techniques. Traditional text representation methods based on the bag-of-words model or term frequency-inverse document frequency often only focus on the frequency of word occurrence, ignoring the word order in a sentence and the context. This leads to an inability to effectively handle issues such as polysemy and synonyms, making it difficult to accurately capture the true semantics of the text. This embodiment, by performing semantic encoding on the complete work order text, can capture the true semantics of words in a specific context and the syntactic relationships between words from a global contextual perspective. The resulting text semantic features can capture the deeper semantic information of the text, rather than just the surface-level vocabulary.

[0074] In one possible implementation, semantic encoding can be achieved using a pre-trained language model, such as the BERT model. Taking the complete work order text as input to the BERT model, the model will output a fixed-dimensional real-valued vector, which represents the text's semantic features.

[0075] Furthermore, considering that structured information is typically discrete and categorical features, these features themselves do not possess numerical magnitude or order relationships. Directly employing simple numerical mapping would artificially introduce a false ordinal relationship, potentially misleading the subsequent model training process and affecting model accuracy. Therefore, this embodiment performs one-hot encoding on the structured information of each hotline work order to obtain the structured features of each hotline work order. This converts these categorical features into mutually independent binary feature vectors, avoiding incorrect learning biases in the model due to the numerical magnitude of feature values.

[0076] One-hot encoding is used to convert categorical variables into numerical forms. Its basic idea is to convert a categorical feature with N possible values ​​into an N-dimensional vector, where only the index position corresponding to the category value is 1, and the rest are 0.

[0077] For example, suppose the work order type includes three possible values: {fault report, business inquiry, complaint / suggestion}. If work order M has a "fault report" type, its one-hot encoding is [1, 0, 0]. If it is a "business inquiry", it is [0, 1, 0]. After performing one-hot encoding on all the structured information fields of the work order, the resulting multiple one-hot encoded vectors can be concatenated to form a longer vector, which is the structured feature of the work order.

[0078] Furthermore, to obtain a comprehensive and accurate representation of hotline work orders, it is necessary to effectively combine textual semantic features that reflect deep semantics with structured features that provide a clear classification context. Therefore, this embodiment performs weighted fusion of the textual semantic features and structured features of each hotline work order to obtain the fused features of each hotline work order.

[0079] In this embodiment, the final fusion features can be calculated using the following formula: V_fused=α×V_text +(1 - α)×V_struct Here, V_fused represents the fused features, V_text represents the textual semantic features, and V_struct represents the structured features. α is a weighting coefficient, typically ranging from 0 to 1, used to balance the importance of textual semantic features and structured features in the final fused features. For example, when processing work orders where content understanding is crucial, the value of α can be set higher (e.g., 0.7); when the classification attributes of the work order are more important, the value of α can be set lower (e.g., 0.3). The value of α can be set manually based on experience as a hyperparameter, or it can be automatically optimized as a learnable parameter during model training.

[0080] Furthermore, considering that V_text and V_struct typically have different dimensions, before performing addition, the lower-dimensional vector needs to be expanded to the same dimension as the higher-dimensional vector by padding with zeros. For example, if V_text is 768-dimensional and V_struct is 100-dimensional, V_struct can be padded with 668 zeros to become 768-dimensional before performing weighted summation. Of course, more complex fusion strategies can also be used, such as concatenating the two and then performing dimensionality reduction fusion through a fully connected layer; this embodiment does not specifically limit this approach.

[0081] Through the above steps, the resulting fused features contain both the deep semantic information of the work order content and its explicit classification attribute information. Compared with features from a single source, these features have stronger information expression capabilities and richer feature dimensions.

[0082] This embodiment encodes the complete text and structured information of the work order separately, and then performs weighted fusion to generate a work order fusion feature with higher information density and more comprehensive expression. This fusion feature takes into account both the deep semantics of the text and the clear classification attributes of the structured data, overcoming the shortcomings of the one-sidedness of information from a single data source. It provides higher quality input for subsequent work order clustering, classification, or intelligent recommendation processing tasks, thereby significantly improving the accuracy and efficiency of the entire work order processing flow.

[0083] Based on any of the above embodiments, and based on the fusion characteristics of each hotline work order, all hotline work orders are clustered to obtain multiple work order clusters, including: Treat each hotline ticket as an independent initial cluster; Based on the fusion features of each hotline work order, the similarity between each initial cluster is calculated. The two initial clusters with the highest similarity are iteratively merged into a new cluster until the final number of clusters meets the preset conditions, resulting in multiple work order clusters.

[0084] At the start of clustering, assuming there are N hotline work orders, these N hotline work orders are treated as N independent initial clusters. At this time, each initial cluster contains only one hotline work order, and the characteristics of the initial cluster are the fusion characteristics of that hotline work order.

[0085] In each iteration, the similarity between all current clusters needs to be calculated. This similarity measures how close two clusters are; a higher value indicates greater similarity. Optionally, the similarity between clusters can be measured by the distance between features of the two clusters.

[0086] It should be noted that when a cluster contains multiple hotline work orders, the characteristics of the cluster can be obtained by averaging the combined characteristics of all hotline work orders within the cluster.

[0087] After calculating the similarity between all current clusters, the two clusters with the highest similarity can be identified. For example, assuming there are M clusters, then among these M×(M-1) / 2 cluster pairs, find the cluster pair with the maximum similarity value (cluster i and cluster j).

[0088] Subsequently, the two clusters with the highest similarity are merged into a new cluster. The merging operation includes: assigning all hotline work orders from clusters i and j to the new cluster, calculating the characteristics of the new cluster, and deleting the original clusters i and j, thus reducing the total number of clusters by one.

[0089] For example, suppose there are 100 clusters in a certain iteration. It is found that the cosine similarity between cluster 25 (containing 3 hotline work orders) and cluster 67 (containing 5 hotline work orders) is the highest, reaching 0.92. Then, these two clusters are merged into a new cluster containing 8 work orders. At this time, the total number of clusters is 99.

[0090] The similarity calculation and cluster merging operations described above are performed iteratively. Each time a merge is performed, the total number of clusters decreases by one. The iterative process continues until a preset condition is met.

[0091] The preset conditions may include: ① Reduce the number of clusters to a preset target number. For example, K=20 can be set, meaning that iteration stops when 20 work order clusters are formed after clustering. This K value can be preset based on actual business needs and historical data experience.

[0092] ② The maximum similarity between the two clusters to be merged is lower than a preset threshold. For example, the preset threshold can be set to 0.5, that is, when the similarity of all remaining cluster pairs is less than 0.5, it is considered that the difference between them is large enough and they should not be merged.

[0093] ③ The number of clusters is reduced to the preset target number and the maximum similarity between the two clusters to be merged is lower than the preset threshold.

[0094] After the above iterative process, the originally large number of hotline tickets were automatically divided into several ticket clusters. The hotline tickets within each ticket cluster are highly related semantically, reflecting the same type or closely related issues. For example, it may eventually form 25 ticket clusters, where the first cluster contains all tickets related to "parking management in the community," the second cluster contains all tickets related to "noise pollution," and so on.

[0095] This embodiment achieves efficient, accurate, and adaptive clustering of hotline work orders by initializing each work order as an independent cluster and iteratively merging the clusters with the highest similarity based on fusion features until a preset stopping condition is met.

[0096] Based on any of the above embodiments, intra-cluster correlation analysis and inter-cluster correlation analysis are performed on the hotline work orders in each work order cluster to obtain the correlation analysis results of each hotline work order, including: Using a large language model, the complete work order text of each hotline work order in each work order cluster is parsed to generate cluster class tags and cluster description text for each work order cluster; Using a large language model, intra-cluster association analysis is performed on the complete work order texts of all hotline work orders within the same work order cluster to extract the distribution of business elements and hotspots within each work order cluster. Using a large language model, we perform inter-cluster association analysis based on the cluster category labels and cluster description text of each work order cluster to identify potential association relationships between different work order clusters. Based on the potential relationships between different work order clusters, the distribution of business elements and hotspots within each work order cluster, the correlation analysis results for each hotline work order are determined.

[0097] Specifically, cluster tags are highly condensed summaries of the topic of a work order cluster, usually a phrase or short sentence, used to quickly identify the core problem attributes of the cluster. For example, the cluster tags for a work order cluster could be noise pollution, parking fee disputes, or inadequate heating. The cluster description text, on the other hand, is a more detailed and comprehensive description of the problems involved in that work order cluster, systematically explaining the specific circumstances reflected in the work orders within the cluster, the geographical scope involved, and the characteristics of the problems.

[0098] In one specific implementation, cluster tags and cluster description text can be generated in the following way: First, extract the complete work order text of all hotline work orders in each work order cluster. Then, using the powerful text understanding and generation capabilities of the large language model, perform a comprehensive analysis on these complete work order texts to identify recurring keywords, common themes, and semantically overlapping parts in multiple work order texts within the cluster. Based on these identified parts, generate cluster tags and cluster description text.

[0099] In another implementation, when calling the large language model, several examples are provided in the prompt words to inform the model what cluster tags and cluster description text are, as well as their standard formats, so that the large language model can extract cluster tags and cluster description text. For example, the prompt words could be "Please refer to the following example [Example 1: Cluster tag = 'Community property management dispute', Cluster description text = 'This cluster of work orders mainly reflects problems in the maintenance of public areas and transparency of fee collection by a certain community property company, with the demands concentrated in October and November 2023'], now please generate the corresponding cluster tags and cluster description text for the following work order cluster...".

[0100] In another implementation, to improve accuracy, each hotline work order can be parsed individually. First, the large language model extracts the core requests of each hotline work order, and then the core requests of all work orders in the same cluster are summarized and concluded to finally generate a unified cluster label and cluster description text for that work order cluster.

[0101] Next, using a large language model, intra-cluster association analysis is performed on the complete work order texts of all hotline work orders within the same work order cluster to extract the distribution of business elements and hotspots within each work order cluster.

[0102] The business elements refer to key dimensions of information related to the problem reflected in the hotline work order. Business elements may include, but are not limited to, time elements (such as the specific date and time period of the complaint), location elements (such as the street, community, and specific address involved), responsible entity elements (such as the government department, company name, property management company, etc. involved), problem type elements (such as noise type, type of environmental pollution, etc.), and scope of impact elements (such as the number of affected households and duration). The distribution of business elements is a statistical description of the frequency, proportion, or distribution characteristics of these elements within the work order cluster. Figure 3 This is a statistical diagram illustrating the distribution of business elements provided by the present invention.

[0103] Hotspot distribution refers to which specific sub-issues, phenomena, or demands are mentioned most frequently within this work order cluster, representing the focus of public concern and concentrated feedback. Among these, Figure 4 This is a schematic diagram of hotspot distribution statistics provided by the present invention.

[0104] In one embodiment, a structured prompt word can be constructed to guide the large language model to complete this task. For example, the prompt word can be provided to the model: "Please analyze all the work order texts in the following work order clusters, and count and output: (1) Time distribution, count the number of work orders by date; (2) Location distribution, list the top 5 locations with the highest frequency and their work order counts; (3) Responsible entity distribution, count the number of work orders corresponding to each department involved; (4) Hot issues, summarize the top 3 specific demands with the most frequent occurrences."

[0105] In another embodiment, the information extraction capabilities of a large language model can be utilized to predefine patterns for business elements, such as defining them in JSON format as {"Time":"YYYY-MM-DD","Location":"XXX Street XXX Community","Responsible Entity":"XXX Department","Problem Type":"XXX"}. Then, the model parses the complete work order text one by one, extracting structured data that conforms to the pattern. Finally, these structured data are statistically summarized to obtain the distribution of business elements.

[0106] In another embodiment, the extraction of hotspot distribution can be achieved by combining frequency statistics and semantic aggregation methods. First, a large language model is used to extract and standardize the demands of all work orders within a cluster. Then, the frequency of each demand is counted, and finally, the hotspot distribution is obtained by sorting by frequency. For example, the hotspot distribution extracted from a certain work order cluster might be: {1. Nighttime construction noise (occurred 45 times); 2. Dump truck horns disturbing residents (occurred 32 times); 3. Construction time not in accordance with regulations (occurred 28 times)}.

[0107] In addition, using a large language model, inter-cluster association analysis is performed based on the cluster category labels and cluster description text of each work order cluster to identify potential association relationships between different work order clusters.

[0108] Potential correlations refer to the relevance or causal relationship between the problems reflected in different work order clusters in terms of root cause, impact chain, responsible party, or spatiotemporal distribution. These correlations are often not directly apparent and require deep semantic reasoning to identify. For example, a work order cluster about "frequent traffic congestion on a certain road section" may have a potential correlation with another work order cluster about "illegal parking of vehicles during school hours near the school," both potentially stemming from unreasonable traffic planning or insufficient enforcement in the area.

[0109] In one implementation, the cross-text reasoning capabilities of a large language model can be used to perform inter-cluster association analysis. Specifically, the cluster labels and cluster description texts of all work order clusters are used as input to construct an analysis prompt, such as "The following are the labels and descriptions of several work order clusters: [Cluster A: Label = 'Community heating temperature not up to standard', Description = '...'; Cluster B: Label = 'Aging and leaking heating pipe network', Description = '...'; Cluster C: Label = 'Slow service response from heating company', Description = '...']. Please analyze whether there are potential associations between these work order clusters. If so, please explain the type of association (e.g., causal association, common root cause, mutual influence, etc.) and the specific reasons."

[0110] In another implementation, for N work order clusters, they are paired to form N×(N-1) / 2 cluster pairs. Then, a large language model is used to determine the correlation of each cluster pair and output whether there is a correlation and the strength of the correlation, such as strong correlation, weak correlation or no correlation, so as to identify the correlation relationship in a refined manner.

[0111] In another implementation, correlation analysis can be performed by combining temporal and spatial dimensions. For example, if the work orders in two work order clusters highly overlap in time (e.g., both concentrated within a certain week) or are spatially close (e.g., both occurring on the same street), there may be a potential correlation between them. The large language model can comprehensively determine the correlation based on the extracted temporal and location element distributions, combined with semantic information.

[0112] Finally, based on the potential relationships between different work order clusters, the distribution of business elements within each work order cluster, and the distribution of hotspots, the correlation analysis results for each hotline work order are determined.

[0113] The correlation analysis results of each hotline work order are a structured, multi-dimensional data set. It not only includes intra-cluster analysis information of the cluster to which a single work order belongs, such as the distribution of business elements and hotspots in that cluster, but also includes the correlation information between the cluster to which the work order belongs and other clusters, thus giving each work order a richer context and deeper insights.

[0114] In one specific implementation, the results of association analysis can be organized using a hierarchical data structure.

[0115] In another implementation, these analysis results can be used to generate a visualized association graph. For example, each work order cluster can be used as a node, and potential associations can be used as edges to construct a network graph. The size of the nodes can reflect the number of work orders within a cluster, and the thickness of the edges can reflect the association strength, thereby intuitively showing the problem structure and association network in the overall work order data.

[0116] In another implementation, for scenarios requiring rapid decision-making, priority information can be automatically extracted from the correlation analysis results. For example, based on the frequency of hotspot distribution within clusters, the strength of inter-cluster correlations, and the scope of influence involved in the distribution of business elements, a large language model or preset scoring rules can be used to score each work order cluster or each work order, identifying the most urgent, critical, or representative work orders for priority processing.

[0117] Based on any of the above embodiments, and based on the correlation analysis results of each hotline work order, an intelligent summary of each hotline work order is generated, including: Based on the correlation analysis results of each hotline work order, a briefing outline for each hotline work order and a description of the content to be written for each chapter in the briefing outline are generated. Based on the descriptions of the content to be written in each chapter of the briefing outline of each hotline work order, the briefing materials required for each chapter are retrieved and generated from the corresponding hotline work orders and the results of the correlation analysis. Based on the briefing materials of each hotline work order, the content of each chapter is written and integrated using a large language model to generate intelligent briefings for each hotline work order.

[0118] Specifically, the briefing outline is the skeleton structure of the intelligent briefing, which clarifies the chapters included in the intelligent briefing, the titles of each chapter, the order of the chapters, and the core information type that each chapter should present.

[0119] The description of the content to be written is an explanation of the specific writing requirements for each chapter in the briefing outline. It specifies which key information points the chapter should include, what expression style should be used, and whether it needs to be accompanied by pictures, etc.

[0120] In one specific implementation, a structured prompt word can be constructed. The association analysis results obtained in step 130 are used as input and fed into a large language model for processing, generating briefing outlines for each hotline work order and descriptions of the content to be written for each section of the briefing outline. This prompt word needs to clearly inform the large language model of the current task's objective, the meaning of the input data, and the expected output format.

[0121] In another implementation, several briefing templates can be predefined. For different types of work orders or different analysis objectives, the appropriate template framework can be selected, and then the large language model can personalize and adjust the templates according to the specific correlation analysis results.

[0122] For example, for a high-frequency, rising hot issue work order, the briefing outline might include chapters such as an overview of the issue, development trends, scope of impact, root cause analysis, and response recommendations. The development trend chapter might describe the content to be written as follows: "Based on work order volume changes over the past 30 days, create a trend chart and provide a textual interpretation, highlighting the severity and urgency of the issue."

[0123] Next, based on the descriptions of the content to be written in each chapter of the briefing outline of each hotline work order, the briefing materials required for each chapter are retrieved from the corresponding hotline work orders and the results of the correlation analysis.

[0124] Presentation materials refer to the specific data, text snippets, chart codes, and other raw materials used to populate and support the content of each chapter. Presentation materials can be retrieved using the following methods: First, read the description of the content to be written for a certain section in the generated briefing outline. For example, if the description of a certain section is "It is necessary to show the trend of the number of work orders in the past 7 days for the problem category to which this work order belongs, accompanied by a line graph", then it is necessary to identify the key requirement elements, such as the time range (past 7 days), data type (number of work orders), and display method (line graph).

[0125] Based on the parsed demand elements, the corresponding data is retrieved from the association analysis results generated in step 130. These association analysis results are typically organized and stored according to different dimensions, such as time series, cluster categories, and statistical indicators, allowing for quick location and retrieval through structured queries.

[0126] For example, if the correlation analysis results contain a field named time_series_data that records the number of work orders each day, then data points from the most recent 7 days can be extracted from it [{day: "2025-10-08", count: 15}, {day:"2025-10-09", count: 18}, ...].

[0127] For data that requires visualization, chart generation tools or the code generation capabilities of large language models can be used to convert the data into chart code. After execution, this code will generate a trend chart image, which will serve as the image material for this section.

[0128] For text-based materials, they can be directly extracted from the complete work order text or the textual description of the correlation analysis results. For example, if a chapter needs to cite typical work order cases, representative work orders can be selected from the corresponding work order cluster, and key information such as their work order number and user description can be extracted as text materials.

[0129] In an optimized implementation, the description of the chapter's content to be written can be transformed into a query vector, and vector similarity search can be performed in the original work order database or a structured analysis result database to retrieve the most relevant material fragments.

[0130] Finally, based on the briefing materials of each hotline work order, the content of each chapter is written and integrated using a large language model to generate an intelligent briefing for each hotline work order.

[0131] Specifically, each chapter can be processed sequentially according to the order of the chapters in the briefing outline. For each chapter, a targeted prompt word is constructed, and the description of the content to be written for that chapter, along with the corresponding briefing materials, are input into the large language model to obtain the descriptive text for that chapter. This descriptive text is the main content of the corresponding chapter in the intelligent briefing. For chapters containing visual charts, the chart code is generated simultaneously with or after the large language model generates the descriptive text for that chapter, or an already generated chart image is directly called.

[0132] Subsequently, following the predetermined report format specifications, the descriptive text of each chapter, along with charts and images, were formatted and integrated to produce a smart briefing.

[0133] In one implementation, the generated charts and images from each chapter, along with their descriptive text, can be input into a large language model. The model will then automatically interpret and explain the charts, generating analytical text that better reflects the meaning of the charts, thus achieving a deep integration of text and images to obtain the content of each chapter.

[0134] After all chapter content is generated, a global optimization process can be performed. Specifically, all chapter content can be re-input into the large language model to optimize the model, ensuring natural transitions between chapters, clear logical arguments, avoiding information duplication or conflict, maintaining a consistent language style throughout the text, and bolding, highlighting, or summarizing key conclusions and data in the report.

[0135] Based on any of the above embodiments, according to the description of the content to be written in each chapter of the briefing outline of each hotline work order, the briefing materials required for each chapter are retrieved and generated from the corresponding hotline work orders and the correlation analysis results, including: A hybrid retrieval strategy combining embedded vector similarity calculation and text semantic matching is adopted. Based on the description of the content to be written in each chapter of the hotline work order's briefing outline, targeted material retrieval is performed in the hotline work order database to generate the briefing materials required for each chapter. If any chapter type is a brief overview, a single-round search is performed; if any chapter type is an in-depth analysis, a large language model is used to autonomously initiate multiple rounds of iterative search until the semantic similarity between the new briefing material obtained in the current iteration and all the briefing materials obtained is higher than a threshold, or the number of iterations reaches a preset number, then all the briefing materials obtained are used as the briefing materials required for the corresponding chapter; each chapter type is determined based on the description of the content to be written in the corresponding chapter.

[0136] Understandably, after generating the intelligent briefing outline, it's necessary to find and prepare specific briefing materials for each section. Briefing materials refer to the raw data and information used to support the writing of the briefing content, including but not limited to specific work order texts, statistical data, chart codes, and analytical conclusions. Since the hotline work order database may contain thousands or even tens of thousands of work order records, efficiently and accurately retrieving the most relevant materials for each section becomes a key issue.

[0137] This embodiment uses a hybrid retrieval strategy to solve this problem. This strategy combines two complementary techniques: embedding vector similarity calculation and text semantic matching.

[0138] The embedding vector similarity calculation first uses a text embedding model to convert the description of the content to be written in a certain chapter of the briefing outline into a high-dimensional query vector. Simultaneously, each work order and each analysis result in the work order database is pre-converted into a corresponding text semantic vector. Then, by calculating the cosine similarity or Euclidean distance between the query vector and each vector in the database, several candidate materials that are semantically closest are quickly selected. The advantage of this method is that it can capture deep semantic relationships; even if the expressions are different, as long as the meaning is similar, they can be retrieved.

[0139] Text semantic matching focuses more on literal matching. For example, it uses techniques such as keyword extraction, fuzzy matching, and regular expressions to further filter candidate materials for content containing specific terms or meeting specific format requirements. For instance, if the description of the work order explicitly requires the inclusion of the keyword "noise disturbance," text matching technology can be used to accurately locate these cases.

[0140] By combining the two methods—first using embedding vector similarity calculation for coarse screening to quickly narrow down the scope, and then using text semantic matching for fine screening—the final search results are both semantically relevant and meet specific requirements, thus achieving efficient and accurate targeted material retrieval.

[0141] In this embodiment, during the retrieval process, the appropriate retrieval mode is selected based on the chapter type. The chapter type is a classification identifier for the functional positioning of each chapter in the briefing outline, determined based on the description of the content to be written for that chapter. In one implementation, a large language model can automatically identify and label the chapter type based on the description of the content to be written. Common chapter types include brief overviews and in-depth analyses.

[0142] Brief overview sections typically only require high-level, summary information. For these sections, the required material is relatively clear and limited, so a single-round search is sufficient. Specifically, based on the description of the content to be written for the section, a hybrid search strategy is used to perform a single query in the work order database to retrieve relevant statistical data or summary information, which serves as the briefing material needed for that section. This method is efficient and fast, enabling the acquisition of sufficient material in a short time.

[0143] Chapters involving in-depth analysis require a multi-dimensional and multi-layered analysis of a specific issue, such as analyzing the root causes of a particular type of complaint and its potential connections to other issues. For these chapters, a single search often fails to cover all necessary information dimensions and may miss some implicit or related material. Therefore, this embodiment employs a multi-round iterative search mechanism for chapters involving in-depth analysis.

[0144] Among them, multi-round iterative retrieval is a dynamic and intelligent material expansion process, the specific process of which is as follows: First round of retrieval: Based on the description of the content to be written in the chapter, an initial mixed retrieval is performed to obtain a batch of initial briefing materials, which are marked as obtained briefing materials.

[0145] The second and subsequent rounds of retrieval: The acquired briefing materials are input into the large language model, which analyzes whether the current materials are sufficient to support in-depth writing of the chapter. The large language model has powerful reasoning and judgment capabilities; it can identify potential information gaps or related directions that need further exploration in the current materials. For example, the model may find that "the current materials lack information about the distribution of the time period in which the problem occurred," and then autonomously generate a new query description, such as "retrieve the distribution of work orders for this problem in different time periods," and initiate a new round of hybrid retrieval to obtain new briefing materials.

[0146] In this embodiment, the multi-round iterative retrieval will not proceed indefinitely. Two parallel termination conditions are set, as follows: In each iteration, the semantic similarity between the newly acquired presentation materials and all previously acquired presentation materials is calculated. This similarity can be measured using the cosine similarity of the text embedding vectors. If the similarity is higher than a preset threshold, it indicates that the new materials highly overlap with existing materials, and the marginal benefit of continuing the search is minimal. In this case, the iteration is terminated, and all previously acquired presentation materials are used as the presentation materials required for that chapter, ensuring the sufficiency and non-redundancy of the materials.

[0147] In addition, to prevent the retrieval from getting stuck in a loop due to consistently low semantic similarity in certain situations, a preset number of iterations is set as a hard upper limit. When the preset number of iterations is reached, the iteration is forcibly terminated regardless of the semantic similarity, and all the obtained briefing materials are used as the briefing materials required for the corresponding chapter.

[0148] The hotline briefing generation apparatus provided by the present invention is described below. The hotline briefing generation apparatus described below and the hotline briefing generation method described above can be referred to in correspondence.

[0149] Based on any of the above embodiments Figure 5 This is a schematic diagram of the hotline briefing generation device provided by the present invention, as shown below. Figure 5 As shown, the device includes: The fusion unit 510 is used to convert the image content contained in each hotline work order into descriptive text, and fuse the descriptive text with the text content contained in the corresponding hotline work order to obtain the complete work order text of each hotline work order. Clustering unit 520 is used to cluster all hotline work orders based on the semantic information of each complete work order text to obtain multiple work order clusters; Analysis unit 530 is used to perform intra-cluster correlation analysis and inter-cluster correlation analysis on hotline work orders in each work order cluster, and obtain the correlation analysis results of each hotline work order. The generation unit 540 is used to generate intelligent briefings for each hotline work order based on the correlation analysis results of each hotline work order.

[0150] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions stored in the memory 630 to execute a hotline briefing generation method.

[0151] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program that can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute the hotline briefing generation method provided by the above methods.

[0153] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the hotline briefing generation methods provided by the above methods.

[0154] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown 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 this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0155] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hotwire briefing generation method, characterized by, The method comprises the following steps: Converting image content contained in each hotline ticket into descriptive text, and fusing the descriptive text with text content contained in the corresponding hotline ticket to obtain complete ticket text of each hotline ticket; Based on the semantic information of each complete ticket text, clustering all hotline tickets to obtain multiple ticket clusters; Performing intra-cluster correlation analysis and inter-cluster correlation analysis on the hotline tickets in each ticket cluster to obtain correlation analysis results of each hotline ticket; Based on the correlation analysis results of each hotline ticket, generating an intelligent briefing of each hotline ticket.

2. The hotline brief generation method of claim 1, wherein, The method comprises the following steps: Based on the semantic information of each complete ticket text, extracting structured information of the corresponding hotline ticket to obtain structured information of each hotline ticket; Fusing the complete ticket text and the structured information of each hotline ticket to obtain fusion features of each hotline ticket; Based on the fusion features of each hotline ticket, clustering all hotline tickets to obtain multiple ticket clusters.

3. The hotline brief generation method of claim 2, wherein, The method comprises the following steps: Using a large language model to extract structured information of the corresponding hotline ticket according to the semantic information of each complete ticket text to obtain structured information of each hotline ticket, and classifying each structured information.

4. The hotline brief generation method of claim 2, wherein, The method comprises the following steps: Performing semantic coding on the complete ticket text of each hotline ticket to obtain text semantic features of each hotline ticket; Performing one-hot coding on the structured information of each hotline ticket to obtain structured features of each hotline ticket; Weighted fusion of the text semantic features and the structured features of each hotline ticket to obtain fusion features of each hotline ticket.

5. The hotline brief generation method of claim 2, wherein, The method comprises the following steps: Regarding each hotline ticket as an independent initial cluster; Based on the fusion features of each hotline ticket, calculating the similarity between each initial cluster, iteratively merging the two initial clusters with the highest similarity into a new cluster, until the number of clusters formed finally meets the preset condition, to obtain multiple ticket clusters.

6. The hotwire report generation method according to any one of claims 1 to 5, characterized by, The method comprises the following steps: Using a large language model to analyze the complete ticket text of each hotline ticket in each ticket cluster to generate cluster class labels and cluster description texts of each ticket cluster; Using the large language model to perform intra-cluster correlation analysis on the complete ticket text of all hotline tickets in the same ticket cluster, and extract business element distribution and hotspot distribution in each ticket cluster; Using the large language model to perform inter-cluster correlation analysis according to the cluster class labels and the cluster description texts of each ticket cluster, and identify potential correlation relationships between different ticket clusters; According to the potential correlation relationships between different ticket clusters, the business element distribution and the hotspot distribution in each ticket cluster, determine the correlation analysis results of each hotline ticket.

7. The hotwire briefing generation method according to any one of claims 1 to 5, characterized by, Based on the correlation analysis results of each hotline work order, an intelligent summary is generated for each hotline work order, including: Based on the correlation analysis results of each hotline work order, a briefing outline for each hotline work order and a description of the content to be written for each chapter in the briefing outline are generated. Based on the descriptions of the content to be written in each chapter of the briefing outline of each hotline work order, the briefing materials required for each chapter are retrieved and generated from the corresponding hotline work orders and the results of the correlation analysis. Based on the briefing materials of each hotline work order, the content of each chapter is written and integrated using a large language model to generate intelligent briefings for each hotline work order.

8. The hotwire briefing generation method of claim 7, wherein, The process involves retrieving and generating the necessary briefing materials for each chapter from the corresponding hotline work orders and the correlation analysis results, based on the descriptions of the content to be written in each chapter of the briefing outline, including: A hybrid retrieval strategy combining embedded vector similarity calculation and text semantic matching is adopted. Based on the description of the content to be written in each chapter of the hotline work order's briefing outline, targeted material retrieval is performed in the hotline work order database to generate the briefing materials required for each chapter. If any chapter type is a brief overview, a single-round search is performed; if any chapter type is an in-depth analysis, a large language model is used to autonomously initiate multiple rounds of iterative search until the semantic similarity between the new briefing material obtained in the current iteration and all the briefing materials obtained is higher than a threshold, or the number of iterations reaches a preset number, then all the briefing materials obtained are used as the briefing materials required for the corresponding chapter; each chapter type is determined based on the description of the content to be written in the corresponding chapter.

9. A hotwire briefing generation apparatus characterized by comprising: include: The fusion unit is used to convert the image content contained in each hotline work order into descriptive text, and to fuse the descriptive text with the text content contained in the corresponding hotline work order to obtain the complete work order text of each hotline work order. Clustering units are used to cluster all hotline work orders based on the semantic information of each complete work order text, resulting in multiple work order clusters; The analysis unit is used to perform intra-cluster correlation analysis and inter-cluster correlation analysis on the hotline work orders in each work order cluster, and obtain the correlation analysis results of each hotline work order. The generation unit is used to generate intelligent briefings for each hotline work order based on the correlation analysis results.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the hotline briefing generation method as described in any one of claims 1 to 8. 11.A non-transitory computer-readable storage medium having stored thereon a computer program. When the computer program is executed by a processor, it implements the hotline briefing generation method as described in any one of claims 1 to 8.

Images