Dynamic hotspot clue accurate extraction method based on multi-source text

Abstract

The invention discloses a multi-source text-based dynamic hotspot clue accurate extraction method, which realizes real-time adaptation and efficient training of a sensitive intelligence recognition model through dynamic label configuration and Prompt design in combination with a lightweight fine tuning technology (RsLoRA), and can accurately recognize in a changeable intelligence scene without re-training. Through human feedback reinforcement learning (PPO) and sorting training, clue texts output by the generative large model are optimized, event subjects and contents are accurately extracted, and a generated result is structured and highly summarized. A semantic vector model and a dynamic multi-layer DBSCAN clustering algorithm are combined, the problem that an existing clustering algorithm is insufficient in adaptability to semantic density difference scenes is solved, and high-precision clue aggregation and popularity statistics are achieved. The overall scheme reduces resource consumption and training time through the RsLoRA technology, comprehensively improves intelligence processing efficiency and analysis ability, and has significant technical advantages and wide application prospects in the fields of intelligent public security, government management and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of data management and analysis, and in particular to a method for accurately extracting dynamic hot spot clues based on multi-source texts. Background Art

[0002] The analysis and clue extraction of multi-source intelligence texts have important applications in the fields of smart public security and smart cities. The core of this is to assist decision-making and improve management efficiency by screening, extracting and aggregating sensitive information from multi-channel and massive unstructured text data. In the existing technology, the key issue in this field is how to efficiently process intelligence texts and quickly extract valuable information from them.

[0003] Traditional intelligence processing methods are usually based on fixed-rule keyword matching or traditional natural language processing (NLP) models. The basic working principle of such methods is to filter sensitive texts through predefined keywords or templates, and then manually analyze the filtered texts. Although these methods can partially achieve the goal of intelligence extraction, they have the following problems in practical applications: 1. Poor scenario adaptability: Traditional keyword matching methods and fixed rule models are difficult to flexibly adapt to the ever-changing actual scenario needs. For example, when the user's focus shifts from "traffic accidents" to "environmental pollution", the rules and models must be redesigned and trained, resulting in a long response cycle and poor adaptability.

[0004] 2. Low processing efficiency: Faced with massive amounts of intelligence data, the manual analysis process is lengthy and easily affected by subjective judgment. At the same time, the performance of traditional NLP models is often insufficient in large-scale data processing scenarios, especially in dynamic label classification tasks, which cannot quickly update or efficiently process new labels.

[0005] 3. Inaccurate results: Even if some systems can automatically generate intelligence clues, the output content often lacks structured expression and cannot meet the actual needs of specific businesses. For example, the event subject (such as place names, names of people) and event content are often covered by redundant information, which further reduces the efficiency of intelligence analysis.

[0006] In order to overcome the above problems, existing technologies attempt to optimize the intelligence processing process through deep learning models. For example, intelligence analysis technology based on generative large models has been widely studied, which can generate more diverse and accurate intelligence clues by using the language understanding ability of large models. However, this type of technology has introduced new problems, such as the training and reasoning process of generative large models requires a lot of computing resources, and has high requirements for deployment and application environment, especially in the batch processing scenario of massive texts, where slow reasoning speed becomes a key bottleneck. In addition, although large models have excellent general capabilities, their generated results often lack field specificity. When directly applied to specific intelligence fields, the generated content may not meet actual needs and needs further optimization. Or cluster analysis of intelligence clues is a key step in realizing intelligence hotspot analysis. However, existing clustering algorithms have poor adaptability to intelligence texts with large differences in semantic density, and cannot simultaneously meet the aggregation needs of large-scale low-similarity texts and small amounts of high-similarity texts, affecting the accuracy of intelligence heat analysis.

[0007] In addition, the industry usually adopts methods to solve the above problems, including fine-tuning all parameters of large models or increasing data training in specific fields. However, this method not only requires additional resources and time costs, but may also lead to unstable results due to the complexity of the model. At the same time, in order to improve the clustering effect, some technologies introduce multi-stage clustering algorithms, but due to the lack of dynamic adjustment mechanism, it is still difficult to obtain the best clustering results in intelligence data of different scales.

[0008] Therefore, how to efficiently identify dynamic intelligence, accurately generate structured clue texts and achieve accurate aggregation of hot clues has become a technical problem to be solved by the present invention. Summary of the invention

[0009] The technical problem solved by the present invention is to provide a method for accurately extracting dynamic hot spot clues based on multi-source texts in response to the defects existing in the above-mentioned prior art, so as to solve the problems of insufficient flexibility in sensitive intelligence processing, inaccurate generation results and limited clustering analysis effect raised in the above-mentioned background technology.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows: A method for accurately extracting dynamic hot spot clues based on multi-source texts, the method comprising the following steps: Step 1, obtaining multi-source intelligence text data; Step 2: dynamically adjust the labels according to the sensitive domain scenarios configured by the user, and build a sensitive intelligence text recognition model, which is trained by the following method: Using a large generative model and Prompt design, multiple sets of positive and negative samples are generated based on the sensitive scene labels that users are concerned about. Based on low-rank adaptive optimization technology, the generative large model is lightly fine-tuned to optimize its ability to recognize dynamic labels; Step 3: extract text belonging to sensitive areas from multi-source intelligence text data through a sensitive intelligence text recognition model; Step 4: Optimize the generative large model using a reinforcement learning algorithm for the extracted sensitive text data to generate structured clue text, where the clue text includes the event subject and event content; Step 5: Input the generated clue text into the semantic vector model for vectorization processing, and combine the multi-layer DBSCAN clustering algorithm to achieve clue aggregation; Step 6: Based on the clustering results, count the popularity of clue texts and generate a hot clue list for intelligence analysis and decision support.

[0011] As a further solution of the present invention, the training of the sensitive intelligence text recognition model specifically includes the following steps: Dynamically generate training samples based on user-configured sensitive domain scene labels to construct positive and negative samples; Dynamically insert the scene labels that users are concerned about into the prompt of the generative large model, and determine the sensitivity of the intelligence text through the text classification task; Using RsLoRA technology, the linear layer of the generative large model is decomposed into low rank, which enables lightweight fine-tuning by reducing the number of parameters, while improving the training efficiency and stability of the model.

[0012] As a further solution of the present invention, the process of generating clue text includes the following steps: Utilize the event summarization capability of the generative big model to generate event subjects and event content through prompts; For the generated clue text results, the model output results are sorted by combining the LAC named entity recognition tool and manual verification; Based on the sorted training data set, the generative large model is optimized through reinforcement learning to make the clue text it generates more in line with actual needs.

[0013] As a further solution of the present invention, the semantic vector model adopts the bge-zh-base open source model to map the clue text sequence into a 768-dimensional semantic vector for representing the semantic features of the text.

[0014] As a further solution of the present invention, the parameters of the multi-layer DBSCAN clustering algorithm are dynamically adjusted to the following three groups: 5.1 eps=0.15, MmPts=4; 5.2 eps=0.17, MmPts=2; 5.3 eps=0.1,MmPts=1; Through multi-level clustering processing, it can adapt to intelligence texts with low semantic similarity but large number of aggregations, and intelligence texts with high semantic similarity but small number of aggregations.

[0015] As a further solution of the present invention, the clustered clue text heat statistics are achieved by the following method: Calculate the average similarity between the text points and the center point in each cluster to evaluate the clustering quality; Generate a hot clue list based on the number of text clusters and the semantic relevance of the center point.

[0016] As a further solution of the present invention, the hot clue list is automatically updated at preset time intervals, supporting users to perform trend analysis on clue popularity changes within different time ranges.

[0017] As a further solution of the present invention, the output results of the method can be used in scenarios including but not limited to smart public security and government management to improve the efficiency of intelligence analysis and support quick decision-making by relevant departments.

[0018] As a further solution of the present invention, the multi-source intelligence text data includes but is not limited to data from police information, Internet public opinion, and government hotline sources.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. The use of dynamic label configuration and prompt design combined with lightweight fine-tuning technology (RsLoRA) has achieved efficient training and real-time adaptability of sensitive intelligence recognition models. In the ever-changing actual intelligence scenarios, users can dynamically adjust the focus areas according to their needs and obtain high-precision predictions without retraining the model, which significantly shortens the sensitive clue discovery cycle and combines flexibility and efficiency.

[0020] 2. Combine human feedback reinforcement learning (PPO) and sorting training to optimize the clue text output of the generative large model, so that it can accurately extract "event subject + event content" and meet the actual needs of public security intelligence analysis. This optimization method ensures that the generated results are both structured and highly summarized, significantly improving the accuracy and application value of intelligence clue extraction.

[0021] 3. By introducing a dynamically adjusted multi-layer DBSCAN algorithm, the present invention overcomes the limitations of existing clustering algorithms in scenarios with different semantic density, and can simultaneously process text collections with low similarity but large scale and intelligence texts with high similarity but small quantity. Combining the semantic vector model and the semantic judgment of the generative large model, the accuracy of the clustering results is further improved, providing a solid foundation for the heat statistics and trend analysis of intelligence clues.

[0022] 4. The RsLoRA technology effectively reduces the resource consumption and training time of the generative large model, while retaining the excellent performance of the large model in text understanding and generation. The efficient processing of multi-source intelligence has achieved rapid implementation in the fields of smart public security and government management, significantly enhancing the efficiency and operability of intelligence analysis. At the same time, it takes into account the efficiency of sensitive intelligence identification, the accuracy of clue generation, the depth of cluster analysis and the intuitiveness of result display.

[0023] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0025] Figure 1 It is a flow chart of the method for accurately extracting dynamic hot spot clues based on multi-source texts of the present invention. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] See also Figure 1 In an embodiment of the present invention, a method for accurately extracting dynamic hot spot clues based on multi-source texts includes the following steps: Step 1, obtaining multi-source intelligence text data; Step 2: dynamically adjust the labels according to the sensitive domain scenarios configured by the user, and build a sensitive intelligence text recognition model, which is trained by the following method: Using a large generative model and Prompt design, multiple sets of positive and negative samples are generated based on the sensitive scene labels that users are concerned about. Based on low-rank adaptive optimization technology, the generative large model is lightly fine-tuned to optimize its ability to recognize dynamic labels; Step 3: extract text belonging to sensitive areas from multi-source intelligence text data through a sensitive intelligence text recognition model; Step 4: Optimize the generative large model using a reinforcement learning algorithm for the extracted sensitive text data to generate structured clue text, where the clue text includes the event subject and event content; Step 5: Input the generated clue text into the semantic vector model for vectorization processing, and combine the multi-layer DBSCAN clustering algorithm to achieve clue aggregation; Step 6: Based on the clustering results, count the popularity of clue texts and generate a hot clue list for intelligence analysis and decision support.

[0028] As a further solution of the present invention, the training of the sensitive intelligence text recognition model specifically includes the following steps: Dynamically generate training samples based on user-configured sensitive domain scene labels to construct positive and negative samples; Dynamically insert the scene labels that users are concerned about into the prompt of the generative large model, and determine the sensitivity of the intelligence text through the text classification task; Using RsLoRA technology, the linear layer of the generative large model is decomposed into low rank, which enables lightweight fine-tuning by reducing the number of parameters, while improving the training efficiency and stability of the model.

[0029] The process of generating clue text includes the following steps: Utilize the event summarization capability of the generative big model to generate event subjects and event content through prompts; For the generated clue text results, the model output results are sorted by combining the LAC named entity recognition tool and manual verification; Based on the sorted training data set, the generative large model is optimized through reinforcement learning to make the clue text generated by it more in line with actual needs; the semantic vector model uses the bge-zh-base open source model to map the clue text sequence into a 768-dimensional semantic vector to represent the semantic features of the text; the parameters of the multi-layer DBSCAN clustering algorithm are dynamically adjusted to the following three groups: 5.1 eps=0.15, MmPts=4; 5.2 eps=0.17, MmPts=2; 5.3 eps=0.1,MmPts=1; Through multi-level clustering processing, it can adapt to intelligence texts with low semantic similarity but large number of aggregations, and intelligence texts with high semantic similarity but small number of aggregations.

[0030] The heat statistics of clue texts after clustering are achieved in the following ways: the average similarity between the text points and the center point in each cluster is calculated to evaluate the clustering quality; the hot clue list is generated according to the number of text clusters and the semantic relevance of the center point.

[0031] The hot clue list is automatically updated at preset time intervals, supporting users to conduct trend analysis on clue popularity changes within different time ranges; the output results of the method can be used in scenarios including but not limited to smart public security and government management to improve intelligence analysis efficiency and support quick decision-making by relevant departments; multi-source intelligence text data include but are not limited to data from police information, Internet public opinion, and government hotline sources.

[0032] See also Figure 1 , showing the overall process of a dynamic hot spot clue accurate extraction method based on multi-source text, including the main steps of sensitive intelligence identification, clue text generation, clue clustering and heat statistics. Through dynamic label configuration and generative big model, the system first screens the multi-source intelligence text for sensitive areas, and then uses the optimized generative big model to convert the screened text into a structured clue text of "event subject + event content".

[0033] Subsequently, the generated clue text is processed by semantic vectorization and a multi-layer DBSCAN clustering algorithm is used to achieve hot clue aggregation, and finally a list of intelligence clues is generated by combining semantic similarity analysis to provide users with intuitive hot event analysis and decision support. This process fully reflects the efficiency, accuracy and adaptability of the technical solution.

[0034] Embodiment 1: The overall algorithm process of this patent mainly includes the following three stages: sensitive intelligence text recognition in custom scenarios; event text generation based on sensitive intelligence; clustering and heat statistics based on clue text.

[0035] 1. Sensitive intelligence text recognition in custom scenarios.

[0036] Compared with traditional text classification tasks, sensitive intelligence text recognition in the field of intelligence analysis is a text binary classification task with real-time changing label rules, that is, the definition of "sensitive" labels can be dynamically adjusted according to different key areas configured by users. The prompt design of the application of large models combined with the efficient fine-tuning technology of RsLoRA parameters can well meet this requirement. The specific process is as follows: 20 sensitive scenarios of user concern are selected, such as sudden traffic accidents, property owners' rights protection, environmental pollution, etc., and recorded as , where i ranges from [1, 20] to represent 20 sensitive key areas. According to each area, the corresponding keywords are configured to search based on the full amount of intelligence text, and manually classified and labeled to determine whether each piece of data is sensitive data related to government, society, and people's livelihood. The sensitive data corresponding to the 20 sensitive field scenarios are recorded as , where i ranges from [1, 20] to represent 20 sensitive key areas; , indicating non-sensitive data.

[0037] according to , , Construct a large model fine-tuning dataset. First, randomly sample a value in the range [1, 20], denoted as k, and Randomly sample k labels and concatenate them into text denoted as {label_phrase}; Randomly extract one data belonging to these k labels to form a positive sample, recorded as {text}, and construct the input prompt in the following format: "Judge whether the following intelligence text belongs to the sensitive data of the government, people's livelihood, and society in the field of {label_phrase}: {text}", and the corresponding output is "yes"; Randomly extract one data that does not belong to these k labels to form a negative sample, and construct the input prompt in the same format, and the corresponding output is "no"; Randomly extract one piece of data from the training set to form a negative sample, and construct the input prompt in the same format. The corresponding output is "No". Repeat the above process to form 2,000 training samples.

[0038] Based on the qwen2-7B-instruct open source large model, the RsLoRA method is used to fine-tune the model with 10,000 training samples. qwen2-7B-instruct is an open source 7B parameter large model supporting 32k contexts from Alibaba. It adopts a Transformer-based decoder architecture and applies a group attention mechanism and a sliding window attention mechanism. LoRA refers to low-rank adaptive technology. Compared with fine-tuning the large model of all parameters, LoRA replaces the large model linear layer with a low-rank decomposition matrix for training, which can effectively reduce training resources and time. RsLoRA is a rank-stabilized optimized LoRA technology that can be more stable during model training.

[0039] During training, RsLoRA uses a rank of 64, a learning rate of 0.0001, and two epochs. During model inference, users can freely configure the scene labels in the prompt according to the sensitive field scenarios they are concerned about, and the model can produce excellent prediction results.

[0040] 2. Generate clue text based on sensitive intelligence.

[0041] Based on the sensitive intelligence text identified in the previous step, it is necessary to apply the big model to further generate clue text for each piece of intelligence, so as to convert the multi-channel and multi-style text into the form of event subject + event content. The event subject includes key entities such as people, place names, company names, and organization names, and the event content is a refined description of the intelligence event.

[0042] Since the large model itself has excellent event summarization capabilities, we hope to train the model so that the model output content is more aligned with our business preferences, without having to train the model's clue time extraction capabilities from scratch. Therefore, this step uses human feedback reinforcement learning technology to train model preferences and uses the PPO algorithm to optimize the model's output content. The specific steps are: First, 500 sensitive intelligence texts were randomly selected and input into the qwen2-7B-instruct model in turn. The model was asked to output 10 different event clues. The prompt used was "Based on the following intelligence text, use concise language to extract event clues related to the government, society, and people's livelihood reflected in the text. Event clues should include specific organizations, place names and other event subjects. The text format of the answer is event subject plus event content. If there is no related event, output "no event clues"". Based on the 10 different output results, the LAC named entity recognition tool combined with manual verification is used to rank the results and the adaptability to the intelligence analysis business. That is, if the output event clue does not contain a person's name, place name, or organization name, or the output length is too long, or the output text does not accurately summarize the event in the original text, then the corresponding ranking needs to be lowered. The sorting data is thus obtained. , where i ranges from [1, 500] to represent a ranking data set generated from 500 different intelligence texts.

[0043] according to , randomly sample k from [1, 500], and Randomly select two texts from the dataset and construct a triplet (text1, text2, label), where label is 1 if text1 is ranked higher than text2, otherwise label is -1. Repeat the above process until 10,000 text triplet training data are formed.

[0044] The RsLoRA fine-tuning qwen2-7B-instruct method is used to train a 7B parameter scoring model, denoted as The specific approach is to superimpose a linear model on the token position of the last output of the model, convert the output embedding into a score to represent the preference score of the text. During training, two texts in a set of data are input into the model in turn and the corresponding scores are obtained. The pairwise loss function is used to optimize the text model weights through reverse gradient propagation.

[0045] According to the obtained , the PPO algorithm is applied to fine-tune the RsLoRA of qwen2-7B-instruct, making the model's event clue summary ability closer to the public security intelligence analysis business field.

[0046] 3. Clustering and heat statistics based on clue text.

[0047] Based on the generated batch of clue texts, further association clustering is performed to form an intelligence clue popularity list to assist intelligence analysts in conducting statistical grading and processing. The algorithm is implemented by combining semantic vector encoding with a large model.

[0048] The clustering algorithm based on semantic vector encoding is divided into two steps: text vectorization and clustering. The text vector model uses the open source bge-zh-base model to map the text sequence into a 768-dimensional output vector to represent the semantic features of the text. The clustering algorithm uses the multi-layer DBSCAN algorithm and applies the cosine similarity algorithm as the distance metric, which has a good aggregation effect on text sets of different sizes. The multi-layer DBSCAN algorithm is designed as follows: dynamically adjust the algorithm's eps (neighborhood radius at density) and MmPts (threshold when defining core points, the minimum number of core points required to form a cluster), and take eps and MmPts as (0.15, 4), (0.17, 2), and (0.1, 1) three groups of data, and perform multi-level clustering to ensure that texts that are slightly far away but have a large number of clusters, or very close but have a small number of clusters in the semantic space can be clustered into one category.

[0049] For each text cluster, the sum of the distances between each point and other points is calculated, and the point with the smallest distance is selected as the center point of the cluster, which serves as the aggregate intelligence clue of the entire text cluster.

[0050] Then, the big model is used to perform secondary clustering on the aggregated intelligence clues obtained by semantic vector encoding. The center point intelligence clue texts of the two classes are input into the big model for semantic similarity judgment. The prompt is designed as "Judge whether the following two intelligence texts are semantically similar. You can only answer yes or no, and the length should not exceed 1 word: \n1.{text 1}\n2.{text 2}".

[0051] Based on the results of secondary clustering, the popularity statistics of intelligence clues can be realized, and the intelligence clue popularity list can be generated regularly.

[0052] Furthermore, the present invention successfully applies AI algorithms such as large models, semantic vectors, and clustering in the field of intelligence clue mining, combines the advantages of various algorithms, achieves a balance between effect and efficiency, and is successfully applied in multiple actual scenarios.

[0053] Embodiment 2: In the field of smart public security, a large amount of intelligence data from Internet public opinion, government hotlines and police records need to be analyzed in a timely manner to assist decision-making. These data often come from diverse sources and complex formats, and contain sensitive information and potential clues. Traditional processing methods are difficult to meet actual needs due to their lack of flexibility and efficiency. This embodiment specifically demonstrates the application of a dynamic hot spot clue accurate extraction method based on multi-source text in a smart public security scenario.

[0054] First, intelligence data comes from a variety of channels, such as public opinion texts in Internet forums, 12345 citizen service hotline records, and public security police reports. These data are generated hundreds of thousands of times every day, covering areas such as traffic accidents and environmental pollution. In order to extract key intelligence clues from them, this embodiment adopts a dynamic sensitive intelligence recognition model.

[0055] 1. Dynamic sensitive intelligence identification: In the intelligence screening stage, users need to dynamically adjust sensitive scene labels according to the current social hot spots. For example, they may pay special attention to "environmental pollution" in a certain period. Through the generative big model combined with prompt design, users can configure keywords in the interface, such as "chemical plant". This embodiment uses lightweight fine-tuning technology (RsLoRA) to train the big model so that it can quickly adapt to user needs without retraining the entire model.

[0056] For example, when the user configures the "environmental pollution" scene label, the system will filter out texts that may be related to the scene from the massive intelligence text in a short time. The screening results such as "pollution from a chemical plant" are marked as sensitive intelligence as a further solution of the present invention. This dynamic label configuration avoids the problem of missing identification caused by the untimely update of the traditional rule model, and significantly shortens the clue discovery cycle.

[0057] 2. Generation of sensitive intelligence clues: After extracting the intelligence text, it is necessary to simplify the complex original text into structured event clues to facilitate rapid analysis by the public security department. This embodiment optimizes the generative large model through the reinforcement learning algorithm (PPO), and the generated clue text is presented in the form of "event subject + event content".

[0058] For example, for an intelligence text about "pollution in a chemical plant", the clue text output by the model is: Event subject: pollution from a chemical plant; Event content: Wastewater pollution.

[0059] This structured expression avoids the problem of lengthy and inaccurate results generated by traditional large models, and is particularly suitable for the needs of public security intelligence analysis. At the same time, the clue text is verified and sorted through named entity recognition (LAC tool) to ensure that the output clue content fits the actual application scenario.

[0060] 3. Intelligence clustering and hot spot analysis: To improve the analysis efficiency, this embodiment further clusters the generated clue text to form a hot spot clue list. The system uses a semantic vector model to map the clue text into a 768-dimensional semantic vector and clusters these semantic vectors using a multi-layer DBSCAN algorithm.

[0061] For example, if multiple clue texts mention "pollution from a chemical plant", these clues are aggregated into the same category through the multi-layer DBSCAN algorithm, representing environmental pollution-related events. At the same time, the algorithm dynamically adjusts the density parameters (eps and MmPts) to adapt to the different needs of high-density and low-density data.

[0062] The clustering results are used to generate a list of hot clues, such as: Environmental pollution (hotspot index: 85); Environmental pollution (hotspot index: 73).

[0063] This list format clearly shows the priority of different hot events, assisting public security departments in quickly handling them in a graded manner.

[0064] This embodiment achieves the following significant improvements through the coordinated application of technologies such as dynamic tag recognition, reinforcement learning to optimize clue generation, and multi-layer clustering analysis: In dynamic intelligence scenarios, the system can quickly adapt to user needs without retraining, significantly improving the efficiency of sensitive intelligence identification. The generated clue text is structured and precise, making it easy to use directly in actual analysis. The clustering algorithm optimizes the aggregation and statistics of hot events, and the generated results are more intuitive and practical. The overall system has low resource usage and fast response speed, providing efficient and accurate technical support for the daily decision-making of public security departments.

[0065] Embodiment 3: In the actual application of smart city environmental monitoring, a large amount of intelligence data comes from citizen reports, environmental monitoring equipment and Internet public opinion. Such data not only has diverse sources and complex formats, but also often contains unstructured text and redundant information. In order to quickly respond to sudden environmental events (such as industrial emission pollution), an efficient multi-source intelligence hot spot clue extraction method is needed. The following is a description of the implementation of the present invention in conjunction with specific scenarios.

[0066] For example, a city's environmental protection bureau receives more than tens of thousands of complaints every day, including pollution incidents reported by citizens through the 12345 hotline, real-time data collected by environmental monitoring equipment, and public opinion information on air and water quality issues on Internet platforms. This information may involve multiple pollution sources, such as chemical plants and industrial parks, but traditional methods cannot quickly locate key pollution incidents, resulting in delayed regulatory response.

[0067] The system first integrates the 12345 hotline text records, equipment data logs and Internet public opinion texts. Combining the generative big model with dynamic label configuration, the Environmental Protection Bureau can set sensitive field labels such as "sewage gas" and "industrial wastewater" according to current environmental hotspot needs (such as "industrial emissions" and "river pollution").

[0068] The labels entered by the user are embedded into the model tasks through prompts, and the model automatically filters out text that may contain relevant intelligence. For example, the text "Citizens report that a chemical plant emits pungent gas at night" or "The Internet exposes that industrial park sewage is directly discharged into the river" is marked as sensitive intelligence.

[0069] Use RsLoRA fine-tuning technology to optimize the screening process, ensuring that the model is adapted to different pollution scenarios and can be quickly adjusted without full retraining.

[0070] For the sensitive texts that have been screened, the system uses a generative large model optimized by reinforcement learning to extract event clues and generate a clear description of the environmental event in the structured form of "pollution source + pollution behavior". For example: Original text: "Some citizens reported that a chemical factory in the industrial park has been discharging sewage into the nearby river for a long time." Generate clue text: Pollution source: a chemical plant in an industrial park; Polluting behavior: long-term discharge of sewage into rivers.

[0071] The accuracy of the generated results is further verified through named entity recognition tools (such as LAC), and the model generation results are sorted to ensure that high-priority clues are processed first.

[0072] The system uses a semantic vector model to semantically vectorize all generated clue texts, mapping the text sequence into a high-dimensional vector representation. Subsequently, the clues are clustered using a multi-layer DBSCAN algorithm.

[0073] For example, multiple pollution clues about "industrial parks" were grouped together, indicating that similar problems existed in multiple factory sites.

[0074] The clustering parameters (e.g., eps=0.15, MmPts=4) are dynamically adjusted to adapt to pollution event data of different densities, which can aggregate large-scale similar events as well as capture a small number of highly similar events.

[0075] Based on the clustering results, the system generates a list of environmental hotspot events to provide a basis for the Environmental Protection Bureau to prioritize its handling. For example: Hotspot 1: Excessive emissions from industrial parks (heat index: 92), involving 5 factories.

[0076] Hotspot 2: River pollution (heat index: 88), involving 3 sewage discharge points.

[0077] Improved response speed: Dynamic tag configuration combined with generative large models enables rapid screening of sensitive intelligence. The EPA can adjust the model according to real-time needs and avoid the delays of traditional static rule models.

[0078] Precision information: The generated structured clue text highlights key pollution sources and behaviors, providing an intuitive basis for quickly locating problems.

[0079] Improved analysis accuracy: Through semantic vectors and dynamic multi-layer DBSCAN clustering algorithms, hotspot clues are clearly classified and closely related, effectively avoiding low-quality clustering results.

[0080] Strong implementation: This solution combines multiple data sources to provide a practical pollution clue mining and analysis process, which is suitable for real-time application in the field of environmental supervision.

[0081] In summary, this embodiment fully demonstrates the application value of the present invention in the smart city environmental monitoring scenario, which can significantly improve the response efficiency and judgment accuracy of sudden pollution incidents, and solve the problem of insufficient flexibility and accuracy of traditional methods.

[0082] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "screw-on" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0083] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.

Claims

1. A method for accurately extracting dynamic hot spot clues based on multi-source texts, characterized by: The method comprises the following steps: Step 1, obtaining multi-source intelligence text data; Step 2: dynamically adjust the labels according to the sensitive domain scenarios configured by the user, and build a sensitive intelligence text recognition model, which is trained by the following method: Using a large generative model and Prompt design, multiple sets of positive and negative samples are generated based on the sensitive scene labels that users are concerned about. Based on low-rank adaptive optimization technology, the generative large model is lightly fine-tuned to optimize its ability to recognize dynamic labels; Step 3: extract text belonging to sensitive areas from multi-source intelligence text data through a sensitive intelligence text recognition model; Step 4: Optimize the generative large model using a reinforcement learning algorithm for the extracted sensitive text data to generate structured clue text, where the clue text includes the event subject and event content; Step 5: Input the generated clue text into the semantic vector model for vectorization processing, and combine the multi-layer DBSCAN clustering algorithm to achieve clue aggregation; Step 6: Based on the clustering results, count the popularity of clue texts and generate a hot clue list for intelligence analysis and decision support.

2. According to the method of claim 1, the method is characterized in that: The training of the sensitive intelligence text recognition model specifically includes the following steps: Dynamically generate training samples based on user-configured sensitive domain scene labels to construct positive and negative samples; Dynamically insert the scene labels that users are concerned about into the prompt of the generative large model, and determine the sensitivity of the intelligence text through the text classification task; Using RsLoRA technology, the linear layer of the generative large model is decomposed into low rank, which enables lightweight fine-tuning by reducing the number of parameters, while improving the training efficiency and stability of the model.

3. The method for accurately extracting dynamic hot spot clues based on multi-source text according to claim 1 is characterized in that: The process of generating clue text comprises the following steps: Utilize the event summarization capability of the generative big model to generate event subjects and event content through prompts; For the generated clue text results, the model output results are sorted by combining the LAC named entity recognition tool and manual verification; Based on the sorted training data set, the generative large model is optimized through reinforcement learning to make the clue text it generates more in line with actual needs.

4. The method for accurately extracting dynamic hot spot clues based on multi-source text according to claim 1 is characterized in that: The semantic vector model adopts the bge-zh-base open source model to map the clue text sequence into a 768-dimensional semantic vector for representing the semantic features of the text.

5. The method for accurately extracting dynamic hot spot clues based on multi-source text according to claim 1 is characterized in that: The parameters of the multi-layer DBSCAN clustering algorithm are dynamically adjusted to the following three groups: 5.1 eps=0.15, MmPts=4; 5.2 eps=0.17, MmPts=2; 5.3 eps=0.1,MmPts=1; Through multi-level clustering processing, it can adapt to intelligence texts with low semantic similarity but large number of aggregations, and intelligence texts with high semantic similarity but small number of aggregations.

6. The method for accurately extracting dynamic hot spot clues based on multi-source text according to claim 1 is characterized in that: The clustered clue text heat statistics are achieved by: Calculate the average similarity between the text points and the center point in each cluster to evaluate the clustering quality; Generate a hot clue list based on the number of text clusters and the semantic relevance of the center point.

7. The method for accurately extracting dynamic hot spot clues based on multi-source text according to claim 1 is characterized in that: The hot clue list is automatically updated at preset time intervals, supporting users to perform trend analysis on clue popularity changes within different time ranges.

8. The method for accurately extracting dynamic hot spot clues based on multi-source text according to claim 1 is characterized in that: The output results of the method can be used in scenarios including but not limited to smart public security and government management to improve the efficiency of intelligence analysis and support rapid decision-making by relevant departments.

9. The method for accurately extracting dynamic hot spot clues based on multi-source text according to claim 1, characterized in that: The multi-source intelligence text data includes but is not limited to data from police reports, Internet public opinion, and government hotline sources.

Images