Traffic public opinion analysis method and device based on large model and electronic equipment

Abstract

The embodiment of the invention provides a traffic public opinion analysis method and device based on a large model and electronic equipment. The traffic public opinion analysis method and device are used for analyzing traffic public opinions and determining attributions and solutions. According to the embodiment of the invention, the electronic equipment identifies the traffic control or facility abnormity of the specific type through the plurality of preset intelligent agents, determines the corresponding identification information, determines the semantic similarity between the identification information output by each intelligent agent and the public opinion information, and inputs the identification information with high semantic similarity into the large language model. And the big language model is used for determining the problem attribution and solution, so that the public opinion analysis can be accurately and effectively carried out, and the problem attribution and solution can be determined.

Description

Technical Field

[0001] This application relates to the intersection of intelligent transportation systems and natural language processing, and in particular to a traffic sentiment analysis method, device, and electronic device based on a large model. Background Technology

[0002] With the acceleration of urbanization and the continuous development of intelligent transportation systems, the public's demands for the quality of traffic management services are increasing. At the same time, a large amount of public opinion information regarding traffic conditions is emerging through social media, government platforms, and hotline feedback channels, such as "unreasonable traffic light timings," "insufficient pedestrian crossing time," and "severe intersection congestion." These unstructured texts contain rich clues about abnormal traffic facility operations and have become an important data source for discovering potential traffic problems.

[0003] Currently, handling traffic-related public opinion mainly relies on manual screening and experience-based judgment, with manual determination of the cause set and solutions corresponding to the public opinion information. However, this method suffers from problems such as low response efficiency, strong subjectivity, and incomplete coverage. Although related technologies have attempted to introduce natural language processing techniques for keyword matching or simple classification, they generally suffer from insufficient semantic understanding capabilities, inability to accurately locate the cause of the fault, and lack of multi-dimensional collaborative analysis mechanisms. Summary of the Invention

[0004] This application provides a traffic public opinion analysis method, device, and electronic device based on a large model, which are used to analyze traffic public opinion and determine the causes and solutions.

[0005] In a first aspect, embodiments of this application provide a traffic public opinion analysis method based on a large model, the method comprising: Obtain traffic-related public opinion information; input the public opinion information into multiple preset intelligent agents, each intelligent agent is used to identify a specific type of traffic control or facility anomaly, and obtain the identification information output by each intelligent agent; wherein, the identification information includes at least the intersection involved in the public opinion information, the corresponding time period, and the traffic control information or traffic facility information associated with the intersection and time period; The semantic similarity between the public opinion information and the recognition information output by each agent is determined by using a pre-trained Sentence-Bi-Encoder Representations (SBERT) model. Obtain the identification information of each target with a semantic similarity higher than a threshold, input the identification information of each target into a large language model, and obtain the target answer output by the analysis of the large language model; wherein, the target answer includes at least the cause of the problem and the solution of the public opinion information determined based on the identification information of each target.

[0006] Secondly, embodiments of this application also provide a traffic public opinion analysis device based on a large model, the device comprising: The acquisition module is used to acquire traffic-related public opinion information; the public opinion information is input into multiple preset intelligent agents, each intelligent agent is used to identify a specific type of traffic control or facility anomaly, and the identification information output by each intelligent agent is acquired; wherein, the identification information includes at least the intersection involved in the public opinion information, the corresponding time period, and the traffic control information or traffic facility information associated with the intersection and time period; The processing module is used to determine the semantic similarity between the public opinion information and the recognition information output by each agent using a pre-trained SBERT model; obtain the recognition information of each target with a semantic similarity higher than a threshold; input the recognition information of each target into a large language model; and obtain the target answer output by the analysis of the large language model; wherein the target answer includes at least the cause of the problem and the solution of the public opinion information determined based on the recognition information.

[0007] Thirdly, embodiments of this application also provide an electronic device, which includes at least a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the steps of the traffic sentiment analysis method based on the large model as described.

[0008] In this embodiment, the electronic device identifies specific types of traffic control or facility anomalies through multiple preset intelligent agents, determines the corresponding identification information, and determines the semantic similarity between the identification information output by each intelligent agent and the public opinion information. The identification information with high semantic similarity is input into a large language model, which then determines the cause of the problem and the solution. This allows for accurate and effective public opinion analysis and the determination of the cause of the problem and the solution. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic diagram illustrating the process of a traffic public opinion analysis method based on a large model, provided in an embodiment of this application; Figure 2 A detailed schematic diagram illustrating a public opinion analysis process provided in this application embodiment; Figure 3This is a schematic diagram illustrating the process of acquiring target identification information provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a traffic public opinion analysis device based on a large model provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] To make the objectives and implementation methods of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments.

[0012] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0013] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0014] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0015] The term "module" refers to any known or subsequently developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware and / or software code that is capable of performing the functions associated with that element.

[0016] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0017] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the above exemplary discussion is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the principles and practical applications, thereby enabling those skilled in the art to better utilize the described embodiments and various different variations of embodiments suitable for specific use considerations.

[0018] To conduct accurate and effective public opinion analysis, embodiments of this application provide a traffic public opinion analysis method, apparatus, and electronic device based on a large model.

[0019] This traffic public opinion analysis method based on a large model includes: acquiring traffic-related public opinion information; inputting the public opinion information into multiple pre-set agents, each agent being used to identify specific types of traffic control or facility anomalies, and acquiring the identification information output by each agent; wherein the identification information includes at least the intersection involved in the public opinion information, the corresponding time period, and the traffic control information or traffic facility information associated with the intersection and time period; determining the semantic similarity between the public opinion information and the identification information output by each agent using a pre-trained SBERT model; acquiring the identification information of each target with a semantic similarity higher than a threshold, inputting each target identification information into a large language model, and obtaining the target answer output by the large language model analysis; wherein the target answer includes at least the cause of the problem in the public opinion information and the solution determined based on each identification information.

[0020] Figure 1 A schematic diagram illustrating a traffic public opinion analysis method based on a large model, provided for embodiments of this application, includes the following steps: S101: Obtain traffic-related public opinion information; input the public opinion information into multiple preset intelligent agents, each intelligent agent is used to identify a specific type of traffic control or facility anomaly, and obtain the identification information output by each intelligent agent; wherein, the identification information includes at least the intersection involved in the public opinion information, the corresponding time period, and the traffic control information or traffic facility information associated with the intersection and time period.

[0021] The traffic sentiment analysis method based on a large model provided in this application is applied to electronic devices, such as PCs or servers.

[0022] To conduct accurate and effective public opinion analysis, electronic devices can first acquire traffic-related public opinion information. This information includes text data from social media, news reports, public complaint platforms, or traffic radio broadcasts, containing descriptive content about traffic conditions. For example, the public opinion information could be "Congestion at the intersection of Street A and Street B".

[0023] After acquiring public opinion information, the electronic device can input the information in parallel into multiple pre-configured agents. Each agent performs specialized modeling for specific types of traffic anomalies, including but not limited to: traffic light malfunctions, temporary traffic control, road closures, road construction, missing signs and markings, damaged guardrails, or sudden congestion. Each agent, based on its corresponding semantic recognition model and domain knowledge base, performs fine-grained analysis of the input public opinion information, extracts key elements related to the target anomaly, and outputs corresponding identification information. The output identification information may include: the name or geographical coordinates of the intersection involved in the public opinion information, the time period of the event, the anomaly type classification, and traffic control information or traffic facility information associated with the intersection and time period. Traffic control information includes signal timing schemes, release modes, or control status; traffic facility information includes facility type, deployment status, or damage condition.

[0024] S102: Using a pre-trained SBERT model, determine the semantic similarity between the public opinion information and the recognition information output by each agent.

[0025] In this embodiment, the electronic device can determine the semantic similarity between public opinion information and the identification information output by each intelligent agent using a pre-trained SBERT model. This semantic similarity characterizes the degree of semantic relevance between public opinion information and specific traffic anomaly events, reflecting whether the public opinion supports, describes, or corroborates the existence of a certain type of traffic control anomaly or facility anomaly.

[0026] In one possible implementation, the SBERT model can be fine-tuned based on a large-scale traffic domain corpus to enhance its ability to understand traffic-specific terminology, colloquial expressions, and contextual information, thereby improving the accuracy and robustness of semantic matching.

[0027] S103: Obtain the identification information of each target with a semantic similarity higher than a threshold, input the identification information of each target into a large language model, and obtain the target answer output by the analysis of the large language model; wherein, the target answer includes at least the cause of the problem and the solution of the public opinion information determined based on the identification information of each target.

[0028] The electronic device can acquire target identification information with semantic similarity higher than a preset threshold. This target identification information with semantic similarity higher than the preset threshold represents candidate results of abnormal events highly correlated with the public opinion information. For example, the preset threshold can be dynamically adjusted according to the actual application scenario to filter low-relevance identification results and ensure that the information input to the large language model has sufficient semantic confidence. That is, in this embodiment, the electronic device uses a fine-tuned SBERT model to calculate text similarity and filter irrelevant information. For example, the threshold can be 0.3.

[0029] Each target identification information is input into a large language model, which uses its contextual understanding and knowledge reasoning capabilities to conduct in-depth analysis of abnormal events. The large language model generates and outputs the target answer, which includes at least the root cause of the public opinion problem and the solution determined based on the target identification information. In one possible implementation, the target answer may also include the possible scope of the public opinion information and the relevant responsible units or recommended response procedures.

[0030] In one possible implementation, the identification information can be sorted by combining the LambdaRankMultipleAdditiveRegressionTrees (LambdaMART) model with public opinion information, thereby improving the reliability of the large language model's response. This method particularly relates to a dynamic screening technique for multi-source data for traffic public opinion response.

[0031] In public opinion analysis within related technologies, general-purpose large language models lack knowledge of the transportation domain, making it difficult to identify key information from massive amounts of data, leading to redundant responses or decision-making delays. Furthermore, large language models cannot acquire dynamic data (such as traffic flow, congestion index, large-scale events, etc.), limiting their application effectiveness. Transportation system data is complex in dimensions, and the acquired data may not be directly related to public opinion information. Traditional keyword matching methods cannot dynamically adapt to the semantic context of public opinion issues, while existing public opinion systems mostly rely on static rules and cannot dynamically respond to real-time changes in traffic conditions. This application proposes a method that combines traditional business processes with the capabilities of large models, using multiple intelligent agents to acquire relevant data and assist human decision-making. This method significantly saves users time in querying system data and avoids the tedious process of manually analyzing causes from massive amounts of information. To address the problem that redundant information extracted by intelligent agents may interfere with the large model's ability to pinpoint the specific causes of public opinion, a re-sorting method combining static text descriptions and dynamic data is proposed. This involves calculating the semantic similarity between the identification information output by each intelligent agent and the public opinion information, and then filtering the identification information based on semantic similarity. The traffic public opinion processing method based on multi-agent collaboration and reordering optimization provided in this application allows a large language model to provide data support for responding to and answering public opinion information in conjunction with professional traffic domain systems, and provides high-quality data through filtering and ranking models. Specifically, this method coordinates multiple agents with the large language model and business system to obtain various indicators when public opinion occurs (such as signal timing, equipment failure, laws and regulations, etc.), and uses an optimized reordering algorithm to filter irrelevant information and rank the importance of the information. Finally, the filtered information is input into the large model to generate response content, assisting users in efficiently handling traffic public opinion complaints or suggestions from citizens.

[0032] In this embodiment, the electronic device identifies specific types of traffic control or facility anomalies through multiple preset intelligent agents, determines the corresponding identification information, and determines the semantic similarity between the identification information output by each intelligent agent and the public opinion information. The identification information with high semantic similarity is input into a large language model, which then determines the cause of the problem and the solution. This allows for accurate and effective public opinion analysis and the determination of the cause of the problem and the solution.

[0033] In order to conduct accurate and effective public opinion analysis, based on the above embodiments, in this application embodiment, the plurality of preset intelligent agents include intersection static parameter intelligent agents, real-time status intelligent agents, duty execution intelligent agents, indicator evaluation intelligent agents and other intelligent agents; The static parameter agent is used to determine whether the intersection involved in the public opinion information is equipped with a safety navigation device, and to detect whether the working status and settings of the safety navigation device comply with relevant national standards. The real-time status agent is used to determine whether there is human intervention or manual adjustment of the signal control system at the intersection involved in the public opinion information during the time period reflected by the public opinion information. The duty execution intelligent agent is used to determine whether the intersection corresponding to the public opinion information is located within the preset traffic coordination control route or regional linkage control range; The indicator evaluation agent is used to evaluate the traffic indicators of the intersections corresponding to the public opinion information.

[0034] The multiple preset intelligent agents in this application embodiment include: intersection static parameter intelligent agent, real-time status intelligent agent, duty execution intelligent agent, indicator evaluation intelligent agent and other extended function intelligent agents. Each intelligent agent performs professional modeling and anomaly identification for traffic management data of a specific dimension.

[0035] For example, the intersection static parameter intelligent agent is used to identify the infrastructure configuration of the intersection involved in the public opinion information. Specifically, it includes determining whether the intersection has deployed safety navigation devices (such as electronic police, traffic lights, guidance screens, radar detectors, etc.), and further detecting whether the working status of the safety navigation devices is normal, whether the installation location is compliant, and whether the parameter settings comply with relevant national or industry standards.

[0036] The real-time status agent is used to analyze the operation mode of the signal control system at the intersection corresponding to the public opinion information within the time period reflected by the public opinion information, and to determine whether there is any unplanned human intervention, manual adjustment operation or temporary release strategy. For example, the real-time status agent can also combine the central control system log to confirm the source and legality of the control command.

[0037] The duty execution intelligent agent is used to determine whether the intersection corresponding to the public opinion information is within the preset traffic coordination control route, key protection route, regional linkage control zone, or special duty control range, and then assess whether the current traffic status conflicts with the established duty plan. Key protection routes include guard routes, emergency lanes, etc., and regional linkage control zones include green wave zones, tidal control zones, etc.

[0038] The indicator evaluation agent is used to calculate and evaluate key traffic operation indicators of the intersection corresponding to the public opinion within a specified time period based on multi-source traffic perception data (such as checkpoints, loop detectors, videos, floating cars, etc.), including but not limited to: traffic delay time, queue length, saturation, average vehicle speed, congestion index and service level, and to determine whether each indicator exceeds the normal threshold range.

[0039] In one possible implementation, other functional extension agents include, but are not limited to, construction road occupation identification agents, weather impact analysis agents, and public feedback aggregation agents, to supplement the identification of potential traffic anomalies caused by external factors.

[0040] Figure 2 This is a detailed schematic diagram illustrating a public opinion analysis process provided in an embodiment of this application.

[0041] Depend on Figure 2 It can be seen that the input public opinion information can be obtained first, and then the public opinion information can be input into the intersection static parameter intelligent agent, real-time status intelligent agent, and duty execution intelligent agent through the intelligent agent scheduling center. The recognition information output by each intelligent agent is put into the data pool, and the recognition information is filtered and reordered. The target recognition information of each target after filtering and reordering is input into the large language model, and the large language model generates the target answer of the response.

[0042] The above-mentioned technical solution has the following advantages or beneficial effects: using multiple intelligent agents to identify specific types of traffic control or facility anomalies can improve the accuracy of identification information determination, thereby accurately and effectively determining the target answer.

[0043] To accurately and effectively determine semantic similarity, based on the above embodiments, in this embodiment, determining the semantic similarity between the public opinion information and the recognition information output by each agent using a pre-trained SBERT model includes: The public opinion information and the recognition information output by each intelligent agent are respectively input into the pre-trained SBERT model to obtain the public opinion vector of the public opinion information output by the SBERT model and the recognition vector of the recognition information output by each intelligent agent. Based on the cosine similarity between the public opinion vector and each identification vector, the semantic similarity between the public opinion information and the identification information output by each intelligent agent is determined.

[0044] In this embodiment, the electronic device can use a pre-trained SBERT model to semantically encode the public opinion information and the identification information output by each agent, mapping the text content to a unified high-dimensional semantic vector space. In one possible implementation, the public opinion information and the identification information output by each agent can be input into the pre-trained SBERT model, and the SBERT model can be used to perform deep semantic encoding on the text. The SBERT model outputs the public opinion semantic vector corresponding to the public opinion information (referred to as the public opinion vector in this embodiment) and the identification semantic vector corresponding to each piece of identification information (referred to as the identification vector in this embodiment).

[0045] In a unified semantic vector space, the cosine similarity between the sentiment vector and each identification vector is calculated as a quantitative indicator to measure their semantic relevance. For example, the cosine similarity satisfies the following formula:

[0046] in, For semantic similarity, As a public opinion vector, For recognition vectors, , These are the magnitudes of the sentiment vector and the identification vector, respectively, where the magnitude represents the length of the vector in space. , where n is the number of components contained in the sentiment vector and semantic vector, and i is the component number of the sentiment vector and semantic vector. Let i be the i-th component of the public opinion vector. For the i-th component of the identification vector, It is the sum of the product of the components of the public opinion vector and the semantic vector in the same direction.

[0047] Based on the calculation results, electronic devices can determine the semantic similarity between public opinion information and the identification information output by each intelligent agent. This semantic similarity reflects the degree of matching between public opinion content and specific traffic anomalies at the semantic level.

[0048] The above technical solution has the following advantages or beneficial effects: it uses a fine-tuned SBERT model to calculate the semantic similarity between public opinion issues and data items, and uses the cosine similarity formula to filter highly relevant information, thereby accurately and effectively filtering out highly relevant identification information and improving the accuracy of public opinion analysis.

[0049] To accurately and effectively determine semantic similarity, based on the above embodiments, in this embodiment, the SBERT model is trained in the following manner: Get any pair of samples in the sample set, and the sample similarity saved for that pair, where the pair contains two texts; Input the two texts contained in the sample pair into the SBERT model respectively, and obtain the two text vectors output by the SBERT model; The recognition similarity is determined based on the cosine similarity of the two text vectors; the SBERT model is trained based on the deviation between the recognition similarity and the sample similarity.

[0050] To accurately and effectively determine semantic similarity, in this embodiment, an SBERT model can be trained. For training the SBERT model, a sample set is pre-stored. This sample set can be stored on an electronic device or on other devices connected to the electronic device. The sample set contains multiple sample pairs, each containing two texts. These sample pairs are generated based on the actual content of responses from on-site tuning personnel to citizens' replies, combined with manual adjustments. Furthermore, the sample similarity score for each sample pair is stored.

[0051] For example, positive and negative samples of a certain public opinion-related issue can be generated by combining the response content generated by the intelligent agent with the response content of the tuning engineer on site. Positive samples indicate high relevance to the issue with a similarity between 0.8 and 1, while negative samples indicate low relevance to the issue with a similarity between 0 and 0.3. For example, the sample similarity between "long waiting time for vehicles crossing the street at night" and "induction control at the intersection is activated at night, but the loop detector at the intersection is malfunctioning" is 0.9; the sample similarity between "yellow flashing traffic light at the intersection" and "traffic signal malfunction - high voltage at the intersection" is 0.95; the sample similarity between "increased congestion during the morning rush hour" and "traffic flow at the intersection from 7:00 to 8:30 is higher than the same period last week" is 0.85; the sample similarity between "traffic congestion on a certain avenue after 8 pm" and "a large event on this road at 8:00 pm" is 0.85; the sample similarity between "long waiting time for pedestrians crossing the street at the intersection" and "detector malfunction - abnormal pedestrian detector data" is 0.8; and the sample similarity between "longer waiting time for red lights at the intersection" and "the intersection is in a high-level traffic control zone at this time" is 0.9. The similarity between the samples "insufficient pedestrian crossing time in the south-to-north direction of the intersection" and "early red light shutdown in the south-to-north direction of the intersection" is 0.8; the similarity between the samples "insufficient pedestrian crossing time in the intersection" and "traffic detector malfunction" is 0.2; the similarity between the samples "traffic light timing problem" and "road construction impact" is 0.1; the similarity between the samples "green wave coordination failure" and "weather-related road congestion" is 0.3; the similarity between the samples "short pedestrian signal timing at the north gate of the intersection during weekday morning rush hour" and "no traffic control intervention during morning rush hour at the intersection" is 0; the similarity between the samples "short pedestrian signal timing at the north gate of the intersection during weekday morning rush hour" and "fixed-cycle mode of morning rush hour execution at the intersection" is 0.

[0052] The electronic device can input the two texts in the sample pair into the initial SBERT model, extract their deep semantic representations through the model's dual-tower encoding structure, and obtain the corresponding text vectors, namely the first text vector and the second text vector. Based on the obtained two text vectors, the cosine similarity between them in the vector space is calculated as the recognition similarity determined by the SBERT model, which reflects the automatic judgment result of the SBERT model on the semantic relevance of the two texts. The deviation between the recognition similarity and the labeled sample similarity is compared. In one possible implementation, the deviation can be quantified by mean squared error (MSE), Kullback-Leibler (KL) divergence, or other loss functions. With the goal of minimizing the deviation, the parameters of the SBERT model are updated by backpropagation, thereby achieving supervised training of the SBERT model.

[0053] Repeat the above process until the semantic matching performance of the SBERT model on the validation set converges or reaches the preset number of training rounds, thus obtaining the SBERT model adapted and optimized for the transportation domain.

[0054] The above technical solution has the following advantages or beneficial effects: by using the method provided in the embodiments of this application, the trained SBERT model can further improve the accuracy of similarity calculation, thereby accurately and effectively performing semantic analysis.

[0055] This embodiment employs the SBERT Siamese network structure and fine-tunes it using traffic-related public opinion corpus to generate text vectors. The SBERT model directly outputs sentence vectors through a shared-parameter encoder and pooling layer, significantly improving computational efficiency (time complexity of Om). Compared to the standard BERT model, which requires pairwise sentence similarity calculation (complexity of Om²), this model is more efficient. Fine-tuning using traffic-related public opinion data further improves the accuracy of similarity calculation. The SBERT model uses the paraphrase-multilingual-MiniLM-L12-v2 pre-trained model as a baseline for fine-tuning. For example, to optimize the model's similarity regression capability, the sample similarity data is normalized to the range of 0 and 1. The purpose of fine-tuning in this embodiment is to make the model's similarity concept in the public opinion domain more explicit, avoiding mismatches between the general model and actual business during classification. For instance, citizens report that the arrow lights at a certain intersection are not lit during off-peak hours, only the circular lights are lit. This is related to the low off-peak traffic flow at the intersection and the use of mixed traffic release to improve efficiency. The training samples were generated based on the actual content of responses from citizens to on-site optimization personnel, combined with manual adjustments.

[0056] To perform semantic analysis accurately and effectively, based on the above embodiments, in this embodiment, the step of inputting the target recognition information of each target into a large language model to obtain the target answer output by the large language model includes: The target identification information is sorted according to semantic similarity from highest to lowest; The sorted target identification information and the corresponding semantic similarity of each target identification information are input into the large language model, so that the large language model performs differential in-depth analysis on each target identification information based on the semantic similarity. If the semantic similarity of the first target identification information is greater than that of the second target identification information, then the analysis depth of the first target identification information is higher than that of the second target information. The target answer output by the analysis of the large language model is obtained.

[0057] In this embodiment, the electronic device can sort the target identification information from high to low semantic similarity to ensure that the identification information with higher relevance to the original public opinion information is given priority in the analysis process. Each sorted target identification information and its corresponding semantic similarity value are then input into a large language model, guiding the model to dynamically adjust the analysis granularity and inference depth based on the similarity values.

[0058] The large language model employs a differentiated deep analysis strategy based on semantic similarity: For target identification information with high semantic similarity, the large language model performs fine-grained analysis, including inferring the root cause of the event, the scope of its associated impact, potential security risks, and evolutionary trends, and generates targeted suggestions by combining historical handling cases; while for information with relatively low semantic similarity, a rapid classification and preliminary judgment mechanism is used, outputting only basic attribute identification results or marking them as low-relevance items to be verified. The model then obtains the target answer output by the large language model after analysis.

[0059] The above-mentioned technical solution has the following advantages or beneficial effects: by using the method provided in the embodiments of this application, the large language model can perform differential analysis based on the similarity of each target recognition information, thereby focusing on analyzing more important recognition information and improving the accuracy of the target answer.

[0060] To accurately and effectively conduct public opinion analysis, based on the above embodiments, in this embodiment, after determining the semantic similarity between the public opinion information and the recognition information output by each agent using a pre-trained SBERT model, and before obtaining the recognition information of each target with a semantic similarity higher than a threshold, the method further includes: If a certain identification information includes time information or location information, then an attenuation factor is determined based on the time interval between the included time information and the occurrence time of the public opinion information, or based on the distance between the included location information and the occurrence location of the public opinion information; wherein, the larger the time interval, the smaller the determined attenuation factor, and the larger the distance, the smaller the determined attenuation factor; the semantic similarity corresponding to the identification information is updated by multiplying the attenuation factor and the semantic similarity corresponding to the identification information. For the updated semantic similarity, perform the subsequent step of obtaining the recognition information of each target with a semantic similarity higher than the threshold.

[0061] Since time and location have a significant impact in traffic scenarios, the determined similarity can be adjusted by combining time and location in this embodiment.

[0062] In one possible implementation, if a certain identification information contains time information or location information, the corresponding attenuation factor can be calculated based on the time interval between the time information and the time when the public opinion information occurred, or the corresponding attenuation factor can be calculated based on the spatial distance between the location information and the location where the public opinion information occurred.

[0063] The larger the time interval in the time dimension, the weaker the temporal correlation between the identification information and the current event, and the smaller the determined decay factor. In one possible implementation, the decay factor can be in the range of (0,1], and can be modeled by exponential decay function, piecewise linear function or Gaussian kernel, etc.

[0064] For example, a time-determined decay factor can satisfy the following formula:

[0065] in, The determined attenuation factor, For time intervals, Using 3600 as the base time, k_t and a_t are preset values, such as k_t = 1.3 and a_t = 0.005, t_0 = 3600. λ_t has different values ​​for different agents; the larger λ_t is, the faster the time decays in the long term. For example, for a 1 × 1 service execution agent... 0.0001, the intersection static parameter agent is 5 × 1 0.00005, for a real-time state agent, is 0.0005 5 × 1 .

[0066] The greater the geographical distance in the spatial dimension, the lower the spatial correlation between the identification information and the location of the incident. The corresponding attenuation factor decreases accordingly and is also mapped to the (0,1] interval. It can be measured by combining road network topology distance or Euclidean distance.

[0067] For example, the attenuation factor determined based on distance satisfies the following formula:

[0068] in, As the attenuation factor, d is the preset baseline distance, d is the distance between the included location information and the location where the public opinion information occurred, and k is the preset attenuation index.

[0069] In one possible implementation, if the identification information includes time information and location information, a target attenuation factor can be determined based on a first attenuation factor determined based on the time information and a second attenuation factor determined based on the location information. For example, the average of the first and second attenuation factors can be used to determine the target attenuation factor.

[0070] After calculating the attenuation factor, the product of the attenuation factor and the original semantic similarity can be deducted, and the semantic similarity of the identified information can be updated using this product. This achieves dynamic calibration of the correlation of information across time and space, and suppresses the risk of mismatch caused by time misalignment or position deviation.

[0071] After completing the similarity update of all identification information, the information is reordered based on the updated semantic similarity, and target identification information with a value higher than the preset threshold is selected as a candidate set for subsequent input into the large language model for in-depth analysis. This ensures that the information ultimately used for evaluation has a high degree of consistency and business credibility in the three dimensions of semantics, time, and space.

[0072] The above-mentioned technical solution has the following advantages or beneficial effects: In this application embodiment, the attenuation factor is determined based on time and / or location, and the semantic similarity is updated based on the attenuation factor, so as to realize the dynamic calibration of the relevance of cross-temporal and spatial information, suppress the risk of mismatch caused by time misalignment or location deviation, and thus accurately and effectively conduct public opinion analysis.

[0073] To accurately and effectively conduct public opinion analysis, based on the above embodiments, in this embodiment of the application, after determining the semantic similarity between the public opinion information and each identified information using a pre-trained SBERT model, and before acquiring each target identified information with a semantic similarity higher than a threshold, the method further includes: The system obtains the parameter values ​​for assessing traffic conditions contained in the identification information output by the indicator assessment agent; wherein the parameter values ​​for assessing traffic conditions include traffic flow, speed, and travel time ratio; and determines the corresponding influencing factor based on the ratio of each type of parameter value to a preset benchmark parameter value, wherein the larger the determined ratio, the larger the influencing factor. The semantic similarity corresponding to the recognition information output by the indicator evaluation agent is updated by using the product of the attenuation factor and the semantic similarity corresponding to the recognition information output by the indicator evaluation agent. For the updated semantic similarity, perform the subsequent step of obtaining the recognition information of each target with a semantic similarity higher than the threshold.

[0074] In this embodiment of the application, the electronic device can obtain the traffic status parameter values ​​contained in the identification information output by the indicator evaluation intelligent agent. The obtained parameter values ​​include traffic flow, speed, and travel time ratio, which are used to reflect the actual operating status of the target intersection within the corresponding time period of public opinion.

[0075] For each type of parameter value, the ratio of that type of parameter value to the preset benchmark parameter value of that type can be calculated, and the corresponding influencing factor can be determined based on the ratio. For the ratio of traffic flow to travel time, the larger the value, the higher the traffic pressure, and the corresponding influencing factor increases accordingly. For speed, the lower the value, the higher the congestion level usually is. Therefore, its reciprocal or normalized value is used in the ratio calculation to ensure that the influence trend is consistent. The influencing factor is used to quantify the degree to which the current operating status of the intersection deviates from the normal level. This influencing factor reflects the severity of traffic anomalies and is introduced as a positive weighting term with a similarity correction mechanism.

[0076] Multiply the aforementioned influencing factors by the obtained semantic similarity to update the semantic similarity corresponding to the recognition information output by the indicator evaluation agent. In this way, the comprehensive relevance score of recognition results that not only have high semantic matching degree but also correspond to significantly abnormal traffic indicators can be improved.

[0077] After the update is completed, a multi-dimensional joint adjustment is performed in conjunction with the spatiotemporal decay factor to form the final comprehensive similarity score. Subsequently, the similarity scores are re-ranked based on the updated scores, and target recognition information with scores above a preset threshold is selected as a high-confidence candidate set for subsequent input into a large language model for in-depth analysis.

[0078] For example, if the identification information output by the indicator evaluation agent contains time information or location information, the attenuation factor can be determined in the above manner. After determining the product of the influencing factor and the corresponding semantic similarity, the target product of the attenuation factor and the product is determined, and the semantic similarity is updated using the target product.

[0079] The above-mentioned technical solution has the following advantages or beneficial effects: In this embodiment of the application, the influencing factors are determined based on the parameter values ​​of traffic conditions, and the semantic similarity is updated based on the influencing factors, thereby improving the comprehensive relevance score of the identification results that not only have high semantic matching degree but also correspond to significantly abnormal traffic indicators, thus accurately and effectively conducting public opinion analysis.

[0080] Figure 3 This is a schematic diagram illustrating the process of acquiring target identification information as provided in an embodiment of this application.

[0081] Depend on Figure 3 It is understood that the raw data output by multiple intelligent agents can be obtained first, namely the recognition information described in the embodiments of this application. Semantic similarity, decay factor and influence factor are determined simultaneously. Specifically, semantic similarity is calculated by SBERT model and cosine similarity. The decay factor is determined by the timeliness and distance weight module and the decay factor is determined by the influence factor module. The updated semantic similarity is determined based on feature fusion. The LambdaMART model is used for re-sorting to select multiple highly relevant datasets with high semantic similarity, namely target recognition information. For example, the Top-K highly relevant datasets can be selected.

[0082] In this embodiment, the LambdaMART model is used to perform global optimal ranking of the filtered information, comprehensively considering multiple dimensions such as text similarity, indicator influence factors, timeliness and distance deviation, to improve the accuracy and reliability of the ranking results.

[0083] To conduct accurate and effective public opinion analysis, based on the above embodiments, in this embodiment, determining the corresponding influence factor according to the ratio of each type of parameter value to a preset benchmark parameter value includes: The determined impact factors satisfy the following formula:

[0084] in, Let be the determined influencing factor, n be the number of traffic state parameter types, and i be the type number of the traffic state parameter. Let be the parameter value for the i-th traffic state parameter. Let be the baseline value of the i-th traffic state parameter. and These are the preset values ​​corresponding to the i-th traffic state parameter. is the parameter threshold for the i-th traffic state parameter.

[0085] To conduct accurate and effective public opinion analysis, electronic devices can determine the impact of traffic parameter deviations on public opinion information. For example, if a user inputs public opinion related to the reduced efficiency of green wave traffic, the indicator evaluation agent will output information such as traffic flow deviation, travel time ratio, and speed indicators for that artery.

[0086] The influence factor of this agent's information is calculated as follows:

[0087] in, Let be the determined influencing factor, n be the number of traffic state parameter types, and i be the type number of the traffic state parameter. The parameter values ​​for the i-th type of traffic state (such as flow rate, travel time ratio, and speed) are the parameter values ​​in the recognition information output by the intelligent agent. This is the baseline value for the i-th traffic state parameter (such as the travel time under design flow and free-flow speed). If there is no design value, the historical average value can be used. The parameter threshold for the i-th traffic state parameter is determined by its historical average value; The basic weight of the i-th traffic state parameter (reflecting the importance of the parameter, calibrated, such as flow rate, speed, the sum of the weights of all indicators is 1); This is a nonlinear adjustment coefficient (to control the amplification effect when the threshold is exceeded, such as γ=0.2).

[0088] In one possible implementation, It can be dynamically updated according to a preset cycle, which is no more than 24 hours. This can be determined based on a threshold: the 85th percentile of historical data (congestion trigger boundary). For example, positive indicators (i.e., traffic flow, travel time ratio, etc.) can be used. =1.2 Negative indicators (i.e., speed) are taken =0.8 For example, traffic state parameters such as flow rate. The value is 0.4, and the travel time is higher than that of this type of traffic state parameter. The value is 0.3, which is a type of traffic state parameter called average speed. The value is 0.3; positive indicators (i.e., traffic volume, travel time ratio, etc.). =0.2~0.5, negative indicator (i.e., speed) =0.3~0.6. It is understandable that the influencing factor for a negative indicator (speed) satisfies the following formula:

[0089] It is understandable that for traffic state parameters such as flow rate, exceeding the design flow rate positively amplifies the congestion effect. Therefore, when determining influencing factors, based on... To determine; for this type of traffic state parameter, the travel time ratio, The larger the value, the lower the traffic efficiency. Therefore, when determining the influencing factor, based on... To determine this, for traffic state parameters such as speed, a lower speed generally indicates more severe congestion, i.e., an inverse relationship. Based on this, when determining influencing factors, [the following factors should be considered]. To be confirmed.

[0090] The above-mentioned technical solution has the following advantages or beneficial effects: the method provided in the embodiments of this application can accurately and effectively determine the influencing factors, thereby improving the accuracy of public opinion analysis.

[0091] In this embodiment, after semantic similarity calculation using the BERT model and weighted adjustment using a decay factor constructed from time intervals and spatial distances, the recognition results output by each agent have been transformed into quantifiable numerical features. To further improve the ranking accuracy of highly correlated abnormal events, this scheme introduces a LambdaMART-based learning ranking model to comprehensively re-rank candidate recognition information. LambdaMART is a ranking learning algorithm based on gradient boosting trees, which can effectively integrate multi-dimensional heterogeneous features and support the handling of missing feature values. It is suitable for the actual situation in the transportation field where some agents' output information is incomplete or lacks quantitative indicators. Model training uses the LightGBM framework to achieve efficient iteration, with Normalized Discounted Cumulative Gain (NDCG) as the core evaluation metric to measure the degree of matching between the ranking results and the manual annotation standard.

[0092] The input features in the training samples include the following dimensions: Cosine similarity: represents the degree of semantic matching between public opinion text and identified information, with a value range of [0,1], denoted as semantic_similarity; Does it contain indicator information? This indicates whether the identification information contains quantifiable traffic operation indicators (such as traffic flow, congestion index, signal timing deviation, etc.). It is a binary variable, where 0 indicates no and 1 indicates yes. The corresponding field is has_indicator_metric. Does it have a time context? This indicates whether it contains explicit time information (such as the time period, duration, etc.) and is used to determine spatiotemporal consistency. 0 indicates no, and 1 indicates yes. The corresponding field is has_time_context. Does it have a distance context? This indicates whether it contains geographic location or spatial association information with the intersection where the incident occurred. 0 indicates no, and 1 indicates yes. The corresponding field is has_distance_context. The comprehensive influencing factor of the indicators reflects the degree of deviation of current traffic state parameters (such as the rate of decrease in average speed and the increase in travel time ratio) from the benchmark value. It is generated by normalization or nonlinear mapping and reflects the severity of the anomaly. It is denoted as indicator_deviation. The influence weight of the decay influence factor of a single indicator, calculated separately, can be used for fine-grained attribution analysis and can be used as an extended feature for future reference. Relevance labels: Manual annotations of each sample by experts or on-site managers based on practical business experience, using a four-level scoring system: 0 indicates no relevance, 1 indicates weak relevance, 2 indicates moderate relevance, and 3 indicates strong relevance, which constitute the target ranking labels for supervised learning.

[0093] For cases where certain agents do not output specific numerical indicators (e.g., only returning "signal light malfunctions" without fault probability or performance data), the corresponding feature fields are represented by NA (NaN, Not a Number) to indicate missing values. The LightGBM model has a built-in default value handling mechanism that can automatically learn the optimal splitting direction to ensure model robustness.

[0094] By combining the aforementioned multi-dimensional feature system with the ranking learning framework, the system can further integrate key factors such as traffic operation status, spatiotemporal context integrity, and anomaly severity on top of semantic matching, generating priority ranking results that are closer to actual management needs. Finally, only the top-ranked targets with a comprehensive similarity higher than a preset threshold are retained and input into a large language model for in-depth attribution analysis and response suggestions, achieving intelligent judgment throughout the entire process from "coarse screening" to "fine ranking" and then to "in-depth analysis."

[0095] Figure 4 A schematic diagram of a traffic public opinion analysis device based on a large model, provided as an embodiment of this application, is shown below. Figure 4 As shown: The device includes an acquisition module 401 and a processing module 402.

[0096] The acquisition module 401 is used to acquire traffic-related public opinion information; input the public opinion information into multiple preset intelligent agents, each intelligent agent is used to identify a specific type of traffic control or facility anomaly, and acquire the identification information output by each intelligent agent; wherein, the identification information includes at least the intersection involved in the public opinion information, the corresponding time period, and the traffic control information or traffic facility information associated with the intersection and time period; Processing module 402 is used to determine the semantic similarity between the public opinion information and the recognition information output by each agent through a pre-trained SBERT model; obtain the recognition information of each target with a semantic similarity higher than a threshold, input the recognition information of each target into a large language model, and obtain the target answer output by the analysis of the large language model; wherein, the target answer includes at least the cause of the problem and the solution of the public opinion information determined based on the recognition information of each agent.

[0097] In one possible implementation, the processing module 402 is specifically used to input the public opinion information and the recognition information output by each intelligent agent into a pre-trained SBERT model, respectively, to obtain the public opinion vector of the public opinion information output by the SBERT model and the recognition vector of the recognition information output by each intelligent agent; and to determine the semantic similarity between the public opinion information and the recognition information output by each intelligent agent based on the cosine similarity between the public opinion vector and each recognition vector.

[0098] In one possible implementation, the processing module 402 is further configured to train the SBERT model by: acquiring any sample pair in the sample set, and the sample similarity stored for the sample pair, wherein the sample pair contains two texts; inputting the two texts contained in the sample pair into the SBERT model respectively, and acquiring two text vectors output by the SBERT model; determining the recognition similarity based on the cosine similarity of the two text vectors; and training the SBERT model based on the deviation between the recognition similarity and the sample similarity.

[0099] In one possible implementation, the processing module 402 is specifically configured to sort each target identification information according to semantic similarity from largest to smallest; input each sorted target identification information and its corresponding semantic similarity into a large language model, so that the large language model performs differentiated in-depth analysis on each target identification information based on the semantic similarity, wherein if the semantic similarity of the first target identification information is greater than that of the second target identification information, then the analysis depth of the first target identification information is higher than that of the second target information; and obtain the target answer output by the analysis of the large language model.

[0100] In one possible implementation, the processing module 402 is further configured to, if a certain identification information includes time information or location information, determine an attenuation factor based on the time interval between the included time information and the occurrence time of the public opinion information, or based on the distance between the included location information and the occurrence location of the public opinion information; wherein, the larger the time interval, the smaller the determined attenuation factor, and the larger the distance, the smaller the determined attenuation factor; update the semantic similarity corresponding to the identification information by using the product of the attenuation factor and the semantic similarity corresponding to the identification information; and for the updated semantic similarity, execute the subsequent step of obtaining each target identification information with a semantic similarity higher than a threshold.

[0101] In one possible implementation, the processing module 402 is further configured to acquire parameter values ​​for assessing traffic conditions included in the identification information output by the indicator assessment agent; wherein the parameter values ​​for assessing traffic conditions include traffic flow, speed, and travel time ratio; and determine a corresponding influencing factor based on the ratio of each type of parameter value to a preset benchmark parameter value, wherein the larger the determined ratio, the larger the influencing factor; update the semantic similarity corresponding to the identification information output by the indicator assessment agent by using the product of the attenuation factor and the semantic similarity corresponding to the identification information output by the indicator assessment agent; and for the updated semantic similarity, execute the subsequent step of acquiring identification information for each target with a semantic similarity higher than a threshold.

[0102] Figure 5 This application provides a schematic diagram of an electronic device structure based on an embodiment of the present application. In addition to the above embodiments, this application also provides an electronic device, such as... Figure 5 As shown, it includes: processor 501, communication interface 502, memory 503 and communication bus 504, wherein processor 501, communication interface 502 and memory 503 communicate with each other through communication bus 504. The memory 503 stores a computer program, which, when executed by the processor 501, causes the processor 501 to perform any of the above method steps.

[0103] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0104] Communication interface 502 is used for communication between the above-mentioned electronic device and other devices.

[0105] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0106] The processors mentioned above can be general-purpose processors, including central processing units, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0107] This application also provides a computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform any of the above method steps.

[0108] This application provides a computer program product, which includes an executable program that, when executed by a processor, implements the method described herein.

[0109] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0110] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A traffic public opinion analysis method based on a large model, characterized in that, The method includes: Obtain traffic-related public opinion information; input the public opinion information into multiple preset intelligent agents, each intelligent agent is used to identify a specific type of traffic control or facility anomaly, and obtain the identification information output by each intelligent agent; wherein, the identification information includes at least the intersection involved in the public opinion information, the corresponding time period, and the traffic control information or traffic facility information associated with the intersection and time period; The semantic similarity between the public opinion information and the recognition information output by each agent is determined by using a pre-trained sentence-bidirectional encoder to represent the SBERT model. Obtain the identification information of each target with a semantic similarity higher than a threshold, input the identification information of each target into a large language model, and obtain the target answer output by the analysis of the large language model; wherein, the target answer includes at least the cause of the problem and the solution of the public opinion information determined based on the identification information of each target.

2. The method according to claim 1, characterized in that, The plurality of pre-defined intelligent agents include real-time status intelligent agents, duty execution intelligent agents, indicator evaluation intelligent agents, and other intelligent agents; The static parameter agent is used to determine whether the intersection involved in the public opinion information is equipped with a safety navigation device, and to detect whether the working status and settings of the safety navigation device comply with relevant national standards. The real-time status agent is used to determine whether there is human intervention or manual adjustment of the signal control system at the intersection involved in the public opinion information during the time period reflected by the public opinion information. The duty execution intelligent agent is used to determine whether the intersection corresponding to the public opinion information is located within the preset traffic coordination control route or regional linkage control range; The indicator evaluation agent is used to evaluate the traffic indicators of the intersections corresponding to the public opinion information.

3. The method according to claim 1, characterized in that, The SBERT model, pre-trained, determines the semantic similarity between the public opinion information and the recognition information output by each agent, including: The public opinion information and the recognition information output by each intelligent agent are respectively input into the pre-trained SBERT model to obtain the public opinion vector of the public opinion information output by the SBERT model and the recognition vector of the recognition information output by each intelligent agent. Based on the cosine similarity between the public opinion vector and each identification vector, the semantic similarity between the public opinion information and the identification information output by each intelligent agent is determined.

4. The method according to claim 1 or 3, characterized in that, The SBERT model is trained in the following manner: Get any pair of samples in the sample set, and the sample similarity saved for that pair, where the pair contains two texts; Input the two texts contained in the sample pair into the SBERT model respectively, and obtain the two text vectors output by the SBERT model; The recognition similarity is determined based on the cosine similarity of the two text vectors; the SBERT model is trained based on the deviation between the recognition similarity and the sample similarity.

5. The method according to claim 1, characterized in that, The step of inputting the target recognition information of each target into the large language model and obtaining the target answer output by the large language model includes: The target identification information is sorted according to semantic similarity from highest to lowest; The sorted target identification information and the corresponding semantic similarity of each target identification information are input into the large language model, so that the large language model performs differential in-depth analysis on each target identification information based on the semantic similarity. If the semantic similarity of the first target identification information is greater than that of the second target identification information, then the analysis depth of the first target identification information is higher than that of the second target information. The target answer output by the analysis of the large language model is obtained.

6. The method according to claim 1, characterized in that, After determining the semantic similarity between the public opinion information and the recognition information output by each agent using the pre-trained SBERT model, and before acquiring the recognition information of each target with a semantic similarity higher than a threshold, the method further includes: If a certain identification information includes time information or location information, then an attenuation factor is determined based on the time interval between the included time information and the occurrence time of the public opinion information, or based on the distance between the included location information and the occurrence location of the public opinion information; wherein, the larger the time interval, the smaller the determined attenuation factor, and the larger the distance, the smaller the determined attenuation factor; the semantic similarity corresponding to the identification information is updated by multiplying the attenuation factor and the semantic similarity corresponding to the identification information. Based on the updated semantic similarity, perform the subsequent step of obtaining the identification information of each target with a semantic similarity higher than the threshold.

7. The method according to claim 2, characterized in that, After determining the semantic similarity between the public opinion information and each identified information using the pre-trained SBERT model, and before acquiring each target identified information with a semantic similarity higher than a threshold, the method further includes: The system obtains the parameter values ​​for assessing traffic conditions contained in the identification information output by the indicator assessment agent; wherein the parameter values ​​for assessing traffic conditions include traffic flow, speed, and travel time ratio; and determines the corresponding influencing factor based on the ratio of each type of parameter value to a preset benchmark parameter value, wherein the larger the determined ratio, the larger the influencing factor. The semantic similarity corresponding to the recognition information output by the indicator evaluation agent is updated by using the product of the attenuation factor and the semantic similarity corresponding to the recognition information output by the indicator evaluation agent. For the updated semantic similarity, perform the subsequent step of obtaining the recognition information of each target with a semantic similarity higher than the threshold.

8. The method according to claim 7, characterized in that, The step of determining the corresponding influence factor based on the ratio of each type of parameter value to a preset benchmark parameter value includes: The determined impact factors satisfy the following formula: in, Let be the determined influencing factor, n be the number of traffic state parameter types, and i be the type number of the traffic state parameter. Let be the parameter value for the i-th traffic state parameter. Let be the baseline value of the i-th traffic state parameter. and These are the preset values ​​corresponding to the i-th traffic state parameter. is the parameter threshold for the i-th traffic state parameter.

9. A traffic public opinion analysis device based on a large model, characterized in that, The device includes: The acquisition module is used to acquire traffic-related public opinion information; the public opinion information is input into multiple preset intelligent agents, each intelligent agent is used to identify a specific type of traffic control or facility anomaly, and the identification information output by each intelligent agent is acquired; wherein, the identification information includes at least the intersection involved in the public opinion information, the corresponding time period, and the traffic control information or traffic facility information associated with the intersection and time period; The processing module is used to determine the semantic similarity between the public opinion information and the recognition information output by each agent by using a pre-trained sentence-bidirectional encoder to represent the SBERT model; to obtain the recognition information of each target with a semantic similarity higher than a threshold, and to input the recognition information of each target into a large language model to obtain the target answer output by the analysis of the large language model; wherein, the target answer includes at least the cause of the problem and the solution of the public opinion information determined based on the recognition information of each target; 10. An electronic device, characterized in that, The electronic device includes at least a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the steps of the traffic sentiment analysis method based on a large model as described in any one of claims 1-8.

Images