Urban transportation evaluation method and system fused with large-scale language model

Abstract

The invention relates to an urban traffic travel evaluation method and system fused with a large-scale language model. The method comprises the following steps: establishing a multi-dimensional index library for urban traffic evaluation, performing semantic mining on social media data by using the large-scale language model, identifying emerging evaluation dimensions, and generating a dynamically updated comprehensive evaluation index system; classifying the social media data by using a pre-training language model, and generating a text set corresponding to each evaluation index; performing sentiment intensity and polarity analysis on the text set by using a dictionary-based sentiment analysis model; inputting the emotion intensity and polarity analysis result and the traffic sensor data into a large language model for multi-modal causal reasoning, and generating a causal association report; and finally, verifying the evaluation result based on the sentiment analysis, and generating an urban traffic intelligent evaluation report. Compared with the prior art, the method has the advantages that the multi-dimensional emotion information in the social media data can be fully explored, and the accuracy of urban traffic travel evaluation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of urban traffic operation management, and in particular to an urban traffic travel evaluation method and system integrating a large-scale language model. Background Art

[0002] Existing urban transportation intelligence assessment methods in China mostly focus on quantitative analysis based on physical data (such as traffic flow, vehicle speed, and delay time), neglecting the critical factor of the travel experience of traffic participants. However, the ultimate goal of a transportation system is to improve the user experience. Therefore, relying solely on physical data often fails to fully reflect the actual effectiveness of the transportation system and user perception.

[0003] While some studies have attempted to incorporate user travel experience into their evaluations, these studies are often limited to a single mode of transportation (such as public transportation or urban rail), lacking a comprehensive evaluation framework that can comprehensively and systematically assess all modes of transportation and related factors. Furthermore, traditional data collection methods, such as questionnaires, suffer from small sample sizes, low response rates, and high subjectivity. Social media data, as an emerging data source, can provide a large amount of real-time, extensive, and representative user feedback.

[0004] Chinese patent application CN202111087045.3 discloses a method for evaluating emergency transportation strategies based on social media data. This approach uses a text sentiment pre-training model and an entropy weighting method to optimize a latent Dirichlet allocation (LDA) model to address the issues of changing transportation strategies and prioritizing demand during public emergencies. However, existing research has not fully utilized social media data to evaluate transportation service quality. In particular, the analysis of sentiment intensity and multi-dimensional sentiment in social media data remains exploratory. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an urban transportation travel evaluation system based on travel experience emotion analysis.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] As a first aspect of the present invention, a method for evaluating urban transportation trips by integrating a large language model is provided, comprising the following steps:

[0008] Collect text data related to urban transportation from social media platforms and combine it with traffic sensor data and questionnaire data to form a structured dataset;

[0009] Establish a multidimensional indicator library for urban transportation evaluation, use large-scale language models to conduct semantic mining on text data from social media platforms and user feedback, identify emerging evaluation dimensions, and compare and integrate them with the multidimensional indicator library to generate a dynamically updated comprehensive evaluation indicator system;

[0010] Use a pre-trained language model to classify the collected text data from social media platforms and generate a text set corresponding to each evaluation indicator in the comprehensive evaluation index system;

[0011] Use a dictionary-based sentiment analysis model to analyze the sentiment intensity and polarity of text collections, map the results to user ratings, and input the sentiment intensity and polarity analysis results and traffic sensor data into a large language model. Use the large language model to perform multimodal causal reasoning and generate a causal association report.

[0012] The questionnaire data were processed using the analytic hierarchy process (AHP) to obtain the evaluation results, which were then combined with the evaluation results based on sentiment analysis to verify the evaluation results, ultimately generating an urban traffic intelligence evaluation report.

[0013] As a preferred technical solution, the large language model is used to generate a dynamically updated comprehensive evaluation index system, as follows:

[0014] Collecting travel-related text data from multiple social media platforms, and after cleaning and processing, using a large language model to deeply embed the corpus;

[0015] Use topic modeling and semantic clustering algorithms to identify semantically similar text clusters and extract topic tags that are frequently mentioned but not yet included in the evaluation framework;

[0016] After comparing the extracted topic tags with the original multidimensional indicator library, they are further classified into new dimensions or refined supplements to existing dimensions, and ranked and evaluated in combination with quantitative features including frequency weights to construct an indicator system that is dynamically updated with semantic evolution.

[0017] As a preferred technical solution, the pre-trained language model uses the BERT model to perform multi-category text classification and generate a text set corresponding to each evaluation indicator in the comprehensive evaluation index system as follows:

[0018] Using a pre-trained BERT model, public datasets and a large-scale unlabeled text corpus are used to generate initial parameters as the basic parameters of the model to capture the semantics and contextual relevance of language;

[0019] The text data in the dataset is annotated according to predefined urban transportation evaluation indicators to form an annotated dataset. The annotated data is input into the pre-trained BERT model, and the model parameters are optimized through a fine-tuning process so that it can accurately classify text into multiple categories.

[0020] By constructing a text classification task, the input text data is mapped to a predefined set of categories. The BERT model outputs the classification result, that is, the probability that the text data belongs to the indicator, and finally classifies the text data into the category with the highest probability.

[0021] As a preferred technical solution, the BERT model adopts the cross entropy loss function as the objective function, updates the model parameters through the stochastic gradient descent method, and iteratively optimizes until convergence, thereby obtaining a BERT model for text classification.

[0022] As a preferred technical solution, the dictionary-based sentiment analysis model performs sentiment intensity and polarity analysis on a text collection, as follows:

[0023] The sentiment analysis model is used to extract the sentiment polarity and intensity of the text. The value range of the sentiment intensity δ(x) is [0,1], and the polarity is represented by the parameter s(x):

[0024] s(x)=1 indicates positive sentiment

[0025] s(x)=0 indicates neutral sentiment

[0026] s(x)=-1 indicates negative sentiment

[0027] According to the sentiment analysis results, the dataset X corresponding to each evaluation index is i Divided into three subsets Corresponding to positive, neutral and negative text data respectively;

[0028] In view of the difference between the emotional intensity and the traditional scoring system, the intensity is standardized and mapped, the emotional intensity of each subset is normalized, and the standardized intensity is calculated.

[0029]

[0030] Among them, δ k is the sentiment intensity of the kth evaluation index in the subset, and are the minimum and maximum values of the sentiment intensity in the subset, respectively;

[0031] Convert the normalized intensity to a percentile-based score:

[0032]

[0033] Among them, g k is the sentiment intensity score of the kth evaluation indicator, s k For the emotion category.

[0034] As a preferred technical solution, the method corrects the emotional polarity distribution deviation through resampling technology, as follows:

[0035] Calculate the data skew width indicator to quantify the imbalance of the data set, data skew width ρ i The calculation is as follows:

[0036]

[0037] Among them, 90th SI, 50th SI, and 10th SI represent the 90th, 50th, and 10th percentile values of the sentiment intensity in the dataset, respectively;

[0038] Based on the skew width ρ i , for each indicator dataset X i Resampling is performed to balance the data proportions of different emotion categories. The resampling ratio is determined based on the following formula:

[0039]

[0040] Among them, b i,s is the indicator X i Medium sentiment categories k The amount of text data, b i is the total amount of text data for the indicator;

[0041] After resampling, according to each indicator X i The emotional data is used to calculate its comprehensive score g i :

[0042] g i =v i,-1 g i,-1 +v i,0 g i,0 +ρ i v i,1 g i,1

[0043] Among them, v i,s is the indicator X i Medium sentiment categories k The resampling ratio, g i,s is the indicator X i Medium sentiment categories k Rating of emotional intensity;

[0044] Combine the scores of all indicators to calculate the overall evaluation score of urban transportation:

[0045]

[0046] Among them, v i Indicates the comprehensive score of the indicator g i The weight ratio in the overall evaluation.

[0047] As a preferred technical solution, the multimodal causal reasoning method using a large language model specifically involves using a large language model to uniformly encode and semantically associate the results of sentiment intensity and polarity analysis with heterogeneous traffic sensor data. By constructing a causal reasoning chain, the potential causes of specific traffic problems are identified, and a logically interpretable causal report is generated.

[0048] Perform unified spatiotemporal alignment on multi-source data, constructing traffic event-time-space triples with minute-level time granularity. Traffic events come from sensor monitoring, spatial information comes from GPS trajectory cluster analysis, and the time dimension serves as the data synchronization anchor point.

[0049] Perform sentiment recognition and polarity analysis on social media texts based on sentiment dictionaries and pre-trained language models to extract users' subjective feelings associated with specific events;

[0050] Combining the co-occurrence patterns between traffic state changes and users' subjective feelings, we use prompt engineering to build an input framework for causal reasoning, guiding large-scale language models to perform cross-modal logic induction.

[0051] Large language models generate structured text explanations based on contextual memory and logical reasoning capabilities, and further visualize causal chains as causal graphs.

[0052] As a second aspect of the present invention, a system for evaluating urban traffic trips integrating a large language model is provided. The system executes the method for evaluating urban traffic trips integrating a large language model as described above, comprising:

[0053] The data acquisition module collects and preprocesses multi-source heterogeneous data, including questionnaire data, social media text, traffic perception data, and vehicle GPS trajectory data. After preprocessing, the data is formed into a structured data set for analysis;

[0054] The text classification and sentiment analysis module uses a pre-trained BERT model to semantically understand social media text, categorizing the text content into established traffic evaluation dimensions. It then combines this with a Chinese sentiment dictionary model to extract the sentiment polarity and intensity of each text, and further maps the sentiment intensity into a score.

[0055] A multimodal causal reasoning module, which uses a large language model to jointly model social media text sentiment and traffic sensor data. This module achieves multimodal alignment through semantic fusion, identifies potential causal relationships between public sentiment changes and specific traffic events, and constructs a causal reasoning chain.

[0056] The result verification and report generation module models the questionnaire survey results based on the hierarchical analysis method to generate subjective scores and weight distributions for each evaluation dimension; the hierarchical analysis scores are compared and verified with the system scores based on sentiment analysis, and finally an intelligent transportation evaluation report is generated, which includes score charts for each dimension, a comprehensive score radar chart, a summary of typical problem analysis, and corresponding optimization suggestions.

[0057] As a preferred technical solution, the text classification and sentiment analysis module introduces a resampling mechanism during the processing to correct the sentiment polarity distribution deviation, as follows:

[0058] Calculate the data skew width indicator to quantify the imbalance of the data set, data skew width ρ i The calculation is as follows:

[0059]

[0060] Among them, 90th SI, 50th SI, and 10th SI represent the 90th, 50th, and 10th percentile values of the sentiment intensity in the dataset, respectively;

[0061] Based on the skew width ρ i , for each indicator dataset X i Resampling is performed to balance the data proportions of different emotion categories. The resampling ratio is determined based on the following formula:

[0062]

[0063] Among them, b i,s is the indicator X i Medium sentiment categories k The amount of text data, b i is the total amount of text data for the indicator;

[0064] After resampling, according to each indicator X i The emotional data is used to calculate its comprehensive score g i :

[0065] g i =v i,-1 g i,-1 +v i,0 g i,0 +ρ i v i,1 g i,1

[0066] Combine the scores of all indicators to calculate the overall evaluation score of urban transportation:

[0067]

[0068] Among them, v i Indicates the comprehensive score of the indicator g i The weight ratio in the overall evaluation.

[0069] As a preferred technical solution, the multimodal causal reasoning module performs multimodal causal reasoning through a large language model. Specifically, the large language model is used to uniformly encode and semantically associate the sentiment intensity and polarity analysis results with heterogeneous traffic sensor data. By constructing a causal reasoning chain, the potential causes of specific traffic problems are identified and a logically interpretable causal report is generated.

[0070] Perform unified spatiotemporal alignment on multi-source data, constructing traffic event-time-space triples with minute-level time granularity. Traffic events come from sensor monitoring, spatial information comes from GPS trajectory cluster analysis, and the time dimension serves as the data synchronization anchor point.

[0071] Perform sentiment recognition and polarity analysis on social media texts based on sentiment dictionaries and pre-trained language models to extract users' subjective feelings associated with specific events;

[0072] Combining the co-occurrence patterns between traffic state changes and users' subjective feelings, we use prompt engineering to build an input framework for causal reasoning, guiding large-scale language models to perform cross-modal logic induction.

[0073] Large language models generate structured text explanations based on contextual memory and logical reasoning capabilities, and further visualize causal chains as causal graphs.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] 1) The present invention has the characteristics of strong semantic expansion and dynamic updating capabilities of the indicator system. With the help of a large language model, it can deeply understand the semantics of social media and user feedback, explore potential traffic evaluation dimensions of public concern, realize automatic discovery and updating of evaluation indicators, solve the problems of fixed and lagging traditional indicator systems, and significantly enhance the flexibility and foresight of the system.

[0076] 2) The present invention utilizes the Large Language Model (LLM) to understand the context and perform semantic abstraction on complex natural languages. It can accurately identify the traffic scenarios, emotional tendencies, and service evaluations involved in the text, and transform the travel experience into a quantifiable structured score, thereby improving the quantification of user experience and providing objective support for subjective evaluation.

[0077] 3) This paper addresses the problem of uneven distribution of social media data by using resampling technology to correct the bias in the distribution of sentiment polarity. The imbalance of the data set is quantified by calculating the data skew width index. Based on the skew width, for each index X i The dataset is resampled to balance the data proportions of different sentiment categories (positive, neutral, and negative), and the sentiment data is recalculated to calculate its comprehensive score, thereby improving the accuracy of urban transportation travel evaluation.

[0078] 4) This invention utilizes the multimodal causal reasoning capability of the Large Language Model (LLM) to perform unified modeling and logical reasoning on data from different modalities (text, sensors, trajectories, etc.), enhance the interpretability of traffic problems, and assist in generating travel problem explanation reports with a causal chain structure, which helps traffic managers identify the root causes behind traffic problems and improve the scientific nature of decision-making.

[0079] In summary, the present invention provides a new solution for the evaluation of urban transportation travel quality by integrating large-scale model technology with an evaluation system based on user experience. It is significantly superior to existing technologies in terms of method innovation, result accuracy, and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is a flow chart of an urban traffic travel evaluation method integrating a large-scale language model according to the present invention;

[0081] Figure 2 This is a functional diagram of the text classification and sentiment analysis module in the urban transportation travel evaluation system that integrates a large-scale language model in the present invention. DETAILED DESCRIPTION

[0082] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0083] Example 1

[0084] In order to overcome the problems of the existing technology in urban traffic intelligence evaluation, such as lack of dynamics, weak subjectivity, and low data utilization, this paper proposes an urban traffic travel intelligence evaluation method that integrates a large language model to achieve an automated, multi-dimensional, and high-precision evaluation of the urban traffic intelligence level with travel experience as the core. Figure 1 As shown in FIG, the integrated urban transportation intelligent evaluation method integrating a large language model includes the following steps:

[0085] Step 1: Multi-source data collection: Use web crawlers to collect text data related to urban transportation in social media, combine traffic sensor data with questionnaire data, and form a structured dataset.

[0086] Step 2: Build a basic evaluation index library: Through literature research and questionnaire surveys, establish an initial index library covering dimensions such as safety, convenience, and comfort.

[0087] Step 3: Dynamic generation of indicators based on LLM: Use large language models to perform semantic mining on social media data and user feedback, automatically identify emerging evaluation dimensions, and integrate them with the initial indicators.

[0088] Step 4: Establish a travel quality evaluation system and use a pre-trained language model (such as BERT) to classify text into various indicator dimensions.

[0089] Step 5: Sentiment analysis processing: The dictionary-based sentiment analysis model evaluates the polarity and intensity of the text, and combines it with traffic sensor data (such as congestion index) for causal analysis.

[0090] Step 6: Multimodal causal reasoning: Integrate social media sentiment analysis results, vehicle GPS data, traffic sensor data, etc. to identify the root causes of travel problems.

[0091] Step 7: Verify the evaluation method: Use the hierarchical analysis method to process the questionnaire data to obtain the evaluation results. Combine the evaluation results of the hierarchical analysis method with the evaluation results based on sentiment analysis to finally generate an urban traffic intelligence evaluation report.

[0092] The web crawler in step 1 obtains review text data containing specified keywords from social media platforms.

[0093] The basic evaluation index library in step 2 is established by summarizing and clustering the evaluation indicators in authoritative traffic evaluation articles. The specific steps include the following:

[0094] 2.1) Literature Research: Through a systematic analysis of existing literature on urban transportation intelligence evaluation, we will preliminarily screen multi-dimensional evaluation indicators that can reflect the level of urban transportation intelligence. This literature research covers research results in the fields of transportation service quality, intelligent transportation system design, and user experience at home and abroad, and determine a preliminary set of indicators.

[0095] 2.2) Questionnaire Survey: Based on the preliminary set of indicators from the literature research, a questionnaire was designed for transportation participants to obtain subjective feedback on transportation service quality and user experience. The questionnaire was distributed online and offline to cover a diverse group of transportation participants.

[0096] The LLM-based indicators in step 3 are dynamically generated as follows:

[0097] 3.1) Indicator Screening and Refinement: Combining literature research and questionnaire survey data, a multi-criteria decision-making approach was used to conduct semantic mining on the preliminary indicator set, ultimately forming a comprehensive evaluation indicator library. This library contains multiple indicators that can comprehensively cover key dimensions such as traffic information accessibility, safety, comfort, and flexibility.

[0098] 3.2) Dynamic update mechanism: To improve the applicability and long-term effectiveness of the index database, a dynamic update mechanism should be established to regularly supplement and adjust the index database based on traffic development trends, user feedback, and emerging research results.

[0099] This step introduces large-scale language models (mature models such as ChatGPT and Qianwen) to construct a semantic mining process based on social media data and user feedback text, which is used to identify and dynamically update "emerging dimensions" in the evaluation index system. Specifically, travel-related text data from multiple social platforms is collected through web crawlers. After basic cleaning, the data is deeply embedded using LLM. Topic modeling and semantic clustering algorithms are then used to identify semantically similar text clusters. Frequently mentioned topic tags that have not yet been incorporated into traditional evaluation frameworks are automatically extracted, such as "shared bike scheduling issues," "accessibility experience," and "public transportation information delays." These tags are then compared with the original index database and further classified into new dimensions or refinements of existing dimensions. Quantitative features such as sentiment intensity and frequency weight are then used for ranking and evaluation, thereby constructing an index system that dynamically updates as semantics evolve. This mechanism enables the system to adapt to changes in social perception, continuously expand evaluation dimensions, and enhance the timeliness and public acceptance of the index system.

[0100] The pre-trained language model in step 4 uses the BERT model for multi-category text classification. The following is the text classification model: F(x,θ) represents a text classifier based on BERT with parameters.

[0101] 4.1) Pre-training and parameter initialization: Using a pre-trained BERT model, we use public datasets and a large-scale unlabeled text corpus to generate initial parameters, which serve as the basic parameters of the model to capture the semantics and contextual relevance of the language.

[0102] 4.2) Data Annotation and Fine-tuning: The text data in the system is annotated according to predefined urban transportation evaluation indicators to form an annotated dataset. This annotated data is fed into the pre-trained BERT model, and the model parameters are optimized through fine-tuning to enable accurate multi-category text classification.

[0103] 4.3) Classification process: By constructing a text classification task, the input text data x is mapped to a predefined set of categories. The model outputs the classification result q k where q k (i) represents the probability that the text data belongs to index i. Finally, the text data is classified into the category with the highest probability.

[0104] q k =F(x k ,θ)

[0105]

[0106] 4.4) Model optimization: The cross entropy loss function is used as the objective function. The model parameters are updated through the stochastic gradient descent method. The optimization is iterated until convergence, thus obtaining an optimized model for text classification.

[0107]

[0108] in, represents the gradient of the loss function L(θ) with respect to θ in the φth round of training, ξ is used to update the step size of θ. By repeating the above fine-tuning process φ times, the BERT model for text classification is finally obtained.

[0109] The dictionary analysis model in step 5 can simultaneously output sentiment polarity (positive, negative, neutral) and sentiment intensity.

[0110] 5.1.1) Sentiment Analysis Model: Using a Chinese language processing model and a large-scale training dataset, we extract the sentiment polarity and intensity of text. The sentiment intensity δ(x) ranges from [0, 1], and the polarity is represented by the parameter s(x), where:

[0111] δk,sk=G(x k )

[0112] s(x)=1 indicates positive sentiment

[0113] s(x)=0 indicates neutral sentiment

[0114] s(x)=-1 indicates negative sentiment

[0115] 5.1.2) Dataset division: According to the sentiment analysis results, the dataset X corresponding to each evaluation indicator is divided into i Divided into three subsets Corresponding to positive, neutral and negative text data respectively.

[0116] 5.1.3) Mapping of sentiment intensity and rating: In view of the difference between sentiment intensity and traditional rating system, the following steps are used to complete the standardization and mapping of intensity. The sentiment intensity of each subset is normalized and the standardized intensity is calculated. in, and are the minimum and maximum values of the sentiment intensity in the subset, respectively. The normalized intensity is then converted into a percentage-based score.

[0117]

[0118] Among them, g k Score the sentiment intensity of the k-th evaluation indicator.

[0119] In step 5, to address the imbalanced distribution of social media data, resampling technology was used to correct the bias in sentiment polarity distribution, thereby improving the accuracy of urban traffic travel evaluation.

[0120] 5.2.1) The degree of imbalance in the dataset is quantified by calculating the skew-width metric, which is defined as follows:

[0121]

[0122] Among them, 90th SI, 50th SI, and 10th SI represent the 90th, 50th, and 10th percentile values of the sentiment intensity in the dataset, respectively.

[0123] 5.2.2) Based on the skew width ρ i , for each indicator X i The dataset is resampled to balance the data proportions of different sentiment categories (positive, neutral, and negative). The resampling ratio is determined based on the following formula:

[0124]

[0125] g i =v i,-1 g i,-1 +v i,0 g i,0 +ρ i v i,1 g i,1

[0126] Among them, b i,s Indicates indicator X i Medium sentiment categories k The amount of text data, b i Indicates the total amount of text data for this indicator. After resampling, according to each indicator X i The comprehensive score is calculated based on the sentiment data.

[0127] 5.2.3) Combine the scores of all indicators to calculate the overall evaluation score of urban transportation:

[0128]

[0129] Among them, v i Indicates the comprehensive score of the indicator g i The weight ratio in the overall evaluation.

[0130] In step 6, a large language model is used to uniformly encode and semantically associate these heterogeneous data. By constructing a causal reasoning chain, the potential causes of specific traffic problems (such as peak-hour congestion, unsafe road sections, low service satisfaction, etc.) are identified, and a logically interpretable causal report is generated to assist traffic managers in accurately locating the root causes of the problems and formulating intervention measures.

[0131] First, multi-source data is aligned in time and space, using minute-level granularity to construct "event-time-space" triples. Traffic events are derived from sensor monitoring (e.g., sudden congestion or accidents), spatial information is derived from GPS trajectory cluster analysis, and the time dimension serves as the data synchronization anchor. Simultaneously, sentiment recognition and polarity analysis are performed on social media texts using a sentiment lexicon and a pre-trained language model (the results of the previous step) to extract user subjective feelings associated with specific events, such as "anxiety," "anger," and "satisfaction." Subsequently, based on the co-occurrence patterns between traffic state changes and user subjective feelings, the system uses prompt engineering to construct an input framework for causal reasoning, guiding the LLM to perform cross-modal logical induction. For example, a prompt template might be constructed as follows: "In XX area, during XX time period, sensors recorded abnormal congestion, user sentiment polarity decreased, and GPS trajectories showed an increase in detour rates. What is the possible cause?" LLM leverages contextual memory and logical reasoning to generate structured text explanations, such as: "Due to untimely incident handling, resulting in traffic delays, delayed bus information releases exacerbated user anxiety." This causal chain can also be further visualized as a causal graph, helping urban transportation managers understand the systemic causes of service issues. Compared to traditional linear statistical or causal regression models, LLM is capable of modeling nonlinear, multidimensional, and highly semantically interactive relationships. It demonstrates enhanced causal identification and interpretation capabilities in complex traffic environments, enabling deep integration and intelligent analysis of subjective evaluations and objective data.

[0132] The evaluation results in step 7 are presented through an automatically generated visual report, including sentiment analysis results and AHP score comparison.

[0133] Example 2

[0134] As another embodiment of the present invention, this embodiment provides an urban transportation travel evaluation system based on travel experience sentiment analysis, specifically including:

[0135] The data acquisition module primarily collects and preprocesses multi-source heterogeneous data, including questionnaire data, social media text, traffic perception data, and vehicle GPS trajectory data. A total of 477 valid questionnaires were collected through online and offline channels. A web crawler was used to capture 3,200 traffic-related user comments from a social platform in the target area. Sensor data such as road congestion index, traffic speed, and accident records were simultaneously accessed from the target city's open transportation platform. Combined with GPS trajectory data from ride-sharing platforms, this module builds a comprehensive, authentic, and representative data foundation. This data is uniformly cleaned, deduplicated, and standardized to form a structured dataset suitable for analysis.

[0136] Text classification and sentiment analysis module, which uses the large language model BERT to perform semantic understanding of social media texts. It first classifies the text content into established traffic evaluation dimensions (such as efficiency, comfort, safety, etc.), and then combines the Chinese sentiment dictionary model to extract the sentiment polarity (positive, neutral, negative) and sentiment intensity (value range [0,1]) of each text. The sentiment intensity is further mapped to a percentage rating range, where positive sentiment is mapped to 65-100 points, neutral sentiment intensity is mapped to 55-65 points, and negative sentiment intensity is mapped to 0-55 points. In order to ensure the balance of the sample sentiment category ratio, the system introduces a resampling mechanism during the processing process to improve the stability and scientificity of the final scoring results, such as Figure 2 shown.

[0137] The multimodal causal reasoning module leverages LLM to jointly model sentiment from social media text, traffic sensor data, and GPS trajectory data, achieving multimodal alignment through semantic fusion. The system automatically identifies potential causal relationships between changes in public sentiment and specific traffic events, such as "frequent accidents on a certain road section → rising negative user sentiment → low overall ratings for that area." By constructing a causal reasoning chain, this module provides explanatory power for locating the root causes of urban traffic problems, helping to identify unseen issues and optimize paths.

[0138] The results verification and report generation module modeled the questionnaire results using the Analytic Hierarchy Process (AHP) to generate subjective scores and weight distributions for each evaluation dimension. The AHP scores were then compared with the system scores based on sentiment analysis, demonstrating high consistency between the two: an AHP score of 69.64 and a sentiment analysis model score of 68.39. The system then automatically generated an intelligent transportation evaluation report, including a score chart for each dimension, a comprehensive score radar chart, a summary of typical issues, and corresponding optimization recommendations to assist management in decision-making.

[0139] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for evaluating urban transportation trips by integrating a large language model, characterized in that the steps include: Collect text data related to urban transportation from social media platforms and combine it with traffic sensor data and questionnaire data to form a structured dataset; Establish a multidimensional indicator library for urban transportation evaluation, use large-scale language models to conduct semantic mining on text data from social media platforms and user feedback, identify emerging evaluation dimensions, and compare and integrate them with the multidimensional indicator library to generate a dynamically updated comprehensive evaluation indicator system; Use a pre-trained language model to classify the collected text data from social media platforms and generate a text set corresponding to each evaluation indicator in the comprehensive evaluation index system; Use a dictionary-based sentiment analysis model to analyze the sentiment intensity and polarity of text collections, map the results to user ratings, and input the sentiment intensity and polarity analysis results and traffic sensor data into a large language model. Use the large language model to perform multimodal causal reasoning and generate a causal association report. The questionnaire data were processed using the analytic hierarchy process (AHP) to obtain the evaluation results, which were then combined with the evaluation results based on sentiment analysis to verify the evaluation results, ultimately generating an urban traffic intelligence evaluation report.

2. The urban traffic evaluation method integrating a large language model according to claim 1 is characterized in that: The large language model is used to generate a dynamically updated comprehensive evaluation index system, as follows: Collecting travel-related text data from multiple social media platforms, and after cleaning and processing, using a large language model to deeply embed the corpus; Use topic modeling and semantic clustering algorithms to identify semantically similar text clusters and extract topic tags that are frequently mentioned but not yet included in the evaluation framework; After comparing the extracted topic tags with the original multidimensional indicator library, they are further classified into new dimensions or refined supplements to existing dimensions, and ranked and evaluated in combination with quantitative features including frequency weights to construct an indicator system that is dynamically updated with semantic evolution.

3. The urban traffic evaluation method integrating a large language model according to claim 1 is characterized in that: The pre-trained language model uses the BERT model to perform multi-category text classification and generate a text set corresponding to each evaluation indicator in the comprehensive evaluation index system as follows: Using a pre-trained BERT model, public datasets and a large-scale unlabeled text corpus are used to generate initial parameters as the basic parameters of the model to capture the semantics and contextual relevance of language; The text data in the dataset is annotated according to predefined urban transportation evaluation indicators to form an annotated dataset. The annotated data is input into the pre-trained BERT model, and the model parameters are optimized through a fine-tuning process so that it can accurately classify text into multiple categories. By constructing a text classification task, the input text data is mapped to a predefined set of categories. The BERT model outputs the classification result, that is, the probability that the text data belongs to the indicator, and finally classifies the text data into the category with the highest probability.

4. The urban traffic evaluation method integrating a large language model according to claim 3 is characterized in that: The BERT model uses the cross entropy loss function as the objective function, updates the model parameters through the stochastic gradient descent method, and iteratively optimizes until convergence, thereby obtaining the BERT model for text classification.

5. The urban traffic evaluation method integrating a large language model according to claim 1 is characterized in that: The dictionary-based sentiment analysis model performs sentiment intensity and polarity analysis on a text collection, as follows: The sentiment analysis model is used to extract the sentiment polarity and intensity of the text. The value range of the sentiment intensity δ(x) is [0,1], and the polarity is represented by the parameter s(x): s(x)=1 indicates positive sentiment s(x)=0 indicates neutral sentiment s(x)=-1 indicates negative sentiment According to the sentiment analysis results, the dataset X corresponding to each evaluation index is i Divided into three subsets Corresponding to positive, neutral and negative text data respectively; In view of the difference between the emotional intensity and the traditional scoring system, the intensity is standardized and mapped, the emotional intensity of each subset is normalized, and the standardized intensity is calculated. Among them, δ k is the sentiment intensity of the kth evaluation index in the subset, and are the minimum and maximum values of the sentiment intensity in the subset, respectively; Convert the normalized intensity to a percentile-based score: Among them, g k is the sentiment intensity score of the kth evaluation indicator, s k For the emotion category.

6. The urban traffic evaluation method integrating a large language model according to claim 1 is characterized in that: The method corrects the emotional polarity distribution deviation through resampling technology, as follows: Calculate the data skew width indicator to quantify the imbalance of the data set, data skew width ρ i The calculation is as follows: Among them, 90th SI, 50th SI, and 10th SI represent the 90th, 50th, and 10th percentile values of the sentiment intensity in the dataset, respectively; Based on the skew width ρ i , for each indicator dataset X i Resampling is performed to balance the data proportions of different emotion categories. The resampling ratio is determined based on the following formula: Among them, b i,s is the indicator X i Medium sentiment categories k The amount of text data, b i is the total amount of text data for the indicator; After resampling, according to each indicator X i The emotional data is used to calculate its comprehensive score g i : g i =v i,-1 g i,-1 +v i,0 g i,0 +ρ i v i,1 g i,1 Among them, v i,s is the indicator X i Medium sentiment categories k The resampling ratio, g i,s is the indicator X i Medium sentiment categories k Rating of emotional intensity; Combine the scores of all indicators to calculate the overall evaluation score of urban transportation: Among them, v i Indicates the comprehensive score of the indicator g i The weight ratio in the overall evaluation.

7. The urban traffic evaluation method integrating a large language model according to claim 1 is characterized in that: The multimodal causal reasoning method using a large language model specifically involves using the large language model to uniformly encode and semantically associate sentiment intensity and polarity analysis results with heterogeneous traffic sensor data. By constructing a causal reasoning chain, the potential causes of specific traffic problems are identified, and a logically interpretable causal report is generated. Perform unified spatiotemporal alignment on multi-source data, constructing traffic event-time-space triples with minute-level time granularity. Traffic events come from sensor monitoring, spatial information comes from GPS trajectory cluster analysis, and the time dimension serves as the data synchronization anchor point. Perform sentiment recognition and polarity analysis on social media texts based on sentiment dictionaries and pre-trained language models to extract users' subjective feelings associated with specific events; Combining the co-occurrence patterns between traffic state changes and users' subjective feelings, we use prompt engineering to build an input framework for causal reasoning, guiding large-scale language models to perform cross-modal logic induction. Large language models generate structured text explanations based on contextual memory and logical reasoning capabilities, and further visualize causal chains as causal graphs.

8. An urban transportation evaluation system integrating a large language model, characterized by: The system executes the urban traffic travel evaluation method integrating a large language model according to any one of claims 1 to 7, comprising: The data acquisition module collects and preprocesses multi-source heterogeneous data, including questionnaire data, social media text, traffic perception data, and vehicle GPS trajectory data. After preprocessing, the data is formed into a structured data set for analysis; The text classification and sentiment analysis module uses a pre-trained BERT model to semantically understand social media text, categorizing the text content into established traffic evaluation dimensions. It then combines this with a Chinese sentiment dictionary model to extract the sentiment polarity and intensity of each text, and further maps the sentiment intensity into a score. A multimodal causal reasoning module, which uses a large language model to jointly model social media text sentiment and traffic sensor data. This module achieves multimodal alignment through semantic fusion, identifies potential causal relationships between public sentiment changes and specific traffic events, and constructs a causal reasoning chain. The result verification and report generation module models the questionnaire survey results based on the hierarchical analysis method to generate subjective scores and weight distributions for each evaluation dimension; the hierarchical analysis scores are compared and verified with the system scores based on sentiment analysis, and finally an intelligent transportation evaluation report is generated, which includes score charts for each dimension, a comprehensive score radar chart, a summary of typical problem analysis, and corresponding optimization suggestions.

9. The urban transportation evaluation system integrating a large language model according to claim 8 is characterized in that: The text classification and sentiment analysis module introduces a resampling mechanism during the processing to correct the sentiment polarity distribution deviation, as follows: Calculate the data skew width indicator to quantify the imbalance of the data set, data skew width ρ i The calculation is as follows: Among them, 90th SI, 50th SI, and 10th SI represent the 90th, 50th, and 10th percentile values of the sentiment intensity in the dataset, respectively; Based on the skew width ρ i , for each indicator dataset X i Resampling is performed to balance the data proportions of different emotion categories. The resampling ratio is determined based on the following formula: Among them, b i,s is the indicator X i Medium sentiment categories k The amount of text data, b i is the total amount of text data for the indicator; After resampling, according to each indicator X i The emotional data is used to calculate its comprehensive score g i : g i =v i,-1 g i,-1 +v i,0 g i,0 +ρ i v i,1 g i,1 Combine the scores of all indicators to calculate the overall evaluation score of urban transportation: Among them, v i Indicates the comprehensive score of the indicator g i The weight ratio in the overall evaluation.

10. The urban transportation evaluation system integrating a large language model according to claim 8 is characterized in that: The multimodal causal reasoning module uses a large language model to perform multimodal causal reasoning. Specifically, it uses the large language model to uniformly encode and semantically associate the results of sentiment intensity and polarity analysis with heterogeneous traffic sensor data. By building a causal reasoning chain, it identifies the potential causes of specific traffic problems and generates a logically interpretable causal report. Perform unified spatiotemporal alignment on multi-source data, constructing traffic event-time-space triples with minute-level time granularity. Traffic events come from sensor monitoring, spatial information comes from GPS trajectory cluster analysis, and the time dimension serves as the data synchronization anchor point. Perform sentiment recognition and polarity analysis on social media texts based on sentiment dictionaries and pre-trained language models to extract users' subjective feelings associated with specific events; Combining the co-occurrence patterns between traffic state changes and users' subjective feelings, we use prompt engineering to build an input framework for causal reasoning, guiding large-scale language models to perform cross-modal logic induction. Large language models generate structured text explanations based on contextual memory and logical reasoning capabilities, and further visualize causal chains as causal graphs.

Images