Intelligent voice dispatching method for transportation service hotline based on large language model

Abstract

This invention discloses an intelligent voice dispatch method for transportation service hotlines based on a large language model. The method includes: S1, constructing a semantic temporal recognition model containing convolutional neural networks and recurrent neural networks; the semantic temporal recognition model combines semantic features and temporal features to obtain a semantic temporal feature sequence; S2, constructing a transportation service knowledge base; using a pre-trained large language model to classify and extract the semantic temporal feature sequence according to consultation and answer services and work order dispatch services; for services identified as consultation and answer services, knowledge content from the transportation service knowledge base is filtered and retrieved, and voice data is generated through a voice encoder for answering and replying; for services identified as work order dispatch services, the content required for work order dispatch is extracted, and tasks are dispatched according to work order dispatch rules. This invention effectively improves the efficiency of interactive control and work order transmission, reduces the workload of business personnel, and significantly improves the service quality of the transportation industry.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice dispatching for the 12328 transportation service hotline, and in particular to an intelligent voice dispatching method for transportation service hotlines based on a large language model. Background Technology

[0002] The 12328 transportation service hotline provides transportation-related services to individuals, legal entities, and other organizations. Currently, the hotline experiences a high volume of inquiries, a large amount of work order data, and numerous subcategories, leading to challenges such as high complexity of business information, heavy information turnover, and difficulties in information interaction for system users. There is an urgent need to develop an intelligent voice dispatch method for the 12328 transportation service hotline to improve the efficiency of work order processing and further enhance the service quality of the transportation industry. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent voice dispatch method for transportation service hotlines based on a large language model. This method extracts semantic and temporal feature sequences from incoming call voice data. A pre-trained large language model is then used to classify and extract these semantic and temporal feature sequences according to whether they are for consultation / answering services or work order dispatching services. For consultation / answering services, knowledge content from the transportation service knowledge base is retrieved and processed by a voice encoder to generate voice data for answering / responding. For work order dispatching services, the required content is extracted and dispatched according to work order dispatching rules.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A method for intelligent voice dispatching of transportation service hotlines based on a large language model, the method comprising:

[0006] S1. Construct a semantic temporal recognition model that includes convolutional neural networks and recurrent neural networks. The convolutional neural network extracts semantic features from incoming call voice data, and the recurrent neural network extracts temporal features from incoming call voice data. The semantic temporal recognition model combines semantic features and temporal features to obtain a semantic temporal feature sequence.

[0007] S2. Construct a transportation service knowledge base. Use a pre-trained language model to classify and extract semantic temporal feature sequences according to consultation and answer services and work order dispatch services. For those identified as consultation and answer services, filter and retrieve knowledge content from the transportation service knowledge base and generate voice data through a speech encoder to provide answers and replies. For those identified as work order dispatch services, extract the content required for work order dispatch and dispatch tasks according to work order dispatch rules.

[0008] To better implement this invention, in method S2, the pre-trained language big data model extracts and identifies consultation and answer services and question items from the semantic temporal feature sequence. The consultation and answer service is a question answer service. It retrieves the knowledge content corresponding to the question item from the transportation service knowledge base and generates voice data through a speech encoder to provide the answer. The semantic temporal recognition model continuously extracts subsequent semantic temporal feature sequences to identify whether they are consultation and answer services and whether they contain supplementary question items. If supplementary question items are included, it retrieves the knowledge content corresponding to the supplementary question item from the transportation service knowledge base and generates voice data through a speech encoder to provide the answer.

[0009] The pre-trained language model retrieves relevant knowledge content from the transportation service knowledge base to confirm the questions identified as related to the consultation and answer service. Based on the knowledge content corresponding to the confirmed questions, it generates voice data to provide answers and responses. If the confirmation fails or the response does not meet the requirements, the consultation and answer service is converted into a work order dispatch service and assigned to the relevant business management unit to provide the answer service.

[0010] Preferably, in method S2, a work order database is constructed according to the work order type and work order element composition hierarchy. The content of the work order dispatch service, including region, personnel identity, keywords, and entity words, is identified and extracted. Based on all the content, the work order type and work order element composition are matched with the corresponding work order type and work order element composition based on the work order database. For work order element composition that is incomplete, the pre-trained language model generates voice questions and collects the missing elements. Work order dispatch rules corresponding to the work order database are constructed. Based on all the content of the obtained work order dispatch service, the corresponding work order type is matched and the task is dispatched according to the work order dispatch rules.

[0011] Preferably, in method S1, the method for extracting and recognizing semantic features using a convolutional neural network is as follows: a sliding window is created, and convolutional kernels are used to extract and recognize features of incoming call voice data. A nonlinear activation function is used to perform nonlinear transformation on the output features. Then, a pooling layer is used to perform feature pooling processing. The pooled features are expanded and processed by a fully connected layer and a softmax function to finally obtain semantic features.

[0012] Preferably, in method S1, the recurrent neural network adopts a long short-term memory network LSTM containing a forget gate, an input gate, and an output gate; the incoming call voice data undergoes preprocessing including signal denoising and filtering before being input into the semantic temporal recognition model.

[0013] Preferably, in method S2, the pre-trained language big model converts the semantic temporal feature sequence into text sequence data, performs semantic, syntactic structure and contextual meaning understanding and analysis on the text sequence data, and classifies and extracts the data according to consultation and answering services and work order delivery services.

[0014] Preferably, the transportation service knowledge base uses structured knowledge graph data learned by a pre-trained language big data model for recognition, and the pre-trained language big data model can quickly retrieve knowledge content from the transportation service knowledge base using the structured knowledge graph data.

[0015] Preferably, the pre-trained language big data model directly converts the semantic temporal feature sequence into text sequence data and stores it as raw text data. The pre-trained language big data model performs error correction and optimization processing on the raw text data to obtain coherent and complete processed text data. The pre-trained language big data model performs classification, recognition and extraction processing on the processed text data. The pre-trained language big data model stores the raw text data and processed text data respectively and continuously updates the training and optimization parameters. The error correction and optimization processing is correspondingly set with threshold and rule data.

[0016] Preferably, the data in the transportation service knowledge base includes policy and regulatory data, business process data, and management and supervision data related to the transportation field. The data in the transportation service knowledge base undergoes standardization processing, including removing invalid data and correcting erroneous information. In method S2, a task dispatch feedback mechanism module is constructed, which obtains dispatch error feedback information and re-dispatches the task.

[0017] A transportation service hotline intelligent voice dispatch system, which implements an intelligent voice dispatch method for transportation service hotlines, includes a semantic temporal recognition model, a transportation service knowledge base, a pre-trained language model, a voice answering model, and a work order dispatching module. The semantic temporal recognition model comprises a convolutional neural network and a recurrent neural network. The convolutional neural network extracts semantic features from the incoming call voice data, while the recurrent neural network extracts temporal features. The semantic temporal recognition model combines semantic and temporal features to obtain a semantic temporal feature sequence. The pre-trained language model classifies and extracts the semantic temporal feature sequence according to consultation and answering services and work order dispatching services. For services identified as consultation and answering services, the pre-trained language model filters and retrieves knowledge content from the transportation service knowledge base and inputs it into the voice answering model. The voice answering model generates voice data through a voice encoder to provide answers and responses. For services identified as work order dispatching services, the pre-trained language model extracts the content required for work order dispatching and dispatches the task according to work order dispatching rules through the work order dispatching module.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0019] (1) This invention extracts semantic and temporal features from incoming voice data to obtain a semantic-temporal feature sequence. A pre-trained language model is used to classify and extract the semantic-temporal feature sequence according to consultation and answer services and work order dispatch services. For consultation and answer services, knowledge content from the transportation service knowledge base is selected and retrieved, and voice data is generated through a voice encoder to provide answers and responses. For work order dispatch services, the required content for work order dispatch is extracted and tasks are dispatched according to work order dispatch rules. This invention covers applications such as semantic recognition of work orders, work order type classification, and accurate work order dispatch, effectively improving interactive control and work order transmission efficiency, and reducing the workload of business personnel. It not only improves the efficiency of business request work order flow but also greatly enhances the service quality of the transportation industry.

[0020] (2) This invention solves the problems of low efficiency of manual answering of traditional 12328 service hotlines, inaccurate identification of work order types, and inaccurate dispatch. It constructs a semantic temporal recognition model and a pre-trained language model to achieve accurate identification and understanding of incoming call requests, automatically and accurately determine the type of work order, and accurately dispatch the work order to the most suitable processing department or personnel, thereby significantly improving the service quality and work efficiency of the 12328 transportation service hotline and enhancing public satisfaction.

[0021] (3) The present invention can quickly and accurately identify incoming voice, automatically determine the type of work order and accurately dispatch it, greatly reducing the time and workload of manual processing, significantly improving the answering efficiency and work order processing efficiency of the 12328 service hotline, so that the public can get a satisfactory answer and solution in a shorter time.

[0022] (4) The pre-trained language model of this invention provides a deeper and more accurate understanding of incoming call content, better meeting the needs of the public. The improved accuracy of work order type identification and precise dispatch ensure that problems are handled professionally and promptly, effectively improving service quality and enhancing public trust and satisfaction with the 12328 transportation service hotline. Work orders are dispatched based on factors such as the professional capabilities and workload of the processing departments and personnel, achieving rational utilization and optimized allocation of human resources, avoiding low work efficiency and resource waste caused by uneven work order distribution; at the same time, this invention can dynamically adjust in real time according to changes in business volume and processing status, further improving resource utilization efficiency.

[0023] (5) This invention reduces the reliance on a large number of manual answering and processing of work orders, reduces labor costs, improves work efficiency, reduces repetitive labor and resource waste caused by untimely or incorrect processing, thereby reducing overall operating costs and improving the economic benefits of the transportation service hotline. Attached Figure Description

[0024] Figure 1 This is a flowchart of the intelligent voice dispatch method for transportation service hotlines according to the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the principle of the semantic temporal recognition model in the embodiment;

[0026] Figure 3 This is a schematic diagram illustrating the principle of retrieving knowledge content to provide answers and responses, as exemplified in the embodiments.

[0027] Figure 4 This is a schematic diagram illustrating the principle of answering questions and dynamically updating the transportation service knowledge base in this embodiment. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to embodiments:

[0029] Example

[0030] like Figure 1 As shown, a method for intelligent voice dispatching of transportation service hotlines based on a large language model is presented, the method comprising:

[0031] S1. Construct a semantic temporal recognition model comprising convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The CNNs extract semantic features from incoming call voice data (achieving speech-to-text processing), while the RNNs extract temporal features from the incoming call voice data (preferably, the incoming call voice data undergoes preprocessing, including signal denoising and filtering, before being input into the semantic temporal recognition model). The semantic temporal recognition model combines semantic and temporal features to obtain a semantic temporal feature sequence. The semantic temporal recognition model is trained using sample data, collecting and labeling a large amount of historical work order data to clarify the request type of each work order. The Natural Language Toolkit (NLTK) is used to extract features (including keywords, parts of speech, named entities, and caller roles) from the labeled work order text. Preferably, a deep learning algorithm can be selected to construct a work order type classification model based on a convolutional neural network (CNN), converting text features into vector form and inputting them into the network for training. During model training, methods such as cross-validation are used to evaluate the model's performance, continuously adjusting parameters and optimizing the model structure to improve the model's accuracy and generalization ability.

[0032] In some embodiments, the method for extracting and recognizing semantic features using a convolutional neural network (to achieve speech-to-text processing) is as follows: A sliding window is created, and a convolutional kernel is used to extract and recognize features of the incoming call voice data (semantic features, such as the caller's identity, the type of request, and the content of the request), as shown in the following expression: Where M and N are the kernel sizes, b is the bias term, and i and j are the location points of the feature data points (such as the location coordinates of the position encoding). This indicates convolution processing; it involves performing convolution operations on the input speech to extract local semantic features.

[0033] A non-linear activation function is used to perform a non-linear transformation on the output features. Then, a pooling layer is used for feature pooling. The pooled features are expanded and processed through a fully connected layer and a softmax function to finally obtain semantic features. The softmax function maps the result to the final category label or numerical value, and its expression is: That is, the output = weight matrix × flattened eigenvector + bias vector (i.e., ), where σ is the activation function. Feature pooling is performed using a pooling layer. Taking an input size of W×H, a pooling kernel of k×k, and a stride of s as an example, the output is: .

[0034] In some embodiments, the recurrent neural network employs a Long Short-Term Memory (LSTM) network, which includes a forget gate, an input gate, and an output gate. The LSTM network controls information retention and forgetting. Assume the current time step is t, and the input vector is... The hidden state of the previous time step was Cell state is The hidden state at the current time step is Cell state is The weight matrices and biases of each gate are as follows: and Forget gate: determines how much past information to discard. Input gate: determines how much new information should be added to the cell state. , , To control the write ratio, These are new candidate memories. Cell state update: with the forgetting gate. and input gate and candidate cell status This updates the old cell state. To obtain a new cell state The expression is: , It represents the cell state at the current time step; This represents element-wise multiplication. Output gate: The output gate determines the output from the current cell state. What information is output as the hidden state of the current time step? First, a sigmoid layer determines which parts of the cell state will be output; then, the cell state... Through a The function processes the value (normalizing it to between -1 and 1) and performs element-wise multiplication with the sigmoid output of the output gate. , ,in, Control the output ratio.

[0035] S2. Construct a transportation service knowledge base. Preferably, the data in the transportation service knowledge base includes policy and regulatory data, business process data, and management and supervision data related to the transportation field. The data in the transportation service knowledge base undergoes standardization processing, including removing invalid data and correcting erroneous information. Preferably, such as... Figure 2 As shown, a task dispatch feedback mechanism module is constructed. This module acquires dispatch error feedback information and re-dispatches the task. It also monitors and evaluates the results of voice recognition-based work orders in real time. When the recognition accuracy falls below a set threshold, it automatically triggers secondary recognition or manual intervention to ensure the accuracy of voice-to-text conversion of work orders. In some embodiments, the transportation service knowledge base (or 12328 hotline knowledge base) primarily targets relevant groups in the transportation sector. By integrating policies, regulations, business processes, and service information in the transportation field, it provides comprehensive services such as consultation, complaint handling guidance, and information inquiry. A knowledge base framework that meets the core needs of service recipients is constructed to ensure that knowledge content accurately matches service scenarios, laying the foundation for subsequent knowledge organization and updates. See also... Figure 3 This embodiment is illustrated below: Data such as policies, regulations, business processes, and service information in the transportation sector are integrated and standardized for corresponding knowledge content. Data preprocessing: Due to differences in the expression habits of truck drivers across the country and variations in the recording methods of 12328 operators, different work orders exhibit variations in their wording, making it difficult to directly extract truck driver issues from the complaint work orders. Therefore, preprocessing of the complaint work orders is necessary. Large-scale model analysis: The preprocessed work orders are analyzed using a large-scale model to identify the complainant and the object of the complaint, accurately selecting work orders where the complainant is a truck driver. Simultaneously, duplicate records, multiple complaints, and invalid requests are removed to ensure data validity and relevance. Standardization processing: Invalid data is removed and errors are corrected. The style of problem expression is standardized and unified, and the complaint items are summarized to ensure consistency and standardization of knowledge content, facilitating subsequent storage, retrieval, and application. Figure 4As shown, the transportation service knowledge base, in addition to its function of providing answers and responses, can also be dynamically updated. The knowledge update content and process mainly involve periodically analyzing and extracting hot issues related to public demands, and supplementing the knowledge base with response guidelines, handling opinions, or typical case verification opinions provided by relevant business management units. (During the operation of this invention, newly generated work order data is collected in real time. After preprocessing and labeling the new work order data, it is incorporated into the training dataset. The stochastic gradient descent (SGD) online learning algorithm is used to update and optimize the work order type classification model in real time. When new work order data arrives, the model adjusts its parameters according to the characteristics of the new data to adapt to the new work order types and changes in language expression.) Based on relevant policy changes, updated demands, and handling feedback results, the knowledge base content that needs to be updated is identified in a timely manner, and the update process is initiated to ensure that the knowledge base content keeps pace with changes in actual needs.

[0036] In some embodiments, the transportation service knowledge base uses structured knowledge graph data learned by a pre-trained language model. The pre-trained language model quickly retrieves knowledge content from the transportation service knowledge base using this structured knowledge graph data. The pre-trained language model is used to classify and extract semantic temporal feature sequences according to consultation and answer services and work order dispatch services. In this embodiment, the pre-trained language model is trained using DeepSeek-R1 with sample data. It performs in-depth semantic understanding and analysis on the text converted from the voice of a call to the 12328 service hotline. The text is formatted and encoded, and then input into the pre-trained language model for semantic analysis. The pre-trained language model can understand the semantics, grammatical structure, and contextual meaning of the text. During the access process, a secure and stable communication connection is established to ensure the confidentiality and integrity of data transmission. Simultaneously, based on actual business needs, the output of the pre-trained language model is filtered and parsed to extract key information related to work order processing. For services identified as consultation and answering, the system filters and retrieves knowledge content from the transportation service knowledge base, generates voice data through a speech encoder, and provides answers and responses. For services identified as work order dispatching, the system extracts the content required for work order dispatching and dispatches the task according to the work order dispatching rules. In some embodiments, the pre-trained language model DeepSeek-R1 in this embodiment supports knowledge retrieval from the existing knowledge base via a dialogue interface. Retrieval via the dialogue interface: Simply enter the result "[12328 request work order voice-to-text]" in the DeepSeek-R1 input box, and the model will automatically perform grammatical parsing, knowledge matching detection, and other operations. If structured data from a relevant knowledge graph is matched in the existing knowledge base, a standard answer will be generated directly; if the match fails, the generation model will be activated, generating an answer through methods such as combination and innovation. Finally, after answer verification and language style adaptation, the final result is output.

[0037] In some embodiments, the pre-trained language big data model extracts and identifies consultation and answer services and question items from semantic temporal feature sequences. The consultation and answer service is a question answer service, which retrieves knowledge content corresponding to the question item from the transportation service knowledge base, such as... Figure 2 As shown, a speech encoder generates speech data for answering questions and responding. The semantic temporal recognition model continuously extracts subsequent semantic temporal feature sequences to identify whether it is a consultation and answering service and whether it contains supplementary questions (a pre-trained language model generates intelligent guiding questions to conduct multiple rounds of dialogue with the caller, maintaining the coherence and logic of the dialogue during these rounds, and adjusting the guiding questions in a timely manner based on the caller's answers to ensure accurate confirmation of the caller's intent). If supplementary questions are included, the corresponding knowledge content from the transportation service knowledge base is retrieved, and speech data is generated by the speech encoder to provide the answer and response.

[0038] The pre-trained language model retrieves relevant knowledge content from the transportation service knowledge base to confirm the questions identified as related to the consultation and answer service. Based on the knowledge content corresponding to the confirmed questions, it generates voice data to provide answers and responses. If the confirmation fails or the response does not meet the requirements, the consultation and answer service is converted into a work order dispatch service and assigned to the relevant business management unit to provide the answer service.

[0039] In some embodiments, a work order database is constructed according to work order type and work order element hierarchy. Work order elements include several elements, such as region, keywords, and / or entity words. The work order dispatch service identifies and extracts content including region, personnel identity, keywords, and entity words. Based on all content, the work order database is matched with the corresponding work order type and work order element composition. For incomplete work order element compositions, a pre-trained language model generates missing elements to supplement the voice questions and collect the missing elements. Work order dispatch rules corresponding to the work order database are constructed. Based on all content obtained from the work order dispatch service, the corresponding work order type is matched, and tasks are dispatched according to the work order dispatch rules. In some embodiments, the pre-trained language model converts semantic temporal feature sequences into text sequence data. The text sequence data undergoes semantic, syntactic structure, and contextual meaning analysis. It is then classified and its identification and extraction are performed according to the consultation and answering service and the work order dispatch service.

[0040] In some embodiments, the pre-trained language model directly converts semantic temporal feature sequences into text sequence data, storing it as raw text data. The pre-trained language model then performs error correction and optimization on the raw text data to obtain coherent and complete processed text data. The pre-trained language model uses the processed text data for classification, recognition, and extraction. The pre-trained language model stores the raw text data and processed text data respectively and continuously updates training and optimization parameters. The error correction and optimization processes are configured with thresholds and rules. In some embodiments, the text generation capability of the pre-trained language model can be utilized to correct and optimize the content of the speech-to-text version of a work order. The speech recognition result is used as input, allowing the pre-trained language model to generate an optimized text version. By comparing the original text and the optimized text, potential errors and irregularities are identified and automatically corrected. During the error correction process, reasonable thresholds and rules are set to avoid over-correction that alters the original meaning of the work order. Simultaneously, the corrected text and the original text are recorded together for subsequent training and optimization of the work order semantic recognition model, continuously improving the effectiveness of semantic error correction and optimization.

[0041] A transportation service hotline intelligent voice dispatch system, which implements an intelligent voice dispatch method for transportation service hotlines, includes a semantic temporal recognition model, a transportation service knowledge base, a pre-trained language model, a voice answering model, and a work order dispatching module. The semantic temporal recognition model comprises a convolutional neural network and a recurrent neural network. The convolutional neural network extracts semantic features from the incoming call voice data, while the recurrent neural network extracts temporal features. The semantic temporal recognition model combines semantic and temporal features to obtain a semantic temporal feature sequence. The pre-trained language model classifies and extracts the semantic temporal feature sequence according to consultation and answering services and work order dispatching services. For services identified as consultation and answering services, the pre-trained language model filters and retrieves knowledge content from the transportation service knowledge base and inputs it into the voice answering model. The voice answering model generates voice data through a voice encoder to provide answers and responses. For services identified as work order dispatching services, the pre-trained language model extracts the content required for work order dispatching and dispatches the task according to work order dispatching rules through the work order dispatching module. According to the rules in Part 4 of the "12328 Transportation Service Supervision Hotline System: Business Classification and Coding," incoming voice work orders are classified and coded. Detailed work order dispatch rules are formulated, taking into account factors such as work order type, business area, geographical scope, processing department, and staff on-duty status. These rules cover multiple factors including work order type, business area, geographical scope, processing department, and staff on-duty status. Based on the work order type identification results and dispatch rules, the system uses intelligent algorithms to automatically and accurately match and dispatch work orders to the most suitable processing department or personnel. After work order type identification, the system intelligently matches the work order with the processing department and personnel database based on specific information such as work order type, business area, and involved region, combined with the dispatch rules. After matching, the system accurately dispatches the work order to the corresponding processing department or personnel through internal communication mechanisms. During the dispatch process, the dispatch path and time of the work order are recorded for subsequent tracking and querying.

[0042] After receiving a work order, the system can track its processing progress in real time and receive feedback. The system monitors the processing progress in real time and issues an alert if a work order is not completed within the specified time. Simultaneously, if processing personnel discover errors in work order assignment or issues beyond their capabilities, they can use the system's feedback function to report the problem to the work order assignment module. The assignment module then re-evaluates and assigns the work order based on the feedback. If assignment errors or problems arise during processing, the system can promptly reassign the work order or adjust the processing flow based on the feedback, ensuring that the work order is properly resolved.

[0043] In the field of the 12328 transportation service hotline, this invention uses convolutional neural networks (CNN) to accurately capture local semantic features of work order text, recurrent neural networks (RNN) to effectively process the temporal information of incoming voice, and combines large-scale pre-trained language models to achieve full-process automation and intelligence from speech to text to semantic understanding, intelligent identification of work order type, and accurate delivery. This not only significantly improves the hotline response efficiency and work order processing accuracy, but also optimizes the system by continuously learning new data, providing the public with more efficient and personalized transportation services and reshaping the industry's service model. This paper innovatively utilizes the Natural Language Processing (NLTK) tool to identify caller ID types for the 12328 transportation service hotline. It deeply integrates NLTK's rich text processing capabilities (such as part-of-speech tagging, named entity recognition, and syntactic analysis) with transportation business scenarios. By quickly extracting key semantic features such as urban passenger transport, highways, and ordinary roads from caller ID text, and combining this with deep learning algorithms to build a classification model, it efficiently parses the caller's intent. Simultaneously, leveraging NLTK's text corpus and pre-trained models, it accurately understands complex expressions and industry terminology. Based on NLTK's flexible customization capabilities, it can quickly adapt to new caller ID types and language changes in the transportation field, achieving high efficiency, accuracy, and dynamic optimization in caller ID type identification. This provides a low-cost, highly adaptable technical path for the intelligent upgrade of hotline services. The 12328 work order precision dispatch rule integrates multiple factors such as work order type, business area, geographical scope, processing department, and real-time workload of personnel. It uses intelligent optimization algorithms to achieve optimal matching between work orders and processing resources. At the same time, by establishing a real-time feedback and dynamic adjustment mechanism, the dispatch route is promptly corrected based on abnormal feedback during the processing. This ensures that complex and ever-changing transportation service requests can be accurately matched with professional resources, significantly improving the efficiency of work order circulation and the quality of problem solving, and providing an innovative paradigm for the efficient collaborative management of transportation service hotline resources.

[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent voice dispatching of transportation service hotlines based on a large language model, characterized in that: The methods include: S1. Construct a semantic temporal recognition model that includes convolutional neural networks and recurrent neural networks. The convolutional neural network extracts semantic features from incoming call voice data, and the recurrent neural network extracts temporal features from incoming call voice data. The semantic temporal recognition model combines semantic features and temporal features to obtain a semantic temporal feature sequence. S2. Construct a transportation service knowledge base. Utilize a pre-trained language model to classify and extract semantic temporal feature sequences according to consultation and answer services and work order delivery services. The pre-trained language model converts the semantic temporal feature sequences into text sequence data, performs semantic, syntactic structure, and contextual meaning analysis on the text sequence data, and classifies and extracts data according to consultation and answer services and work order delivery services. For those identified as consultation and answer services, knowledge content from the transportation service knowledge base is selected and retrieved, and voice data is generated through a speech encoder for answering and replying. The pre-trained language model extracts and identifies consultation and answer services and question items from the semantic temporal feature sequences. Consultation and answer services are question answering services; knowledge content corresponding to the question item is retrieved from the transportation service knowledge base, and voice data is generated through a speech encoder for answering and replying. The semantic temporal recognition model continuously extracts subsequent semantic temporal feature sequences to identify whether they are consultation and answer services and whether they contain supplementary question items. If supplementary question items are included, knowledge content corresponding to the supplementary question item is retrieved from the transportation service knowledge base, and voice data is generated through a speech encoder for answering and replying. The pre-trained language model retrieves relevant knowledge from the transportation service knowledge base to confirm questions identified as related to the consultation and answer service. Based on the confirmed knowledge content, it generates voice data to provide answers. If confirmation fails or the answer response is unsuitable, the consultation and answer service is converted into a work order dispatch service and assigned to the relevant business management unit for assistance. For work order dispatch services, the required content is extracted and dispatched according to work order dispatch rules. A work order database is constructed based on work order type and work order element hierarchy. Content including region, personnel identity, keywords, and entity words is identified and extracted for work order dispatch services. All content is matched against the work order database to the corresponding work order type and elements. For incomplete work order elements, the pre-trained language model generates supplementary voice questions and collects the missing elements. Work order dispatch rules corresponding to the work order database are constructed. All content in the work order dispatch service is matched against the corresponding work order type, and tasks are dispatched based on these rules.

2. The intelligent voice dispatching method for transportation service hotlines based on a large language model according to claim 1, characterized in that: In method S1, the convolutional neural network extracts and recognizes semantic features as follows: a sliding window is created, and convolutional kernels are used to extract features from incoming call voice data. A non-linear activation function is used to perform non-linear transformation on the output features. Then, a pooling layer is used to perform feature pooling. The pooled features are expanded and processed by a fully connected layer and a softmax function to finally obtain semantic features.

3. The intelligent voice dispatching method for transportation service hotlines based on a large language model according to claim 1, characterized in that: In method S1, the recurrent neural network uses a long short-term memory network LSTM that includes a forget gate, an input gate, and an output gate; the incoming call voice data undergoes preprocessing, including signal denoising and filtering, before being input into the semantic temporal recognition model.

4. The intelligent voice dispatching method for transportation service hotlines based on a large language model according to claim 1, characterized in that: The transportation service knowledge base uses structured knowledge graph data learned by a pre-trained language big data model for recognition. The pre-trained language big data model can quickly retrieve knowledge content from the transportation service knowledge base using the structured knowledge graph data.

5. The intelligent voice dispatching method for transportation service hotlines based on a large language model according to claim 1, characterized in that: The pre-trained language big data model directly converts semantic temporal feature sequences into text sequence data and stores it as raw text data. The pre-trained language big data model performs error correction and optimization on the raw text data to obtain coherent and complete processed text data. The pre-trained language big data model performs classification, recognition and extraction processing on the processed text data. The pre-trained language big data model stores the raw text data and processed text data respectively and continuously updates the training and optimization parameters. The error correction and optimization processing is set with threshold and rule data.

6. The intelligent voice dispatching method for transportation service hotlines based on a large language model according to claim 1, characterized in that: The data in the transportation service knowledge base includes policy and regulatory data, business process data, and management and supervision data related to the transportation field. The data in the transportation service knowledge base undergoes standardization processing, including removing invalid data and correcting erroneous information. In method S2, a task dispatch feedback mechanism module is constructed. The feedback mechanism module obtains dispatch error feedback information and re-dispatches the task.

7. A transportation service hotline intelligent voice dispatch system implementing the transportation service hotline intelligent voice dispatch method of claim 1, characterized in that: It includes a semantic temporal recognition model, a transportation service knowledge base, a pre-trained language model, a voice answering model, and a work order dispatching module. The semantic temporal recognition model includes a convolutional neural network and a recurrent neural network. The convolutional neural network extracts semantic features from incoming call voice data, and the recurrent neural network extracts temporal features from incoming call voice data. The semantic temporal recognition model combines semantic features and temporal features to obtain a semantic temporal feature sequence. The pre-trained language big data model classifies and extracts semantic temporal feature sequences according to consultation and answer services and work order dispatch services. For consultation and answer services, the pre-trained language big data model filters and retrieves knowledge content from the transportation service knowledge base and inputs it into the voice answer model. The voice answer model generates voice data through a voice encoder to provide answers and replies. For work order dispatch services, the pre-trained language big data model extracts the content required for work order dispatch and dispatches the task according to the work order dispatch rules through the work order dispatch scheduling module.

Images