Natural language container multimodal transport interaction method and system based on large language model

Abstract

The invention belongs to the field of container multimodal transport, and particularly relates to a natural language container multimodal transport interaction method and system based on a large language model. The method comprises the steps of obtaining a container multimodal transport query request submitted by a user in a natural language form; and inputting a preset prompt word and a user query request into the large language model together to generate a system query text, and evaluating the integrity of the container multimodal transport query request and a question proposed in a natural language form when the container multimodal transport query request is insufficient, so that the system query text is more perfect. Compared with manual service query and self-service query in the traditional container multimodal transport industry, the query request proposed by the user in a natural language form can be directly processed, and the usability and accessibility of the system are improved. And meanwhile, rich and accurate feedback information is provided for the user by adopting a retrieval enhancement generation technology and combining a large language model and a professional knowledge base.

Description

Technical Field

[0001] The present invention belongs to the field of container multimodal transport, and in particular relates to a natural language container multimodal transport interaction method and system based on a large language model. Background Art

[0002] With the booming development of global trade, container multimodal transport has gradually become the key to trade circulation with its high efficiency and standardized form of cargo transportation. Container multimodal transport involves the combination of different modes of transportation, such as rail, road, sea and air, aiming to achieve the transfer of goods from the departure point to the destination. However, the continuous expansion of the supply chain and the increasing complexity of the transportation process have brought challenges to traditional information exchange, status inquiry and transportation coordination. At present, users often need to enter a lot of information through manual customer service or use complex forms of self-service systems, such as cargo number, origin, destination, mode of transportation, container specifications, type of goods, etc., in order to place an order or obtain logistics information. These methods have reduced efficiency and increased costs.

[0003] At the same time, as the global division of labor continues to refine, trade participants are often distributed all over the world, and language and cultural differences may lead to misunderstandings and communication barriers, increasing the complexity of the interaction process. In addition, with the rise of cross-border e-commerce, the threshold for global trade participants continues to decrease, but the traditional operation interface is too complicated for non-professional users, information input is cumbersome and error-prone, and often requires rich cross-border logistics-related experience, so communication can only be carried out with the help of large freight forwarders and carriers. For integrated logistics service providers that provide container multimodal transport, maintaining the traditional user service system requires a lot of human resources, especially in manual customer service. This high operating cost causes logistics service providers to often have to give up small and medium-sized customer groups with relatively high profit margins, thereby limiting business expansion and profit growth. Therefore, how to use artificial intelligence technology to provide users with more friendly, accurate and efficient container multimodal transport interactive services is an urgent problem to be solved. Summary of the invention

[0004] In order to solve the technical problems in the prior art that manual customer service for container intermodal transport requires a large amount of human resources and high operating costs limit the profit growth of container intermodal transport interactive services, the present invention proposes a natural language container intermodal transport interactive method and system based on a large language model. The system aims to achieve natural language interaction with users through advanced artificial intelligence technology, especially a large language model, and greatly improve the efficiency and accuracy of users' query of container intermodal transport information.

[0005] The technical solution claimed in the present invention is as follows:

[0006] A natural language container multimodal transport interaction method based on a large language model, comprising:

[0007] S1. Obtaining a query request: obtaining a container multimodal transport query text submitted by a user in a natural language form, that is, an original query text;

[0008] S2. Generate system query text: Input the container multimodal transport query text and preset prompt words into the large language model to generate parsed and automatically completed system query text, wherein the prompt words are generated by a fixed template method and a prompt word library method, including stipulating the data format and field requirements returned by the large language model, evaluating the completeness of the container multimodal transport query request, and asking questions in natural language when the container multimodal transport query request is insufficient;

[0009] S3, supplement and confirm the complete system query text: determine whether the system query text in S2 is sufficient according to the evaluation result of the large language model on the completeness of the container multimodal transport query request; if it is complete, display the system query text to the user as a complete system query text for confirmation and use it as a complete system query text after confirmation; if it is incomplete, ask the user in natural language form to supplement the information to the container multimodal transport query text in S1, and repeat step S2;

[0010] S4. Retrieving container multimodal transport routes and prices using a container multimodal transport query database: using the complete system query text confirmed by the user in S3, calling the container multimodal transport query database, and obtaining container multimodal transport route and price query results;

[0011] S5. Initiate a search in the container multimodal transport knowledge base through the large language model search enhancement module: obtain user related information, use the user related information, the route and price query results and the complete system query text to construct a complete search text, and input it into the large language model again; use the large language model search enhancement module to initiate a search in the container multimodal transport knowledge base to obtain matching information fragments, and merge them with the container multimodal transport route and price query results obtained in S4 according to the search category to obtain the final query result, and the final query result is displayed according to the search category;

[0012] S6. The system displays the query results: The system displays the final query results described in S5 and collects feedback from system users for future query optimization.

[0013] Preferably, S1 specifically includes:

[0014] The natural language request includes a natural language request submitted directly by the user in the form of text and a natural language request submitted in the form of voice or picture;

[0015] The natural language request includes query requests in multiple languages ​​that can be set, including internationally used English, Chinese and other languages.

[0016] Preferably, S2 specifically includes:

[0017] S21, large language model multimodal conversion: When the user inputs voice and pictures, the large language model converts the user's input voice and pictures into text query requests;

[0018] S22. Use prompt words to agree on the data format and field requirements returned by the large language model, including but not limited to: detailed addresses of the origin and destination of cargo transportation, shipping date, type, volume, and weight of the cargo;

[0019] S23. The query text language returned by the prompt word agreement includes both the language used by the user when querying and the language supported by the container multimodal transport database query.

[0020] Preferably, S3 specifically includes:

[0021] S31. Evaluate the completeness of the system query text based on the large language model. If it is determined that the information is insufficient or needs to be further clarified, the system will ask the user a follow-up question in natural language, where the follow-up question is a natural language question generated by step S2; and add the result of the follow-up question to the container multimodal transport query text described in step S1, and then input it into the large language model of step S2 in combination with the preset prompt words;

[0022] S32: The system displays the parsed query text in the language of the natural language request input by the user for confirmation. If the user confirms, it will be used as the complete system query text. If not, the system returns to step S31 to question the user in natural language.

[0023] Preferably, S5 specifically includes:

[0024] S51. Obtaining user-related information, including but not limited to user history query, order information, and preference settings;

[0025] S52, constructing a complete search text using the user information, the aforementioned route and price query results, and the system query text, and inputting the complete search text into the large language model again;

[0026] S53: Using the large language model retrieval enhancement generation technology, a search is initiated to the container multimodal transport knowledge base. The search content includes domestic and international transportation regulations, policies, tariffs, insurance, routes, environmental protection, service provider information and evaluation. The retrieval enhancement generation technology includes: BM25 scoring formula or semantic search technology.

[0027] Preferably, the BM25 scoring formula in S53 is as follows

[0028]

[0029] Where D represents the document; Q represents the query, which consists of multiple query terms q i Composition; f(q i ,D) represents the query word q i The frequency in document D, IDF(q i ) represents the query word q i The inverse document frequency; |D| represents the length of the document; avgdl represents the average length of the document in the document collection; k1, b represent free parameters, which are used to adjust the saturation of word frequency and the influence of document length, usually k1∈[1.2,2.0] and b=0.75; In addition, the IDF calculation formula is:

[0030]

[0031] Where N represents the total number of documents in the document collection; n(q i ) represents the query word q i The number of documents.

[0032] Preferably, the steps of semantic search in S53 are as follows:

[0033] S531: Using a pre-trained neural network model: Converting queries and documents into embedding vectors in a high-dimensional space can be summarized as:

[0034]

[0035] Among them, EmbeddingModel is the selected pre-trained model, and text is the query and document text data;

[0036] S532: Calculate the cosine similarity between the query vector and the text vector, and determine the most semantically relevant documents by sorting. The cosine similarity is specifically calculated by the following formula:

[0037]

[0038] in, and are two vectors, representing the embedding vectors of the query and document respectively. The dot product in the numerator computes the inner product of the two vectors, while the denominator is the product of the norms (or lengths) of the two vectors.

[0039] Preferably, S6 specifically includes:

[0040] S61: The system collects user feedback on the query results, including but not limited to satisfaction scores, specific opinions and suggestions;

[0041] S62: Analyze user feedback data and identify improvement points, such as query accuracy and user interface optimization;

[0042] S63: Integrate feedback into future query optimization and system iterations, including model training, algorithm adjustment, and user experience optimization.

[0043] A natural language container multimodal transport interaction system based on a large language model, comprising:

[0044] The user interface module includes: a container query submodule for obtaining a container multimodal transport query request submitted by a user in natural language, and an interactive feedback submodule for feeding back the query result processed by the large language model or a follow-up question raised in natural language when the container multimodal transport query request is insufficient to the user; the query result includes: container multimodal transport route and price query result;

[0045] The natural language processing module includes: a large language model query submodule that inputs the container multimodal transport query request and preset prompt words described in the user interface module into a large language model to generate a system query text, wherein the prompt words include the data format, field requirements, query text integrity assessment, and a follow-up questioning mechanism when the information is insufficient.

[0046] The data query module includes: a query text input submodule for inputting the system query text described in the natural language processing module into the container multimodal transport database query interface for query, and a result output submodule for outputting the container multimodal transport route and price query results;

[0047] Container intermodal transport knowledge base, used to store basic knowledge related to container intermodal transport, including but not limited to domestic and international transportation regulations, policies, tariffs, insurance, routes, environmental protection, service provider information and evaluation;

[0048] The retrieval enhancement generation module, combined with the RAG technology, initiates an enhanced retrieval request to the container multimodal transport knowledge base with a large language model, obtains matching information fragments and merges them with the container multimodal transport route and price query results in the data query module to obtain the final query result;

[0049] A result display module, including a result display submodule that displays the final query result obtained by the search enhancement generation module in natural language, wherein the query result includes container transportation options, timetables, cost estimates, and special tips or instructions;

[0050] The user feedback module is used to collect user feedback and improve the response speed, accuracy and satisfaction of future queries.

[0051] Preferably, the user interface module further includes:

[0052] An information confirmation submodule, used to confirm to the user the user query request extracted and parsed by the large language model;

[0053] The information inquiry submodule is used to inquire about information when the large language model evaluates that the user query request is insufficient.

[0054] Preferably, the natural language processing module further includes:

[0055] The information extraction submodule is used to convert the query request in non-text form such as user voice or picture into a natural language query request in text form;

[0056] The language conversion submodule is used to convert a user's query request in a language not supported by the system into a query request in a language supported by the system.

[0057] Preferably, the container multimodal transport knowledge base includes:

[0058] The transportation regulations and policies database submodule is used to store domestic and international transportation regulations, policies, tariffs and other information;

[0059] Route and service provider database submodule: used to contain various transportation routes, service provider information, evaluation and cost;

[0060] Environmental protection submodule: used to include carbon emission regulations and calculation related information.

[0061] Preferably, the search enhancement generation module further includes:

[0062] Knowledge base retrieval submodule: using the acquired information, perform search queries in the container intermodal transport knowledge base to find matching information fragments;

[0063] Information fusion submodule: fuses the information retrieved by the knowledge base retrieval submodule with the container multimodal transport route and price query results in the data query module, and constructs an information set;

[0064] Generate optimization submodule: Based on the fused information, use a large language model to generate optimized and adjusted natural language responses.

[0065] In the present invention, a natural language container multimodal transport interaction method and system based on a large language model is proposed, which obtains a container multimodal transport query request submitted by a user in the form of natural language; then the preset prompt words and the user query request are input into the large language model together to generate a system query text. At the same time, based on the preset prompt words, the system query text can be improved by evaluating the completeness of the container multimodal transport query request and the follow-up questions raised in the form of natural language when the container multimodal transport query request is insufficient; the complete system query text confirmed by the user is used to call the container multimodal transport query database to obtain the multimodal transport route and price query results; the retrieval enhancement generation technology is used to initiate a search to the container multimodal transport knowledge base to enhance the feedback results generated by natural language; and the user feedback is collected for future query optimization. Compared with the traditional manual service query and self-service query in the container multimodal transport industry, it can directly process the query request raised by the user in the form of natural language, which improves the ease of use and accessibility of the system. At the same time, the retrieval enhancement generation technology is adopted, combined with the large language model and the professional knowledge base, to provide users with rich and accurate feedback information, which greatly reduces the threshold for use. In addition, by analyzing user feedback, the system continuously optimizes the intelligent query algorithm, which can continuously improve the speed, accuracy and user satisfaction of future queries.

[0066] Users only need to make a query request in a simple natural language form, such as: "I need to send a batch of goods from Zhangjiang Hi-Tech Park in Pudong New Area, Shanghai to Williamsburg in Brooklyn, New York next Monday. The goods include 200 electronic devices, each measuring 30cm x 20cm x 15cm, with a total weight of approximately 500kg. Please provide transportation options, estimated arrival time, and cost estimates." Or "I plan to transport a batch of photovoltaic panels from the photovoltaic manufacturing plant in Donggang Town, Wuxi, Jiangsu to the solar power station in Aschim Town, east of Munich, Germany in about 20 days. This batch of goods contains 100 photovoltaic panels, each measuring 165cm x 99cm x 4cm, with a total weight of approximately 2,000kg. Considering the fragility of photovoltaic panels, please provide transportation options suitable for such goods, including insurance coverage, estimated arrival time, and cost estimates."

[0067] The system proposed by the present invention can understand such natural language requests and automatically extract key information, such as the detailed addresses of the origin and destination, the shipping date, and the description of the goods (type, volume, weight), so that the user does not need to manually enter complex details. For information that the user does not mention but the system needs to know, the system can also ask the user in the form of natural language to ensure the integrity of the information. For the query information confirmed by the user, the system can provide detailed query results including transportation method selection, estimated shipping and arrival dates, and cost estimates, and attach precautions and instructions for special transportation requirements of the goods for further operation by the customer. In addition, the system can also prompt the user with additional information such as insurance options, possible tariffs, and import and export restrictions to support customers in making comprehensive transportation decisions.

[0068] The interactive method and system proposed in the present invention not only provide users with more efficient and accurate query services, but also greatly reduce the user's operating difficulty and improve the user experience through natural language interaction. In addition, the system can also optimize its response capabilities through continuous learning and further improve the service quality for users. By solving the problems of low efficiency, high cost, poor user experience, etc. in the prior art, the present invention provides an innovative solution for integrated logistics service providers in the container multimodal transport industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A flowchart of a natural language container multimodal transport interaction method based on a large language model proposed by the present invention;

[0070] Figure 2 A detailed workflow diagram of a natural language container multimodal transport interaction method based on a large language model proposed by the present invention in an embodiment;

[0071] Figure 3 It is a schematic diagram of the architecture of a natural language container multimodal transport interaction system based on a large language model proposed by the present invention in an embodiment;

[0072] Figure 4 It is an architecture diagram of another natural language container multimodal transport interaction system based on a large language model proposed by the present invention in the embodiment;

[0073] Figure 5 The present invention is a schematic diagram of the architecture of a retrieval enhancement generation module in a natural language container multimodal transport interactive system based on a large language model proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0074] Reference Figure 1-2 The present invention proposes a natural language container multimodal transport interaction method based on a large language model, comprising:

[0075] S1. Obtaining a container multimodal transport query text submitted by a user in natural language form;

[0076] Specifically, in this implementation example, the user can enter a specific query in the form of text, voice or image, such as: "I need to send a batch of goods from Zhangjiang Hi-Tech Park in Pudong New Area, Shanghai to Williamsburg in Brooklyn, New York next Monday. The goods include 200 electronic devices, each measuring 30cm x 20cm x 15cm, with a total weight of approximately 500kg. Please provide transportation options, estimated arrival time and cost estimates." Or "I plan to transport a batch of photovoltaic panels from the photovoltaic manufacturing plant in Donggang Town, Wuxi, Jiangsu to the solar power station in Aschim Town, east of Munich, Germany in about 20 days. This batch of goods contains 100 photovoltaic panels, each measuring 165cm x 99cm x 4cm, with a total weight of approximately 2,000kg. Considering the fragility of photovoltaic panels, please provide transportation options suitable for such goods, including insurance coverage, estimated arrival time and cost estimates."

[0077] S2. Inputting the container multimodal transport query request and preset prompt words described in S1 into the large language model to generate a parsed and automatically completed system query text, wherein the preset prompt words are generated by a fixed template method and a prompt word library method, including stipulating the data format and field requirements returned by the large language model, evaluating the completeness of the container multimodal transport query request, and asking questions in natural language when the container multimodal transport query request is insufficient;

[0078] “S2” specifically includes:

[0079] S21. Using the multimodal capabilities of the large language model, when the user inputs voice and pictures, the voice and pictures input by the user are converted into text;

[0080] S22. The data format and field requirements returned by the large language model include but are not limited to the detailed addresses of the origin and destination, the date of shipment, the type, volume, and weight of the goods;

[0081] S23. The query text language returned by the prompt word agreement includes both the language used by the user when querying and the language supported by the container multimodal transport database query.

[0082] Specifically, the large model in the example is a multimodal general model. A general model is a type of model designed to handle a variety of different tasks or data types. Such models are usually trained based on a wide range of data sets that cover a rich range of language usage, knowledge domains, and task types. Multimodality means that the model has the ability to understand and process multimodal data such as text, images, audio, etc. Multimodal general models usually exhibit high adaptability and can adapt to specific tasks through fine-tuning. They can also demonstrate processing capabilities on tasks that have not been trained. GPT-3, Llama 3, etc. are generally considered to be multimodal general models.

[0083] For large language models, prompts play a role in guiding model content generation, constraining model output, stimulating model creativity, and achieving zero-shot or few-shot learning. Prompts not only help the model understand user intentions and needs, but also guide the model to produce accurate, relevant, and diverse outputs, providing a bridge for effective communication between users and large language models.

[0084] The prompt words include guiding the large model to understand the voice and image query requests input by the user, the data format and field requirements parsed after the large language model understands and captures the user's query requests, and the language requirements for the returned results.

[0085] Specifically, the fields required for the parsing results returned by the prompt word agreement large model include the origin address, destination address, shipping date, type of goods, volume, weight and other supplementary requirements.

[0086] Here is an example of a prompt word:

[0087] [1. Convert all non-text input, such as voice or pictures, into text format;

[0088] 2. Extract the following key information in detail to form a complete query request. The returned data should be strictly presented in JSON object format, excluding any text or explanation that is not in JSON object format;

[0089] 3. The required field 'data' indicates the content of the input data.

[0090] 4. The required fields 'departure' and 'destination' represent the place of shipment (origin) and the place of receipt (destination) respectively. If there is no information, fill in 'NULL'.

[0091] 5. Optional fields 'shipmentDate' and 'expArrivalDate', which represent the shipping date and expected delivery date respectively. If it is "next Monday", etc., it needs to be converted in combination with the current query date. The format must follow the YY-MM-DD format. If there is no specific information, fill in 'NULL'

[0092] 6. The optional fields 'cargoType', 'weight', 'volume' and 'supplement' indicate the cargo type, cargo weight, cargo volume and other supplementary information respectively. ]

[0093] S3. Display the parsed and automatically completed system query text to the user for confirmation, and ask additional questions in the form of natural language to supplement information based on the evaluation results of the integrity of the container multimodal transport query request by the aforementioned large language model; 14. S3 specifically includes:

[0094] S31. Evaluate the completeness of the system query text based on the large language model. If it is determined that the information is insufficient or needs to be further clarified, the system will ask the user a follow-up question in natural language, where the follow-up question is a natural language question generated by step S2; and add the result of the follow-up question to the container multimodal transport query text described in step S1, and then input it into the large language model of step S2 in combination with the preset prompt words;

[0095] S32: The system displays the parsed query text in the language of the natural language request input by the user for confirmation. If the user confirms, it will be used as the complete system query text. If not, the system returns to step S31 to question the user in natural language.

[0096] S4. Using the complete system query text confirmed by the user, calling the container multimodal transport query database, and obtaining multimodal transport route and price query results;

[0097] S5. Obtain user-related information, use the user-related information, the aforementioned route and price query results, and the system query text, and combine the retrieval enhancement generation (RAG) technology to initiate a search in the container multimodal transport knowledge base to construct a complete search text, and input it into the large language model again;

[0098] “S5” specifically includes:

[0099] S51. Obtaining user related information, including but not limited to user history query, order information, preference settings, etc.;

[0100] S52. Using the user information, the aforementioned route and price query results, and the system query text, and combining the retrieval enhancement generation technology, the most relevant information to the user query is retrieved from the container multimodal transport knowledge base to construct a complete retrieval text.

[0101] The search enhancement generation technology described in S53 and S52 specifically includes the BM25 scoring formula or semantic search technology. The BM25 (Best Matching 25) formula is as follows:

[0102]

[0103] Where D represents the document; Q represents the query, which consists of multiple query terms q i Composition; f(q i ,D) represents the query word q i The frequency in document D, IDF(q i ) represents the query word q i The inverse document frequency; |D| represents the length of the document; avgdl represents the average length of the document in the document collection; k1, b represent free parameters, which are used to adjust the saturation of word frequency and the influence of document length, usually k1∈[1.2,2.0] and b=0.75; In addition, the IDF calculation formula is:

[0104]

[0105] Where N represents the total number of documents in the document collection; n(q i ) represents the query word q i The number of documents.

[0106] The steps of semantic search are as follows:

[0107] (1) First, use a pre-trained neural network model such as BERT or GPT to convert queries and documents into embedding vectors in a high-dimensional space, which can be summarized as follows:

[0108]

[0109] Among them, EmbeddingModel is the selected pre-trained model, and text is the query and document text data.

[0110] (2) Calculate the cosine similarity between the query vector and the text vector, and determine the most semantically relevant documents by sorting. The cosine similarity is calculated using the following formula:

[0111]

[0112] in, and are two vectors, representing the embedding vectors of the query and document respectively. The dot product in the numerator computes the inner product of the two vectors, while the denominator is the product of the norms (or lengths) of the two vectors.

[0113] Here is an example of constructing a complete search text:

[0114] [1. Get user information:

[0115] -User ID: USER_12345

[0116] - Historical query: ['Shanghai to Rotterdam', '40-foot container price', 'marine insurance']

[0117] - Common routes: ['Shanghai-Rotterdam', 'Ningbo-Hamburg']

[0118] - Preference: 'Price first'

[0119] 2. Construct a search query:

[0120] -Merge the user's current query, historical query and preference settings

[0121] -Add relevant professional terms and synonyms

[0122] 3. Search the container multimodal transport knowledge base

[0123] 4. Apply BM25 scoring formula or semantic search technology to sort relevance:

[0124] 5. Fusion search results:

[0125] - Combine results from different data sources

[0126] - Remove duplicates and sort by relevance

[0127] 6. Personalized result processing:

[0128] - Secondary sorting of results based on user preferences

[0129] - Highlight information related to the user's frequently used routes

[0130] 7. Construct complete search text:

[0131] -Integrate the Top N search results

[0132] -Add metadata (e.g. source of information, credibility score)

[0133] 8. Output the search-enhanced text in JSON format:

[0134] {

[0135] "userQuery":"Container transport from Shanghai to Rotterdam",

[0136] "enhancedResults":[

[0137] {

[0138] "content":"The container shipping route from Shanghai to Rotterdam usually takes 15-18 days...",

[0139] "source":"Container Multimodal Transport Knowledge Base",

[0140] "relevanceScore":0.92,

[0141] "lastUpdated":"2024-07-25"

[0142] },

[0143] {

[0144] "content":"The current average price for a 40-foot container from Shanghai Port to Rotterdam Port is US$2,000...",

[0145] "source":"Real-time shipping database",

[0146] "relevanceScore":0.88,

[0147] "lastUpdated":"2024-07-30"

[0148] },

[0149] {

[0150] "content":"The Port of Rotterdam may experience mild congestion in the coming week. It is recommended to reserve 1-2 days of buffer time...",

[0151] "source":"Port Status Database",

[0152] "relevanceScore":0.75,

[0153] "lastUpdated":"2024-07-30"

[0154] }

[0155] ],

[0156] "userPreference":"Price is the priority",

[0157] "recommendedAction":"It is recommended to choose shipping company A with lower current prices. It is estimated that the total freight can be saved by about 8%"

[0158] }

[0159] 9. Input the constructed retrieval-enhanced text into the large language model for subsequent processing]

[0160] S6. The system collects user feedback for future query optimization.

[0161] “S6” specifically includes:

[0162] S61: The system collects user feedback on the query results, including but not limited to satisfaction scores, specific opinions and suggestions;

[0163] S62: Analyze user feedback data and identify improvement points, such as query accuracy, user interface optimization, etc.

[0164] S63: Integrate feedback into future query optimization and system iterations, including model training, algorithm adjustment, and user experience optimization, to achieve continuous service improvement.

[0165] like Figure 3-5 As shown, a natural language container multimodal transport interaction system based on a large language model includes:

[0166] The user interface module includes: a container query submodule for obtaining a container multimodal transport query request submitted by a user in natural language and providing interactive feedback;

[0167] The natural language processing module includes: a large language model query submodule that inputs the container multimodal transport query request and preset prompt words described in the user interface module into the large language model to generate a system query text, wherein the prompt words include the data format, field requirements, query text integrity assessment, and a follow-up questioning mechanism when the information is insufficient, which are agreed upon by the large language model;

[0168] The data query module includes: a query text input submodule for inputting the system query text described in the natural language processing module into the container multimodal transport database query interface for query, and a result output submodule for outputting the container multimodal transport route and price query results;

[0169] Container intermodal transport knowledge base, used to store basic knowledge related to container intermodal transport, including but not limited to domestic and international transportation regulations, policies, tariffs, insurance, routes, environmental protection, service provider information and evaluation;

[0170] The retrieval enhancement generation module, combined with RAG technology, uses a large language model to initiate a retrieval request to the container multimodal transport knowledge base, obtains matching information fragments, and integrates them with the container multimodal transport route and price query results to obtain the final query results;

[0171] A result display module, including a result display submodule that displays the final query result obtained by the search enhancement generation module in natural language, wherein the query result includes container transportation options, timetables, cost estimates, and special tips or instructions;

[0172] The user feedback module is used to collect user feedback and improve the response speed, accuracy and satisfaction of future queries.

[0173] Preferably, the user interface module further includes:

[0174] An information confirmation submodule, used to confirm to the user the user query request extracted and parsed by the large language model;

[0175] The information inquiry submodule is used to inquire about information when the large language model evaluates that the user query request is insufficient.

[0176] Preferably, the natural language processing module further includes:

[0177] The information extraction submodule is used to convert the query request in non-text form such as user voice or picture into a natural language query request in text form;

[0178] The language conversion submodule is used to convert a user's query request in a language not supported by the system into a query request in a language supported by the system.

[0179] Preferably, the container multimodal transport knowledge base includes:

[0180] The transportation regulations and policies database submodule is used to store domestic and international transportation regulations, policies, tariffs and other information;

[0181] Route and service provider database submodule: used to contain various transportation routes, service provider information, evaluation and cost;

[0182] Environmental protection submodule: used to include carbon emission regulations and calculation related information.

[0183] Preferably, the search enhancement generation module further includes:

[0184] Knowledge base retrieval submodule: using the acquired information, perform search queries in the container intermodal transport knowledge base to find matching information fragments;

[0185] Information fusion submodule: fuses the information retrieved by the knowledge base retrieval submodule with the original query text to construct an information set;

[0186] Generate optimized submodule: Based on the fused information, a large language model is used to generate optimized and adjusted natural language responses. This module may also include further optimization and adjustment of the generated text to ensure its readability and accuracy.

[0187] The above specific implementation modes are preferred embodiments of the present invention and cannot be used to limit the present invention. Any other changes or other equivalent replacement methods that do not deviate from the technical solution of the present invention are included in the protection scope of the present invention.

Claims

1. A natural language container multimodal transport interaction method based on a large language model, It is characterized in that include: S1. Obtaining a query request: obtaining a container multimodal transport query text submitted by a user in a natural language form, that is, an original query text; S2. Generate system query text: Input the container multimodal transport query text and preset prompt words into the large language model to generate parsed and automatically completed system query text, wherein the prompt words are generated by a fixed template method and a prompt word library method, including stipulating the data format and field requirements returned by the large language model, evaluating the completeness of the container multimodal transport query request, and asking questions in natural language when the container multimodal transport query request is insufficient; S3, supplement and confirm the complete system query text: determine whether the system query text in S2 is sufficient according to the evaluation result of the large language model on the completeness of the container multimodal transport query request; if it is complete, display the system query text to the user as a complete system query text for confirmation and use it as a complete system query text after confirmation; if it is incomplete, ask the user in natural language form to supplement the information to the container multimodal transport query text in S1, and repeat step S2; S4. Retrieving container multimodal transport routes and prices using a container multimodal transport query database: using the complete system query text confirmed by the user in S3, calling the container multimodal transport query database, and obtaining container multimodal transport route and price query results; S5. Initiate a search in the container multimodal transport knowledge base through the large language model search enhancement module: obtain user related information, use the user related information, the route and price query results and the complete system query text to construct a complete search text, and input it into the large language model again; use the large language model search enhancement module to initiate a search in the container multimodal transport knowledge base to obtain matching information fragments, and merge them with the container multimodal transport route and price query results obtained in S4 according to the search category to obtain the final query result, and the final query result is displayed according to the search category; S6. The system displays the query results: The system displays the final query results described in S5 and collects feedback from system users for future query optimization.

2. The natural language container multimodal transport interaction method based on a large language model according to claim 1, It is characterized in that S1 specifically includes: The natural language request includes a natural language request submitted directly by the user in the form of text and a natural language request submitted in the form of voice or picture; The natural language request includes query requests in multiple languages ​​that can be set, including internationally used English, Chinese and other languages.

3. The natural language container multimodal transport interaction method based on a large language model according to claim 1, It is characterized in that S2 specifically includes: S21. Multimodal conversion of large language model: When the user inputs voice and pictures, the multimodal large language model converts the user's input voice and pictures into text query requests; S22. Use prompt words to agree on the data format and field requirements returned by the large language model, including but not limited to: detailed addresses of the origin and destination of cargo transportation, shipping date, type, volume, and weight of the cargo; S23. The query text language returned by the prompt word agreement includes both the language used by the user when querying and the language supported by the container multimodal transport database query.

4. The natural language container multimodal transport interaction method based on a large language model according to claim 1, It is characterized in that S3 specifically includes: S31. Evaluate the completeness of the system query text based on the large language model. If it is determined that the information is insufficient or needs to be further clarified, the system will ask the user a follow-up question in natural language, where the follow-up question is a natural language question generated by step S2; and add the result of the follow-up question to the container multimodal transport query text described in step S1, and then input it into the large language model of step S2 in combination with the preset prompt words; S32: The system displays the parsed query text in the language of the natural language request input by the user for confirmation. If the user confirms, it will be used as the complete system query text. If not, the system returns to step S31 to question the user in natural language.

5. The natural language container multimodal transport interaction method based on a large language model according to claim 1, It is characterized in that S5 specifically includes: S51. Obtaining user-related information, including but not limited to user history query, order information, and preference settings; S52, constructing a complete search text using the user information, the aforementioned route and price query results, and the system query text, and inputting the complete search text into the large language model again; S53: Using the large language model retrieval enhancement generation technology, a search is initiated to the container multimodal transport knowledge base. The search content includes domestic and international transportation regulations, policies, tariffs, insurance, routes, environmental protection, service provider information and evaluation. The retrieval enhancement generation technology includes: BM25 scoring formula or semantic search technology.

6. The natural language container multimodal transport interaction method based on a large language model according to claim 5, It is characterized in that The BM25 scoring formula in S53 is as follows Where D stands for document; Q represents a query, which consists of multiple query words q i Composition; f(q i ,D) represents the query word q i The frequency in document D, IDF(q i ) represents the query word q i The inverse document frequency; |D| represents the length of the document; avgdl represents the average length of the document in the document collection; k 1 ,b represents a free parameter, which is used to adjust the saturation of word frequency and the influence of document length. Usually k 1 ∈[1.2,2.0] and b=0.75; in addition, the IDF calculation formula is: Where N represents the total number of documents in the document collection; n(q i ) represents the query word q i The number of documents.

7. The natural language container multimodal transport interaction method based on a large language model according to claim 5, It is characterized in that The steps of semantic search in S53 are as follows: S531: Using a pre-trained neural network model: Converting queries and documents into embedding vectors in a high-dimensional space can be summarized as: Among them, EmbeddingModel is the selected pre-trained model, and text is the query and document text data; S532: Calculate the cosine similarity between the query vector and the text vector, and determine the most semantically relevant documents by sorting. The cosine similarity is specifically calculated by the following formula: in, and are two vectors, representing the embedding vectors of the query and document respectively. The dot product in the numerator computes the inner product of the two vectors, while the denominator is the product of the norms (or lengths) of the two vectors.

8. The natural language container multimodal transport interaction method based on a large language model according to claim 1, It is characterized in that S6 specifically includes: S61: The system collects user feedback on the query results, including but not limited to satisfaction scores, specific opinions and suggestions; S62: Analyze user feedback data and identify improvement points, such as query accuracy and user interface optimization; S63: Integrate feedback into future query optimization and system iterations, including model training, algorithm adjustment, and user experience optimization.

9. A natural language container multimodal transport interaction system based on a large language model, It is characterized in that include: The user interface module includes: a container query submodule for obtaining a container multimodal transport query request submitted by a user in natural language, and an interactive feedback submodule for feeding back the query result processed by the large language model or a follow-up question raised in natural language when the container multimodal transport query request is insufficient to the user; the query result includes: container multimodal transport route and price query result; The natural language processing module includes: a large language model query submodule that inputs the container multimodal transport query request and preset prompt words described in the user interface module into a large language model to generate a system query text, wherein the prompt words include the data format, field requirements, query text integrity assessment, and a follow-up questioning mechanism when the information is insufficient. The data query module includes: a query text input submodule for inputting the system query text described in the natural language processing module into the container multimodal transport database query interface for query, and a result output submodule for outputting the container multimodal transport route and price query results; Container intermodal transport knowledge base, used to store basic knowledge related to container intermodal transport, including but not limited to domestic and international transportation regulations, policies, tariffs, insurance, routes, environmental protection, service provider information and evaluation; The retrieval enhancement generation module, combined with the RAG technology, initiates an enhanced retrieval request to the container multimodal transport knowledge base with a large language model, obtains matching information fragments and merges them with the container multimodal transport route and price query results in the data query module to obtain the final query result; A result display module, including a result display submodule that displays the final query result obtained by the search enhancement generation module in natural language, wherein the query result includes container transportation options, timetables, cost estimates, and special tips or instructions; The user feedback module is used to collect user feedback and improve the response speed, accuracy and satisfaction of future queries.

10. The natural language container multimodal transport interaction system based on a large language model according to claim 9, It is characterized in that The user interface module also includes: An information confirmation submodule, used to confirm to the user the user query request extracted and parsed by the large language model; The information inquiry submodule is used to inquire about information when the large language model evaluates that the user query request is insufficient.

11. The natural language container multimodal transport interaction system based on a large language model according to claim 9, It is characterized in that The natural language processing module also includes: The information extraction submodule is used to convert the query request in non-text form such as user voice or picture into a natural language query request in text form; The language conversion submodule is used to convert a user's query request in a language not supported by the system into a query request in a language supported by the system.

12. The natural language container multimodal transport interaction system based on a large language model according to claim 9, It is characterized in that The container intermodal transport knowledge base includes: The transportation regulations and policies database submodule is used to store domestic and international transportation regulations, policies, tariffs and other information; Route and service provider database submodule: used to contain various transportation routes, service provider information, evaluation and cost; Environmental protection submodule: used to include carbon emission regulations and calculation related information.

13. The natural language container multimodal transport interaction system based on a large language model according to claim 9, It is characterized in that The search enhancement generation module also includes: Knowledge base retrieval submodule: using the acquired information, perform search queries in the container intermodal transport knowledge base to find matching information fragments; Information fusion submodule: fuses the information retrieved by the knowledge base retrieval submodule with the container multimodal transport route and price query results in the data query module, and constructs an information set; Generate optimization submodule: Based on the fused information, use a large language model to generate optimized and adjusted natural language responses.

Images