An air freight forwarding multi-modal inquiry understanding and airline recommendation method and system

Abstract

The application discloses an air freight forwarding multi-modal inquiry understanding and route recommendation method and system, receives text, voice, picture, screenshot or table inquiry information, generates a field candidate set through pretreatment and industry fine-tuning large model; performs multi-source fusion, conflict identification and evidence chain generation on the same field, and carries out field standardization, integrity and rule checking; calculates the follow-up question priority for the missing, abnormal or conflict field and interacts with the completion; queries the route, price, cabin and rule database based on the standardized parameters, filters the candidate routes and generates the unavailable reason coding, and then dynamically scores and sorts according to the user preference, and outputs the recommended scheme, the alternative scheme and the recommendation reason. The application improves the inquiry understanding, route matching and recommendation explanation ability.

Description

Technical Field

[0001] This invention relates to the fields of intelligent information processing, natural language processing, multimodal information recognition, knowledge extraction, logistics information processing, and air freight auxiliary decision-making technology, specifically to a method and system for understanding multimodal price inquiries and recommending routes for air freight forwarding. Background Technology

[0002] In air freight forwarding, inquiry processing is typically a prerequisite for quoting, loading, booking, and route selection. Air freight forwarding companies need to determine the feasible air transport solution based on information provided by the user, such as the origin, destination, cargo name, quantity, weight, volume, dimensions, cargo attributes, time requirements, and price preferences, combined with the route, price, cargo space, and the acceptance rules of the airline and supplier.

[0003] In actual business operations, the sources of user-submitted inquiry information are complex. Inquiries may be submitted in natural language text, or as voice messages, images of goods lists, chat screenshots, email screenshots, spreadsheets, or a combination of these formats. Different users use inconsistent expressions, and inquiry content often suffers from problems such as colloquial language, numerous industry abbreviations, incomplete fields, inconsistent units, implicit shipping conditions, and contradictions between information from different sources.

[0004] For example, a user might simply state, "Shanghai to Germany, three items, approximately 180 kg, electronic components, preferably flying tomorrow." In this inquiry, the destination city or airport for "Germany" is not specified, it's unclear whether the "electronic components" are battery-powered, and the dimensions and volume of the goods are not provided. This information could directly impact the airline's cargo acceptance rules, the supplier's cargo acceptance rules, and the assessment of route feasibility.

[0005] Current air freight forwarding inquiry processing mainly relies on manual quotation personnel to read, extract, verify, inquire, query databases, and recommend routes based on user information. This method is highly dependent on personnel experience, has low processing efficiency, and it is difficult to unify judgment standards among different personnel. When the inquiry information is complex or there are conflicting fields, key constraints are easily overlooked.

[0006] Existing form-based query systems typically require users to fill in structured fields according to a preset format. Since most of the information actually submitted by users is unstructured or semi-structured, users still need to manually organize and repeatedly fill in the information, making it impossible to directly convert voice, images, screenshots, or natural language queries into executable query parameters.

[0007] While general-purpose natural language processing models or large general-purpose models can identify some entity information, they struggle to reliably identify key fields such as weight, volume, dimensions, number of pieces, cargo type, electrical attributes, dangerous goods attributes, cargo space requirements, timeliness requirements, and price preferences due to a lack of specialized adaptations to the air freight forwarding industry's corpus, terminology, receiving rules, and business constraints. Furthermore, when fields are missing, abnormal, or conflicting, general-purpose models also struggle to determine which fields will affect subsequent route queries and which information should be prioritized for follow-up requests to the user.

[0008] Furthermore, existing systems typically only connect technologies such as speech-to-text, OCR recognition, field extraction, database querying, and ranking recommendations, lacking a multimodal candidate fusion mechanism for the same field. When the values ​​of the same field recognized in text, speech, or images are inconsistent, the system cannot comprehensively determine which field value should be adopted, nor can it retain the original source and judgment basis of the field value.

[0009] Existing technologies lack field evidence chains, field conflict markers, follow-up question priorities, and unavailability reason coding mechanisms, making it difficult for the system to explain the source of field values, determine the degree of conflict between multiple sources, generate effective follow-up questions for key fields affecting route queries, and explain to users the specific reasons why candidate routes were not recommended.

[0010] Meanwhile, existing route recommendation systems typically only perform queries based on origin, destination, price, or timeliness, failing to link with airline cargo acceptance rules, supplier cargo acceptance rules, cargo space rules, transit rules, and cargo attribute constraints. This can easily lead to recommending routes that cannot actually carry electronic goods, dangerous goods, or oversized cargo.

[0011] Therefore, it is necessary to provide a technical solution that can achieve unified understanding of multimodal air freight forwarding inquiries, field candidate fusion, field conflict identification, field evidence tracing, structured processing, business rule verification, interactive completion, database linkage query, and route recommendation ranking, so as to improve the automation level of air freight forwarding inquiry processing, field identification accuracy, route matching accuracy, and interpretability of recommendation results. Summary of the Invention

[0012] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for understanding multimodal quotations and recommending routes in air freight forwarding, so as to solve the problems of difficulty in unified understanding of multimodal quotations, difficulty in converting unstructured information into query parameters, difficulty in handling conflicts between multiple modal source fields, invalid queries caused by missing or abnormal fields, insufficient accuracy of candidate route screening, and lack of explanation of recommendation results.

[0013] To achieve the above objectives, the present invention provides a method for understanding multimodal price inquiries and recommending routes in air freight forwarding, comprising the following steps: S1: Receive air freight forwarding inquiry information input by the user, wherein the inquiry information includes at least one of text, voice, image, screenshot, and table; S2: Preprocess the inquiry information according to the input modality of the inquiry information to generate a sequence of candidate text fragments with modality source, source location and session context identifier; S3: Input the candidate text fragment sequence into a large model fine-tuned with air freight forwarding industry corpus to identify inquiry intent and generate a candidate set of fields; S4: When the same field corresponds to multiple candidate values ​​or multiple modal sources, the candidate values ​​are fused from multiple sources based on field confidence, modal source reliability, source priority, context consistency and business rule adaptability to determine the target field value and generate field conflict markers and field evidence chains. S5: Convert the field information after multi-source fusion into a unified structured field object, and standardize the unified structured field object by location coding, cargo type, weight unit, volume unit, size unit and preference label; S6: Perform integrity checks, numerical anomaly checks, consistency conflict checks, and business rule checks on the standardized fields, and generate field check results; S7: When missing, abnormal, or conflicting fields affect route queries, calculate the follow-up question priority based on field importance, field confidence, conflict coefficient, and rule impact, and generate supplementary questions according to the follow-up question priority for multi-round interactive completion; when a field meets the query conditions, generate standardized query parameters. S8: Query the route database, price database, cabin class database, airline cargo receiving rule database, and supplier rule database according to the standardized query parameters to obtain a set of candidate routes; S9: Filter candidate routes based on route basic conditions, cargo attribute constraints, airline cargo receiving rules, supplier cargo receiving rules, cargo space rules, and transit rules, and generate unavailability reason codes for the filtered candidate routes; S10: Score and sort the filtered candidate routes, dynamically adjust the scoring weights based on the user's preference for price or timeliness, and output the recommended route, alternative routes, reasons for recommendation, and reasons for unavailability.

[0014] Further, step S2 includes: When the input is text, perform sentence segmentation, noise reduction, error correction, terminology segmentation, and context concatenation on the text; When the input is speech, speech is transcribed and corrected for homophones and industry abbreviations using an air freight forwarding industry dictionary. When the input is an image or screenshot, perform text recognition, layout analysis, table area recognition, field adjacency recognition, and key field area extraction on the image or screenshot; When the input is a table, column name recognition, cell parsing, and field mapping are performed on the table.

[0015] Furthermore, in step S3, the training corpus of the large model fine-tuned by the air freight forwarding industry corpus includes at least air freight forwarding inquiry samples, standard field annotation samples, rule text samples, and feedback samples. The rule text samples include airline receiving rule texts or supplier rule texts, and the feedback samples include historical inquiry samples or manually corrected samples. The training objectives of the large model include at least the field extraction objective, the candidate value ranking objective, and the confidence calibration objective, and further include one or more of the following: inquiry intent identification objective, missing field identification objective, or conflicting field identification objective.

[0016] Furthermore, the fields in the candidate set of fields generated in step S3 include at least some of the following fields: origin, destination, cargo name, cargo type, number of pieces, weight, volume, dimensions, receiving type, electrical attributes, dangerous goods attributes, oversized attributes, timeliness requirements, cargo space requirements, and price preferences.

[0017] Further, step S4 includes: Generate multiple candidate values ​​and their candidate source identifiers for the same field; The comprehensive score of each candidate value is calculated based on field confidence, modality source reliability, source priority, context consistency, and business rule adaptability. Select the candidate value with the highest overall score as the target field value; Generate field conflict flags based on the differences between unadopted candidate values ​​and target field values; And generate a chain of field evidence for the target field value.

[0018] Furthermore, the reliability of the modality source is determined based on one or more of the following: source type, identification method, historical accuracy, and manual confirmation status. The business rule adaptability is determined based on whether the candidate field values ​​meet one or more of the following: airline receiving rules, supplier receiving rules, battery restrictions, dangerous goods restrictions, oversize restrictions, or special receiving type restrictions.

[0019] Furthermore, the field evidence chain includes at least the original input fragment, modality source, source location, and field confidence, and further includes one or more of the following: candidate interpretation, adopted candidate value, and unadopted candidate value.

[0020] Further, step S5 includes: Map city names, airport names, or abbreviations to airport three-letter codes, city codes, or preset node codes; Map natural language cargo descriptions to general cargo, cargo with batteries, dangerous goods, fresh goods, oversized cargo, or pre-defined cargo category codes; The weight is standardized to kilograms, the volume to cubic meters, and the dimensions to centimeters or millimeters. Map timeliness descriptions to timeliness levels, and price descriptions to price priorities; Calculate volumetric weight, maximum single-piece size, or oversized risk label based on weight, volume, number of pieces, and dimensions.

[0021] Further, step S6 includes: Perform integrity checks on the origin, destination, and basic cargo attributes; Perform numerical reasonableness and consistency checks on weight, volume, number of pieces, and dimensions; Perform business risk assessment based on cargo type, electrical properties, hazardous materials properties, and receiving type; Perform conflict identification on the same field values ​​from different modalities.

[0022] Furthermore, in step S7, the rule impact degree is determined based on the degree of impact of missing, abnormal, or conflicting fields on route database queries, airline receiving rule matching, supplier receiving rule matching, electric limit judgment, dangerous goods limit judgment, or oversize limit judgment.

[0023] Further, in step S7, the fields satisfying the query conditions include: The standardized field object must include at least the origin code, destination code, and weight field; When the cargo name, cargo type, or risk label indicates that the cargo may involve electrical components, dangerous goods, oversized items, or special receiving types, the standardized field object further includes one or more of the following: size field, cargo type code, electrical component label, or dangerous goods label.

[0024] Furthermore, in step S8, the route, price, cabin class, and rule data are parametrically retrieved and matched based on the standardized query parameters, rather than being retrieved directly based on the original natural language.

[0025] Further, step S9 includes: Filter candidate routes based on origin, destination, schedule, and timeliness requirements; Candidate routes are filtered based on weight restrictions, volume restrictions, size restrictions, electrical restrictions, dangerous goods restrictions, and cargo type restrictions; Candidate routes are filtered based on airline receiving rules, supplier receiving rules, cabin space rules, and transit rules.

[0026] Furthermore, the unavailability reason code includes at least one of the following: non-acceptance of goods with batteries, dangerous goods restrictions, single piece exceeding length, no current space available, time limit not met, weight restrictions, volume restrictions, or supplier not accepting the goods.

[0027] Further, step S10 includes: The comprehensive score of candidate routes is calculated based on two or more of the following factors: timeliness matching, price competitiveness, availability of cabin space, cargo attribute suitability, rule compliance, stability, and user preference satisfaction. The scoring weights are dynamically determined based on two or more of the following: user preference fields, cabin availability, rule satisfaction, and unavailability reason coding distribution. When users prioritize price, increase the weight corresponding to price competitiveness; When users prioritize timeliness, increase the weight corresponding to timeliness matching.

[0028] Furthermore, when the distribution of unavailability reason codes shows that more than a preset proportion of candidate routes are eliminated due to electrical restrictions, dangerous goods restrictions, or oversize restrictions, the weights corresponding to cargo attribute suitability or rule satisfaction are increased.

[0029] This invention also provides a multimodal inquiry understanding and route recommendation system for air freight forwarding, comprising: A multimodal input receiving module is used to receive air freight forwarding inquiry information input by the user, wherein the inquiry information includes at least one of text, voice, image, screenshot, and table; The preprocessing module is used to preprocess the inquiry information according to the input modality of the inquiry information to generate a sequence of candidate text fragments with modality source, source location and session context identifier; The industry fine-tuning large model understanding module is used to input the candidate text fragment sequence into the large model fine-tuned by the air freight forwarding industry corpus, identify the inquiry intent and generate a set of field candidates. The multi-source field fusion module is used to perform multi-source fusion of candidate values ​​based on field confidence, modality source reliability, source priority, context consistency, and business rule adaptability when the same field corresponds to multiple candidate values ​​or multiple modal sources, in order to determine the target field value. The field conflict identification module is used to generate field conflict flags based on the differences between unadopted candidate values ​​and target field values; The field evidence chain generation module is used to generate field evidence chains that include the original input fragment, modality source, source location, and field confidence. The structured field object generation module is used to convert field information after multi-source fusion into a unified structured field object; The field standardization module is used to standardize the location coding, cargo type, weight unit, volume unit, size unit, and preference label of the unified structured field object; The field validation module is used to perform integrity checks, numerical anomaly checks, consistency conflict checks, and business rule checks on the standardized fields. The follow-up question priority calculation module is used to calculate the follow-up question priority based on field importance, field confidence, conflict coefficient, and rule impact when missing, abnormal, or conflicting fields affect route queries. A multi-round interactive completion module is used to generate supplementary questions according to the priority of follow-up questions and to receive supplementary information from the user. The query parameter generation module is used to generate standardized query parameters when a field meets the query conditions. The route database query module is used to query the route database, price database, cabin class database, airline cargo receiving rule database and supplier rule database according to the standardized query parameters to obtain a set of candidate routes; The candidate route filtering module is used to filter candidate routes based on basic route conditions, cargo attribute constraints, airline cargo receiving rules, supplier cargo receiving rules, cargo space rules, and transfer rules. The unavailability reason coding module is used to generate unavailability reason codes for the filtered candidate routes; The recommendation and ranking module is used to score and rank the filtered candidate routes, and dynamically adjust the scoring weights based on the user's preference for price or timeliness. The recommendation reason generation module is used to output recommended route options, alternative options, reasons for recommendation, and reasons for unavailability.

[0030] Furthermore, the preprocessing module includes a text processing unit, a speech-to-text unit, an OCR recognition unit, a table parsing unit, and a multi-source information alignment unit; The text processing unit is used to perform sentence segmentation, noise reduction, error correction, terminology segmentation, and context concatenation on the text. The speech transcription unit is used to perform speech transcription and, in conjunction with an air freight forwarding industry dictionary, correct homophones and industry abbreviations. The OCR recognition unit is used to perform text recognition, layout analysis, table area recognition, field adjacency recognition, and key field area extraction on images or screenshots; The table parsing unit is used to perform column name recognition, cell parsing, and field mapping on the table. The multi-source information alignment unit is used to align the candidate field values ​​corresponding to different modalities according to the field name, source location, and session context.

[0031] Furthermore, the field standardization module includes a location coding standardization unit, a cargo type coding standardization unit, a unit conversion unit, a size structuring unit, and a preference label normalization unit; The location coding standardization unit is used to map city names, airport names or abbreviations to airport three-letter codes, city codes or preset node codes; The cargo type coding standardization unit is used to map natural language cargo descriptions to cargo category codes and risk labels; The unit conversion unit is used to convert weight, volume, and size into a preset uniform unit; The size structured unit is used to calculate the total volume, volumetric weight, chargeable weight, or oversized risk label based on the number of pieces and the size of each piece. The preference label normalization unit is used to map timeliness descriptions or price descriptions to timeliness levels or price priorities.

[0032] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for understanding multimodal quotations and recommending routes for air freight forwarding.

[0033] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the above-mentioned method for understanding multimodal quotations and recommending routes for air freight forwarding.

[0034] Compared with the prior art, the present invention has at least the following beneficial effects.

[0035] 1. This invention improves the consistency and traceability of query information processing under different input methods by converting multimodal query information such as text, voice, images, screenshots and tables into candidate text fragment sequences with modality source, source location and session context identifiers.

[0036] 2. This invention improves the accuracy of identifying abbreviations, colloquial expressions, and implicit transportation conditions in the air freight forwarding industry by using a large model fine-tuned from the air freight forwarding industry corpus to identify inquiry intent and business fields. Through multi-source field fusion, field conflict identification, and field evidence chain generation, when there are multiple sources or multiple candidate values ​​for the same field, it can comprehensively determine the target field value and retain its original source and judgment basis.

[0037] 3. This invention standardizes unstructured inquiry information by using location coding, cargo type coding, weight units, volume units, size units, and preference tags, converting it into standardized parameters that can be queried in the database and matched according to rules. Through integrity checks, numerical anomaly checks, consistency conflict checks, and business rule checks, it reduces invalid queries caused by missing fields, numerical errors, field conflicts, or unclear business attributes.

[0038] 4. This invention calculates the priority of follow-up questions based on field importance, field confidence, conflict coefficient, and rule impact, and can prioritize generating supplementary questions for key fields that affect route queries and rule matching, thereby reducing invalid and duplicate follow-up questions.

[0039] 5. This invention links route databases, price databases, cabin space databases, airline cargo receiving rule databases, and supplier rule databases to perform parameterized queries and rule filtering on candidate routes, thereby improving the feasibility of route recommendations under complex cargo conditions.

[0040] 6. This invention generates unavailability reason codes for filtered candidate routes and generates recommendation reasons based on the scoring items, thereby improving the interpretability of the recommendation results; it can identify the main reasons why candidate results are empty or a large number of candidate routes are filtered based on the distribution of unavailability reason codes, and conduct reverse inquiry based on this to improve the hit rate of subsequent queries; it dynamically adjusts the scoring weights according to the user's price priority or time priority preference, so that the recommendation results can more accurately match the user's needs. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0042] Figure 2 This is a schematic diagram of the system of the present invention.

[0043] Figure 3 This is a schematic diagram of the multimodal inquiry information processing and field standardization process in this invention.

[0044] Figure 4 This is a schematic diagram of the field validation and multi-round interactive completion process in this invention.

[0045] Figure 5 This is a schematic diagram of the candidate route filtering, scoring and sorting, and result output process in this invention. Detailed Implementation

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions, simple transformations, or combinations made by those skilled in the art to the present invention without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0047] In this embodiment, the air freight forwarding inquiry information refers to business information submitted by users in order to obtain information such as air cargo transportation routes, prices, cargo space, delivery time, or acceptance conditions. The air freight forwarding inquiry information may consist of one or more of the following: text, voice, images, screenshots, tables, email bodies, PDF documents, web page content, or API messages.

[0048] The fields referred to in this embodiment are business parameters identified from air freight forwarding inquiry information and used for subsequent rule verification, route query, candidate filtering, and recommendation ranking. These fields may include origin, destination, cargo name, cargo type, number of pieces, weight, volume, dimensions, receiving type, electrical attributes, dangerous goods attributes, oversized attributes, timeliness requirements, cargo space requirements, and price preferences, etc.

[0049] System Overall Structure like Figure 2 As shown, the air freight forwarding multimodal inquiry understanding and route recommendation system based on an industry fine-tuning large model provided by this invention can be deployed on servers, cloud platforms, local business systems, or in a hybrid architecture consisting of cloud platforms and local business systems.

[0050] The system includes a multimodal input receiving module, a preprocessing module, an industry fine-tuning large model understanding module, a multi-source field fusion module, a field conflict identification module, a field evidence chain generation module, a structured field object generation module, a field standardization module, a field validation module, a follow-up question priority calculation module, a multi-round interactive completion module, a query parameter generation module, a route database query module, a candidate route filtering module, an unavailability reason encoding module, a recommendation ranking module, and a recommendation reason generation module.

[0051] The multimodal input receiving module is used to receive air freight forwarding inquiries submitted by users through chat windows, web pages, applications, voice upload interfaces, image upload interfaces, table import interfaces, or business system interfaces. The multimodal input receiving module also records the modality type, data source, timestamp, source location, and session identifier for each piece of input data, so as to trace the source of candidate values ​​for each field during subsequent processing.

[0052] The preprocessing module performs corresponding preprocessing based on the modality type of the input information and generates a sequence of candidate text segments with modality source, source location, and session context identifiers. The preprocessing module includes a text processing unit, a speech-to-text unit, an OCR recognition unit, a table parsing unit, and a multi-source information alignment unit.

[0053] The industry fine-tuning large model understanding module is used to perform semantic understanding on candidate text fragment sequences, identify whether the current user has business intentions such as inquiring about prices, supplementing information, modifying information, or confirming information, and generate a set of candidate fields based on a preset field system.

[0054] The multi-source field fusion module is used to calculate the comprehensive score of candidate values ​​based on field confidence, modal source reliability, source priority, context consistency and business rule adaptability when there are multiple candidate values ​​or multiple modal sources for the same field, and select the candidate value that meets the preset conditions as the target field value.

[0055] The field conflict identification module is used to determine whether there are numerical conflicts, semantic conflicts, or rule conflicts between different candidate values ​​of the same field, and to generate corresponding field conflict markers.

[0056] The field evidence chain generation module is used to record one or more of the following: the original input fragment of the target field value, modality source, source location, field confidence, candidate interpretation, adopted candidate value, and unadopted candidate value.

[0057] The structured field object generation module is used to convert field information fused from multiple sources into a unified structured field object. This unified structured field object can be represented as a collection of key-value pairs, a JSON object, a database record, or other data structures that can be read by a computer.

[0058] The field standardization module is used to standardize location information, cargo type, weight unit, volume unit, size unit, and preference labels in a unified structured field object.

[0059] The field validation module is used to perform integrity checks, numerical anomaly checks, consistency conflict checks, and business rule checks.

[0060] The follow-up question priority calculation module is used to calculate the follow-up question priority of each field to be supplemented based on field importance, field confidence, conflict coefficient, and rule influence.

[0061] The multi-round interactive completion module is used to generate supplementary questions according to the priority of follow-up questions, receive supplementary or confirmed information from users, and feed the newly input information from users back to the preprocessing module, the industry fine-tuning large model understanding module, the multi-source field fusion module, or the field validation module.

[0062] The query parameter generation module is used to convert standardized and validated fields into standardized query parameters that can be recognized by the route database, price database, cabin class database, and rule database when the fields meet the query conditions.

[0063] The route database query module is used to query the route database, price database, cabin class database, airline cargo receiving rule base and supplier rule base based on standardized query parameters, and obtain a set of candidate routes.

[0064] The candidate route filtering module is used to filter candidate routes based on basic route conditions, cargo attribute constraints, airline cargo receiving rules, supplier cargo receiving rules, cabin space rules, and transit rules.

[0065] The unavailability reason coding module is used to generate unavailability reason codes for candidate routes that do not meet the corresponding filtering conditions.

[0066] The recommendation ranking module is used to calculate the comprehensive score of candidate routes based on timeliness matching, price competitiveness, availability of cabin space, cargo attribute suitability, rule compliance, stability, and user preference satisfaction, and dynamically adjusts the corresponding score weights according to user preferences.

[0067] The recommendation reason generation module generates recommendation reasons based on the candidate route's rating items, rule matching, cargo attributes, and user preferences, and converts the unavailability reasons into natural language explanations that users can understand.

[0068] Key data object definition In this embodiment, a price inquiry request is represented as a multimodal input set X, denoted as: (1) Where, x i This represents the i-th input segment, where n is the number of input segments. Input segments can be text strings, speech-to-text transcriptions, OCR recognition results, or table fields. Preferably, each input segment also includes one or more of the following: modality source, source location, timestamp, and session identifier.

[0069] Let F represent the set of inquiry fields, denoted as: (2) Among them, f j This represents the j-th field, and m is the total number of preset fields. Preferably, the fields include at least the origin, destination, cargo name, cargo type, number of pieces, weight, volume, dimensions, receiving type, electrical attribute (whether it is electrical), dangerous goods attribute (whether it is dangerous), oversize attribute (whether it is oversize), timeliness requirement, space requirement, and price preference.

[0070] For any field f jIts structured representation is defined as: (3) Among them, name j Indicates the field name, value j Represents the field value, source j Indicates the source of the field, conf j Indicates the confidence level of the field, unit j Indicates the unit of the field.

[0071] The standardized set of query parameters output by the system is denoted as Q, and is further denoted as: (4) Where, q k For parameters used in database queries, standardized query parameters preferably include one or more of the following: origin code, destination code, total weight, total volume, maximum single piece size, cargo type code, electronic label, dangerous goods label, timeliness class, cargo space preference, and price priority.

[0072] Detailed explanation of the method and process like Figure 1 As shown, this embodiment provides a method for understanding multimodal price inquiries and recommending routes for air freight forwarders based on an industry-fine-tuned large model.

[0073] S1: Receive air freight forwarding inquiry information input by the user.

[0074] The system receives air freight forwarding inquiry information input by the user, which includes at least one of the following: text, voice, image, screenshot, and table.

[0075] Users can directly input natural language text in the chat window, send voice messages, and upload images of goods lists, chat screenshots, email screenshots, or spreadsheets.

[0076] When receiving air freight forwarding inquiries, the system records the modality type, source identifier, timestamp, source location, and session identifier for each input item.

[0077] The source location can be a character position in text, a coordinate region in an image, a row or column position in a table, or a time interval in audio.

[0078] S2: Perform preprocessing on different modal inputs.

[0079] The inquiry information is preprocessed based on the input modality of the air freight forwarding inquiry information to generate a sequence of candidate text fragments with modality source, source location and session context identifier.

[0080] like Figure 2As shown, when the input is text, sentence segmentation, noise reduction, error correction, terminology segmentation, and context concatenation are performed.

[0081] When the input is speech, speech transcription is performed first, and then homophones and industry abbreviations are corrected using an industry dictionary.

[0082] When the input is an image or screenshot, it performs text recognition, layout analysis, table area recognition, field adjacency recognition, and key field area extraction.

[0083] Field adjacency recognition is used to determine the positional correspondence between field names and field values ​​in an image. For example, values ​​located to the right or below the "weight" field can be considered candidate values ​​for the weight field.

[0084] When the input is a table, column name recognition, cell parsing, and field mapping are performed. For example, "GW", "GrossWeight", or "Gross Weight" are mapped to the weight field, and "DIM" or "Size" is mapped to the size field.

[0085] After preprocessing, preferably, the candidate text segment after speech or OCR recognition can be represented as sequence T, denoted as: (5) Among them, t r Let represent the r-th text segment. The preprocessing module normalizes the text segment according to the air freight forwarding industry dictionary D, resulting in a normalized text sequence T': (6) Here, Normalize represents the normalization mapping function.

[0086] S3: Call the large model, which has been fine-tuned with the air freight forwarding industry corpus, to perform semantic understanding and generate a set of candidate fields.

[0087] The candidate text fragment sequence is input into a large model fine-tuned with an air freight forwarding industry corpus to identify inquiry intent and generate a set of candidate fields.

[0088] The large model was trained on inquiry expressions in the air freight forwarding industry and can recognize industry terms, abbreviations, colloquial expressions, and implicit transportation conditions. For example, "flying tomorrow" can be identified as having high timeliness requirements, "cheaper is fine" can be identified as having a price preference, and "electronic components with batteries" can be identified as goods named as electronic components with batteries.

[0089] In this embodiment, the preprocessed normalized text sequence T' is input into the industry fine-tuning model M to obtain the field extraction result E, denoted as: (7) Here, E includes field name, field value, field confidence level, field source fragment, and candidate interpretation set.

[0090] For any field f j Its confidence level can be represented by the model output probability p(valuej|T'), that is: (8) If there are multiple candidate values ​​for the same field, a candidate set C can be constructed. j : (9) Where s represents the s-th candidate value.

[0091] S4: Perform multi-source field fusion, conflict identification, and field evidence chain generation.

[0092] When the same field corresponds to multiple candidate values ​​or multiple modal sources, the candidate values ​​are fused from multiple sources based on field confidence, modal source reliability, source priority, context consistency and business rule adaptability to determine the target field value and generate field conflict markers and field evidence chains.

[0093] The system determines candidate values ​​based on contextual semantics, source priority, source reliability, and rule constraints. Preferably, source reliability is determined based on one or more of the following: source type, recognition method, historical accuracy, and manual confirmation status; for example, the source reliability of a manually confirmed field is higher than that of a table parsing field, which in turn is higher than that of an OCR recognition field, and vice versa. Business rule compatibility is determined based on whether the candidate field meets one or more of the following: airline receiving rules, supplier receiving rules, battery restrictions, dangerous goods restrictions, oversize restrictions, or special receiving type restrictions.

[0094] For candidate value c u Its overall score can be expressed as: (10) Where Conf represents the field confidence level, SourceRel represents the modal source reliability, SourcePriority represents the source priority, Context represents context consistency, RuleFit represents the business rule adaptability, and d1 to d5 are the corresponding weight coefficients. The system selects the candidate value with the highest comprehensive score as the target field value. (11) The system also generates a chain of evidence for the target field value. j : (12) Among them, source_span j This represents the original input fragment, modality. j Indicates the modal origin, position j Indicates the source location, reason j Indicates candidate interpretation, conf j This indicates the confidence level of the field. Preferably, the field evidence chain includes at least the original input fragment, modality source, source location, and field confidence level, and further records the adopted candidate values ​​and the unadopted candidate values, so as to show the basis for the field value when there is a field conflict, follow-up question completion, or recommendation reason generation.

[0095] S5: Convert the field information after multi-source fusion into a unified structured field object and perform field standardization.

[0096] In this embodiment, the structured field object is represented as a collection of key-value pairs or a JSON object. Preferably, a structured object for a price inquiry can be represented as follows: (13) The fields correspond to the place of origin, destination, cargo name, cargo type, number of pieces, weight, volume, dimensions, receiving type, electrical attributes, dangerous goods attributes, oversized attributes, time requirements, cargo space requirements, and price preferences, respectively.

[0097] The field information after multi-source fusion is converted into a unified structured field object, and the unified structured field object is standardized by location coding, cargo type, weight unit, volume unit, size unit and preference label.

[0098] Location standardization includes mapping city names, airport names, or abbreviations to airport three-letter codes, city codes, or preset node codes.

[0099] Standardization of cargo types includes mapping natural language cargo descriptions to general cargo, cargo with batteries, dangerous goods, fresh goods, oversized cargo, or pre-defined cargo category codes.

[0100] Unit standardization includes converting weight to kilograms, volume to cubic meters, and dimensions to centimeters or millimeters.

[0101] Map the expressions "Shanghai Pudong", "Pudong", and "PVG Airport" to a unified three-letter airport code; Natural language descriptions such as "electronic components", "lithium battery components", and "general cargo with batteries" are mapped to system cargo category codes and attribute tags; Information such as "180 kg", "0.8 cubic meters", and "120x80x75 cm" will be uniformly converted into preset units; Calculate the volumetric weight and risk of exceeding limits for a single item based on the number of pieces, weight, volume, and dimensions.

[0102] Preferably, location standardization can be expressed as: (14) (15) If the input weight unit is lb, ton, or kg, it will be uniformly converted to kg and can be represented as: (16) Wherein, Wraw represents the original weight value identified from the inquiry information, unitw represents the weight unit corresponding to the original weight value, Wkg represents the converted kilogram weight, and ConvertWeight represents the weight unit conversion function that converts Wraw to kilograms based on unitw.

[0103] If the input volume unit is CBM, cubic centimeters, or cubic meters, it will be uniformly converted to cubic meters and can be expressed as: (17) If the dimension field includes length L, width B, and height H, then the volume of a single piece can be expressed as: (18) When L, B, and H are in centimeters, 10 -6 Used to convert to cubic meters; if millimeters are used, use 10. -9 .

[0104] If the number of items is N, then the total volume can be expressed as: (19) In air freight, volumetric weight is often used as a billing reference. Preferably, volumetric weight W v It can be represented as: (20) Where K is the volumetric weight conversion factor, preferably taken as 167 kg / m³. 3 Alternatively, it can be configured to other preset values ​​according to business rules.

[0105] The system can further generate the billing weight W. c : (twenty one) S6: Perform integrity checks, exception checks, and business rule checks.

[0106] Perform integrity checks, numerical anomaly checks, consistency conflict checks, and business rule checks on the standardized fields, and generate field check results.

[0107] Integrity checks are used to determine whether the current field meets the minimum query conditions required for a route query. The minimum set of query fields includes at least the origin code, destination code, and basic cargo attributes.

[0108] Numerical anomaly verification is used to determine whether there are numerical anomalies, consistency anomalies, or multimodal conflicts in weight, volume, number of pieces, and dimensions; Business rule verification is used to determine whether goods with batteries, dangerous goods, oversized goods, or special receiving types trigger airline receiving restrictions or supplier restrictions.

[0109] Preferably, the minimum set of query fields is defined as Rmin: (twenty two) Wherein, the field set Y follows the unified structured field object in equation (13), representing the field set obtained after multi-source fusion and standardization in this inquiry; if Y does not cover all the required fields in Rmin, or if any field in Rmin is missing a valid value, it is marked as incomplete, denoted as: (twenty three) otherwise: (twenty four) Dimensional and volume consistency can be verified using an error function. If the user also provides the total volume V... input Calculate volume V based on dimensions total Then the consistency error e v for: (25) Where ε is a very small positive constant to prevent the denominator from being zero. If e v If the value exceeds the preset threshold thetav, the volume is determined to be abnormal.

[0110] If multiple modalities give different values ​​for the same field, a conflict coefficient (Conflict) can be defined. j for: (26) Where freq(valuej*) represents the number of times the candidate optimal value appears in each source, n j This indicates the total number of sources for field j. If Conflict j If the value exceeds the threshold thetac, then a clarification inquiry is triggered.

[0111] For cargo risks, a risk label set G can be defined: (27) Wherein, G represents the set of cargo risk labels, gbattery represents risk labels related to electricity or batteries, gdg represents dangerous goods risk labels, governorsize represents oversized risk labels, and gspecial represents risk labels for special receiving types or special operational requirements; each risk label can be determined based on the cargo name, attribute field, size field, and business rule matching results, and is used for subsequent rule restriction judgments.

[0112] If any risk label triggers the corresponding rule restriction, then: (28) otherwise: (29) S7: Generate multiple rounds of supplementary questions or query parameters.

[0113] If the origin, destination, weight, dimensions, or cargo attributes are missing, the system automatically generates supplementary questions based on field priority. For example, the system calculates follow-up question priority based on field importance, field confidence, conflict coefficient, and rule impact. For instance, the system might say, "Please provide the maximum dimensions of a single item so we can determine if there are any oversize restrictions." After the user provides the information, the system re-executes the field update and validation process.

[0114] Preferably, the field follow-up priority Pask can be determined by the field importance w. j Field confidence conf j and the influence of the rule r j Determined jointly, and expressed as: (30) Among them, w j Indicates field importance, conf j Indicates the confidence level of the field, r j Indicates the degree of influence of the rule, Conflict j α represents the conflict coefficient. j β j γ j and η j These are weighting coefficients. The system prioritizes P. ask The field with the higher value generates follow-up questions.

[0115] If a field meets the query criteria, a standardized query parameter Q is generated. Preferably, the standardized field object includes at least an origin code, a destination code, and a weight field; when the cargo name, cargo type, or risk label indicates that it may involve electrical components, dangerous goods, oversized items, or special receiving types, it further includes one or more of the following: a size field, a cargo type code, an electrical component label, or a dangerous goods label. The parameter generation function can be expressed as: (31) Where R represents the set of rules and G represents the set of risk labels.

[0116] S8: Query multiple business databases.

[0117] The system uses standardized query parameters to query the route database, price database, cabin class database, airline cargo receiving rule database, and supplier rule database, resulting in a candidate route set A. This set is denoted as: (32) Among them, QueryRoute represents the route retrieval function, QueryPrice represents the price retrieval function, QuerySpace represents the cabin class retrieval function, MatchCarrierRule represents the airline rule matching function, and MatchVendorRule represents the supplier rule matching function.

[0118] Each route a in the candidate route set i It can be represented as: (33) Where oi represents the origin, di represents the destination, ci represents the airline, si represents cabin class information, ti represents timeliness information, pi represents price information, ui represents transit information, and ri represents rule constraint information.

[0119] S9: Filter candidate routes and generate unavailable reason codes.

[0120] Candidate routes are filtered based on basic route conditions, cargo attribute constraints, airline cargo acceptance rules, supplier cargo receiving rules, cargo space rules, and transfer rules. Specifically: First, perform basic filtering based on the origin, destination, schedule, and time requirements; Then, filter the goods based on weight, volume, size, number of pieces, electrical properties, dangerous goods properties, and receiving type; Then, business rules are filtered based on airline receiving rules, supplier receiving rules, cargo space rules, and transit rules.

[0121] For any candidate route ai, a feasibility function Feasible(ai,Y) can be defined: (34) Among them, I base Indicates whether the basic conditions are met, I cargo Indicates whether the cargo attribute constraints are satisfied, I rule This indicates whether the business rules are met. When Feasible(ai,Y)=1, the candidate route is retained; otherwise, it is discarded.

[0122] Preferably, if the maximum length L of a single piece max Exceeding the allowed length threshold L for candidate routes limit ,but: (35) If the goods are electronically charged and the corresponding airline does not support the acceptance of electronically charged goods, then: (36) For routes that are rejected, the system retains the rejection reason CodeReject(ai), such as "Do not accept cargo with batteries", "Single piece is too long", "No space available at present", "Timeliness not met", etc., which are used to generate explanations of the reasons for unavailability later.

[0123] S10: Score and sort the filtering results and output the recommended results.

[0124] The filtered candidate routes are scored and ranked, and the scoring weights are dynamically adjusted based on the user's preference for price or timeliness.

[0125] In this embodiment, for any reserved route a i Its overall score is S i It can be represented as: (37) Among them, S time (i) represents the timeliness matching score, S price (i) represents the price competitiveness score, S space (i) represents the cabin availability score, S fit (i) represents the goods attribute fit score, S rule (i) represents the rule satisfaction score, S stable (i) represents the stability score, S prefer (i) represents the user preference satisfaction score, where w1 to w7 are the corresponding weights, and satisfy: (38) Preferably, the price score can be expressed as a normalized reciprocal: (39) Among them, P i Indicates route a i Price, P min P represents the lowest price in the candidate set. max This indicates the highest price in the candidate set.

[0126] Timeliness score can be expressed as: (40) Among them, T i Indicates route ai The corresponding transportation time, T min and T max These represent the best and worst timeliness in the candidate set, respectively.

[0127] When determining scoring weights, the system dynamically determines the weights based on two or more of the following: user preference fields, cabin availability, rule fulfillment, and unavailability reason code distribution. The unavailability reason code distribution refers to the number or proportion of filtered candidate routes under various unavailability reason codes. If the user preference is price priority, the weight corresponding to price competitiveness is increased; if the user preference is timeliness priority, the weight corresponding to timeliness matching is increased; if the unavailability reason code distribution shows that a large number of candidate routes are eliminated due to battery restrictions, dangerous goods restrictions, or oversize restrictions, the weight corresponding to cargo attribute suitability or rule fulfillment is increased. The system determines the weights based on S... i Sort the candidate routes in descending order to obtain the recommended sequence A*: (41) Where A represents the set of candidate routes retained after filtering, S represents the set of comprehensive scores or score vectors corresponding one-to-one with each candidate route in the set of candidate routes A, and A* represents the recommended route sequence arranged in descending order of comprehensive score Si.

[0128] After sorting, the system sorts the candidate routes in descending order based on the comprehensive score, outputs the candidate route with the highest comprehensive score as the primary recommended route, and outputs one or more alternative routes, simultaneously generating recommendation reasons and reasons for unavailability. For filtered candidate routes, the system generates an unavailability reason code CodeReject(ai) based on the filtering process, and maps the unavailability reason code to a natural language unavailability reason; for retained candidate routes, the system generates recommendation reasons based on the contribution of the scoring items. The recommendation reason generation function can be expressed as: (42) The unavailability reason generation function can be represented as: (43) Recommendations can be presented in the form of natural language descriptions, structured tables, or visual cards.

[0129] Training and role of industry-specific fine-tuning large models In this invention, the large model is not only used for general text recognition, but is specifically trained or fine-tuned based on the corpus of the air freight forwarding industry. Its training corpus preferably includes at least air freight forwarding inquiry samples, standard field annotation samples, rule text samples, and feedback samples; wherein, the rule text samples include airline receiving rule text or supplier rule text, and the feedback samples include historical follow-up samples or manually corrected samples, and may further include quotation records, route descriptions, cargo type descriptions, and supplier operation instructions.

[0130] Preferably, the training samples can be represented as: Z = input i label i (44) Among them, input i Represents the i-th industry input sample, label i This indicates the corresponding field annotation result, intent label, or rule label.

[0131] Model training objectives can include field extraction loss L field Intent recognition loss L intent Loss L for identifying missing fields miss Conflict field identification loss L conflict Candidate value ranking loss L rank And confidence calibration loss L conf The total loss can be expressed as: L total = λ1×L field + λ2×L intent + λ3×L miss + λ4×L conflict + λ5×L rank + λ6×L conf (45) Where λ1 to λ6 are the loss weight coefficients.

[0132] After industry-specific fine-tuning, the model can achieve the following functions: Identify colloquial and abbreviation expressions in air freight forwarding scenarios; Identify explicit and implicit fields in the inquiry text; Integrate complementary information from multimodal sources; Output candidate explanations and confidence levels for conflicting fields; Provides a basis for ranking candidate values ​​for multi-source field fusion; To generate supplementary questions related to the business scenario through multiple rounds of follow-up questioning. Specific Implementation Example 1: Text Inquiry Processing The user entered: "Shanghai to Frankfurt, 3 pieces, 180 kg, electronic components, with batteries, preferably flying tomorrow." The system identifies the place of origin, destination, number of pieces, weight, name of goods, electrical properties, and time requirements.

[0134] After verification, the system determined that a size field was missing and generated a supplementary question: "Please provide the maximum size of a single item so that we can determine whether the airline can accept it." After the user adds "maximum size 120×80×75 cm", the system calculates the single item volume, total volume, volumetric weight and chargeable weight according to formulas (18) to (21); then queries the relevant database according to formula (32), performs route filtering according to formula (34), and performs comprehensive sorting according to formulas (37) to (41). The system filters out candidate routes that do not accept goods with batteries, do not meet size restrictions, have no cabin space or do not meet timeliness requirements, and adjusts the scoring weight according to high timeliness requirements, outputs the main recommended scheme, alternative schemes, reasons for recommendation and reasons for unavailability.

[0135] Example 2: Voice Inquiry Processing The user sent a voice message: "Guangzhou to Los Angeles, two pallets, about 500 kilograms, the shipment is urgent, direct flight is preferred." The system performs speech transcription and industry dictionary correction, recognizing the origin, destination, delivery form, weight, time requirements, and direct flight preference.

[0136] If the system determines that the type of goods, the electrical attributes, and the single-item size are missing, it will calculate the priority of follow-up questions according to formula (30) and generate supplementary questions first: "Please confirm whether the goods are electrical and supplement the pallet size." After the user adds information, the system completes rule matching, route query, candidate filtering, and recommendation sorting.

[0137] Example 3: Processing price inquiries based on images or screenshots The user uploads a screenshot of the goods list. The system extracts information such as "Product Name: Lithium Battery Accessories", "Quantity: 5", "Weight: 260kg", and "Dimensions: 95×70×60cm" through text recognition, and then generates a unified structured field object by combining the destination and time requirements in the chat text. When there is a conflict between the image recognition result and the text recognition result, the system calculates the conflict coefficient according to formula (26) and automatically initiates a clarification inquiry. After the user confirms, the system re-executes the query and recommendation process.

[0138] Example 4: Handling when fields do not match The user entered the weight of the goods as 180 kg in the text, but the uploaded goods list identified the weight as 185 kg.

[0139] The system generates two candidate values ​​for the weight field and calculates the combined score and conflict coefficient of the two candidate values.

[0140] The system generates a conflict flag for the weight field and a chain of evidence for the field, and asks the user: "The weight entered in the text is 180 kg, but the packing list shows 185 kg. Please confirm the actual weight." After user confirmation, the system updates the weight field and its status.

[0141] Example 5: Completion of empty results The system did not obtain any candidate routes to retain after performing route query and filtering.

[0142] The system statistical analysis of the unavailability reason code distribution revealed that most candidate routes were filtered out due to single-item size exceeding the limit.

[0143] The system generated a reverse supplementary question: "Currently, most routes cannot accept shipments due to exceeding single-item size limits. Please confirm whether the goods can be split or repackaged." After the user confirms that the item can be split, the system updates the number of items and size fields, and re-executes standardization, validation, query, filtering, and sorting.

[0144] Electronic devices and storage media The present invention also provides an electronic device, which includes a processor, a memory, a communication interface, and a communication bus.

[0145] The processor, memory, and communication interface are interconnected via a communication bus.

[0146] The memory stores a computer program, and when the processor executes the computer program, it performs the above steps S1 to S10.

[0147] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above steps S1 to S10.

[0148] The above are merely preferred embodiments of the present invention and are 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 understanding multimodal price inquiries and recommending routes in air freight forwarding, characterized in that, Includes the following steps: S1: Receive air freight forwarding inquiry information input by the user, wherein the inquiry information includes at least one of text, voice, image, screenshot, and table; S2: Preprocess the inquiry information according to the input modality of the inquiry information to generate a sequence of candidate text fragments with modality source, source location and session context identifier; S3: Input the candidate text fragment sequence into a large model fine-tuned with air freight forwarding industry corpus to identify inquiry intent and generate a candidate set of fields; S4: When the same field corresponds to multiple candidate values ​​or multiple modal sources, the candidate values ​​are fused from multiple sources based on field confidence, modal source reliability, source priority, context consistency and business rule adaptability to determine the target field value and generate field conflict markers and field evidence chains. S5: Convert the field information after multi-source fusion into a unified structured field object, and standardize the unified structured field object by location coding, cargo type, weight unit, volume unit, size unit and preference label; S6: Perform integrity checks, numerical anomaly checks, consistency conflict checks, and business rule checks on the standardized fields, and generate field check results; S7: When missing, abnormal, or conflicting fields affect route queries, calculate the follow-up question priority based on field importance, field confidence, conflict coefficient, and rule impact, and generate supplementary questions according to the follow-up question priority for multi-round interactive completion; when a field meets the query conditions, generate standardized query parameters. S8: Query the route database, price database, cabin class database, airline cargo receiving rule database, and supplier rule database according to the standardized query parameters to obtain a set of candidate routes; S9: Filter candidate routes based on route basic conditions, cargo attribute constraints, airline cargo receiving rules, supplier cargo receiving rules, cargo space rules, and transit rules, and generate unavailability reason codes for the filtered candidate routes; S10: Score and sort the filtered candidate routes, dynamically adjust the scoring weights based on the user's preference for price or timeliness, and output the recommended route, alternative routes, reasons for recommendation, and reasons for unavailability.

2. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 1, characterized in that, Step S2 includes: When the input is text, perform sentence segmentation, noise reduction, error correction, terminology segmentation, and context concatenation on the text; When the input is speech, speech is transcribed and corrected for homophones and industry abbreviations using an air freight forwarding industry dictionary. When the input is an image or screenshot, perform text recognition, layout analysis, table area recognition, field adjacency recognition, and key field area extraction on the image or screenshot; When the input is a table, column name recognition, cell parsing, and field mapping are performed on the table.

3. The method for understanding multimodal price inquiries and recommending routes for air freight forwarding according to claim 1, characterized in that, In step S3, the training corpus of the large model, which has been fine-tuned by the air freight forwarding industry corpus, includes at least air freight forwarding inquiry samples, standard field annotation samples, rule text samples, and feedback samples. The rule text samples include airline receiving rule texts or supplier rule texts, and the feedback samples include historical inquiry samples or manually corrected samples. The training objectives of the large model include at least the field extraction objective, the candidate value ranking objective, and the confidence calibration objective, and further include one or more of the following: inquiry intent identification objective, missing field identification objective, or conflicting field identification objective.

4. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 1, characterized in that, The candidate set of fields generated in step S3 includes at least some of the following fields: origin, destination, cargo name, cargo type, number of pieces, weight, volume, dimensions, receiving type, electrical attributes, dangerous goods attributes, oversized attributes, timeliness requirements, cargo space requirements, and price preferences.

5. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 1, characterized in that, Step S4 includes: Generate multiple candidate values ​​and their candidate source identifiers for the same field; The comprehensive score of each candidate value is calculated based on field confidence, modality source reliability, source priority, context consistency, and business rule adaptability. Select the candidate value with the highest overall score as the target field value; Generate field conflict flags based on the differences between unadopted candidate values ​​and target field values; And generate a chain of field evidence for the target field value.

6. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 5, characterized in that, The modality source reliability is determined based on one or more of the following: source type, identification method, historical accuracy, and manual confirmation status. The business rule adaptability is determined based on whether the candidate field values ​​meet one or more of the following: airline receiving rules, supplier receiving rules, battery restrictions, dangerous goods restrictions, oversize restrictions, or special receiving type restrictions.

7. The method for understanding multimodal price inquiries and recommending routes for air freight forwarding according to claim 1, characterized in that, The field evidence chain includes at least the original input fragment, modality source, source location, and field confidence, and further includes one or more of the following: candidate interpretation, adopted candidate value, and unadopted candidate value.

8. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 1, characterized in that, Step S5 includes: Map city names, airport names, or abbreviations to airport three-letter codes, city codes, or preset node codes; Map natural language cargo descriptions to general cargo, cargo with batteries, dangerous goods, fresh goods, oversized cargo, or pre-defined cargo category codes; The weight is standardized to kilograms, the volume to cubic meters, and the dimensions to centimeters or millimeters. Map timeliness descriptions to timeliness levels, and price descriptions to price priorities; Calculate volumetric weight, maximum single-piece size, or oversized risk label based on weight, volume, number of pieces, and dimensions.

9. The method for understanding multimodal price inquiries and recommending routes for air freight forwarding according to claim 1, characterized in that, Step S6 includes: Perform integrity checks on the origin, destination, and basic cargo attributes; Perform numerical reasonableness and consistency checks on weight, volume, number of pieces, and dimensions; Perform business risk assessment based on cargo type, electrical properties, hazardous materials properties, and receiving type; Perform conflict identification on the same field values ​​from different modalities.

10. The method for understanding multimodal price inquiries and recommending routes for air freight forwarding according to claim 1, characterized in that, In step S7, the rule impact degree is determined based on the degree of impact of missing, abnormal, or conflicting fields on route database queries, airline receiving rule matching, supplier receiving rule matching, electric limit judgment, dangerous goods limit judgment, or oversize limit judgment.

11. The method for understanding multimodal price inquiries and recommending routes for air freight forwarding according to claim 1, characterized in that, In step S7, the fields satisfying the query conditions include: The standardized field object must include at least the origin code, destination code, and weight field; When the cargo name, cargo type, or risk label indicates that the cargo may involve electrical components, dangerous goods, oversized items, or special receiving types, the standardized field object further includes one or more of the following: size field, cargo type code, electrical component label, or dangerous goods label.

12. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 1, characterized in that, In step S8, the route, price, cabin class and rule data are parametrically retrieved and matched based on the standardized query parameters, instead of being retrieved directly based on the original natural language.

13. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 1, characterized in that, Step S9 includes: Filter candidate routes based on origin, destination, schedule, and timeliness requirements; Candidate routes are filtered based on weight restrictions, volume restrictions, size restrictions, electrical restrictions, dangerous goods restrictions, and cargo type restrictions; Candidate routes are filtered based on airline receiving rules, supplier receiving rules, cabin space rules, and transit rules.

14. The method for understanding multimodal price inquiries and recommending routes for air freight forwarding according to claim 1, characterized in that, The unavailability reason code includes at least one of the following: non-acceptance of goods with batteries, dangerous goods restrictions, single piece exceeding length, no current space available, time limit not met, weight restrictions, volume restrictions, or supplier not accepting the goods.

15. The method for understanding multimodal price inquiries and recommending routes in air freight forwarding according to claim 1, characterized in that, Step S10 includes: The comprehensive score of candidate routes is calculated based on two or more of the following factors: timeliness matching, price competitiveness, availability of cabin space, cargo attribute suitability, rule compliance, stability, and user preference satisfaction. The scoring weights are dynamically determined based on two or more of the following: user preference fields, cabin availability, rule satisfaction, and unavailability reason coding distribution. When users prioritize price, increase the weight corresponding to price competitiveness; When users prioritize timeliness, increase the weight corresponding to timeliness matching.

16. The method for understanding multimodal price inquiries and recommending routes for air freight forwarding according to claim 15, characterized in that, When the distribution of unavailability reason codes shows that more than a preset proportion of candidate routes are eliminated due to power restrictions, dangerous goods restrictions, or oversize restrictions, the corresponding weights of cargo attribute suitability or rule satisfaction are increased.

17. A multimodal inquiry understanding and route recommendation system for air freight forwarding, characterized in that, include: A multimodal input receiving module is used to receive air freight forwarding inquiry information input by the user, wherein the inquiry information includes at least one of text, voice, image, screenshot, and table; The preprocessing module is used to preprocess the inquiry information according to the input modality of the inquiry information to generate a sequence of candidate text fragments with modality source, source location and session context identifier; The industry fine-tuning large model understanding module is used to input the candidate text fragment sequence into the large model fine-tuned by the air freight forwarding industry corpus, identify the inquiry intent and generate a set of field candidates. The multi-source field fusion module is used to perform multi-source fusion of candidate values ​​based on field confidence, modality source reliability, source priority, context consistency, and business rule adaptability when the same field corresponds to multiple candidate values ​​or multiple modal sources, in order to determine the target field value. The field conflict identification module is used to generate field conflict flags based on the differences between unadopted candidate values ​​and target field values; The field evidence chain generation module is used to generate field evidence chains that include the original input fragment, modality source, source location, and field confidence. The structured field object generation module is used to convert field information after multi-source fusion into a unified structured field object; The field standardization module is used to standardize the location coding, cargo type, weight unit, volume unit, size unit, and preference label of the unified structured field object; The field validation module is used to perform integrity checks, numerical anomaly checks, consistency conflict checks, and business rule checks on the standardized fields. The follow-up question priority calculation module is used to calculate the follow-up question priority based on field importance, field confidence, conflict coefficient, and rule impact when missing, abnormal, or conflicting fields affect route queries. A multi-round interactive completion module is used to generate supplementary questions according to the priority of follow-up questions and to receive supplementary information from the user. The query parameter generation module is used to generate standardized query parameters when a field meets the query conditions. The route database query module is used to query the route database, price database, cabin class database, airline cargo receiving rule database and supplier rule database according to the standardized query parameters to obtain a set of candidate routes; The candidate route filtering module is used to filter candidate routes based on basic route conditions, cargo attribute constraints, airline cargo receiving rules, supplier cargo receiving rules, cargo space rules, and transfer rules. The unavailability reason coding module is used to generate unavailability reason codes for the filtered candidate routes; The recommendation and ranking module is used to score and rank the filtered candidate routes, and dynamically adjust the scoring weights based on the user's preference for price or timeliness. The recommendation reason generation module is used to output recommended route options, alternative options, reasons for recommendation, and reasons for unavailability.

18. The air freight forwarding multimodal inquiry understanding and route recommendation system according to claim 17, characterized in that, The preprocessing module includes a text processing unit, a speech-to-text unit, an OCR recognition unit, a table parsing unit, and a multi-source information alignment unit; The text processing unit is used to perform sentence segmentation, noise reduction, error correction, terminology segmentation, and context concatenation on the text. The speech transcription unit is used to perform speech transcription and, in conjunction with an air freight forwarding industry dictionary, correct homophones and industry abbreviations. The OCR recognition unit is used to perform text recognition, layout analysis, table area recognition, field adjacency recognition, and key field area extraction on images or screenshots; The table parsing unit is used to perform column name recognition, cell parsing, and field mapping on the table. The multi-source information alignment unit is used to align the candidate field values ​​corresponding to different modalities according to the field name, source location, and session context.

19. The air freight forwarding multimodal inquiry understanding and route recommendation system according to claim 17, characterized in that, The field standardization module includes a location coding standardization unit, a cargo type coding standardization unit, a unit conversion unit, a size structuring unit, and a preference label normalization unit; The location coding standardization unit is used to map city names, airport names or abbreviations to airport three-letter codes, city codes or preset node codes; The cargo type coding standardization unit is used to map natural language cargo descriptions to cargo category codes and risk labels; The unit conversion unit is used to convert weight, volume, and size into a preset uniform unit; The size structured unit is used to calculate the total volume, volumetric weight, chargeable weight, or oversized risk label based on the number of pieces and the size of each piece. The preference label normalization unit is used to map timeliness descriptions or price descriptions to timeliness levels or price priorities.

20. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the air freight forwarding multimodal inquiry understanding and route recommendation method according to any one of claims 1 to 16.

21. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the air freight forwarding multimodal inquiry understanding and route recommendation method according to any one of claims 1 to 16.

Images