Flight service system based on knowledge graph and large model

Abstract

The invention discloses a flight service system based on a knowledge graph and a large model. Comprising an ontology construction module used for predefining related information of a flight service system based on user demands; the data acquisition module is used for acquiring open source data related to the flight service system from multiple channels and performing deduplication and normalization processing on the open source data to obtain standard open source data; the knowledge graph construction module is used for extracting triads from standard open source data based on the related information to obtain a triad set of the knowledge graph; the prompt generation module is used for generating a corresponding target prompt based on a user demand; the large model fine tuning module is used for training a pre-constructed large model based on the target prompt and the triple set to obtain a trained large model; the trained large model is used for generating a result matched with the target task according to an input user prompt and the triple. According to the invention, diversified services can be provided for users.

Description

Technical Field

[0001] The present invention relates to the technical field of service system optimization, and in particular to a flight service system based on knowledge graph and big model. Background Art

[0002] With the rapid development of the air transportation industry, flight service systems play a vital role in the operations of airlines and airports. Traditional flight service systems mainly rely on rule-driven database management systems. Although they can provide basic flight query, ticket booking, service and other functions, they have obvious limitations in handling complex user needs, personalized service recommendations and responding to emergencies.

[0003] Therefore, there is an urgent need for a flight service system that can better meet the diverse needs of users. Summary of the invention

[0004] The present invention provides a flight service system based on knowledge graph and big model, which can provide diversified services to users. The technical solution is as follows:

[0005] The present invention provides a flight service system based on knowledge graph and big model.

[0006] An ontology building module, used to predefine relevant information of the flight service system based on user needs, wherein the relevant information includes concepts related to flight services, attributes associated with each of the concepts, and relationships between the concepts;

[0007] A data acquisition module, used to acquire open source data related to the flight service system from multiple channels, and perform deduplication and normalization processing on the open source data to obtain standard open source data;

[0008] A knowledge graph construction module is used to extract triples from the standard open source data based on the relevant information to obtain a triple set of the knowledge graph; the entities and relationships in each triple are used to represent the concepts, attributes and relationships in the relevant information respectively;

[0009] A prompt generation module, used to generate corresponding target prompts based on user needs;

[0010] The large model fine-tuning module is used to train the pre-built large model based on the target prompt and the triple set to obtain a trained large model; the trained large model is used to generate a result matching the target task according to the input user prompt and triple.

[0011] The embodiment of the present invention provides a flight service system based on a knowledge graph and a large model. By constructing a knowledge graph in the field of flight services and integrating multi-dimensional data such as flight information, passenger needs, and service resources, the standardization and intelligent management of flight service information is achieved. At the same time, the system introduces a large-scale pre-trained language model, and then generates a target prompt according to user needs, and fine-tunes the large model under the target prompt to obtain a trained model. The trained model can generate results that match the target task based on the input user prompts and triples. This application improves the system's natural language processing capabilities and user interaction experience by fine-tuning specific tasks. The system can efficiently handle tasks such as passenger inquiries, flight recommendations, and question answering, provide accurate personalized services, and have adaptive optimization and intelligent decision-making capabilities. It is suitable for airlines, airport management and other fields, and can significantly improve the efficiency of flight services and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 It is a flow chart of a flight service system based on knowledge graph and big model provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0015] As mentioned above, the existing flight service system cannot meet the diverse needs of customers. The inventor found in his work that this is mainly because: on the one hand, the existing flight service system lacks the ability to process a large amount of unstructured data (such as user interaction history, natural language feedback, etc.), and cannot accurately understand user intentions and provide flexible intelligent services. On the other hand, the existing flight service system lacks in-depth use of domain knowledge and fails to give full play to the potential of knowledge graphs.

[0016] Based on this, the inventor proposed that we can make full use of open source data from multiple channels to obtain the latest and most comprehensive flight service information and obtain multi-dimensional data of flight services. We can also extract entities and relationships from the multi-dimensional data, effectively combine the structured information management capabilities of the knowledge graph and the reasoning capabilities of the large model, and obtain a flight service system that meets the requirements.

[0017] The specific implementation of the above concept is described below.

[0018] Please refer to Figure 1 , an embodiment of the present invention provides a flight service system based on knowledge graph and big model, including:

[0019] An ontology building module is used to predefine relevant information of the flight service system based on user needs. The relevant information includes concepts related to flight services, attributes associated with each concept, and relationships between concepts.

[0020] The data acquisition module is used to obtain open source data related to the flight service system from multiple channels, and perform deduplication and normalization on the open source data to obtain standard open source data;

[0021] The knowledge graph construction module is used to extract triples from standard open source data based on relevant information to obtain a set of triples in the knowledge graph; the entities and relationships in each triple are used to represent the concepts, attributes, and relationships in the relevant information respectively;

[0022] A prompt generation module, used to generate corresponding target prompts based on user needs;

[0023] The large model fine-tuning module is used to train the pre-built large model based on the target prompts and triplet sets to obtain a trained large model; the trained large model is used to generate results that match the target task based on the input user prompts and triplet sets.

[0024] In this embodiment, by constructing a knowledge graph in the field of flight services and integrating multi-dimensional data such as flight information, passenger needs, and service resources, the standardization and intelligent management of flight service information is achieved. At the same time, the system introduces a large-scale pre-trained language model, and then generates target prompts according to user needs, and fine-tunes the large model under the target prompts to obtain a trained model. The trained model can generate results that match the target task based on the input user prompts and triples. This application improves the system's natural language processing capabilities and user interaction experience by fine-tuning specific tasks. The system can efficiently handle tasks such as passenger inquiries, flight recommendations, and question answering, provide accurate personalized services, and have adaptive optimization and intelligent decision-making capabilities. It is suitable for airlines, airport management and other fields, and can significantly improve the efficiency of flight services and user satisfaction.

[0025] In some embodiments, for an ontology building module:

[0026] Concepts related to flight services include at least: flights, airlines, aircraft types and airports;

[0027] The attributes associated with each concept include at least: flight time, fare, punctuality, food service quality, user praise rate, airline reputation and service rating, and aircraft seat configuration and comfort.

[0028] In addition, the relationships between concepts can be determined based on actual conditions, such as "(flight) aircraft type (model)", "(flight) departure point (airport)", "(flight) affiliated airline (airline company)", etc., to ensure that the relationships between different concepts can be semantically expressed through the knowledge graph, further supporting the system's intelligent recommendations.

[0029] In this application, the ontology building module can not only represent the basic information of flights, but also expand the user's personalized features in the knowledge graph by analyzing the user's historical preferences and purchasing behaviors, such as the airlines that users frequently choose, the preferred aircraft models, the commonly used airports, and the price range preferences. Through these predefined concepts, attributes, and relationships, the system can more accurately understand and process user needs, thereby providing semantic and structural support for subsequent large model fine-tuning and flight recommendations.

[0030] In some embodiments, the data acquisition module can automatically crawl information related to flight services from a large number of open source data sources such as official websites of major airlines, ticket booking websites, travel platforms, and online encyclopedias. The data acquisition module uses web crawler technology and API interfaces to capture multi-dimensional data of flight services in real time to ensure that the flight recommendation system obtains the latest and most comprehensive flight service information. The open source data crawled by the data acquisition module includes but is not limited to the following:

[0031] Flight information: Get detailed information about flights around the world from major airline websites, including flight number, take-off and landing time, take-off and landing airports, flight status (such as delay, cancellation, etc.) and aircraft type. For example, flight information crawled from the official website of Air China can show "Flight CA123, take-off time 10:00, origin Beijing Capital International Airport, destination Shanghai Pudong International Airport, aircraft type Airbus A320".

[0032] Route and fare information: Obtain route information and dynamic fares provided by airlines from air ticket booking websites (such as Ctrip, Expedia, etc.). The data includes the starting and ending points of the route, flight time, fare changes, cabin classification (economy class, business class, first class), etc. The system can understand price fluctuations through real-time data updates, for example: "International route from Beijing to New York, the current economy class fare is ￥5000, and the departure time is tomorrow 14:00" obtained from Ctrip.

[0033] Model information: crawling flight-related model data from online encyclopedias (such as Wikipedia) or airline websites, including model specifications, seat configuration, fuel efficiency, range, etc. For example, the system can crawl from Wikipedia to obtain "detailed information about the Boeing 777 model: passenger capacity 350 people, range of about 14,000 kilometers, seat layout 3-4-3".

[0034] Airport information: Get information about major airports from travel websites or official airport websites, including airport facilities, terminal services, transportation information, boarding procedures, etc. For example, information crawled from Qunar.com can show "the terminal layout, rest area, catering services, and security check time reminders of Beijing Capital International Airport."

[0035] Airline information: crawl information such as the airline's service level, customer reviews, fleet size, route coverage, etc. from the airline's official website or related industry reports. For example, the system can obtain "United Airlines has 800 aircraft covering multiple international routes in North America, Asia and Europe" from the airline's official website.

[0036] In addition, the data acquisition module can automatically process and dynamically update the above data, and ensure the diversity and real-time nature of the information. For example, when the status of a flight changes (such as delay or cancellation), the system can promptly capture the latest information from the data source and update the flight recommendation system to ensure that the flight recommendations provided to users are always the latest and in line with the actual status. In addition, the data acquisition module also has efficient data cleaning and deduplication functions, ensuring that duplicate data crawled from different sources can be processed, and seamlessly connecting structured data with the knowledge graph, further improving the training effect and recommendation accuracy of the large model.

[0037] It should also be noted that there is a large amount of open source flight data on the Internet, such as journal articles, encyclopedia pages, forum posts, etc. The text format and file format of these corpora are very different, so before the extraction task, the collected open source data needs to be pre-processed and normalized. The processing process is as follows:

[0038] Convert non-text format data into text data;

[0039] Crawl web page data into html file data;

[0040] Convert foreign language corpus data into Chinese corpus through offline translation software or translation model;

[0041] Formulate extraction templates according to different types of corpora.

[0042] Through the above processing, standard open source data can be obtained.

[0043] In some embodiments, the open source data includes structured data and unstructured data;

[0044] For structured data, the template matching method is used to extract triplets from standard open source data.

[0045] This kind of structured and semi-structured text usually has a relatively uniform format, that is, each document is a variety of attribute values ​​of an entity. There will be various titles, which represent the meaning or aspects of the entity attribute values. It also contains tables, and each table has a corresponding title. Therefore, each text paragraph is identified as the tail entity, the title closest to it is found as a relationship, and the main title of the document is used as the head entity to form its basic attribute triples. In addition, for tabular data, its title is used as a relationship. This template matching-based triple extraction method aims to automatically identify and extract entities, relationships, and objects from Word documents to construct a knowledge graph. This method achieves efficient parsing and triple construction of diversified information by analyzing the hierarchical structure and tabular data of the document.

[0046] Template matching methods have attracted attention in information extraction due to their simplicity and efficiency. They are specifically targeted at structured and semi-structured texts in Word documents, and construct knowledge graphs by automatically identifying and extracting entities, relations, and objects. The theoretical basis of template matching methods lies in the semantic role labeling (SRL) technology in the field of natural language processing. SRL aims to identify the arguments of verbs in sentences and their corresponding semantic roles, which is crucial for understanding the deep meaning of sentences. In document analysis, SRL technology can be applied to various parts of documents, such as titles, paragraphs, and tables, treating them as text units with specific semantic functions. Through predefined templates, the system can automatically identify entities, relations, and objects in these units, thereby realizing automatic information extraction. For example, document titles usually contain key entity information, paragraph text provides descriptions or attributes of entities, and tables present comparative information between relations or entities.

[0047] Specifically, we first determine the entity based on the title in the document, and treat the subsequent paragraph text as an object associated with the entity. At the same time, the hierarchical relationship between titles is regarded as a relationship in a triple. For the table content in the document, the title before the table is parsed as a relationship, and the content of each cell in the table is organized into object parts to generate triples. By leveraging the inherent structure of the document, unstructured text, table, and image information are systematized into triples. This method is particularly suitable for scenarios where documents have clear hierarchical relationships and can abstract complex document content into structured data forms.

[0048] For example, for HTML pages crawled from airline websites, ticket booking platforms, etc., the system identifies and matches flight-related entity information by parsing the HTML tag structure of the page. , , By combining the page's hierarchical structure, tag attributes, and CSS styles, the system can accurately locate key entities such as flight number, departure time, price, airline name, and aircraft model. For example, when the HTML page crawled by the system contains a table tag <span class＝"flight-number"> CA123 <span class＝"price">￥5000

[0049]

[0050]

[0051]

[0052]

[0053] n

[0054]

[0055]

[0056]

[0057] i

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] i

[0070]

[0071]

[0072] j i l

[0073] i i i

[0074] n n n i i i

[0075]

[0076] i i i i i i i i i

[0077]

[0078]

[0079]

[0080] i i i i i i

[0081]

[0082]

[0083]

[0084] i i i i i i i i i

[0085]

[0086]

[0087]

[0088]

[0089] i i i i i i

[0090] i i i i

[0091] i j i i i j j j i i i j j j

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] θ i,j i,j θ

[0103] θ

[0104]

[0105]

[0106]

[0107]

[0108] When the system automatically matches the corresponding entities based on the column headers in the table, such as "flight number", "departure time", "price", etc., and extracts specific data from the corresponding cells. This module can identify specific data based on the attribute values ​​of the tag (such as id, class, etc.). For example, the HTML code can be parsed as the flight number "CA123", while the flight price information can be parsed. The system can automatically identify the relevant flight information entities in different page structures through predefined entity rules. For text information collected from online encyclopedias (such as Wikipedia) or other document formats (such as PDF, Word, etc.), the system performs entity matching by analyzing the formatting features of the text, such as font, paragraph format, punctuation, etc. For example, in encyclopedia documents, different concepts and entities are usually represented by bold fonts or specific title formats (such as "Model: Boeing 737"), and the system can extract entities based on these formatting prompts. In addition, the logical relationship between paragraphs (such as the association between the title and the text) can also help the system accurately match entities, such as matching "model" with the subsequent "Boeing 737" in PDF documents. In some embodiments, unstructured text includes but is not limited to news reports, user comments, aviation industry reports, forum posts and discussions on social media. The system uses open source large-scale pre-trained language models to fine-tune these models to adapt to specific task requirements in the field of flight services. The system automatically extracts key entities (such as flight numbers, airlines, aircraft models, airports, etc.) and their relationships in unstructured texts through natural language processing (NLP) technology, combined with named entity recognition (NER), syntactic analysis and semantic understanding. The system uses a large model to perform deep semantic analysis on unstructured texts to identify entities mentioned in the text. Then, the system automatically generates semantic relationships between these entities through relationship extraction technology. For example, from a news report on a flight delay, the system can extract entities such as "Flight CA123", "Beijing Capital International Airport", "Airline Air China", "Delay", and extract the relationships between these entities, such as "Flight CA123's departure point is Beijing Capital International Airport", "Flight CA123 is affiliated to Air China", and "Flight CA123 is delayed due to weather reasons". For unstructured data, triples are extracted from standard open source data based on mature open source large models. For example, the open source large model can use the Baichuan-13B model. Baichuan-13B is an open source commercial large-scale language model with 13 billion parameters developed by Baichuan Intelligence after Baichuan-7B. It has achieved the best results of the same size on authoritative Chinese and English benchmarks. First, you need to import the Baichuan-13B model file into the server. It has been trained with large-scale data, but has not been further adjusted for specific tasks or fields.The basic version of the Baichuan-13B model usually has general language understanding and generation capabilities and can be used for a wide range of natural language processing tasks, including entity extraction and relation extraction tasks. The pre-trained large model can complete entity extraction and relation extraction tasks, mainly due to its strong language understanding ability and training accumulation on a large and diverse corpus. The pre-trained large model, especially the model based on the Transformer structure, can generate context-related representations for each word in the text. This representation can capture the semantic role of the current word in the entire sentence or even paragraph, and can therefore be effectively used to identify entities and relations. Use a context-based prompt method for entity and relation extraction tasks. The following is a detailed description of the triple extraction process: First, given an input text sequence X = {x1, x2, …, x}, construct a prompt to guide the model to identify entities or relations in the text. Let the prompt be P and concatenate the prompt with the input sequence: where represents the concatenation operation, and P is the prompt text designed according to the task, such as "Identify the entities and their relations in the text:", which is used to activate the generation or reasoning ability of the model. The pre-trained model receives the concatenated input and generates a context-related representation: where z is the context vector generated by the model for each input word, and p is the length of the prompt. Under the guidance of the prompt, the model can generate entity extraction results by filling or answering. Assume that after the prompt, it is expected to generate an entity sequence by predicting the corresponding context part in Z to obtain the entity: The generation process can be through the output layer of the language model, generating possible entities by word or by tag. Of course, the entity extraction task can also be converted into a question-answering form, such as asking "What are the entities in the sentence?" through the prompt, and the model generates an entity set based on the context and prompt. For relationship extraction, it can be carried out in a similar way as above, and the model generates relations based on the input: where what is generated is the entity relationship inferred by the model through the context and prompt. In addition, there are also some structured and semi-structured texts in the open source data, which can be extracted using template matching. This structured and semi-structured text usually has a relatively uniform format, that is, each document is a variety of attribute values ​​of an entity. There will be various titles, which represent the meaning or aspects of the entity attribute values. It also contains tables, and each table has a corresponding title. Therefore, each text paragraph is identified as the tail entity, the title closest to it is found as a relation, and the main title of the document is used as the head entity to form its basic attribute triples. In addition, for tabular data, its title is used as a relation. This template matching-based triple extraction method aims to automatically identify and extract entities, relations, and objects from Word documents and build a knowledge graph.This method realizes efficient parsing and triple construction of diversified information by analyzing the hierarchical structure and tabular data of the document. In some embodiments, the target prompt is determined based on a set of relationships, a set of subordinate relationships between relationships, a domain of a relationship, and a range of a relationship. The set of relationships can be represented in the following form: wherein each r is used to represent a type of relationship, i=1,2...k, and k is the total number of relationships. The set of subordinate relationships between relationships can be represented in the following form: wherein each s represents a subordinate or dependent relationship between two relationships r and r, j=1,2...m, l=1,2...k, and i≠l. Assume that the ontology type of the domain and range of the relationship is known: the domain D(r) and range R(r) of each relationship r are known, indicating the ontology type of the head entity and tail entity applicable to the relationship. Relation triple set: The set of triples obtained by relation extraction is T = {(h1, r1, t1), (h2, r2, t2), …, (h, r, t)}, where h is the head entity, r is the relationship, and t is the tail entity. Next, these relationship triples are input into the big model, and prompts are used to match the entity and relationship types in each triple. Each relationship triple (h, r, t) contains a head entity h, a relationship r, and a tail entity t. The task of the big model is to match the ontology types to which h, r, and t belong. In order for the big model to identify the ontology type in the triple, the relationship ontology set and the subordinate relationship set S are used as part of the prompt. Assume that the input prompt P contains the following information: Where D(r) represents the domain of the relationship r, and R(r) represents the value range of the relationship r. These prompts will help the big model match the ontology types to which the head entity h and the tail entity t belong when processing triples. In some embodiments, the pre-built large model is trained in the following manner: determine the constraints and loss function of the large model; under the constraints, take the extracted triple set and the target prompt generated based on user needs as input, and take the ontology type to which each triple belongs as output, and train the pre-built large model until the loss value of the loss function meets the requirements, and obtain a trained large model. In the training process of the large model, the triple set T and the prompt P are used as the input of the large model to generate the ontology type for each triple (h, r, t). The goal of the model is to find the ontology types C(h), C(t), and C(r) that match h, t, and r. Let the model output be: where is the ontology type to which the head entity, relationship, and tail entity in the triple predicted by the model based on the prompt P belong.In some embodiments, based on known domain and range information, the model needs to ensure the following constraints: (1) The ontology type of the head entity h of each triple must match the domain D(r) of its relation r, and the ontology type of the tail entity t must match the range R(r) of its relation r, that is: C(h) = D(r), C(t) = R(r) (2) If the relations r and r in the two triples are subordinate to each other, then the ontology types C(h,r,t) and C(h,r,t) to which the two triples (h,r,t) and (h,r,t) belong also have a subordinate relationship. That is, if the large model generates the ontology type matching result of each triple based on the prompt and the input triple set, and the output is: After the ontology matching of this embodiment, the knowledge graph is initially constructed to form a two-layer network structure consisting of an ontology layer and an entity layer. The two-layer network is mapped through a subordinate relationship (i.e., the mapping between a specific entity and the ontology to which it belongs) and a type relationship (the semantic association between entities), thereby realizing the connection and communication between the ontology network and the entity network. This double-layer structure not only enhances the expressive power of the knowledge graph, but also provides a semantically rich foundation for subsequent reasoning and querying, facilitating the expansion and reasoning of knowledge. In some embodiments, the loss function is: when training the pre-built large model, the parameters of the large model are updated using the gradient descent method shown in the following formula: the target prompt is optimized by the following formula to maximize the log-likelihood of the large model output, ensuring that the output of the large model meets the task requirements: where L(θ) is the loss function; N represents the number of training samples; T represents the output sequence length of each sample; η is the learning efficiency; θ represents the parameters in the large model; is the gradient about the parameter θ; P is the prompt about the parameter θ; x is the input of the large model; y is the output of the large model; logP represents the log-likelihood. In this embodiment, by introducing the prompt p, the model is guided to generate the required output according to a specific task, and the design of the prompt p directly affects the conditional probability distribution of the model. The ideal prompt word design should maximize the log-likelihood logP(y|p,x) of the target output. By adjusting the content and form of the prompt word, the model can more easily generate outputs consistent with the task objectives. In some embodiments, when training a pre-built large model, LoRA technology is used to accelerate the convergence speed of the large model. LoRA (Low-Rank Adaptation of Large Language Models) is a lightweight fine-tuning method specifically designed for efficient parameter fine-tuning on large-scale pre-trained models. The purpose of LoRA is to adapt to downstream tasks without significantly increasing model parameters, thereby reducing the demand for computing and storage resources. The key idea of ​​LoRA is to simplify weight updates during fine-tuning through low-rank decomposition. Specifically, LoRA assumes that during the fine-tuning process, some parameter changes in the pre-trained model have a low-rank structure.Through this low-rank constraint, LoRA is able to significantly reduce the number of parameters that need to be fine-tuned while maintaining model performance. By limiting the fine-tuning process to low-rank matrices, LoRA greatly reduces the number of parameters that need to be adjusted, usually reducing the number of parameters by more than an order of magnitude. LoRA only introduces additional low-rank matrices at specific layers and does not need to modify the main parameters of the original model, so it can be fine-tuned without changing the model architecture. Due to the small number of updated parameters, LoRA can significantly reduce the computing resources required for fine-tuning. This makes fine-tuning large pre-trained models more feasible, especially for scenarios with limited computing resources. In the fine-tuned model, LoRA only inserts low-rank matrices at specific layers, which means that very little additional storage is required for deployment. Compared with traditional fine-tuning, LoRA is more lightweight. Through large model fine-tuning training, the large model has knowledge in related fields, forming a basic question-answering model that can complete tasks related to flight services. Finally, it should be noted that, in this article, relational terms such as first, second, third and fourth are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The above is only a preferred embodiment of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principles of the present application, several improvements and modifications can also be made, and these improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A flight service system based on knowledge graph and big model, characterized in that: include: An ontology building module, used to predefine relevant information of the flight service system based on user needs, wherein the relevant information includes concepts related to flight services, attributes associated with each of the concepts, and relationships between the concepts; A data acquisition module, used to acquire open source data related to the flight service system from multiple channels, and perform deduplication and normalization processing on the open source data to obtain standard open source data; A knowledge graph construction module is used to extract triples from the standard open source data based on the relevant information to obtain a triple set of the knowledge graph; the entities and relationships in each triple are used to represent the concepts, attributes and relationships in the relevant information respectively; A prompt generation module, used to generate corresponding target prompts based on user needs; The large model fine-tuning module is used to train the pre-built large model based on the target prompt and the triple set to obtain a trained large model; the trained large model is used to generate a result matching the target task according to the input user prompt and triple.

2. The system according to claim 1, characterized in that Concepts related to flight services include at least: flights, airlines, aircraft types and airports; The attributes associated with each of the concepts include at least: flight time, fare, punctuality, food service quality, user praise rate, airline reputation and service rating, and seat configuration and comfort of the aircraft model.

3. The system according to claim 1, characterized in that The standard open source data at least includes: Flight information, route and fare information, aircraft type information, airport information and airline information.

4. The system according to claim 1, characterized in that The open source data is normalized using the following method: Convert non-text format data into text data; Crawl web page data into html file data; Convert foreign language corpus data into Chinese corpus through offline translation software or translation model; Formulate extraction templates according to different types of corpora.

5. The system according to claim 1, characterized in that The open source data includes structured data and unstructured data; For the structured data, a template matching method is used to extract triples from the standard open source data; For the unstructured data, triples are extracted from the standard open source data based on a mature open source big model.

6. The system according to claim 2, characterized in that The target prompt is determined based on a set of relations, a set of subordinate relations between relations, a domain of a relation and a range of a relation.

7. The system according to claim 6, characterized in that The pre-built large model is trained as follows: Determine the constraints and loss function of the large model; Under the constraints, the pre-constructed large model is trained with the extracted triple set and the target prompts generated based on user needs as input and the ontology type to which each triple belongs as output until the loss value of the loss function meets the requirements, thereby obtaining a trained large model.

8. The system according to claim 7, characterized in that The constraints of the large model are: The ontology type of the head entity of each triple must match the definition domain of its relationship, and the ontology type of the tail entity must match the value range of its relationship; If the relationship between two triples exists in a subordinate relationship, then the ontology types to which the two triples belong also exist in a subordinate relationship.

9. The system according to claim 7, characterized in that The loss function is: When training the pre-built large model, the parameters of the large model are updated using the gradient descent method shown in the following formula: The target prompt is optimized by the following formula to maximize the log-likelihood of the large model output and ensure that the output of the large model meets the task requirements: Where L(θ) is the loss function; N is the number of training samples; T is the length of the output sequence of each sample; η is the learning efficiency; θ is the parameter in the large model; θ (t) and θ (t+1) are the parameters at the tth and t+1th iterations respectively; is the gradient with respect to the parameter θ; P θ is a hint about the parameter θ; x i,j is the input of the large model; y i,j is the output of the large model; log P θ Represents a natural number.

10. The system according to claim 6, characterized in that When training the pre-built large model, LoRA technology is used to accelerate the convergence speed of the large model.

Images