Hotel cross-platform intelligent mapping method and system based on large language model and dynamic case library

Abstract

The invention provides a hotel cross-platform intelligent mapping method based on a large language model and a dynamic case library. The hotel cross-platform intelligent mapping method comprises the following steps: (1) performing multi-dimensional feature extraction on to-be-mapped hotel data from an external platform and standard hotel data in a standard hotel library; (2) constructing a dynamic mapping decision case library based on historical mapping records, wherein cases of the case library at least comprise input features, judgment results and judgment bases; (3) searching similar historical cases from the dynamic mapping decision case library for the current hotel pair to be judged, and constructing prompt information for guiding the large language model to perform analogy reasoning based on a search result; (4) inputting the prompt information into a general large language model, and obtaining a structured reasoning result which is output by the general large language model and contains the judgment of whether the same entity exists or not; and (5) executing a mapping decision based on the reasoning result, and feeding back a new judgment result generated in the decision process and a basis thereof to the dynamic mapping decision case library.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent matching and entity alignment of hotel data, which is particularly applicable to scenarios in online travel agency (OTA) distribution businesses where hotel information from multiple heterogeneous platforms is efficiently and accurately mapped to a unified standard database. Background Technology

[0002] In the hotel distribution industry chain, distributors or channel managers need to obtain massive amounts of dynamic hotel room data (including prices, inventory, hotel information, etc.) from multiple upstream OTA platforms such as Ctrip, Tongcheng, and Qunar, and then map this data to their own standard hotel master data (MasterHotelData). This mapping process is the core data foundation supporting multi-channel price synchronization, unified inventory management, and accurate order distribution. Traditionally, this process relies on a combination of automated rule matching and manual review.

[0003] Currently, the technical solutions commonly used in the industry can be mainly divided into the following categories, but all of them have significant limitations:

[0004] For example, structured field matching based on rule engines is the most widely used existing technology. Its core is to compare structured fields such as hotel names, phone numbers, latitude and longitude coordinates, and address text by writing predefined rules. This method uses string similarity algorithms (such as edit distance and Jaccard similarity) to match hotel names; geofencing algorithms (such as determining whether the distance between two points is within a threshold) to match location information; and performs precise or fuzzy comparisons of phone numbers and addresses.

[0005] However, this existing method has the following drawbacks: First, naming ambiguity and inconsistency exist. The same hotel can have vastly different names on different platforms (e.g., All Seasons Hotel (Shanghai Nanjing East Road Branch) vs. All Seasons·Nanjing East Road Pedestrian Street Branch), making it difficult to accurately judge based on simple string similarity. Second, unstructured information is underutilized. The rule engine cannot effectively understand and utilize unstructured text such as proximity to landmarks in addresses, brand change history in hotel descriptions (e.g., the former XX Hotel was upgraded to YY), and semantic information in user reviews. Third, there is low tolerance for missing or incorrect fields. When key fields (such as phone numbers or precise coordinates) are empty or contain significant errors, rule matching immediately fails, leading to a large number of records being transferred to manual processing.

[0006] In addition, the following drawbacks exist: matching results typically only provide a score, lacking human-understandable reasoning for the judgment, which hinders manual review and experience accumulation. Due to these drawbacks, approximately 20%-30% of hotels require manual intervention for mapping, with each transaction taking several minutes, resulting in low efficiency and the inability to effectively translate expert experience into system capabilities.

[0007] To overcome the semantic understanding limitations of rule engines, some research has attempted to construct knowledge graphs for the hotel industry, treating hotels as entities and associating them with various attributes and relationships, using graph algorithms for entity disambiguation and alignment. However, this approach is extremely costly to build and maintain. Constructing a knowledge graph requires a large amount of high-quality, structured seed data, and it needs continuous updates as new brands and hotels emerge and hotel information changes, resulting in a heavy operational burden. Furthermore, knowledge graphs are ineffective for emerging brands or niche hotels lacking prior knowledge. They also struggle to quickly adapt to new OTA data sources with vastly different information formats.

[0008] Recently, with the development of machine learning, solutions have emerged that use labeled hotel data to train binary classification models (such as deep neural networks) or fine-tune pre-trained large language models to determine whether two hotels are the same entity. However, model training relies on a large amount of labeled data, requiring significant manpower for data labeling, resulting in high costs and long cycles. Secondly, the fine-tuning process is complex and iterative: model fine-tuning requires a professional technical team and computing resources, and each change in business logic or data distribution may necessitate re-fine-tuning, leading to poor agility.

[0009] Existing cross-platform mapping technologies for hotels either rely on rigid rules, resulting in weak generalization capabilities and excessive human intervention; or they pursue complex models, leading to high implementation barriers, poor flexibility, and difficulty in continuously accumulating knowledge. The fundamental problem lies in the failure to transform the interpretable mapping decision-making experience accumulated by human experts in long-term practice into the core intelligence of the system in a low-cost and iterative manner. Summary of the Invention

[0010] A key advantage of this invention lies in providing a cross-platform intelligent mapping method and system for hotels based on a large language model and a dynamic case library. Addressing the core shortcomings of current cross-platform data mapping technologies for hotels, such as rigid rules, lack of semantic understanding, high implementation costs, and the inability of the system to continuously learn, this invention offers a novel intelligent solution based on general artificial intelligence and closed-loop feedback of domain knowledge. Specifically, this invention aims to achieve the following sub-objectives:

[0011] A key advantage of this invention is that it provides a cross-platform intelligent mapping method and system for hotels based on a large language model and a dynamic case library. This overcomes the limitations of traditional rule engines that rely solely on surface-level string and coordinate matching. By introducing a general large language model (LLM) with powerful natural language understanding and contextual reasoning capabilities, it achieves accurate interpretation of deep semantic information such as unstructured hotel descriptions, fuzzy addresses, and brand associations, thereby handling complex matching scenarios such as inconsistent naming and colloquial expressions.

[0012] A key advantage of this invention lies in providing a cross-platform intelligent mapping method and system for hotels based on a large language model and a dynamic case library. It abandons the traditional approach of requiring large amounts of labeled data to train dedicated models or constructing and maintaining highly costly knowledge graphs, instead creating a dynamic mapping decision case library. This case library transforms each manual review decision (regardless of whether the initial source is a rule failure or a low-confidence model result) and its reasoning logic into structured, searchable case knowledge. This enables the system to make decisions by analogy with historical experience, much like human experts, and to continuously accumulate and optimize its experience base as business operations progress.

[0013] A key advantage of this invention lies in providing a cross-platform intelligent mapping method and system for hotels based on a large language model and a dynamic case library. This changes the current situation of black-box decision-making in existing models, requiring the large language model to provide logically sound reasoning alongside the matching results. This not only makes the automated decision-making process more transparent, facilitating manual review and auditing, but more importantly, these reasonings themselves are accumulated as high-quality knowledge in the case library, forming a reinforcing cycle of decision-interpretation-learning.

[0014] A key advantage of this invention is that it provides a cross-platform intelligent mapping method and system for hotels based on a large language model and a dynamic case library. It proposes an integration method that requires no parameter fine-tuning or retraining of the basic large language model. Through carefully designed prompt engineering and dynamic case retrieval injection, it guides a general-purpose large model to complete specialized domain tasks, significantly reducing the technical threshold, computational resource consumption, and deployment cycle, enabling small and medium-sized enterprises to quickly apply cutting-edge AI capabilities.

[0015] A key advantage of this invention lies in providing a cross-platform intelligent mapping method and system for hotels based on a large language model and a dynamic case library. It designs a complete closed-loop system encompassing automatic triggering, human intervention, and feedback learning. The system can not only handle routine matching but also identify its own uncertain states (low confidence) and proactively seek human assistance. It then absorbs the authoritative human judgment, enabling continuous self-optimization and self-adaptation in real business flows, ultimately significantly reducing reliance on continuous manual labor.

[0016] Another advantage of this invention is that it provides a cross-platform intelligent mapping method for hotels based on a large language model and a dynamic case library, characterized by the following steps:

[0017] (1) Perform multi-dimensional feature extraction on the hotel data to be mapped from external platforms and the standard hotel data in the standard hotel database;

[0018] (2) Construct a dynamic mapping decision case library based on historical mapping records. The cases in the case library shall at least include input features, judgment results and judgment basis.

[0019] (3) For the current hotel pair to be determined, retrieve similar historical cases from the dynamic mapping decision case library, and construct prompts to guide the large language model to perform analogical reasoning based on the retrieval results;

[0020] (4) Input the prompt information into a general large language model that has not been specifically trained for this mapping task, and obtain its output structured reasoning result containing the determination of whether they are the same entity;

[0021] (5) Perform mapping decision based on the reasoning result, and feed back the new judgment result and its basis generated in the decision-making process to the dynamic mapping decision case library.

[0022] According to one embodiment of this application, the multi-dimensional features in step (1) include basic fields, geographic information, and unstructured text;

[0023] The basic fields include hotel name, telephone number, star rating, and affiliated group or brand;

[0024] The geographic information includes latitude and longitude coordinates, administrative divisions, and landmark keywords;

[0025] The unstructured text includes hotel introductions, descriptions of the surrounding environment, or user reviews.

[0026] According to one embodiment of this application, in step (2), each case in the dynamic mapping decision case library also includes metadata tags for assisting retrieval and management, the metadata tags including city, brand or error type.

[0027] According to one embodiment of this application, in step (3), the method of retrieving similar historical cases is any one or a combination of the following:

[0028] (a) Weighted comprehensive similarity retrieval based on hotel name similarity and geographical distance;

[0029] (b) Semantic similarity retrieval based on text vector embedding model;

[0030] (c) Hybrid rule retrieval based on brand, city and geographical distance.

[0031] According to one embodiment of this application, the top K historical cases are selected as the similar historical cases in descending order of similarity, where K is a preset positive integer.

[0032] According to one embodiment of this application, K is 3.

[0033] According to one embodiment of this application, in step (3), when constructing the prompt information, it is possible to select to inject only positive historical cases where the judgment result is the same entity.

[0034] According to one embodiment of this application, in step (4), the general large language model is a cloud model called via API or an open-source model deployed locally.

[0035] According to one embodiment of this application, in step (4), the structured reasoning result further includes confidence level and human-readable reasoning.

[0036] According to one embodiment of this application, step (5) specifically includes: comparing the confidence level with a preset threshold; if the confidence level is not lower than the threshold, then automatic mapping is performed; otherwise, the process is transferred to manual review.

[0037] According to one embodiment of this application, the preset threshold is a configurable parameter that can be dynamically adjusted based on the city where the hotel is located or the brand it belongs to.

[0038] According to one embodiment of this application, the preset threshold is set to a first higher value in the early stage of the launch of a new business area, and is automatically adjusted to a second lower value after a certain number of local cases are accumulated.

[0039] According to one embodiment of this application, a candidate pair generation step is included before step (3): for hotels that do not pass the preliminary rule matching, a candidate pair of <standard hotel, hotel to be mapped> is generated based on geographical proximity, brand or star rating consistency.

[0040] According to one embodiment of this application, the dynamic mapping decision case library is initialized in the initial stage of the system by importing historical high-quality manual mapping records.

[0041] According to one embodiment of this application, the number of seed cases imported during initialization is no less than 2,000, with priority given to high-frequency brands and core cities.

[0042] According to one embodiment of this application, in step (5), the new cases fed back to the case library are subject to quality control, including marking conflicting cases, deduplicating duplicate cases, or setting different search weights based on the validity of the cases.

[0043] According to another aspect of this application, this application further provides a hotel cross-platform intelligent mapping system, characterized in that it includes:

[0044] The feature extraction module is used to extract multi-dimensional features from hotel data;

[0045] The case library management module is used to build and maintain the dynamic mapping decision case library;

[0046] The case retrieval and suggestion construction module is used to retrieve similar historical cases and construct suggestion information;

[0047] The large language model reasoning module is used to call a general large language model and obtain structured reasoning results;

[0048] The decision and feedback control module is used to execute mapping decisions and manage feedback loops.

[0049] According to one embodiment of this application, the feature extraction module extracts feature dimensions including basic fields, geographic information, and unstructured text;

[0050] The basic fields include hotel name, telephone number, star rating, and affiliated group or brand;

[0051] The geographic information includes latitude and longitude coordinates, administrative divisions, and landmark keywords;

[0052] The unstructured text includes hotel introductions, descriptions of the surrounding environment, or user reviews.

[0053] According to one embodiment of this application, each case in the dynamic mapping decision case library also includes metadata tags for assisting retrieval and management, the metadata tags including city, brand or error type.

[0054] According to one embodiment of this application, the case retrieval and suggestion construction module uses one or a combination of the following methods to retrieve similar historical cases:

[0055] (a) Weighted comprehensive similarity retrieval based on hotel name similarity and geographical distance;

[0056] (b) Semantic similarity retrieval based on text vector embedding model;

[0057] (c) Hybrid rule retrieval based on brand, city and geographical distance.

[0058] According to one embodiment of this application, a rule pre-screening module is also included, which is used to perform fast matching based on rule combinations of name, telephone number and coordinates before the feature extraction module works, so as to divert high-confidence matching results.

[0059] According to one embodiment of this application, the case library management module further includes a quality control submodule, which is used to perform automatic case deduplication, quality grading, or lifecycle management.

[0060] According to one embodiment of this application, a cold start initialization module is also included for loading a seed case library during the system deployment phase.

[0061] According to one embodiment of this application, a data security and compliance module is also included for encrypting sensitive hotel information and implementing role-based access control.

[0062] According to another aspect of the invention, the invention further provides

[0063] The further objects and advantages of the invention will become fully apparent from the following description and accompanying drawings.

[0064] These and other objects, features and advantages of the present invention will become fully apparent from the following detailed description and accompanying drawings. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating a method for a hotel cross-platform intelligent mapping method based on a large language model and a dynamic case library according to a first preferred embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram of the system framework of a hotel cross-platform intelligent mapping system based on a large language model and dynamic case library according to a first preferred embodiment of the present invention.

[0067] Figure 3 This is a full-process multi-source OTA hotel data mapping method based on a large language model and dynamic case library according to a first preferred embodiment of the present invention. Detailed Implementation

[0068] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0069] Those skilled in the art should understand that, in the disclosure of this invention, the terms longitudinal, transverse, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention.

[0070] It is understood that the term "one" should be understood as at least one or more, that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple. The term "one" should not be understood as a limitation on the number.

[0071] Refer to the accompanying drawings in this application specification. Figures 1 to 3 As shown, a hotel cross-platform intelligent mapping method and system based on a large language model and dynamic case library, according to a first preferred embodiment of this application, will be described below. The hotel cross-platform intelligent mapping method based on a large language model and dynamic case library includes the following steps:

[0072] (1) Perform multi-dimensional feature extraction on the hotel data to be mapped from external platforms and the standard hotel data in the standard hotel database;

[0073] (2) Construct a dynamic mapping decision case library based on historical mapping records. The cases in the case library shall at least include input features, judgment results and judgment basis.

[0074] (3) For the current hotel pair to be determined, retrieve similar historical cases from the dynamic mapping decision case library, and construct prompts to guide the large language model to perform analogical reasoning based on the retrieval results;

[0075] (4) Input the prompt information into a general large language model that has not been specifically trained for this mapping task, and obtain its output structured reasoning result containing the determination of whether they are the same entity;

[0076] (5) Perform mapping decision based on the reasoning result, and feed back the new judgment result and its basis generated in the decision-making process to the dynamic mapping decision case library.

[0077] This method first performs deep fusion and extraction of structured and unstructured information from hotel data from standard libraries and any OTA platform. This is not just about simply extracting fields, but about building a unified, machine-understandable digital profile of the hotel.

[0078] In step (1) of this application, multi-dimensional hotel features are extracted, namely, name, telephone number, star rating, and group / brand. This layer processes the consistency representation of the data, such as normalizing the brand core of All Seasons Hotel (Shanghai Nanjing East Road Store) and All Seasons·Nanjing East Road Pedestrian Street Store.

[0079] Multi-dimensional hotel feature extraction not only includes precise latitude and longitude coordinates, but also uses natural language processing (NLP) technology to parse administrative divisions (such as Huangpu District of Shanghai) and landmark keywords (such as the Bund and People's Square subway stations) from address and description text. This constructs a multi-scale spatial representation from precise points to fuzzy regions.

[0080] Multi-dimensional hotel feature extraction utilizes text analysis technology to extract key semantic information from the hotel's official introduction, descriptions of the surrounding environment, and high-frequency words in aggregated user reviews. For example, based on the information that the hotel is located at the western end of Nanjing East Road Pedestrian Street and is about a 10-minute walk from the Bund, a semantic triple can be extracted, consisting of the attributes: western end of pedestrian street, relationship: proximity-Bund, and measure: approximately 10 minutes.

[0081] Understandably, traditional methods can only process name strings and coordinate numbers, while this invention, through deep extraction of unstructured text layers, successfully captured key relative positional relationships and semantic descriptions such as "west end" and "approximately 10 minutes walk to the Bund." This gives the system additional, human-language-based judgment criteria when faced with coordinate offsets or colloquial addresses, fundamentally compensating for the lack of information in purely structured data.

[0082] By introducing landmark keywords and semantic triples, even if the hotel phone number is missing or the coordinates have an error of hundreds of meters, the system can still make inferences by judging whether the same landmark is mentioned or whether the spatial relationship is consistent, which significantly improves the robustness of feature expression and tolerance to noisy data.

[0083] Step (2) of the above method in this application is to construct a dynamic mapping decision case library. When constructing and initializing (cold start), the system does not need to be trained from scratch. Instead, the historically accumulated and verified high-quality manual mapping records are formatted through the above structure and imported as seed cases. For example, a seed case may record why the auditor determined that Hanting (Lujiazui Store) and Hanting Youjia (Pudong Lujiazui Store) are different hotels (reason: different brand lines, and the address distance is more than 1.5 kilometers).

[0084] This invention requires absolutely no parameter fine-tuning of the basic large language model. Importing the seed case library is equivalent to directly encapsulating and injecting years of business expert experience into the system. This avoids the cost and time involved in collecting massive amounts of labeled data and performing expensive GPU training, achieving a very low initial technical barrier and rapid deployment.

[0085] The natural language reasoning underlying each case transforms knowledge from a black box hidden within model weights into a plaintext, readable, auditable, and manageable asset. Metadata tags enable efficient classification and retrieval of this knowledge (e.g., quickly finding all matching cases related to the Shanghai-All Seasons brand due to name variations), laying the foundation for subsequent intelligent retrieval.

[0086] Step (3) of the above method is the construction of similar case retrieval and prompt, which is a key bridge to applying historical experience to the current problem.

[0087] It is worth mentioning that in this preferred embodiment of the application, a multi-level hybrid retrieval strategy is adopted. First, metadata tags (such as city and brand) are used for rapid filtering to narrow the search scope. Second, within the candidate set, semantic vector similarity calculation is used as the core. Specifically, models such as Sentence-BERT are used to encode the feature text of the current hotel pair to be judged (such as a combination of name, address, and description) and the input feature text of cases in the case library into high-dimensional vectors, and the semantic similarity is measured by calculating cosine similarity. Finally, the Top-K (e.g., K=3) cases with the highest similarity are returned.

[0088] Design a structured prompt template that uses the complete information (input, output, and reasoning basis) of the K retrieved cases as a few-shot examples. This information, along with the detailed features of the hotel pair to be judged, is organized into an instruction for the larger model. For example: Please refer to the following expert judgment logic to analyze whether the new pair of hotels is the same. Case 1: ... Case 2: ... Case 3: ... Now please judge: Hotel X (feature) vs. Hotel Y (feature).

[0089] Understandably, through semantic vector retrieval, the system can accurately find historical decisions that are semantically most similar to the current context, rather than simply matching literal meanings. This enables large models to perform analogical reasoning: that is, if experts made this judgment in similar situations historically, then I should follow similar logic now. This approach gives the system powerful generalization capabilities, enabling it to handle novel naming conventions or new OTA platform data formats never seen in the training data.

[0090] The Few-shotPrompt constructed in this application serves as a thought demonstration and contextual constraint. It clearly informs the large model of the task's objectives and format, and provides high-quality reasoning examples, greatly guiding the model's thinking direction and focusing its output on business logic. This effectively reduces the illusions or irrelevant outputs that may arise from general large models, improving the accuracy and reliability of decision-making.

[0091] Step (4) of the above method involves calling a general-purpose large language model for inference. Specifically, the constructed prompt text is sent to a powerful general-purpose large language model service interface (such as the APIs of GPT-4, Qwen-Max, and GLM-4) that has not been specifically fine-tuned for this hotel mapping task. The model is required to strictly follow the example format and output a structured JSON object, which should contain at least: is_same (boolean value), confidence (confidence score, 0-1), and reasoning (detailed natural language inference process).

[0092] By directly leveraging the out-of-the-box capabilities of a general-purpose large language model, this system integrates the most advanced natural language understanding and generation technologies currently available. The model can deeply understand complex semantics such as "upgraded to," "renamed," "adjacent," and "original location," and can generate clear and logical reasoning chains that align with human expert thinking. This solves the pain point of traditional rule engines and simple machine learning models being unable to handle complex semantics.

[0093] The key to achieving trustworthy AI in this invention lies in requiring the model to output natural language reasoning. This not only enables operations personnel to understand and trust automated decision-making, but more importantly, when disputes arise, the traceable reasoning process provides direct evidence for manual review, meeting the requirements of business compliance and quality control.

[0094] Step (5) of the above method is a mapping decision and feedback loop. Specifically, dynamic threshold decision: the system receives the confidence output by the model. Instead of using a fixed threshold, a dynamic threshold strategy is adopted. For example, for first-tier cities with high data quality and abundant cases, the threshold can be set to 0.82; for newly developed cities or brands with sparse data, the threshold can be temporarily set to 0.90 to control risks. If the confidence level is greater than or equal to the threshold, the mapping is completed automatically; otherwise, it flows into the manual review queue.

[0095] The closed-loop feedback and self-evolving case library ensure that regardless of whether the final decision is made automatically or manually, all information from this task (input features, final output, and complete reasoning) will automatically generate a new case. Before being added to the library, quality control measures are implemented: such as deduplication comparison with existing cases; if a new case conflicts with an old case, it may trigger a downgrade or flagging of the old case's weight. Finally, the new case is stored in the dynamic case library.

[0096] The feedback loop is a revolutionary step in this method. Each mapping task is a learning opportunity for the system. Successful experiences are consolidated, and lessons learned from mistakes are corrected. This allows the case library to evolve from a static knowledge base into a dynamically growing organic body of experience. The overall judgment capability of the system shows a continuously improving curve as business operations progress, with the ultimate goal of infinitely approaching or even surpassing the level of expert teams.

[0097] The dynamic threshold mechanism achieves a fine balance between business risk and automation efficiency. The system can automatically identify scenarios with and without certainty of success, accurately guiding human resources to the most needed and complex issues. At the same time, quality control of the case library ensures the health of the knowledge base, preventing low-quality or outdated cases from interfering with judgment, and guaranteeing the long-term stability and efficiency of the system.

[0098] Furthermore, the multi-dimensional features mentioned in step (1) include basic fields, geographic information, and unstructured text;

[0099] The basic fields include hotel name, telephone number, star rating, and affiliated group or brand;

[0100] The geographic information includes latitude and longitude coordinates, administrative divisions, and landmark keywords;

[0101] The unstructured text includes hotel introductions, descriptions of the surrounding environment, or user reviews.

[0102] The system uses a data parser to extract key identifiers such as hotel name, phone number, star rating, and affiliated group or brand from the raw data. Among these, brand identification is particularly important. The system has a built-in brand dictionary and uses fuzzy matching to normalize and extract core brand words from the names (e.g., extracting "All Seasons" from "All Seasons Hotel (Beijing Guomao Branch)"). This lays the foundation for subsequent rapid filtering and reasoning based on the brand.

[0103] Beyond simple latitude and longitude coordinates, the system invokes geocoding services to convert text addresses into standardized administrative divisions (province, city, district, street). Simultaneously, it utilizes named entity recognition technology to extract landmark keywords (such as Shanghai Railway Station and Tianhe City Shopping Center) from the address and description text, constructing a network linking the hotel to well-known surrounding locations.

[0104] The system preprocesses aggregated text from hotel introductions, surrounding area descriptions, and user reviews (removing stop words and segmenting words), and extracts key phrases and representative sentences. For example, from reviews indicating convenient travel and proximity to subway stations, it extracts "convenient transportation" and "close to subway station" as feature words. Essentially, this transforms subjective human descriptions into machine-processable semantic tags.

[0105] By extracting brand data through normalization and standardizing administrative divisions, raw data from different sources and with varying expressions are unified under the same semantic framework, solving the problem of data heterogeneity and creating a prerequisite for accurate comparison.

[0106] The extraction of landmark keywords and unstructured text features constructs a rich semantic context network for each hotel. When coordinates are inaccurate or addresses are ambiguous, the system can perform strong logical reasoning by judging the overlap of the semantic networks of two hotels (e.g., whether they are both associated with the Bund and Nanjing East Road subway stations), greatly enhancing the matching ability in cases of incomplete information.

[0107] Furthermore, in step (2), each case in the dynamic mapping decision case library also includes metadata tags for assisting retrieval and management, and the metadata tags include city, brand or error type.

[0108] When cases are added to the database, the system automatically extracts these tags from the case content or generates them according to rules. For example, the city is automatically labeled based on the hotel address: Shanghai; the brand is labeled based on the brand field: Marriott; and the error type is labeled by algorithm or manual annotation based on the judgment result and reasoning: name abbreviation confusion. These tags are indexed along with the case content.

[0109] When performing case retrieval, efficient coarse-grained filtering can be performed first using metadata tags. For example, when processing a hotel pair between Shanghai and All Seasons Hotels, the system can prioritize fine-grained vector similarity calculations within the subset of cases where the city is Shanghai and the brand is All Seasons, thereby significantly improving retrieval efficiency and reducing unnecessary global calculations.

[0110] Metadata tags enable the analysis of the case library. Operations personnel can quickly identify which city has the most cases, which brand has the highest error rate, and what percentage of errors are "coordinate drift," thereby pinpointing business bottlenecks and providing data insights for optimizing data source quality or adjusting strategies.

[0111] Furthermore, in step (3), the method for retrieving similar historical cases is any one or a combination of the following:

[0112] (a) Weighted comprehensive similarity retrieval based on hotel name similarity and geographical distance;

[0113] (b) Semantic similarity retrieval based on text vector embedding model;

[0114] (c) Hybrid rule retrieval based on brand, city and geographical distance.

[0115] The system uses preset weights (e.g., name similarity weight 0.6, inverse geographical distance weight 0.4). Name similarity is calculated using an optimized edit distance algorithm, and geographical distance is calculated using spherical distance. This method is fast and suitable for scenarios requiring moderate accuracy and rapid response.

[0116] Using a pre-trained Sentence-BERT model, the feature text of hotel pairs (such as name, address, and core description) is encoded into 768-dimensional vectors. The input features of all cases in the case library are also pre-encoded as vectors. During retrieval, the cosine similarity between the current hotel pair vector and the case vector library is calculated. This method provides a deep understanding of the semantic similarity between hotels near train stations and those within a 5-minute walk of high-speed rail stations, achieving the highest accuracy.

[0117] This is a fast funnel based on business rules. For example, the rule could be: same brand AND same city AND distance < 2km. It can quickly filter out the most relevant candidate case pool and is often used as a pre-step for vector retrieval.

[0118] The system allows for the configuration of retrieval strategies based on different scenarios (e.g., online services use the (c)->(b) combination, while offline batch tasks use the (a) method), achieving a balance between flexibility and performance optimization.

[0119] Hybrid rule retrieval ensures strong relevance of search results across core business dimensions (brand, city, distance), avoiding semantically relevant but business-irrelevant interference cases that may occur with pure vector retrieval, and ensuring that the examples provided to the large model are always in the correct business context.

[0120] Furthermore, the top K historical cases are selected from high to low similarity as the similar historical cases, where K is a preset positive integer. Preferably, K is 3.

[0121] After performing the retrieval, the system sorts candidate cases by similarity score and selects the top K cases. Returning a fixed number of Top-K cases makes the number of input tokens for each large model call relatively controllable, thus predicting and managing API call costs. Simultaneously, the K=3 design provides the model with a diverse range of examples sufficient for pattern recognition and analogical reasoning within a limited context window, achieving an optimal balance between performance and cost.

[0122] Furthermore, in step (3), when constructing the prompt information, it is possible to select only injecting positive historical cases where the judgment result is the same entity. After retrieving the Top-K cases, the system adds a filtering step, retaining only the cases whose output is "yes" (i.e., judged as the same entity) for constructing the Few-shot example. This approach is based on the assumption that for matching tasks, learning what constitutes a match is more direct and efficient than learning what constitutes a mismatch.

[0123] Providing only positive examples is equivalent to giving the large model a clear set of successful matching patterns, guiding it to focus on learning and reproducing these successful logics. This can avoid interference from the complex and varied reasons for mismatches in negative examples, and can sometimes improve the model's accuracy and confidence in judging positive examples. It is an effective cueing engineering technique for specific tasks.

[0124] Furthermore, in step (4), the general-purpose large language model is either a cloud-based model accessed via API or an open-source model deployed locally. In terms of system architecture, the large model access module is abstracted into a unified interface. The service endpoint can be specified in the configuration file: if a cloud API (such as OpenAIGPT-4) is selected, it is accessed via the network; if a local model (such as Qwen-72B-Chat) is selected, it is accessed via intranet RPC. This achieves pluggability of the inference engine.

[0125] Offering two options means this approach is highly adaptable. Enterprises seeking optimal performance and maintenance-free operation can choose the cloud API; those with stringent data security requirements or who wish to control long-term costs can choose to deploy the local open-source model in their own data center. This makes the invention suitable for diverse customer groups, from large OTAs to small and medium-sized distributors.

[0126] Furthermore, in step (4), the structured reasoning result also includes confidence level and human-readable reasoning.

[0127] Further, step (5) specifically includes: comparing the confidence level with a preset threshold; if the confidence level is not lower than the threshold, then automatic mapping is performed; otherwise, the process is transferred to manual review.

[0128] Understandably, the confidence mechanism is key to achieving seamless integration between fully automated and semi-automated processes in this method. It delegates deterministic tasks to machines and uncertain tasks to humans, forming a highly efficient collaborative model where AI handles the majority and humans handle the exceptions, significantly improving the throughput and reliability of the overall business process.

[0129] Furthermore, the preset threshold is a configurable parameter and can be dynamically adjusted according to the city where the hotel is located or the brand it belongs to.

[0130] Furthermore, the preset threshold is set to a first higher value in the initial stage of launching a new business area, and is automatically adjusted to a second lower value after accumulating a certain number of local cases.

[0131] The decision threshold T is not globally fixed but stored in the configuration center and can be bound to dimensions such as city and brand. For example: Shanghai / All Seasons: 0.82, Lhasa / *: 0.90. For the new city of Lhasa, the initial threshold is set to a relatively high 0.90 (the first relatively high value) to ensure that only cases with extremely high probability are automatically mapped, thus controlling risk. As the system processes more data in Lhasa, when the number of local cases exceeds 50, the threshold is automatically lowered to 0.85 (the second relatively low value) to improve the automation rate.

[0132] The dynamic threshold mechanism endows the system with risk awareness and learning capabilities. A conservative strategy is adopted in new, inexperienced regions, gradually shifting to a more proactive strategy as local experience (case studies) accumulates. This ensures stability and a smooth transition during business expansion, avoiding the problems of high initial misjudgment rates or insufficient automation later on that might result from applying thresholds across the board.

[0133] Furthermore, before step (3), a candidate pair generation step is also included: for hotels that fail to pass the initial rule matching, a candidate pair <standard hotel, hotel to be mapped> is generated based on geographical proximity, brand or star rating consistency.

[0134] Furthermore, the dynamic mapping decision case library is initialized in the initial stage of the system by importing historical high-quality manual mapping records.

[0135] Furthermore, the number of seed cases imported during the initial setup should be no less than 2,000, with priority given to high-frequency brands and core cities.

[0136] During the initial system deployment, export mapping records from the business database for a past period (e.g., 6 months). Filter high-quality records using rules (e.g., completed by senior auditors, no complaints about changes after mapping), ensuring at least 2000 records. During import, prioritize coverage of high-frequency brands (e.g., brands under Huazhu and Jinjiang) and core cities (e.g., Beijing, Shanghai, Guangzhou, and Shenzhen) based on brand popularity (frequency of appearance) and the business importance of the city.

[0137] By importing no fewer than 2,000 high-quality seed cases, the system was capable of handling mainstream scenarios from day one, achieving high availability during a cold start. Prioritizing coverage of high-frequency scenarios ensured that the system could immediately resolve 80% of common problems, allowing customers to experience significant efficiency improvements from the initial deployment stage and accelerating return on investment.

[0138] Furthermore, in step (5), quality control is performed on the new cases fed back to the case library, including marking conflicting cases, deduplicating duplicate cases, or setting different search weights based on the validity of the cases.

[0139] When the judgment result of a new returned case conflicts with a highly similar old case in the case library, the system automatically marks both cases as conflicting and notifies the administrator for review to prevent contamination by erroneous knowledge.

[0140] Calculate the feature similarity between the new case and the existing case. If it exceeds an extremely high threshold (such as 98%), it is considered a duplicate. Only the metadata of the original case (such as the last time it was used) is updated, and no redundant records are added.

[0141] Each case is assigned a dynamic weight. Cases that are retrieved multiple times and ultimately determined to be correct have their weight increased; cases marked as conflicting or subsequently overturned manually have their weight decreased. During retrieval, the similarity score is combined with the case weights for ranking.

[0142] The quality control mechanism acts as the immune system of the case library as an organism. It automatically identifies and isolates erroneous knowledge, eliminates information redundancy, and dynamically evaluates the value of knowledge based on practical experience. This ensures that the knowledge quality and retrieval effectiveness of the case library are maintained or even improved during its continuous expansion, guaranteeing the system's long-term stable and reliable evolutionary capabilities.

[0143] According to another aspect of this application, this application further provides a hotel cross-platform intelligent mapping system for implementing the above-described method, comprising: a feature extraction module 10, a case library management module 20, a case retrieval and prompt construction module 30, a large language model reasoning module 40, and a decision and feedback control module 50; wherein the feature extraction module 10 is used to perform multi-dimensional feature extraction on the hotel data to be mapped from an external platform and the standard hotel data in a standard hotel library; the case library management module 20 constructs a dynamic mapping decision case library based on historical mapping records, wherein the cases in the case library at least include input features, judgment results, and... The judgment criteria are as follows: The case retrieval and prompt construction module 30 retrieves similar historical cases from the dynamic mapping decision case library for the current hotel pair to be judged, and constructs prompt information to guide the large language model to perform analogical reasoning based on the retrieval results; The large language model reasoning module 40 inputs the prompt information into a general large language model that has not been specifically trained for this mapping task, and obtains its output structured reasoning results containing the judgment of whether they are the same entity; The decision and feedback control module 50 executes the mapping decision based on the reasoning results, and feeds back the new judgment results and their basis generated during the decision-making process to the dynamic mapping decision case library.

[0144] Specifically, the hotel cross-platform intelligent mapping system described in this invention adopts a layered and decoupled microservice architecture design. Its core consists of five modules with clearly defined functions that communicate through standardized data interfaces, working together to complete a complete closed loop from data input to intelligent decision-making and knowledge updates.

[0145] Raw hotel data first flows into the feature extraction module, where it is converted into standardized multi-dimensional feature objects. These feature objects are then passed to the case retrieval and suggestion construction module. This module interacts with the case library management module to retrieve similar historical cases. Suggestions constructed based on the retrieval results are sent to the large language model inference module for intelligent reasoning. The inference results are returned to the decision and feedback control module, which executes the final mapping decision (automatically or manually) and drives the newly generated decision knowledge to flow back to the case library management module.

[0146] The modular design allows each core function to be deployed, maintained, and upgraded independently. For example, upgrading the feature extraction algorithm or changing the supplier of the large language model only requires adjusting the corresponding module, without affecting the overall stable operation of the system, which greatly improves the maintainability and technical agility of the system.

[0147] The feature extraction module extracts features including basic fields, geographic information, and unstructured text.

[0148] The basic fields include hotel name, telephone number, star rating, and affiliated group or brand;

[0149] The geographic information includes latitude and longitude coordinates, administrative divisions, and landmark keywords;

[0150] The unstructured text includes hotel introductions, descriptions of the surrounding environment, or user reviews.

[0151] The feature extraction module 10 extracts structured fields from the raw data through regular expressions and dictionary matching. The core is the brand normalization submodule, which has a built-in brand alias library and fuzzy matching algorithm, enabling it to uniformly map All Seasons Hotel, QJHotel, and All Seasons Hotel to the standard brand code HUAZHU-QJ.

[0152] It integrates third-party geocoding services (such as Amap / Google Maps API) to convert text addresses into latitude and longitude coordinates and standard administrative division codes. Simultaneously, it runs a landmark recognition submodule that uses named entity recognition technology to extract landmarks such as Hongqiao Airport and West Lake Scenic Area from the text and associates them with a standard landmark database.

[0153] After cleaning, segmenting, and removing stop words from the long text, key phrases and representative sentences are extracted using TF-IDF or TextRank algorithms to form a semantic feature summary of the hotel. Through brand normalization and geocoding, this module unifies the messy data source into a clean and standardized internal representation, which is the foundation for all subsequent accurate processing. Landmark recognition and semantic summary generation are equivalent to adding machine-understandable tags and descriptions to the hotel, greatly enriching the comparative dimensions.

[0154] It is worth mentioning that each case in the dynamic mapping decision case library also includes metadata tags for assisting retrieval and management, including city, brand, or error type.

[0155] The case retrieval and suggestion construction module uses one or a combination of the following methods to retrieve similar historical cases:

[0156] (a) Weighted comprehensive similarity retrieval based on hotel name similarity and geographical distance;

[0157] (b) Semantic similarity retrieval based on text vector embedding model;

[0158] (c) Hybrid rule retrieval based on brand, city and geographical distance.

[0159] The hotel cross-platform intelligent mapping system of this preferred embodiment further includes a rule pre-screening module 60, which is used to perform fast matching based on rule combinations of name, telephone number and coordinates before the feature extraction module 10 works, so as to divert high confidence matching results.

[0160] Understandably, the rule pre-screening module 60 loads a series of strong feature rules, such as identical names and phone numbers, latitude and longitude distances less than 50 meters, and name edit distances less than 2. For hotel pairs that trigger any of these rules, a high-confidence matching result is directly output, and all subsequent complex processing flows are short-circuited.

[0161] This module can quickly resolve approximately 30%-40% of simple and straightforward matching tasks with extremely low computational cost. This allows valuable computing resources (such as large model API calls) to be focused on tackling the remaining truly complex and ambiguous cases, significantly reducing overall system processing costs and increasing throughput.

[0162] The case library management module 20 also includes a quality control submodule 21, which is used to perform automatic deduplication, quality grading, or lifecycle management of cases.

[0163] It is worth mentioning that the case library management module 20 uses a database that supports vector retrieval, such as Elasticsearch or Milvus. In addition to storing the complete content of the cases, it also specifically builds inverted indexes and high-dimensional vector indexes for metadata tags (city, brand, etc.) and feature vectors to achieve millisecond-level retrieval.

[0164] Regularly calculate the feature similarity between cases, merge cases with a similarity >95%, and retain the one with higher quality (e.g., manually reviewed). Dynamically calculate case quality scores based on case source (AI / human), usage frequency, and subsequent verification results (whether refuted), classifying them into core cases, ordinary cases, and cases awaiting verification. Automatically archive cold cases that have not been retrieved or verified for a long time (e.g., 2 years) to secondary storage to maintain the efficiency and simplicity of the online index.

[0165] Understandably, proprietary vector indexes enable semantic similarity-based retrieval, which is the technological cornerstone of intelligent retrieval. The quality control submodule ensures that the case library is not a garbage dump that only grows and never diminishes, but a living knowledge base that continuously cleans and optimizes itself, with its knowledge density and retrieval accuracy improving over time.

[0166] The case retrieval and suggestion construction module 30 implements multiple retrieval algorithms. For example, it first uses a hybrid rule retrieval (brand plus city) to quickly locate a candidate pool of tens of thousands, and then uses semantic vector retrieval to perform a fine sorting of billions of candidates, balancing efficiency and accuracy.

[0167] It includes a built-in configurable template library. Templates define the arrangement format of Few-shot examples, system instructions, and output requirements. The engine automatically populates the selected template with the features of the retrieved Top-K cases, the judgment results, the reasoning basis, and the features of the hotel pair to be judged, generating the final Prompt text sent to the large model.

[0168] The hybrid search engine is a perfect combination of business logic and technical performance. The hint engineering engine transforms search results into teaching materials that large models can understand. Its templated design allows non-algorithm experts to adjust the system's thinking style or cope with new task variations by modifying templates, providing valuable business configurability.

[0169] The large language model inference module 40 defines a standard request / response format and internally encapsulates the differences in calling different large model APIs (such as OpenAI, Alibaba Cloud, and Zhipu AI) or local model services. Model providers can be switched through configuration.

[0170] Receive the raw text returned by the large model and force it to be parsed into a preset JSON structure (such as containing fields like is_same, confidence, and reasoning). If parsing fails or a field is missing, trigger a retry or fallback process.

[0171] A unified adaptation interface decouples the system's core business logic from specific large-scale model technologies, avoiding vendor lock-in risks. Result parsing and verification enhance the system's robustness, ensuring that downstream modules always receive structured and processable data.

[0172] The decision and feedback control module 50 reads the confidence level output by the large model, queries the configuration center to obtain the dynamic threshold for the current scenario, and makes a decision on automatic mapping or manual transfer. The result of automatic mapping is directly written to the business database; for cases transferred to manual transfer, a work order is generated and pushed to the approval platform.

[0173] Monitor all decision flows (including the final results of manual review). Once a final decision is made, trigger the feedback process: package the complete information of this task into a new case object, call the API of the case library management module to complete the entry into the library, and may trigger the quality control process.

[0174] This module is crucial for the implementation of business intelligence. The decision engine transforms the uncertainty of AI into clear business actions. The feedback controller ensures that every business interaction, regardless of success or failure, is transformed into experience for system growth, making it the core implementer of the concept of "operations as training."

[0175] The hotel cross-platform intelligent mapping system also includes a cold start initialization module 70, used to load a seed case library during the system deployment phase. Notably, the cold start initialization module 70 connects to the enterprise's historical database, executes SQL queries or reads files to obtain historical mapping records. It filters and samples these records according to preset high-quality rules (such as auditor level and no subsequent complaints) to ensure a specified number (e.g., 2000) of seed cases covering core business operations are obtained. Then, the import interface of the case library management module is called in batches.

[0176] This module transforms the lengthy learning process from scratch into an instant infusion of experience. It ensures that the system possesses the mature capability to handle mainstream business scenarios from day one of deployment, greatly improving customer experience and project delivery efficiency.

[0177] The hotel cross-platform intelligent mapping system also includes a data security and compliance module 80, which is used to encrypt sensitive hotel information and implement role-based access control. The data security and compliance module 80 encrypts sensitive information such as hotel phone numbers stored in the database using algorithms such as AES at the application layer. TLS encrypted channels are used when transmitting sensitive data between modules.

[0178] It integrates with the enterprise's unified identity authentication system to implement role-based access control. For example, ordinary operations personnel can only view and approve cases; administrators can configure thresholds and manage the case library; auditors can view all operation logs.

[0179] This module is a necessary prerequisite for the system to be implemented in sensitive industries such as finance and tourism. It ensures the protection of personal privacy data and meets the requirements of regulations such as the Information Security Compliance Standard and GDPR for data access control and security auditing, thus removing compliance obstacles to the commercialization of the system.

[0180] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

Claims

1. A cross-platform intelligent mapping method for hotels based on a large language model and a dynamic case library, characterized in that, Includes the following steps: (1) Perform multi-dimensional feature extraction on the hotel data to be mapped from external platforms and the standard hotel data in the standard hotel database; (2) Construct a dynamic mapping decision case library based on historical mapping records. The cases in the case library shall at least include input features, judgment results and judgment basis. (3) For the current hotel pair to be determined, retrieve similar historical cases from the dynamic mapping decision case library, and construct prompts to guide the large language model to perform analogical reasoning based on the retrieval results; (4) Input the prompt information into a general large language model that has not been specifically trained for this mapping task, and obtain its output structured reasoning result containing the determination of whether they are the same entity; (5) Perform mapping decision based on the reasoning result, and feed back the new judgment result and its basis generated in the decision-making process to the dynamic mapping decision case library.

2. The method according to claim 1, characterized in that, The multi-dimensional features mentioned in step (1) include basic fields, geographic information, and unstructured text; The basic fields include hotel name, telephone number, star rating, and affiliated group or brand; The geographic information includes latitude and longitude coordinates, administrative divisions, and landmark keywords; The unstructured text includes hotel introductions, descriptions of the surrounding environment, or user reviews.

3. The method according to claim 1 or 2, characterized in that, In step (2), each case in the dynamic mapping decision case library also includes metadata tags for assisting retrieval and management, and the metadata tags include city, brand or error type.

4. The method according to claim 1, characterized in that, In step (3), the method for retrieving similar historical cases is any one or a combination of the following: (a) Weighted comprehensive similarity retrieval based on hotel name similarity and geographical distance; (b) Semantic similarity retrieval based on text vector embedding model; (c) Hybrid rule retrieval based on brand, city and geographical distance.

5. The method according to claim 4, characterized in that, The top K historical cases are selected from high to low similarity as the similar historical cases, where K is a preset positive integer and the value of K is 3.

6. The method according to claim 5, characterized in that, Step (5) specifically includes: comparing the confidence level with a preset threshold; if the confidence level is not lower than the threshold, automatic mapping is performed; otherwise, the process is transferred to manual review.

7. A hotel cross-platform intelligent mapping system for implementing the method of claim 1, characterized in that, include: The feature extraction module is used to extract multi-dimensional features from the hotel data to be mapped from external platforms and the standard hotel data in the standard hotel database; The case library management module constructs a dynamic mapping decision case library based on historical mapping records. The cases in the case library at least include input features, judgment results, and judgment criteria. The case retrieval and prompt construction module retrieves similar historical cases from the dynamic mapping decision case library for the current hotel pair to be judged, and constructs prompt information to guide the large language model to perform analogical reasoning based on the retrieval results; The large language model reasoning module inputs the prompt information into a general large language model that has not been specifically trained for this mapping task, and obtains its output structured reasoning result containing the determination of whether they are the same entity. The decision and feedback control module performs mapping decisions based on the reasoning results and feeds back the new judgment results and their basis generated during the decision-making process to the dynamic mapping decision case library.

8. The system according to claim 7, characterized in that, The feature extraction module extracts features including basic fields, geographic information, and unstructured text. The basic fields include hotel name, telephone number, star rating, and affiliated group or brand; The geographic information includes latitude and longitude coordinates, administrative divisions, and landmark keywords; The unstructured text includes hotel introductions, descriptions of the surrounding environment, or user reviews.

9. The system according to claim 7, characterized in that, Each case in the dynamic mapping decision case library also includes metadata tags for assisting retrieval and management, including city, brand, or error type.

10. The system according to claim 7, characterized in that, The case retrieval and suggestion module uses one or a combination of the following methods to retrieve similar historical cases: (a) Weighted comprehensive similarity retrieval based on hotel name similarity and geographical distance; (b) Semantic similarity retrieval based on text vector embedding model; (c) Hybrid rule retrieval based on brand, city and geographical distance.

Images