A semantic recognition-based travel characteristic product recommendation method

Abstract

This invention relates to the technical field of tourism specialty product recommendation, and discloses a semantic recognition-based method for recommending tourism specialty products. The method includes: using an improved semantic recognition model to perform semantic recognition on tourism demand text to obtain the user's tourism semantic feature vector; dynamically updating the product attributes of tourism specialty products using timeliness evaluation information to construct a dynamic knowledge graph of tourism specialty products; performing hierarchical semantic matching on tourism specialty products in the dynamic knowledge graph to form a candidate product set; and performing personalized filtering on tourism specialty products in the candidate product set to obtain a personalized filtered candidate product set and then recommending tourism specialty products. This invention achieves personalized tourism specialty product recommendation driven by tourism demand by semantically recognizing and matching the product semantics of tourism specialty products with the demand semantics of tourism demand text, effectively improving recommendation quality and user satisfaction.

Description

Technical Field

[0001] This invention relates to the field of tourism product recommendation, and more particularly to a semantic recognition-based method for recommending distinctive tourism products. Background Technology

[0002] With the continuous digitalization and intelligent development of the tourism industry, more and more users tend to express their travel needs through online platforms, mobile applications, and online communities, expecting the system to generate accurate, personalized, and diverse recommendations in real time. This trend not only reflects users' strong demand for convenient services but also drives the evolution of tourism recommendation systems towards intelligence and personalization. As a core application integrating information retrieval and intelligent recommendation technologies, the key to tourism recommendation systems lies in how to comprehensively and accurately identify user needs, efficiently match users' potential intentions with a dynamically updated tourism resource database, and thus generate distinctive tourism products that meet users' interests, budgets, timeliness, and diverse needs.

[0003] However, in practical applications, user travel demand texts often exhibit significant uncertainty and complexity. On the one hand, user expressions may be highly semantically ambiguous; for example, a description like "wanting a relaxing beach vacation" encompasses both scenario requirements and implicit consumption and cultural preferences. On the other hand, users' language expressions are usually not standardized, potentially using dialects, abbreviations, or colloquial descriptions, making traditional keyword-matching-based recommendation methods difficult to effectively parse. Furthermore, due to the diverse nature of user needs, such as simultaneous involvement of family travel, cultural experiences, and dining preferences, recommendation systems must be capable of complex semantic deconstruction and multi-level feature extraction. Traditional rule-driven or shallow semantic methods often fall short in handling such cases. Therefore, many current recommendation results suffer from insufficient relevance, lack of specificity, or severe homogenization, ultimately leading to low user satisfaction and actual adoption rates.

[0004] In existing research, CN107609003B proposed a method and apparatus for visualizing rural tourism recommendation information. This method includes: receiving a pre-calculated list of rural tourism product recommendations and similarity information between the products in the list; and displaying the recommendations in various formats based on the list and similarity information, including direct list display, destination / attraction map display, route map display, and two-dimensional graph display. This method enhances the visualization of recommendation information, enabling users to more intuitively understand the relationships between recommended products, effectively improving the user experience and further stimulating users to convert their tourism needs into actual consumption. However, this method relies on the pre-generated recommendation list and lacks consideration for the timeliness and dynamic updates of tourism resources, making it difficult to guarantee a high degree of matching between the recommendation results and the user's real-time preferences. Furthermore, while improving the user experience, this technology fails to address the issues of homogenized recommendation results, lack of personalized filtering, and insufficient diversity guarantees. Summary of the Invention

[0005] In view of this, the present invention proposes a semantic recognition-based method for recommending tourism specialty products. This method integrates deep semantic recognition technology into the recommendation of tourism specialty products, understands users' explicit preferences, captures users' implicit needs, and achieves more accurate user demand analysis and product recommendations.

[0006] To achieve the above objectives, the present invention provides a semantic recognition-based method for recommending tourism specialty products, comprising the following steps: S1: Collect the user's input travel demand text, and use the improved semantic recognition model to perform semantic recognition on the travel demand text to obtain the user's travel semantic feature vector. S2: Utilize timely evaluation information to dynamically update the product attributes of tourism specialty products, and construct a dynamic knowledge graph of tourism specialty products based on the product attributes of tourism specialty products; S3: Perform hierarchical semantic matching between the tourism semantic feature vector and the tourism featured products in the dynamic knowledge graph of tourism featured products, and select the K tourism featured products with the highest degree of hierarchical semantic matching to form a candidate product set; S4: Perform personalized filtering on the tourism specialty products in the candidate product set to obtain a personalized filtered candidate product set. Then, rank the tourism specialty products in the personalized filtered candidate product set according to the degree of hierarchical semantic matching and make recommendations based on the ranking results.

[0007] As a further improvement of the present invention: Optionally, the improved semantic recognition model includes a word vector embedding layer, a bidirectional LSTM encoding layer, an attention mechanism layer, an adaptive conditional random field decoding layer, and a structured semantic extraction layer; The word embedding layer includes a pre-trained word vector model, a self-built dictionary in the tourism field, and a joint embedding mechanism; The bidirectional LSTM coding layer includes a forward LSTM structure, a backward LSTM structure, and a splicing unit; The adaptive conditional random field decoding layer includes an adaptive constraint unit and a decoding unit.

[0008] Furthermore, the improved semantic recognition model is used to perform semantic recognition on tourism demand texts, including: The text of tourism demand is used as the input text of the improved semantic recognition model. The input text is segmented by word embedding layer. The general word vector and the neighboring word vector of the segmentation result are output by the pre-trained word vector model combined with the self-built dictionary in the tourism field. The general word vector and the neighboring word vector are concatenated based on the joint embedding mechanism to serve as the tourism field information embedded word vector of the segmentation result. The bidirectional LSTM encoding layer models the contextual features of the word vectors embedded with tourism information, and outputs the preceding and following features of the word vectors embedded with tourism information. The concatenation unit concatenates the preceding and following features as the contextual feature encoding of the word vectors embedded with tourism information. The attention mechanism layer calculates the attention weights of the context feature encoding, and performs attention weighting on the context feature encoding to obtain attention-weighted context feature encoding; The adaptive conditional random field decoding layer adopts an adaptive constraint method based on tourism domain information to construct adaptive constraint units and generate decoding labels corresponding to attention-weighted context feature encoding under adaptive constraints. The structured semantic extraction layer selects context feature encodings that match the preset structured labels for decoding tags, and constructs a structured vector from the selected context feature encodings as the user's tourism semantic feature vector.

[0009] Furthermore, the product attributes of tourism specialty products are dynamically updated using time-sensitive evaluation information, including: The product attributes of the tourism specialty products include product name, product type, product price, address, functional tags, cultural tags, popular tags, and product rating. Product types include attractions, catering, accommodation, and cultural and creative products. The popular tags and product ratings are dynamically updated product attributes. The dynamic updating process for the popular tags and product ratings is as follows: Real-time evaluation information of tourism specialty products, including user reviews and user ratings; Calculate the most recent The average of the user ratings is used as the result of dynamic updates to the product rating. Extract the most recent Keywords from user comments are dynamically updated as trending tags.

[0010] Furthermore, a dynamic knowledge graph of tourism specialty products is constructed based on their product attributes, including: Tourism specialty products are used as graph nodes in a dynamic knowledge graph of tourism specialty products. The edge relationships between each graph node are calculated based on the product attribute relationships of tourism specialty products. The edge relationships include spatial proximity relationships, functional and cultural relationships, and dynamic evaluation relationships. The spatial proximity relationship is the normalized distance between the addresses of the tourism specialty products corresponding to the map nodes; The functional-cultural relationship refers to the similarity between the functional and cultural tags of the tourism specialty products corresponding to the graph nodes. The dynamic evaluation relationship is the weighted semantic similarity between the product attributes and product ratings of the tourism specialty products corresponding to the graph nodes. The weighted semantic similarity is calculated as follows: ; in, This represents the rating-weighted semantic similarity between the popular tags and product ratings of the tourism specialty products corresponding to the m-th and c-th graph nodes in the dynamic knowledge graph of tourism specialty products. M represents the number of graph nodes in the dynamic knowledge graph of tourism specialty products. This represents the product rating of the tourism specialty product corresponding to the m-th node in the dynamic knowledge graph of tourism specialty products. This represents the product rating of the tourism specialty product corresponding to the c-th node in the dynamic knowledge graph of tourism specialty products. This represents the semantic vector of the product attributes of the tourism specialty product corresponding to the m-th graph node in the dynamic knowledge graph of tourism specialty products. This represents the semantic vector of the product attributes of the tourism specialty product corresponding to the c-th graph node in the dynamic knowledge graph of tourism specialty products. The cosine similarity algorithm uses the improved semantic recognition model described in step S1 to encode the contextual features of the product attributes of tourism specialty products, which are then used as semantic vectors of the product attributes.

[0011] Furthermore, hierarchical semantic matching is performed between the tourism semantic feature vectors and the tourism specialty products in the dynamic knowledge graph of tourism specialty products, including: We use an edge-relation-based graph convolution method to extract the product semantic vectors of tourism specialty products corresponding to graph nodes in the dynamic knowledge graph of tourism specialty products; A nonlinear projection method is used to project product semantic vectors and tourism semantic feature vectors onto a shared semantic space; The projection results in the shared semantic space are normalized based on the L2 norm to obtain normalized product semantic vectors and normalized tourism semantic feature vectors with consistent length and scale. Calculate the main semantic similarity between the normalized product semantic vector and the normalized tourism semantic feature vector, calculate the bilinear fine-grained semantic similarity between the normalized product semantic vector and the normalized tourism semantic feature vector, and calculate the business penalty term between the tourism demand text and the product attributes of the tourism specialty products. The main semantic similarity and the bilinear fine-grained semantic similarity are the two-layer semantic matching results of the normalized product semantic vector and the normalized tourism semantic feature vector. By combining the main semantic similarity, bilinear fine-grained semantic similarity, and business penalty terms, the hierarchical semantic matching degree of tourism specialty products to tourism demand text is calculated. The K tourism specialty products with the highest hierarchical semantic matching degree are selected to form a candidate product set.

[0012] Furthermore, the tourism-themed products in the candidate product set are individually filtered to obtain a personalized candidate product set, including: Extract the edge relationships between tourism-featured products in the candidate product set within the dynamic knowledge graph of tourism-featured products; The tourism-featured products in the candidate product set are sorted in descending order according to the degree of hierarchical semantic matching, and a personalized filtered candidate product set is constructed. The constructed personalized filtered candidate product set is initially an empty set. Extract tourism specialty products sequentially from the sorted candidate product set, and calculate the maximum edge weight between the graph node corresponding to the extracted tourism specialty product and any graph node corresponding to the tourism specialty product in the personalized filtered candidate product set. If the calculated edge weight is lower than the preset edge weight threshold, the extracted tourism specialty products are added to the personalized filtered candidate product set; otherwise, the extracted tourism specialty products are filtered to increase the diversity of tourism specialty products in the personalized filtered candidate product set. The products are then extracted again from the ranked candidate product set until there are no tourism specialty products in the ranked candidate product set. The edge weight is the product of the spatial proximity relationship, functional and cultural relationship, and dynamic evaluation relationship in the edge relationship.

[0013] Compared with existing technologies, this invention proposes a semantic recognition-based method for recommending tourism specialty products, which has the following beneficial effects: First, this invention utilizes graph convolution to perform edge-weighted aggregation of the semantic vectors of product attributes of neighboring products. This fully captures the spatial proximity, functional and cultural relationships, and dynamic evaluation relationships between products, resulting in more representative aggregated semantic vectors. The aggregated semantic vectors are then nonlinearly mapped to the semantic vectors of product attributes and projected onto a shared semantic space. L2 norm-based normalization ensures consistent length and scale, guaranteeing the comparability of semantic vectors from different sources or dimensions in subsequent similarity calculations. Finally, the invention introduces joint calculation of main semantic similarity and bilinear fine-grained semantic similarity, comprehensively considering both overall semantic matching and local feature matching to achieve a high-precision correspondence between user travel needs and product attributes.

[0014] Meanwhile, this invention effectively avoids the problem of highly homogenized recommendation results by introducing a personalized filtering mechanism into the candidate product set. Specifically, it first extracts edge relationship information based on the dynamic knowledge graph of tourism specialty products and sorts the candidate products according to the hierarchical semantic matching results, ensuring that products semantically closest to user needs are given priority. Based on this, it calculates the maximum edge weight in the knowledge graph between the tourism specialty products in the candidate product set and the selected tourism specialty products in the personalized filtered candidate product set, and compares it with a preset edge weight threshold to achieve a comprehensive judgment on spatial proximity, functional and cultural relationships, and dynamic evaluation relationships among candidate products. When the edge weight is lower than the threshold, the tourism specialty product is included in the personalized filtered candidate product set, thus ensuring sufficient differences among the selected products. This mechanism not only improves the diversity and novelty of the recommendation results and avoids users receiving highly similar tourism specialty products, but also enhances the user's exploration experience and satisfaction while ensuring relevance. Furthermore, the edge weight adopts a multi-relationship product modeling approach, which better reflects the true similarity under the coupling of multi-dimensional factors, improving the refinement and interpretability of the filtering strategy. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a semantic recognition-based method for recommending tourism specialty products, provided as an embodiment of the present invention.

[0016] Figure 2 A dynamic knowledge graph of tourism specialty products provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the extraction process of product semantic vectors according to an embodiment of the present invention. Detailed Implementation

[0017] The realization of the objectives, functional characteristics, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0018] This invention provides a semantic recognition-based method for recommending featured tourism products. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices configured to execute the method provided in this invention: a server, a terminal, etc. In other words, the semantic recognition-based method for recommending featured tourism products can be executed by software or hardware installed on a terminal device or a server device; the software may be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0019] Reference Figure 1 as well as Figure 2 Embodiment 1 of the present invention is as follows: A semantic recognition-based method for recommending tourism specialty products includes the following steps: S1: Collect the user's input travel demand text, and use the improved semantic recognition model to perform semantic recognition on the travel demand text to obtain the user's travel semantic feature vector.

[0020] The improved semantic recognition model includes a word vector embedding layer, a bidirectional LSTM encoding layer, an attention mechanism layer, an adaptive conditional random field decoding layer, and a structured semantic extraction layer. The word embedding layer includes pre-trained word embedding models (such as Word2Vec and BERT models), a self-built dictionary in the tourism domain, and a joint embedding mechanism; The bidirectional LSTM coding layer includes a forward LSTM structure, a backward LSTM structure, and a splicing unit; The adaptive conditional random field decoding layer includes an adaptive constraint unit and a decoding unit.

[0021] Optionally, the loss function of the improved semantic recognition model consists of a CRF sequence labeling loss function and a structured semantic matching loss function, and the Adam optimizer is used for parameter updates and Dropout and regularization are used to prevent overfitting.

[0022] Semantic recognition of tourism demand texts is performed using an improved semantic recognition model, including: The text of tourism demand is used as the input text of the improved semantic recognition model. The input text is segmented by word embedding layer. The general word vector and the neighboring word vector of the segmentation result are output by the pre-trained word vector model combined with the self-built dictionary in the tourism field. The general word vector and the neighboring word vector are concatenated based on the joint embedding mechanism to serve as the tourism field information embedded word vector of the segmentation result. As an embodiment of the present invention, a pre-trained word vector model is used to encode the word segmentation results, and the encoded vector is used as the general word vector of the word segmentation results; The self-built tourism dictionary includes tourism terms, corresponding tourism category types, and word frequencies. Tourism terms include both tourism-specific and extended terms. Tourism category types include scenic spot types, catering types, accommodation types, transportation types, shopping types, festival and event types, natural resource types, cultural experience types, and sports and leisure types. Scenic spot types include terms such as "Forbidden City," "Great Wall," and "Jiuzhaigou." Catering types include terms such as "hot pot," "roast duck," and "sushi." Accommodation types include... Tourism terms include "youth hostels," "resorts," and "hot spring hotels," transportation terms include "bullet trains," "self-driving," and "high-speed rail," shopping terms include "duty-free shops" and "handicrafts," festivals and events terms include "water splashing festival" and "cherry blossom festival," natural resources terms include "deserts," "snow-capped mountains," and "lakes," cultural experiences terms include "intangible cultural heritage," "tea art," and "folk customs," and sports and leisure terms include "diving," "skiing," and "hiking." One-hot coding is used to generate one-hot codes for each type of tourism, where the length of the one-hot code is equal to the number of types of tourism. Extract the tourism domain type corresponding to the word segmentation result in the self-built tourism domain dictionary, and use the one-hot encoding of the tourism domain type as the domain word vector of the word segmentation result; if the self-built tourism domain dictionary does not contain the word segmentation result, generate an all-zero vector with the same length as the one-hot encoding as the domain word vector of the word segmentation result. Specifically, the unique hot codes corresponding to attraction type, catering type, accommodation type, transportation type, shopping type, festival activity type, natural resource type, cultural experience type, and sports and leisure type are as follows:

[0023] ; The bidirectional LSTM encoding layer models the contextual features of the word vectors embedded with tourism information, and outputs the preceding and following features of the word vectors embedded with tourism information. The concatenation unit concatenates the preceding and following features as the contextual feature encoding of the word vectors embedded with tourism information. As an embodiment of the present invention, the bidirectional LSTM encoding layer arranges the word vectors embedded with tourism information in the order of their positions in the tourism demand text to form a sequence of word vectors embedded with tourism information. The forward LSTM structure and the backward LSTM structure receive the word vectors embedded with tourism information in the order and in reverse order, respectively, and calculate the contextual features and contextual features of the word vectors embedded with tourism information. During the sequential reception process, the LSTM internal loop unit is used to receive the contextual features corresponding to the previous tourism domain information embedding word vector and the current tourism domain information embedding word vector. The gating mechanism is used to calculate the forward hidden state of the current tourism domain information embedding word vector, which is used as the contextual feature of the current tourism domain information embedding word vector. During the reverse reception process, the LSTM internal loop unit is used to receive the context features corresponding to the previous tourism domain information embedding word vector and the current tourism domain information embedding word vector. The gating mechanism is used to calculate the backward hidden state of the current tourism domain information embedding word vector, which is used as the context feature of the current tourism domain information embedding word vector. Optionally, the preceding tourism domain information of the first word received in either sequential or reverse order can be embedded into the preceding or following context features corresponding to the word vector and set to an all-zero vector; The attention mechanism layer calculates the attention weights of the context feature encoding, and performs attention weighting on the context feature encoding to obtain attention-weighted context feature encoding; Specifically, the formula for calculating the attention weights of the context feature encoding h is as follows: ; in, The attention weights represent the context feature encoding h. Represents the trainable parameter matrix. This represents an exponential function with the natural constant as its base. This represents the set of context feature codes corresponding to all word segmentation results in a travel demand text. s represents the context feature encoding set Encoding arbitrary contextual features in; The adaptive conditional random field decoding layer adopts an adaptive constraint method based on tourism domain information to construct adaptive constraint units and generate decoding labels corresponding to attention-weighted context feature encoding under adaptive constraints. As an embodiment of the present invention, Conditional Random Fields (CRFs) introduce transition characteristics between labels. When performing sequence labeling, they consider not only the prediction results of individual positions but also the label dependencies of contextual positions, thereby outputting an overall optimal labeled sequence. This avoids problems such as entity boundary misalignment or label inconsistency caused by local predictions. The state transition matrix for sequence labeling in traditional CRFs is expanded from fixed parameters to position-dependent adaptive transition scores. Compared to traditional CRFs, adaptive constraint parameters based on tourism domain information are introduced in the label scoring process of the labeled sequence, enhancing the interpretability of the labeled sequence. The label scoring function for the labeled sequence is: ; ; ; ; Where x represents the attention-weighted context feature encoding sequence, and y represents the annotation sequence corresponding to the attention-weighted context feature encoding sequence x. This represents the nth attention-weighted context feature encoding in the attention-weighted context feature encoding sequence x. This represents attention-weighted context feature encoding. The corresponding decoding tag, N represents the total number of attention-weighted contextual feature encodings; Indicates decoding tag Transfer to decoding tag The transfer score, Represents trainable decoding labels Transfer to decoding tag The base transfer score, Decoding labels based on information constraints in the tourism sector Transfer to decoding tag Adaptive constraint parameters, This represents attention-weighted context feature encoding. Recognized as a decoding tag The probability is calculated using an encoding mapping method. Specifically, the calculation formula is as follows: ,in, Represents the decoded labels in the trainable mapping matrix The corresponding row vector, Represents the decoded labels in the trainable bias matrix The corresponding bias row vector, This represents the Sigmoid function, which outputs a scalar between 0 and 1. Optionally, set adaptive constraint parameters. The constraint weights are , This represents attention-weighted context feature encoding. The corresponding word segmentation results and their word frequencies in a self-built dictionary for the tourism field. This indicates the highest word frequency in a self-built dictionary within the tourism sector. Indicates control parameters, settings =1; Decoding tags based on tourism rule constraints Transfer to decoding tag The rule constraints are defined as follows: the tourism rule is a set of tourism domain types that can be followed by any tourism domain type. This represents attention-weighted context feature encoding. The corresponding word segmentation results are categorized under the tourism field type in the self-built tourism dictionary. This indicates the type of tourism sector within the pre-defined tourism rules. After that, you can follow , This indicates the type of tourism sector within the pre-defined tourism rules. You can't follow ; This represents the decoded tag predicted by an MLP (Multilayer Perceptron). Transfer to decoding tag The transition probability, This represents the rule trust coefficient, where the rule trust coefficient is a scalar between 0 and 1, and the rule trust coefficient can be set. It is 0.4; Optionally, the prediction formula of the MLP is: ; in, These represent the decoding tags respectively. as well as The word vector format is obtained by using a pre-trained word vector model to convert the decoded tags into word vector format. Let represent the trainable weight matrix parameters, and b represent the trainable bias parameters. Represents the hyperbolic tangent function; The decoding unit aims to maximize the label scoring function. It uses the Viterbi algorithm to solve for the label sequence corresponding to the maximized label scoring function. Then, it matches the decoded labels in the label sequence with the attention-weighted context feature codes in the attention-weighted context feature code sequence according to the sequence position, so as to obtain the decoded label of each attention-weighted context feature code. The structured semantic extraction layer selects context feature encodings that conform to preset structure labels for the decoded labels, and constructs a structured vector from these selected context feature encodings as the user's tourism semantic feature vector. It should be noted that, based on the decoded labels of each attention-weighted context feature encoding, the decoded labels are mapped to the corresponding context feature encodings output by the bidirectional LSTM encoding layer, and context feature encodings that conform to preset structure labels for the decoded labels are selected. Specifically, the types of decoding tags include destination tags, budget tags, demand element tags, population and object category tags, preference and constraint category tags, miscellaneous supplementary tags, and other tags; Optionally, the preset structure labels include destination labels, budget labels, and demand element labels.

[0024] It should be noted that by introducing adaptive constraint parameters on the basis of traditional Conditional Random Field (CRF) scoring, dynamic constraints and position-sensitive adjustments can be made to the semantics of word segmentation results in the tourism domain, thereby achieving refined sequence labeling and enhanced contextual dependence. Specifically, the transfer score includes a basic transfer score and adaptive constraint parameters. The adaptive constraint parameters include rule constraint terms based on tourism rules and prediction terms from a multilayer perceptron. By adaptively calculating constraint weights based on word frequencies in a self-built dictionary in the tourism domain, tourism-specific words and extended words with higher word frequencies will have higher constraint weights, incorporating more tourism domain information. Among them, tourism rules can quickly capture specific label sequence constraints, such as the "price" label should not appear after "attractions," while multilayer perceptron predictions can mine potential contextual relationships. The combination of the two significantly improves the labeling accuracy and robustness, resulting in a label scoring function that combines semantic perception and domain information constraints. This enables dynamic modeling of tourism demand text, enhanced constraints on key word segmentation results, and enhanced contextual dependence, thereby significantly improving the accuracy of sequence labeling and semantic recognition.

[0025] S2: Utilize timely evaluation information to dynamically update the product attributes of tourism specialty products, and construct a dynamic knowledge graph of tourism specialty products based on the product attributes of tourism specialty products.

[0026] Dynamically update the product attributes of distinctive tourism products using timely evaluation information, including: The product attributes of the tourism specialty products include product name, product type, product price, address, functional tags (such as "historical and cultural attractions", "homestay", "local snacks", etc.), cultural tags (such as "intangible cultural heritage experience", "traditional handicrafts", etc.), popular tags, and product rating. Product types include attractions, catering, accommodation, and cultural and creative products. Optionally, the address of the cultural and creative products is the address of the store where the cultural and creative products are purchased; The popular tags and product ratings are dynamically updated product attributes. The dynamic updating process for the popular tags and product ratings is as follows: Real-time evaluation information of tourism specialty products is obtained, including user reviews and user ratings; specifically, timeliness evaluation information of tourism specialty products on tourism-related apps such as Meituan, Dianping, and Ctrip is collected through web crawlers or API interfaces. Calculate the most recent The average of the user ratings is used as the result of dynamic updates to the product rating. Optionally, product ratings can be dynamically updated periodically, with an update cycle of 2 days. During the dynamic update process, historical product ratings from the previous cycle are incorporated, and user ratings are weighted over time. The dynamic update formula for product ratings is as follows: ; ; in, This indicates the historical product rating, which was dynamically updated in the previous cycle. This represents the i-th user rating. This represents the time-series rating weight of the i-th user rating. This indicates the timestamp that is dynamically updated in the current period. This represents the timestamp of the i-th user rating. Indicates the time decay coefficient, set It is 0.4. , Indicates the weight of historical ratings, set It is 0.2; By introducing a time-series weighting mechanism, the weight of each user rating is calculated based on the time interval between the current time and the current time, so that the user rating closer to the current time contributes more to the product rating; at the same time, the weighting of historical ratings preserves long-term trends, so as to achieve dynamic updates that balance the timeliness and robustness of tourism product ratings. Extract the most recent Keywords from user comments are dynamically updated as trending tags.

[0027] Optionally, by means of Each user review is segmented into words to obtain product terms from the user reviews. The TF-IDF value of each product term is calculated, and the two product terms with the highest TF-IDF values ​​are selected as the dynamic update results of popular tags. The IDF values ​​are calculated in conjunction with the product terms of historical user reviews to ensure that popular tags reflect the overall trend rather than local noise.

[0028] Optionally, set Set to 50. It is 30.

[0029] A dynamic knowledge graph of tourism specialty products is constructed based on their product attributes, including: Tourism specialty products are used as graph nodes in a dynamic knowledge graph of tourism specialty products. The edge relationships between each graph node are calculated based on the product attribute relationships of tourism specialty products. The edge relationships include spatial proximity relationships, functional and cultural relationships, and dynamic evaluation relationships. Reference Figure 2 As shown, this is a dynamic knowledge graph of tourism specialty products provided in an embodiment of the present invention. Graph node 1, graph node 2, and graph node 3 are graph nodes in the dynamic knowledge graph of tourism specialty products. The edge relationships include spatial proximity relationships, functional and cultural relationships, and dynamic evaluation relationships between graph nodes in sequence. The spatial proximity relationship is the normalized distance between the addresses of the tourism specialty products corresponding to the map nodes; Specifically, the normalization formula for the distance is: ,in This represents the distance between the addresses of the featured tourist products corresponding to the nodes in the graph. This indicates the preset maximum distance (e.g., 5 kilometers). Indicates selection The minimum value in; The functional-cultural relationship refers to the similarity between the functional and cultural tags of the tourism specialty products corresponding to the graph nodes. Specifically, the similarity between the functional tags and cultural tags of the tourism specialty products is the Jaccard similarity, calculated as follows: ; in, This represents the similarity between the functional and cultural tags of the tourism specialty products corresponding to the m-th and c-th graph nodes in the dynamic knowledge graph of tourism specialty products, where M represents the number of graph nodes in the dynamic knowledge graph of tourism specialty products. This represents the functional and cultural tags of the tourism specialty products corresponding to the m-th graph node. This represents the functional and cultural tags of the tourism specialty products corresponding to the c-th graph node, where each functional and cultural tag has multiple sub-tags. express and The intersection of sub-tags express and The union of child tags, Used to calculate the number of sub-labels in a set; The higher the value, the higher the overlap of the functional and cultural tags between the m-th and c-th graph nodes, and the higher the homogeneity of the tourism specialty products corresponding to the m-th and c-th graph nodes. The dynamic evaluation relationship is the weighted semantic similarity between the product attributes and product ratings of the tourism specialty products corresponding to the graph nodes. The weighted semantic similarity is calculated as follows: ; in, This represents the rating-weighted semantic similarity between the popular tags and product ratings of the tourism specialty products corresponding to the m-th and c-th graph nodes in the dynamic knowledge graph of tourism specialty products. M represents the number of graph nodes in the dynamic knowledge graph of tourism specialty products. This represents the product rating of the tourism specialty product corresponding to the m-th node in the dynamic knowledge graph of tourism specialty products. This represents the product rating of the tourism specialty product corresponding to the c-th node in the dynamic knowledge graph of tourism specialty products. This represents the semantic vector of the product attributes of the tourism specialty product corresponding to the m-th graph node in the dynamic knowledge graph of tourism specialty products. This represents the semantic vector of the product attributes of the tourism specialty product corresponding to the c-th graph node in the dynamic knowledge graph of tourism specialty products. The cosine similarity algorithm uses the improved semantic recognition model (word vector embedding layer and bidirectional LSTM encoding layer) described in step S1 to encode the contextual features of the product attributes of tourism specialty products as product attribute semantic vectors.

[0030] Specifically, in the context feature encoding process of the product attributes of the tourism specialty products, the product score is removed from the product attributes in advance.

[0031] It should be noted that the dynamic knowledge graph of tourism specialty products uses tourism specialty products as graph nodes, which can uniformly express heterogeneous information scattered in comments, demand texts and product descriptions, ensuring the integrity and scalability of the semantic information of tourism specialty products.

[0032] Specifically, by utilizing spatial proximity relationships, the geographical location information of tourism specialty products is explicitly modeled, accurately reflecting the spatial distribution and travel convenience among these products, thus supporting itinerary planning. Through Jaccard similarity calculation of functional-cultural relationships, the homogeneity of products in functional and cultural tags is quantitatively measured, automatically identifying resources with similar cultural characteristics or functional services, enhancing semantic-level aggregation capabilities. The introduction of dynamic evaluation relationships, combining product ratings with semantic vectors of product attributes, forms a rating-weighted semantic similarity, merging evaluation information with product features. This not only reflects the static attributes of products but also dynamically reflects users' real-time feedback and preference trends. This mechanism makes the knowledge graph adaptive and timely, dynamically updating with changes in user reviews and ratings. In summary, the comprehensive modeling of spatial, functional-cultural, and dynamic evaluation relationships improves the accuracy and dynamism of the semantic expression of tourism specialty products, providing more comprehensive, accurate, and real-time updated knowledge support for subsequent tourism specialty product recommendations.

[0033] S3: Perform hierarchical semantic matching between the tourism semantic feature vector and the tourism featured products in the dynamic knowledge graph of tourism featured products, and select the K tourism featured products with the highest degree of hierarchical semantic matching to form a candidate product set.

[0034] Hierarchical semantic matching is performed between tourism semantic feature vectors and tourism specialty products in the dynamic knowledge graph of tourism specialty products, including: We use an edge-relation-based graph convolution method to extract the product semantic vectors of tourism specialty products corresponding to graph nodes in the dynamic knowledge graph of tourism specialty products; Reference Figure 3 The diagram shown is a flowchart of a product semantic vector extraction process according to an embodiment of the present invention. The product semantic vector extraction process includes: S301: Obtain the edge relationships between tourism specialty products, and calculate the product of spatial proximity, functional and cultural relationships and dynamic evaluation relationships in the edge relationships as the edge weights between tourism specialty products. S302: Select the edge with the highest weight from the dynamic knowledge graph of tourism specialty products. One distinctive tourist product can be offered as a neighboring product; alternatively, [location] can be set up. It is 5; S303: Obtain the semantic vector of the product attributes of neighboring products, and perform edge weight-based aggregation processing on the semantic vector of the product attributes of neighboring products to obtain the aggregated semantic vector of the neighbors. Specifically, the neighbor aggregation semantic vector of the tourism specialty product corresponding to the m-th graph node in the dynamic knowledge graph of tourism specialty products is: : ; in, Let represent the semantic vector of the product attribute of the j-th neighboring product of the tourism specialty product corresponding to the m-th graph node, where the j-th neighboring product is the j-th tourism specialty product with the highest edge weight between it and the tourism specialty product corresponding to the m-th graph node. This represents the edge weight between the j-th neighboring product and the tourism specialty product corresponding to the m-th graph node. Represents a trainable graph convolution matrix; S304: Summarize the neighborhood aggregation semantic vector and the product attribute semantic vector of the tourism specialty product to obtain the product semantic vector of the tourism specialty product. Specifically, the product semantic vector of the tourism specialty product corresponding to the m-th graph node in the dynamic knowledge graph of tourism specialty products is: : ; in, A summary matrix representing the semantic vectors of product attributes; A nonlinear projection method is used to project product semantic vectors and tourism semantic feature vectors onto a shared semantic space; Optionally, the nonlinear projection method is a multilayer perceptron or a feedforward neural network, used to align the lengths of the product semantic vector and the tourism semantic feature vector; Optionally, the multilayer perceptron or feedforward neural network corresponding to the nonlinear projection method is trained using the InfoNCE contrastive loss function, which maximizes the similarity of positive pairs and minimizes the similarity of negative pairs, thereby learning a more discriminative shared semantic space; The projection results in the shared semantic space are normalized based on the L2 norm to obtain normalized product semantic vectors and normalized tourism semantic feature vectors with consistent length and scale. Calculate the main semantic similarity between the normalized product semantic vector and the normalized tourism semantic feature vector, calculate the bilinear fine-grained semantic similarity between the normalized product semantic vector and the normalized tourism semantic feature vector, and calculate the business penalty term between the tourism demand text and the product attributes of the tourism specialty products. The main semantic similarity and the bilinear fine-grained semantic similarity are the two-layer semantic matching results of the normalized product semantic vector and the normalized tourism semantic feature vector. As an embodiment of the present invention, the main semantic similarity is calculated by combining temperature-scaled cosine similarity. The temperature-scaled cosine similarity introduces a temperature parameter greater than 0 as a divisor on the basis of traditional cosine similarity, and controls the similarity distribution of the normalized product semantic vector and the normalized tourism semantic feature vector to obtain a more interpretable main semantic similarity. The calculation method for bilinear fine-grained semantic similarity involves introducing a tag attribute matrix and calculating a normalized product semantic vector. And the linear similarity of the normalized tourism semantic feature vectors, where the calculation formula is: ,in This represents the tag attribute matrix, where T denotes the transpose. This represents the standardized tourism semantic feature vector of the tourism specialty product corresponding to the m-th graph node; The business penalty item is the normalized result of the difference between the budget and the product price in the tourism demand text. The calculation method for the business penalty item is as follows: ; in, This represents the business penalty item for the tourism specialty product corresponding to the m-th graph node. The budget in the text indicating tourism demand This represents the product price of the tourism specialty product corresponding to the m-th graph node; Indicates selection The maximum value in; By combining the main semantic similarity, bilinear fine-grained semantic similarity, and business penalty terms, the hierarchical semantic matching degree of tourism specialty products to tourism demand text is calculated. The hierarchical semantic matching degree of the tourism specialty products corresponding to the m-th graph node is: : ; in, This represents the main semantic similarity between the standardized tourism semantic feature vector of the tourism specialty product corresponding to the m-th graph node and the standardized product semantic vector; Select the K tourism specialty products with the highest hierarchical semantic matching degree to form a candidate product set. Optionally, K can be set to 15.

[0035] It should be noted that this invention utilizes graph convolution to perform edge-weighted aggregation of the semantic vectors of product attributes of neighboring products. This fully captures the spatial proximity, functional and cultural relationships, and dynamic evaluation relationships between products, resulting in more representative aggregated semantic vectors. The aggregated semantic vectors are nonlinearly mapped to the semantic vectors of product attributes and projected onto a shared semantic space. Then, L2 norm-based normalization ensures consistent length and scale, guaranteeing the comparability of semantic vectors from different sources or dimensions in subsequent similarity calculations. The joint calculation of main semantic similarity and bilinear fine-grained semantic similarity comprehensively considers overall semantic matching and local feature matching, achieving a high-precision correspondence between user travel needs and product attributes. The introduction of business penalty terms incorporates user budget information and product price differences into the constraints, further improving the feasibility and practicality of the recommendation results.

[0036] S4: Perform personalized filtering on the tourism specialty products in the candidate product set to obtain a personalized filtered candidate product set. Then, rank the tourism specialty products in the personalized filtered candidate product set according to the degree of hierarchical semantic matching and make recommendations based on the ranking results.

[0037] Personalized filtering of tourism-related specialty products within the candidate product set yields a personalized candidate product set, including: Extract the edge relationships between tourism-featured products in the candidate product set within the dynamic knowledge graph of tourism-featured products; The tourism-featured products in the candidate product set are sorted in descending order according to the degree of hierarchical semantic matching, and a personalized filtered candidate product set is constructed. The constructed personalized filtered candidate product set is initially an empty set. Extract tourism specialty products sequentially from the sorted candidate product set, and calculate the maximum edge weight between the graph node corresponding to the extracted tourism specialty product and any graph node corresponding to the tourism specialty product in the personalized filtered candidate product set. If the calculated edge weight is lower than the preset edge weight threshold (e.g., 0.3), the extracted tourism specialty products are added to the personalized filtered candidate product set; otherwise, the extracted tourism specialty products are filtered to increase the diversity of tourism specialty products in the personalized filtered candidate product set. The extracted tourism specialty products are then extracted again from the ranked candidate product set until there are no tourism specialty products in the ranked candidate product set. The edge weight is the product of spatial proximity, functional and cultural relationships, and dynamic evaluation relationships in the edge relationship.

[0038] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0039] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0040] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0041] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for recommending tourism specialty products based on semantic recognition, characterized in that, The method includes: S1: Collect the user's input travel demand text, and use the improved semantic recognition model to perform semantic recognition on the travel demand text to obtain the user's travel semantic feature vector. S2: Utilize timely evaluation information to dynamically update the product attributes of tourism specialty products, and construct a dynamic knowledge graph of tourism specialty products based on the product attributes of tourism specialty products; S3: Perform hierarchical semantic matching between the tourism semantic feature vector and the tourism featured products in the dynamic knowledge graph of tourism featured products, and select the K tourism featured products with the highest degree of hierarchical semantic matching to form a candidate product set; S4: Perform personalized filtering on the tourism specialty products in the candidate product set to obtain a personalized filtered candidate product set. Then, rank the tourism specialty products in the personalized filtered candidate product set according to the degree of hierarchical semantic matching and make recommendations based on the ranking results.

2. The semantic recognition-based method for recommending featured tourism products as described in claim 1, characterized in that, The improved semantic recognition model includes a word vector embedding layer, a bidirectional LSTM encoding layer, an attention mechanism layer, an adaptive conditional random field decoding layer, and a structured semantic extraction layer. The word embedding layer includes a pre-trained word vector model, a self-built dictionary in the tourism field, and a joint embedding mechanism; The bidirectional LSTM coding layer includes a forward LSTM structure, a backward LSTM structure, and a splicing unit; The adaptive conditional random field decoding layer includes an adaptive constraint unit and a decoding unit.

3. The semantic recognition-based method for recommending tourism specialty products as described in claim 2, characterized in that, Semantic recognition of tourism demand texts is performed using an improved semantic recognition model, including: The text of tourism demand is used as the input text of the improved semantic recognition model. The input text is segmented by word embedding layer. The general word vector and the neighboring word vector of the segmentation result are output by the pre-trained word vector model combined with the self-built dictionary in the tourism field. The general word vector and the neighboring word vector are concatenated based on the joint embedding mechanism to serve as the tourism field information embedded word vector of the segmentation result. The bidirectional LSTM encoding layer models the contextual features of the word vectors embedded with tourism information, and outputs the preceding and following features of the word vectors embedded with tourism information. The concatenation unit concatenates the preceding and following features as the contextual feature encoding of the word vectors embedded with tourism information. The attention mechanism layer calculates the attention weights of the context feature encoding, and performs attention weighting on the context feature encoding to obtain attention-weighted context feature encoding; The adaptive conditional random field decoding layer adopts an adaptive constraint method based on tourism domain information to construct adaptive constraint units and generate decoding labels corresponding to attention-weighted context feature encoding under adaptive constraints. The structured semantic extraction layer selects context feature encodings that match the preset structured labels for decoding tags, and constructs a structured vector from the selected context feature encodings as the user's tourism semantic feature vector.

4. The semantic recognition-based method for recommending tourism specialty products as described in claim 1, characterized in that, Dynamically update the product attributes of distinctive tourism products using timely evaluation information, including: The product attributes of the tourism specialty products include product name, product type, product price, address, functional tags, cultural tags, popular tags, and product rating. Product types include attractions, catering, accommodation, and cultural and creative products. The popular tags and product ratings are dynamically updated product attributes. The dynamic updating process for the popular tags and product ratings is as follows: Real-time evaluation information of tourism specialty products, including user reviews and user ratings; Calculate the most recent The average of the user ratings is used as the result of dynamic updates to the product rating. Extract the most recent Keywords from user comments are dynamically updated as trending tags.

5. The semantic recognition-based method for recommending featured tourism products as described in claim 4, characterized in that, A dynamic knowledge graph of tourism specialty products is constructed based on their product attributes, including: Tourism specialty products are used as graph nodes in a dynamic knowledge graph of tourism specialty products. The edge relationships between each graph node are calculated based on the product attribute relationships of tourism specialty products. The edge relationships include spatial proximity relationships, functional and cultural relationships, and dynamic evaluation relationships. The spatial proximity relationship is the normalized distance between the addresses of the tourism specialty products corresponding to the map nodes; The functional-cultural relationship refers to the similarity between the functional and cultural tags of the tourism specialty products corresponding to the graph nodes. The dynamic evaluation relationship is the weighted semantic similarity between the product attributes and product ratings of the tourism specialty products corresponding to the graph nodes. The weighted semantic similarity is calculated as follows: ； in, This represents the rating-weighted semantic similarity between the popular tags and product ratings of the tourism specialty products corresponding to the m-th and c-th graph nodes in the dynamic knowledge graph of tourism specialty products. M represents the number of graph nodes in the dynamic knowledge graph of tourism specialty products. This represents the product rating of the tourism specialty product corresponding to the m-th node in the dynamic knowledge graph of tourism specialty products. This represents the product rating of the tourism specialty product corresponding to the c-th node in the dynamic knowledge graph of tourism specialty products. This represents the semantic vector of the product attributes of the tourism specialty product corresponding to the m-th graph node in the dynamic knowledge graph of tourism specialty products. This represents the semantic vector of the product attributes of the tourism specialty product corresponding to the c-th graph node in the dynamic knowledge graph of tourism specialty products. The cosine similarity algorithm uses the improved semantic recognition model described in step S1 to encode the contextual features of the product attributes of tourism specialty products, which are then used as semantic vectors of the product attributes.

6. The semantic recognition-based method for recommending tourism specialty products as described in claim 1, characterized in that, Hierarchical semantic matching is performed between tourism semantic feature vectors and tourism specialty products in the dynamic knowledge graph of tourism specialty products, including: We use an edge-relation-based graph convolution method to extract the product semantic vectors of tourism specialty products corresponding to graph nodes in the dynamic knowledge graph of tourism specialty products; A nonlinear projection method is used to project product semantic vectors and tourism semantic feature vectors onto a shared semantic space; The projection results in the shared semantic space are normalized based on the L2 norm to obtain normalized product semantic vectors and normalized tourism semantic feature vectors with consistent length and scale. Calculate the main semantic similarity between the normalized product semantic vector and the normalized tourism semantic feature vector, calculate the bilinear fine-grained semantic similarity between the normalized product semantic vector and the normalized tourism semantic feature vector, and calculate the business penalty term between the tourism demand text and the product attributes of the tourism specialty products. The main semantic similarity and the bilinear fine-grained semantic similarity are the two-layer semantic matching results of the normalized product semantic vector and the normalized tourism semantic feature vector. By combining the main semantic similarity, bilinear fine-grained semantic similarity, and business penalty terms, the hierarchical semantic matching degree of tourism specialty products to tourism demand text is calculated. The K tourism specialty products with the highest hierarchical semantic matching degree are selected to form a candidate product set.

7. The semantic recognition-based method for recommending featured tourism products as described in claim 6, characterized in that, Personalized filtering of tourism-related specialty products within the candidate product set yields a personalized candidate product set, including: Extract the edge relationships between tourism-featured products in the candidate product set within the dynamic knowledge graph of tourism-featured products; The tourism-featured products in the candidate product set are sorted in descending order according to the degree of hierarchical semantic matching, and a personalized filtered candidate product set is constructed. The constructed personalized filtered candidate product set is initially an empty set. Extract tourism specialty products sequentially from the sorted candidate product set, and calculate the maximum edge weight between the graph node corresponding to the extracted tourism specialty product and any graph node corresponding to the tourism specialty product in the personalized filtered candidate product set. If the calculated edge weight is lower than the preset edge weight threshold, the extracted tourism specialty products are added to the personalized filtered candidate product set; otherwise, the extracted tourism specialty products are filtered to increase the diversity of tourism specialty products in the personalized filtered candidate product set. The products are then extracted again from the ranked candidate product set until there are no tourism specialty products in the ranked candidate product set. The edge weight is the product of the spatial proximity relationship, functional and cultural relationship, and dynamic evaluation relationship in the edge relationship.

Images