Border tourism recommendation system based on large model

Abstract

The application discloses a border cultural travel recommendation system based on a large model and relates to the technical field of cultural travel recommendation.The system comprises a dimension construction module, a dialogue inquiry module, a journey planning module and a recommendation generation module; the dimension construction module acquires cultural value dimensions, each dimension is a continuous value spectrum, the two ends correspond to opposite orientations, and are associated with a situational story pair generated by a large model; the dialogue inquiry module selects a target dimension, presents the situational story pair, collects a user preference degree to obtain a value balance position; the journey planning module reasons a journey arc line and matches an experience sequence according to a multi-dimensional balance vector; and the recommendation generation module generates a narrative recommendation result.The application positions a user's vague intention through a value spectrum, reconciles internal value conflicts through a journey arc line, and is suitable for personalized recommendation in a border cultural travel scene.

Description

Technical Field

[0001] This invention relates to the field of cultural tourism recommendation technology, and more specifically, to a frontier cultural tourism recommendation system based on a large model. Background Technology

[0002] Border regions possess abundant cultural tourism resources, including intangible cultural heritage, ritual culture, oral history, and ecological knowledge. With the upgrading of cultural tourism consumption, users' demand for in-depth cultural experiences in border regions is growing.

[0003] However, in recommendation scenarios involving deep cultural experiences, users often only have vague cultural intentions and find it difficult to accurately express their preferences using keywords or tags. Currently, recommendation solutions based on large models typically use the large model as a dialogue interface or feature extractor, and their underlying recommendation logic still relies on tag matching or vector similarity calculation. Such solutions cannot deduce the potential value orientations from users' vague expressions, let alone handle the value contradictions implicit in users' needs. In addition, existing recommendation results are mostly presented in the form of information lists, lacking sufficient persuasiveness and interpretability.

[0004] Therefore, a frontier cultural tourism recommendation system based on a large model is proposed to address the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a frontier cultural tourism recommendation system based on a large model, in order to solve the problem that existing recommendation systems in frontier cultural tourism scenarios cannot analyze potential value orientations from users' vague cultural intentions, nor can they handle the value contradictions implied in users' deep motivations, resulting in insufficient matching between recommendation results and users' spiritual needs.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A frontier cultural tourism recommendation system based on a large model aims to solve the problems of existing recommendation systems being unable to analyze potential value orientations from users' vague cultural intentions and unable to handle users' internal value contradictions.

[0008] This system includes a dimension building module, a dialogue exploration module, a journey planning module, and a recommendation generation module.

[0009] The dimension construction module acquires at least one cultural value dimension, each cultural value dimension being a continuous value spectrum, with the two ends of the spectrum corresponding to opposing value orientations, providing a structured coordinate system for preference exploration. Each cultural value dimension is associated with a set of situational story pairs generated by a large model. Each situational story pair contains two situational stories, each biased towards one end of the spectrum and using frontier cultural resources as source material.

[0010] The dialogue exploration module selects a target dimension from the cultural value dimensions, presents corresponding situational story pairs to the user, and collects the user's bias towards the situational story pairs to obtain the user's value balance position on the target dimension. Abstract value orientations are thus transformed into concrete scenarios that users can intuitively experience and choose from. Users can reveal their potential value inclinations through comparison and selection without precisely expressing their preferences.

[0011] The journey planning module obtains a balance vector composed of the user's value balance positions across multiple cultural value dimensions, and uses a large model to infer the journey arc based on this balance vector. The journey arc includes an initial value state, at least one value inflection point, and an ending value state, transforming the user's value balance point into an experience path with a structured beginning, development, transition, and conclusion, which can gradually reconcile the user's internal value conflicts.

[0012] The journey planning module matches experience items from the experience library based on the journey arc to form an experience sequence.

[0013] The recommendation generation module obtains the experience sequence and generates a narrative recommendation result by a large model that integrates the experience items in the experience sequence into the storyline. The structure of the narrative recommendation result is that the initial value orientation and the opposing value orientation are eventually integrated, and the experience items are naturally presented as story scenes.

[0014] Furthermore, the dimension construction module includes an entity contradiction extraction unit, a contradiction clustering unit, and a dimension definition unit:

[0015] The entity contradiction extraction unit extracts the descriptive text of each experience item from the knowledge graph, and uses a large model to identify coexisting and mutually exclusive value tendencies within the same experience item from the descriptive text, thereby obtaining entity-level contradictions. The cultural semantics of each experience item are deconstructed into quantifiable value contradictions in a fine-grained manner.

[0016] The contradiction clustering unit summarizes the entity-level contradictions of all experience items, performs semantic clustering on the entity-level contradictions to obtain concept-level contradictions, and allows the cultural value dimension to emerge naturally from the data through bottom-up clustering.

[0017] The dimension definition unit generates a continuous spectrum definition for each conceptual contradiction, with each end of the spectrum corresponding to a value orientation. Typical experience items are selected as anchor points at both ends to obtain the cultural value dimension, transforming discrete opposing concepts into a continuous value coordinate system.

[0018] Furthermore, the knowledge graph includes at least three types of cultural entities from intangible cultural heritage projects, ritual culture, oral history, and ecological knowledge, as well as the affiliation, symbolic relationship, ritual association, or inheritance relationship between entities, thus depicting the cultural connotation of the experience items from multiple cultural levels.

[0019] Furthermore, the scenario stories are generated by the large model for each cultural value dimension. The generation method is as follows: using the frontier cultural resources contained in the anchor point experience at both ends of the dimension as materials, two stories are generated respectively. The two stories are contrasted in terms of atmosphere attributes and interaction method attributes. One story presents the value orientation corresponding to one end of the spectrum, and the other story presents the value orientation corresponding to the other end of the spectrum. During the generation process, the knowledge graph is used to verify the stories.

[0020] Furthermore, when the dialogue inquiry module selects a target dimension, if the user inputs ambiguous intent text, the ambiguous intent text is mapped to the corresponding cultural value dimension as the target dimension; if the user does not input intent text, the target dimension is selected according to a preset dimension selection order.

[0021] Furthermore, the dialogue inquiry module obtains the degree value of the user's bias towards a certain end of the spectrum through a slider, and uses the degree value as the value balance position, so that the user can express the degree of bias between the two ends of the spectrum.

[0022] Furthermore, after obtaining the user's current value balance position in the dialogue inquiry module, it determines whether the degree value falls within a preset value range; if it does, it queries the dimension with the highest semantic relevance to the current dimension from the preset dimension association table as a supplementary target dimension; the dimension association table is pre-established by calculating the semantic similarity defined by the spectrum of any two dimensions.

[0023] Furthermore, when the journey planning module infers the journey arc, the large model determines the initial value state based on the component with the largest absolute value in the balance vector, so that the starting point of the journey is consistent with the user's strongest value preference.

[0024] The large model determines the value inflection point based on the ratio of the second largest absolute value component to the largest component in the balance vector, so that the degree of inflection matches the intensity of the contradiction between the user's intrinsic value.

[0025] The large model generates an endpoint value state, which is the integration of the user's preferred value orientation with its opposite value orientation, thereby reconciling and transcending the user's initial contradictions.

[0026] Furthermore, when the journey planning module matches experience items, it divides the journey arc into arc segments, performs semantic comparison between the pre-labeled entity-level contradiction descriptions of each candidate experience item and the required value state of the arc segment, and outputs a matching score by the large model.

[0027] By selecting a combination of experience items, the sum of the matching score differences between adjacent experience items in the experience sequence is kept below a preset threshold, thereby controlling the value transition of the experience sequence to be smooth and natural.

[0028] Furthermore, the narrative recommendation results generated by the recommendation generation module are structured as follows: the opening presents the user's initial value orientation; the middle section introduces an opposing value orientation through embedded experience items; and the ending integrates the two value orientations into a unified cultural meaning. The experience items appear as story scenes within the narrative. This three-part narrative structure of conflict resolution enables the recommendation results to complete the construction of a spiritual experience from cognitive conflict to meaning integration, enhancing the interpretability and appeal of the recommendations.

[0029] The technical effects and advantages of this invention are as follows:

[0030] This invention extracts the cultural semantics of experience items from a knowledge graph through a dimension construction module. After entity contradiction extraction and semantic clustering, it automatically constructs a cultural value dimension characterized by a continuous value spectrum and generates contextual story pairs anchored to real cultural resources. The dialogue exploration module uses these contextual story pairs as the interaction medium, collecting the user's bias on the spectrum through a slider to obtain a multi-dimensional value balance position. This process transforms the user's originally vague cultural intentions into precise numerical expressions on a continuous value coordinate system, allowing users to reveal their potential value orientations without needing to precisely describe their preferences. This improves upon the problem of existing recommendation systems struggling to obtain user preferences under cold-start conditions.

[0031] After obtaining a multi-dimensional balance vector, the journey planning module uses a large model to determine the initial value state based on the largest component of the vector, and the value inflection point based on the ratio of the second largest component to the largest component, thus generating the final value state and forming a journey arc containing an introduction, development, transition, and conclusion. Based on this arc, experience items are matched from the experience library. By controlling the sum of the differences in the matching scores of adjacent experience items, the experience sequence maintains a smooth transition in value. The recommendation generation module uses this sequence as a framework to generate a narrative recommendation result that presents the initial orientation at the beginning, introduces the opposing orientation in the middle, and achieves a fusion at the end. The above process transforms the user's inherent value contradictions into a gradually reconcilable experience path and completes the recommendation output in a narrative form, which can enhance the interpretability and appeal of the recommendation results. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the system module framework structure of the present invention;

[0033] Figure 2 This is a schematic diagram of the system workflow of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Example 1

[0036] As attached Figures 1 to 2 The system shown is a frontier cultural tourism recommendation system based on a large model. It consists of a dimension construction module, a dialogue exploration module, a journey planning module, and a recommendation generation module. The dimension construction module pre-constructs dimensions and contextual story pairs offline, while the dialogue exploration module, journey planning module, and recommendation generation module work collaboratively in sequence online.

[0037] This system is applied to border cultural tourism scenarios, and the cultural resources it processes include intangible cultural heritage projects, ritual culture, oral history, ecological knowledge and other types.

[0038] The system will now be described in detail with reference to specific implementation methods.

[0039] Dimension building module:

[0040] The dimension construction module is used to obtain at least one pre-constructed cultural value dimension. Each cultural value dimension is a continuous value spectrum, with each end of the spectrum corresponding to a mutually opposing value orientation. For example, one dimension could be "reverence and closeness," with "reverence" at one end of the spectrum and "closeness" at the other; another dimension could be "integration and observation," with "integration" at one end of the spectrum and "observation" at the other.

[0041] The knowledge graph upon which the dimensions are built stores various cultural entities, which are interconnected through affiliation, symbolism, association, and inheritance relationships. Each experience item is associated with a descriptive text. Record the background, activity content, and features of the experience item.

[0042] The entity contradiction extraction unit extracts the descriptive text of each experience item from the knowledge graph one by one. , will text Input the large model. The large model will then be generated from the text according to the prompts. The system identifies two coexisting and mutually exclusive value orientations within the experience item, and the output format is fixed as "First orientation: Second tendency: ". and This constitutes an entity-level contradiction for this experience item. Traversing all experience items yields a set of entity-level contradictions. .

[0043] Contradictory clustering units will be collected Each tendency phrase in the sentence is input into the sentence vector encoding model one by one to obtain the corresponding sentence vector. Calculate all The cosine similarity between each pair of pairs yields the similarity matrix.

[0044] Based on the similarity matrix, perform bottom-up hierarchical clustering. Each time, select the two clusters with the highest similarity and merge them. After merging, recalculate the mean of the cosine similarity between all pairs of vectors in the new cluster, and use it as the average intra-cluster similarity of the new cluster.

[0045] Let the average intra-cluster similarity of the two original clusters before merging be respectively and The average intra-cluster similarity of the merged new cluster is The decrease in average similarity within the cluster for:

[0046] ;

[0047] Set preset threshold ,when The merging in that direction stops when the clustering is complete. After clustering is finished, the following results are obtained: There are clusters, and each cluster corresponds to a conceptual contradiction.

[0048] The dimensional definition unit performs spectral processing on each conceptual contradiction. For a given cluster, it extracts all... and The two phrases located at opposite ends of the semantic distribution within the phrase serve as labels at opposite ends of the spectrum. and .Will , The large model inputs all entity-level contradiction description texts contained in this cluster, and the large model generates spectral definition texts. It records the value orientations and their meanings at each end of the spectrum.

[0049] Subsequently, the dimension definition unit selects anchor points from the experience items belonging to the concept. For each experience item, let its two tendency phrases of entity-level contradiction be... and Calculate separately and cosine similarity and The cosine similarity is used to determine the larger of the two values ​​for the experience item. The similarity; similarly take and cosine similarity and The larger value in the cosine similarity is used as the experience item and The similarity.

[0050] Selection and The experience item with the highest similarity is selected as the first anchor point, and then... The experience item with the highest similarity is used as the second anchor point. If the same experience item ranks highest in both selections, it is used as the first anchor point, and other similar experiences are selected from the remaining items. The one with the highest similarity is used as the second anchor point. A cluster of concepts, along with its spectral definition text, is considered a second anchor point. Together with the two anchor experiences, they constitute a complete dimension of cultural value.

[0051] As an alternative implementation method, dimension construction can be carried out in a semi-automatic manner with the assistance of expert knowledge. Experts propose several sets of candidate contradictory concepts, each set consisting of descriptions of opposing ends.

[0052] The entity contradiction extraction unit uses each set of candidate concepts as retrieval conditions to calculate the descriptive text of each experience item in the knowledge graph. The sum of semantic similarities between the candidate concept and the descriptions at both ends Semantic similarity is calculated using cosine similarity.

[0053] Set preset threshold ,when When an experience item is included in the association set of the candidate concept, the large model is invoked to determine whether it simultaneously reflects the value orientations represented by both ends of the concept. The input is... The output is either "yes" or "no" to the descriptive text at both ends of the candidate concept. If the output is "yes", then the entity-level contradiction of the experience item is marked as the two ends of the candidate concept.

[0054] The automatic annotation results are submitted to experts for item-by-item review, mislabeled items are deleted and missing items are added, forming the final set of entity-level contradictions. The dimension definition unit performs spectral definition and anchor point selection operations on each group of concepts after review, the process being the same as the fully automatic construction method.

[0055] Once each dimension is constructed, a set of contextual story pairs is generated for it, containing two contextual stories, each biased towards one end of the spectrum of that dimension, and both using frontier cultural resources as material.

[0056] In one implementation, a dual-end independent generation strategy is employed. This obtains the first anchor point experience of the dimension. Second Anchor Point Experience , anchor point Inputting entity information and background descriptions associated in the knowledge graph into a large model generates anchor points. A story set against a real-life backdrop and revolving around the values ​​of the first party. Independently anchor the points Inputting its associated information into a large model generates anchor points. A story set against a real-life backdrop and revolving around the values ​​of the second end. .

[0057] Due to anchor points and anchor point Located at opposite ends of the spectrum, the scenes associated with them differ in atmospheric and interactive attributes, resulting in... and A comparison is made between the two attributes mentioned above. After completion, situational story pairs for that dimension are formed and stored in the situational story library using the dimension identifier as an index.

[0058] In another implementation, a contrastive generation strategy is employed. Generation is first performed within the same large-scale model dialogue context. ,generate When As negative reference text, it is required to be placed in the prompt message. In terms of atmosphere and interaction methods, it is similar to This creates the greatest contrast.

[0059] Atmosphere attributes include solemn, lighthearted, calm, and enthusiastic; interaction methods include viewing from afar, touching up close, and participating. The large model selects one of the atmosphere attributes that aligns with the desired interaction. Different values ​​will be selected from the interaction method attributes. Different values ​​are generated based on the selected combination of attributes. .

[0060] Once generated, the knowledge graph is invoked to validate the entity names, process steps, and constraint rules in the story. If the validation fails, the story is regenerated. Stories that pass the validation are stored in the context story library using dimension identifiers as indexes.

[0061] Dialogue Inquiry Module:

[0062] During the online phase, the dialogue inquiry module selects target dimensions from the established dimensions, presents corresponding contextual story pairs to users, collects the degree of bias, and obtains the value balance position.

[0063] When the user enters ambiguous intent text In one implementation, semantic similarity mapping is used. Input sentence vector encoding model obtains Encode the labels at both ends of the spectrum in each dimension to obtain and .

[0064] Calculate cosine similarity and Take the maximum value :

[0065] ;

[0066] The maximum value The corresponding dimension is used as the target dimension. In another implementation, With all dimensions of the identifier and spectral definition text Assemble the input large model, and the target dimension identifier is directly output from the large model.

[0067] When the user does not enter any intent text, selections are made in a preset order. One order prioritizes coverage, sorting by the number of experience items associated with each dimension from most to least. Another order prioritizes historical frequency, sorting by the number of times each dimension has been added as an additional dimension in historical conversations in descending order.

[0068] Once the target dimension is determined, the corresponding scenario story pairs are retrieved from the scenario story library and presented to the user simultaneously. A slider is used in the interface, with each end corresponding to one of the two stories.

[0069] After the user drags the slider, the system converts the slider position into a degree value. This serves as the value balance point for that dimension. Real numbers on the closed interval [0, 1] are used, with 0 corresponding to the first end of the spectrum and 1 corresponding to the second end; or real numbers on the closed interval [-1, 1] are used, with each end corresponding to one end, and 0 representing neutrality.

[0070] Get the current dimension After the value is determined, in one implementation, a threshold-triggered approach is used to supplement the new dimension. The open interval (0.3, 0.7) is set as the interval of contradiction tendency. When... When a value falls within this range, the dimension with the highest semantic relevance to the current dimension is retrieved from the pre-established dimension association table and used as the supplementary target dimension.

[0071] The dimensional association table defines the text by calculating the spectrum of any two dimensions during the offline phase. The cosine similarity is pre-constructed. In another implementation, all the obtained cosine similarities are pre-constructed. The values ​​are arranged in order to form a balance vector. With all dimensions of the identifier and spectral definition text Input them into the large model, and the large model will determine whether to output a supplementary dimension identifier or a termination signal.

[0072] Trip planning module:

[0073] The journey planning module receives data from various dimensions. The balance vector composed of values ​​in the order of exploration The journey arc is inferred using a large model. The journey arc includes an initial value state, at least one value inflection point, and an ending value state. The initial value state aligns with the user's preference orientation, while the ending value state reflects the integration of preference orientation and opposing orientation.

[0074] As one implementation method, the system presets three arc templates. The progressive template starts with a user-led orientation and ends with a progression from the initial state. The transitional template starts with a user-led orientation, introduces an opposing orientation at a turning point, and ends with the integration of the two. The loop template starts with a weakened dominant orientation, returns to and progresses towards a dominant orientation after experiencing the opposing orientation.

[0075] Large model reception After that, identification The orientation corresponding to the component with the largest absolute value is taken as the dominant orientation. If If there is only one component, directly select the progressive template. It contains multiple components, let the component with the largest absolute value be... The second largest component is Calculate the ratio :

[0076] ;

[0077] when When choosing a gradual approach, When choosing a transition type, when When using a loop-type template, the large model fills the starting value state slots with dominant orientation description text and the value inflection point slots with opposing orientation description text. Based on the starting and inflection point states, it generates the ending value state description text, so that the ending reflects the integration of the two orientations.

[0078] As another implementation method, and spectral definition text for all dimensions Input the large model directly, and the large model will automatically generate a sequence of value states containing the start, turning point and end point. The end point must contain an integrated expression of the opposing orientation.

[0079] After obtaining the journey arc, it is divided into several arc segments according to the state transition points, and the required value state description is extracted from each segment. Each candidate experience item in the experience library has a pre-stored entity-level contradiction description.

[0080] In one implementation, for each arc segment, the required value state description and the contradictory description of the experience item are concatenated and input into a large model, and the large model outputs a matching score. , The value is a decimal with one decimal place, ranging from {0.0, 0.1, ..., 1.0}. The matching rating matrix is ​​obtained after iterating through all arc segments and experience items.

[0081] Select one experience item from each arc segment to form a candidate sequence, assuming the sequence contains... The first experience item, the first The rating for each experience item is Overall change for:

[0082] ;

[0083] Preset threshold Segmentation settings:

[0084] hour , hour , hour .

[0085] Select the one that satisfies and The sequence with the smallest value is taken as the final experience sequence. If there are multiple sequences... If the values ​​are equal, one of them will be randomly selected.

[0086] In another implementation, when the experience database is large, a preliminary screening is performed. The value state description of the arc segment and the contradictory description of the experience item are input into a sentence vector encoding model to obtain vectors, and the cosine similarity is calculated as the initial screening score. The highest-scoring segments from each database are retained. One experience item, Take 10.

[0087] The experience items selected in the initial screening are then subjected to semantic comparison using a large model to output a matching score. The following The calculation and sequence selection rules are the same.

[0088] Recommended generation module:

[0089] The recommendation generation module uses an experience sequence as its framework, and a large model generates narrative-style recommendation results that integrate experience items into the storyline. The narrative structure is as follows: the beginning presents the initial orientation, the middle introduces the opposing orientation, and the ending achieves a fusion, with experience items appearing in the form of story scenes.

[0090] In one implementation, the large model is invoked in three stages. The first stage inputs the first experience item, its contradictory description, and the initial orientation text, generating an opening paragraph from a second-person perspective that reflects the initial orientation. The second stage inputs the middle experience item, its contradictory description, and the opposing orientation text, generating a middle section describing the change in experience and introducing the new orientation. The third stage inputs the final item, its contradictory description, and the texts of both orientations, generating an ending that completes the expression of the fusion of orientations. The three stages are pieced together sequentially to form a complete narrative.

[0091] In another implementation, the entire experience sequence, the contradictory descriptions of each experience item, and the structural requirements are integrated into a comprehensive prompt input model. The prompt clearly states the narrative structure and the requirement that experience items must appear in the form of scenes. The complete narrative is generated by the large model through a single inference.

[0092] The collaborative working process of each module:

[0093] The following example, taken in the Tajik cultural region of the Pamir Plateau, illustrates the collaborative working process of each module.

[0094] In the offline phase, the dimension construction module completes the knowledge graph establishment and dimension construction. Taking the dimension of awe and closeness as an example, the first anchor point experience is viewing sacrificial rock paintings from a distance, and the second anchor point experience is an impromptu duet with an eagle's whistle in a pasture. The dimension definition unit generates spectral definition text. The first dimension is described as "silent observation from a distance," and the second dimension is described as "experiences in engaging in everyday interactions." The system generates contextual story pairs for this dimension. and And store it in the situational story library. Corresponding to the distant view of the sacrificial rock paintings, This corresponds to an impromptu eagle flute duet scene in the ranch.

[0095] During the online phase, users input ambiguous intent text. The dialogue inquiry module will Input sentence vector encoding model obtains After encoding the labels at both ends of each dimension, cosine similarity is calculated. The results are then compared. The dimension with the highest similarity to the labels at both ends of the "reverence" and "closeness" dimensions was selected as the first target dimension. (System retrieves...) and Simultaneously, it displays the data collected by the system after the user drags the slider. .

[0096] Falling into the contradiction tendency range (0.3, 0.7), the system retrieves the highest correlation between the integration and observation dimensions from the dimension association table, selecting them as a supplementary target dimension. The system presents situational story pairs for this dimension, and the user provides... The dialogue inquiry module will connect two... The input value is a large model; the large model has sufficient information to determine the outcome and outputs a termination signal. Balance vector. Depend on and Composed in sequence.

[0097] Journey planning module receives The dominant orientation is identified as closeness, and calculations are performed. The template chosen is a transitional one. The large model fills the initial value state with daily interactive experiences, the turning point with feeling the meaning of inheritance through participation in the production process, and the endpoint with a renewed understanding of reverence through intimacy.

[0098] Divide the arc into three segments:

[0099] The first section is an impromptu duet with the ranch eagle flute. ;

[0100] The second segment features a demonstration of eagle flute making by a master craftsman. ;

[0101] The third section matches the sacrificial rock paintings viewed from afar. .

[0102] , , ,satisfy This was selected as the final experience sequence.

[0103] The recommendation generation module calls the main model three times, taking three experience items, their contradictory descriptions, and corresponding preference texts as input, and generating three narrative segments: the opening, the middle, and the ending. These three segments are then concatenated in sequence, with each experience item presented as a story scenario. This final narrative text is then output to the user as the recommendation result.

[0104] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A frontier cultural tourism recommendation system based on a large model, characterized in that, It includes a dimension building module, a dialogue inquiry module, a journey planning module, and a recommendation generation module: The dimension construction module is used to obtain at least one cultural value dimension. Each cultural value dimension is a continuous value spectrum, with the two ends of the spectrum corresponding to opposing value orientations. Each cultural value dimension is associated with a set of situational story pairs generated by a large model. Each situational story pair contains two situational stories, which are biased towards the two ends of the spectrum and use frontier cultural resources as material. The dialogue inquiry module is used to select a target dimension from the cultural value dimensions, present the corresponding situational story pairs to the user, collect the user's bias towards the situational story pairs, and obtain the user's value balance position on the target dimension. The journey planning module is used to obtain a balance vector composed of the user's value balance position in multiple cultural value dimensions, use a large model to infer a journey arc based on the balance vector, and match experience items from the experience library based on the journey arc to form an experience sequence; the journey arc includes a starting value state, at least one value inflection point, and a ending value state. The recommendation generation module is used to obtain the experience sequence and generate a narrative recommendation result by a large model that integrates the experience items in the experience sequence into the storyline. The structure of the narrative recommendation result is that the initial value orientation and the opposing value orientation are ultimately integrated.

2. The frontier cultural tourism recommendation system based on a large model according to claim 1, characterized in that, The dimension construction module includes an entity contradiction extraction unit, a contradiction clustering unit, and a dimension definition unit: The entity contradiction extraction unit is used to extract the descriptive text of each experience item from the knowledge graph, and use a large model to identify coexisting and mutually exclusive value tendencies within the same experience item from the descriptive text to obtain entity-level contradictions. The contradiction clustering unit is used to summarize the entity-level contradictions of all experience items, and perform semantic clustering on the entity-level contradictions to obtain concept-level contradictions; The dimension definition unit is used to generate a continuous spectrum definition for each conceptual contradiction. Each end of the spectrum corresponds to a value orientation, and typical experience items are selected as anchor points at both ends to obtain the cultural value dimension.

3. The frontier cultural tourism recommendation system based on a large model according to claim 2, characterized in that, The knowledge graph includes at least three types of cultural entities: intangible cultural heritage projects, ritual culture, oral history, and ecological knowledge, as well as the relationships of affiliation, symbolism, ritual association, or inheritance between entities.

4. The frontier cultural tourism recommendation system based on a large model according to claim 2, characterized in that, The scenario stories are generated by a large model for each cultural value dimension, and the generation method is as follows: Using the frontier cultural resources contained in the anchor points at both ends of this dimension as material, two stories are generated respectively. The two stories are contrasted in terms of atmosphere and interaction. One story presents the value orientation corresponding to one end of the spectrum, and the other story presents the value orientation corresponding to the other end of the spectrum. During the generation process, the knowledge graph is used to verify the stories.

5. The frontier cultural tourism recommendation system based on a large model according to claim 1, characterized in that, When the dialogue inquiry module selects the target dimension If the user inputs ambiguous intent text, the ambiguous intent text is mapped to the corresponding cultural value dimension as the target dimension; If the user does not enter any intent text, the target dimension will be selected according to the preset dimension selection order.

6. The frontier cultural tourism recommendation system based on a large model according to claim 1, characterized in that, The dialogue inquiry module obtains the degree value of the user's bias towards a certain end of the spectrum through a slider, and uses the degree value as the value balance position.

7. The frontier cultural tourism recommendation system based on a large model according to claim 6, characterized in that, After obtaining the user's current value balance position, the dialogue inquiry module determines whether the degree value falls within a preset value range. If it falls into the category, the dimension with the highest semantic relevance to the current dimension is retrieved from the preset dimension association table and used as the supplementary target dimension; the dimension association table is pre-established by calculating the semantic similarity of the spectral definition of any two dimensions.

8. The frontier cultural tourism recommendation system based on a large model according to claim 1, characterized in that, When the journey planning module infers the journey arc... The initial value state is determined by the large model based on the component with the largest absolute value in the balance vector. The value inflection point is determined by the large model based on the ratio of the second largest absolute value component to the largest component in the balance vector; The large model generates an endpoint value state, which is an integration of the user's preferred value orientation and its opposite value orientation.

9. The frontier cultural tourism recommendation system based on a large model according to claim 1, characterized in that, When the journey planning module matches experience items... The journey arc is divided into arc segments. The pre-labeled entity-level contradiction descriptions of each candidate experience item are semantically compared with the value states required by the arc segments. The large model outputs a matching score. Select a combination of experience items such that the sum of the matching score differences between adjacent experience items in the experience sequence is lower than a preset threshold.

10. The frontier cultural tourism recommendation system based on a large model according to claim 1, characterized in that, The structure of the narrative recommendation results generated by the recommendation generation module is as follows: The opening presents the user's initial value orientation, the middle section introduces another opposing value orientation through embedded experience items, and the ending integrates the two value orientations into a unified cultural meaning; the experience items appear as story scenes in the narrative.

Images