Fine-tuning training method for large model in text travel vertical field

Abstract

The invention discloses a fine tuning training method for a large model in a text travel vertical field, which relates to the technical field of large model analysis and comprises an interactive display feature set determination module, a target text travel interactive feature output module, a key interactive evaluation data analysis module, a response effect evaluation model construction module and a model adjustment module. The interactive display feature set determination module is used for determining an interactive display feature set corresponding to the text travel theme; the target text and travel interaction feature output module is used for analyzing and outputting target text and travel interaction features of each AR interaction scene; the key interaction judgment data analysis module is used for analyzing key interaction judgment data of each interaction user in the same interaction scene; and the response effect evaluation model construction module is used for constructing a response effect evaluation model of the target text travel interaction characteristics in each interaction scene.

Description

Technical Field

[0001] The present invention relates to the technical field of large model analysis, specifically a fine-tuning training method for large models in the cultural and tourism vertical field. Background Art

[0002] With the rapid advancement of informatization and digitalization, high-tech such as 5G, cloud computing, big data, and the Internet of Things are "cross-border" integrated with the cultural and tourism industry, and a modern tourism system centered on smart tourism is accelerating its formation. Cultural and tourism enterprises need to process a large amount of unstructured data, such as images and videos of scenic spots, tourist evaluations, etc., as well as structured data, such as tourism product information, inventory data, etc. The cultural and tourism vertical large model can efficiently process and analyze these multi-source heterogeneous data, provide enterprises with accurate market forecasts, customer demand analysis, optimize products and services, improve operational efficiency, and accelerate the digital transformation of the cultural and tourism industry;

[0003] Moreover, in existing large models, AR technology is often used to achieve tourist interaction and improve the tourist experience. Different interaction models are usually set in different tourist scenarios. For example, in garden scenic areas, images of landscapes are more prominently displayed, while in scenic spots related to cultural relics and cultures, audio explanations are mostly used as the main body of AR interaction. Therefore, diverse display differences will be formed. However, while being rich, it also brings certain adaptability problems. When some scenic spots and scenery without obvious style characteristics are displayed, the large model can only make judgments based on the stored data in the database as a similarity basis, and the interaction feedback it gives cannot directly represent the true feelings of users. Therefore, there will be some interaction perception deviations, which will have a certain impact on the tourist experience. Summary of the Invention

[0004] The purpose of the present invention is to provide a fine-tuning training method for large models in the cultural and tourism vertical field to solve the problems raised in the prior art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A fine-tuning training method for large models in the cultural and tourism vertical field, the method includes:

[0006] Step S1: Mark each AR interaction scenario where the cultural and tourism large model is applied under the same cultural and tourism theme, retrieve the interaction display features in each AR interaction scenario, and determine the interaction display feature set corresponding to the cultural and tourism theme;

[0007] Step S2: Based on the interaction display feature set and the interaction events recorded in each AR interaction scenario; analyze and output the target cultural and tourism interaction features of each AR interaction scenario;

[0008] Step S3: Based on the target cultural and tourism interaction features, extract the behavioral data of users' response interaction events stored in the corresponding AR interaction scenario; analyze the key interaction evaluation data of each interaction user in the same interaction scenario;

[0009] Step S4: Use the key interaction evaluation data as the target data, and combine it with the public opinion data recorded by the cultural and tourism large model application platform to construct a response effect evaluation model for the target cultural and tourism interaction features in each interaction scenario;

[0010] Step S5: Based on the response effect evaluation model, determine whether to adjust the target cultural and tourism interaction features in the model, and re-evaluate the response effect based on the adjusted interaction features until the target cultural and tourism interaction features in each interaction scenario meet the preset evaluation value.

[0011] Furthermore, the process of determining the set of interaction display features corresponding to the cultural and tourism theme includes the following:

[0012] The interaction display feature refers to the display feature recorded in the AR interaction screen scanned through the user terminal device by using the cultural and tourism large model; such as sound, text, and image;

[0013] Traverse all AR interaction scenarios under the same cultural and tourism theme, and regard the same interaction display features in each AR interaction scenario as a type of interaction display feature; take the union to obtain the interaction display feature set of the cultural and tourism theme containing several categories of interaction display features.

[0014] Furthermore, Step S2 includes the following steps:

[0015] Step S21: When the interaction display features recorded in the same AR interaction scenario only contain one category, output the interaction display feature as the target cultural and tourism interaction feature in the corresponding AR interaction scenario; when the interaction display features recorded in the same AR interaction scenario contain multiple categories, extract the action duration T and the display feature interaction times D of each interaction display feature in the interaction event. The display feature interaction times refer to the number of display form changes of the display feature during the interaction period from the initial display to the end; normalize the action duration T and the display feature interaction times D to obtain T0 and D0 respectively; use the formula: C = k1×T0 + k2×D0 to calculate the dynamic index C of each interaction display feature in each AR interaction scenario;

[0016] Step S22: Calculate the average dynamic index C0 of all the interaction display features recorded in each AR interaction scenario, and select the interaction display features corresponding to the dynamic index greater than the average dynamic index C0 as the target cultural and tourism interaction features;

[0017] If the number of target cultural and tourism interaction features in the same AR interaction scenario is greater than or equal to two, combine them to form the target cultural and tourism interaction features in this scenario.

[0018] Furthermore, Step S3 includes the following specific steps:

[0019] Step S31: The behavioral data includes interaction behavioral data, visual perception data, and user feedback data. Mark the behavioral data affected by the target cultural and tourism interaction features as key behavioral data;

[0020] Step S32: Extract the behavioral data of each user stored in the user device under different AR interaction scenarios, and use the AR interaction scenario as the retrieval item for each type of behavioral data to generate the behavioral status index of various behavioral data under the retrieval item; Use the behavioral status index as the target display item under the retrieval item. The analysis of the behavioral status index is specifically as follows:

[0021] Obtain the actual output value P of the corresponding user under each evaluation index of each type of behavioral data, and generate the status score U corresponding to the actual output value according to the scoring basis divided under the evaluation index where the actual output value is located P ; Sum the status scores to obtain the behavioral status index ∑U of each type of behavioral data P .

[0022] Furthermore, step S3 also includes the following specific steps:

[0023] Step S33: Traverse the retrieval items and target display items generated by all AR interaction scenarios of the same user; Extract the target display items corresponding to the same type of behavioral data under different retrieval items to form the data analysis set of the same user, and calculate the characterization fluctuation value Q of the user in each data analysis set;

[0024] Q = {∑[(∑U P ) - U0] 2 / w} 1 / 2 ;

[0025] where U0 represents the average value of the target display items corresponding to the same type of behavioral data recorded under different retrieval items of the same user; w represents the number of retrieval items stored for the corresponding user on the user side;

[0026] Step S34: Set the characterization fluctuation value threshold Q0, and extract the type of behavioral data of the data analysis set corresponding to Q > Q0 as the key interaction evaluation data of the user; If there is no Q > Q0, extract the type of behavioral data of the data analysis set corresponding to the maximum value of Q as the key interaction evaluation data of the user.

[0027] Analyzing the key interaction evaluation data can effectively reduce the data error caused by the user's subjective evaluation method. Extracting the key interaction evaluation data can reflect the behavioral data with different feelings of the user in different interaction scenarios, avoiding the low data accuracy caused by the consistent perception of the behavioral feedback in different interaction scenarios due to the user's subjective randomness, and providing a more accurate selection and preparation direction for the user data basis of each type of scenario.

[0028] Further, step S4 includes the following specific processes:

[0029] Step S41: Extract the monitoring period corresponding to the public opinion data in the same AR interaction scenario being judged as positive and the number of visitors being greater than or equal to the annual average number of visitors as the positive monitoring period of the AR interaction scenario. Obtain the key interaction evaluation data of all users and the corresponding behavior status indices during the positive monitoring period, and generate the positive response interval E for the corresponding AR interaction scenario. E = [(∑U P ) min ,(∑U P ) max ;

[0030] Step S42: Traverse all AR interaction scenarios to generate all positive response intervals E, and calculate the intersection ratio F of the positive response intervals corresponding to each AR interaction scenario. F = L(E1 ∩ E2 ∩... ∩ E w ) / L E , where E1 ∩ E2 ∩... ∩ E w represents the positive response intervals corresponding to the 1st, 2nd,..., wth AR interaction scenarios, L(E1 ∩ E2 ∩... ∩ E w ) represents the difference in behavior status indices of the intersection interval, and L E represents the difference in behavior status indices of the positive response intervals of each AR interaction scenario;

[0031] Step S43: Set the intersection ratio threshold F0, and construct the response effect evaluation model y of the target cultural and tourism interaction characteristics under each interaction scenario.

[0032]

[0033] When F > F0, output y = 1; when F ≤ F0, output y = 0.

[0034] Construct the response effect evaluation model to analyze the differences in user interaction data when the public opinion is evaluated as positive in different interaction scenarios by combining user-based AR interaction data with actual public opinion data. Since the interaction methods are different in different interaction scenarios, this is used to evaluate whether the current interaction method provides a better user interaction experience; when the model output value is 0, it means that although the user experience of the interaction method in this scenario is positive, compared with the interaction modes in other scenarios, the corresponding user behavior status index is not good, so further analysis is required.

[0035] Further, step S5 includes the following specific steps:

[0036] Step S51: Mark the AR interaction scenarios with an output value of 0 from the response effect evaluation model as interaction scenarios to be adjusted; extract the target cultural and tourism interaction features recorded in the interaction scenarios to be adjusted, and based on the interaction display feature set, excluding the target cultural and tourism interaction features, construct a combined form of all the remaining interaction features as the adjusted interaction features.

[0037] Step S52: After applying each type of adjusted interaction feature to the AR interaction scenario and storing the user interaction data, return to Step S4 to calculate the output value of the response effect evaluation model, and select the interaction feature combination row corresponding to when the output value is equal to the preset evaluation value of 1 as the required adjustment display feature for the interaction scenario to be adjusted.

[0038] If the number of required adjustment display features is greater than one, obtain the intersection ratio required for the response effect evaluation model calculation, and select the interaction feature combination corresponding to the maximum intersection ratio of the AR interaction scenario as the required adjustment display feature for the interaction scenario to be adjusted.

[0039] A large model fine-tuning training system for the cultural and tourism vertical domain, the system includes an interaction display feature set determination module, a target cultural and tourism interaction feature output module, a key interaction judgment data analysis module, a response effect evaluation model construction module, and a model adjustment module.

[0040] The interaction display feature set determination module is used to mark each AR interaction scenario applying the cultural and tourism large model under the same cultural and tourism theme, retrieve the interaction display features in each AR interaction scenario, and determine the interaction display feature set corresponding to the cultural and tourism theme.

[0041] The target cultural and tourism interaction feature output module is used to analyze and output the target cultural and tourism interaction features of each AR interaction scenario.

[0042] The key interaction judgment data analysis module is used to analyze the key interaction judgment data of each interaction user in the same interaction scenario.

[0043] The response effect evaluation model construction module is used to construct a response effect evaluation model for the target cultural and tourism interaction features in each interaction scenario.

[0044] The model adjustment module is used to judge whether to adjust the target cultural and tourism interaction features in the model based on the response effect evaluation model, and re-evaluate the response effect based on the adjusted interaction features until the target cultural and tourism interaction features in each interaction scenario meet the preset evaluation value.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. The present invention analyzes the user data stored by the historical cultural tourism large model in the AR interaction scenario to determine the interaction display characteristics in each cultural tourism scenario, and uses these as the initial analysis targets; indexes the user feedback data through the interaction display characteristics to evaluate the user's behavioral feelings towards this AR interaction; thereby improving the accuracy of data analysis.

[0047] 2. The present invention realizes a comprehensive evaluation based on user feedback and actual different scenic spots by adjusting the display characteristics in the model, analyzes the differences in user interaction data when positive public opinions are evaluated in different interaction scenarios, and thus realizes the optimization of the tourist interaction experience through continuous adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic structural diagram of the fine-tuning training method of the large model in the cultural tourism vertical field of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0050] Embodiment: As Figure 1 shown, the present invention provides a fine-tuning training method for a large model in the cultural tourism vertical field, and the method includes:

[0051] Step S1: Mark each AR interaction scenario applying the cultural tourism large model under the same cultural tourism theme, retrieve the interaction display characteristics in each AR interaction scenario, and determine the interaction display characteristic set corresponding to the cultural tourism theme;

[0052] Step S2: Based on the interaction display characteristic set and the interaction events recorded in each AR interaction scenario; analyze and output the target cultural tourism interaction characteristics of each AR interaction scenario;

[0053] Step S3: Based on the target cultural tourism interaction characteristics, extract the behavioral data of the user's response interaction events stored in the corresponding AR interaction scenario; analyze the key interaction evaluation data of each interacting user in the same interaction scenario;

[0054] Step S4: Use the key interaction evaluation data as the target data, and combine with the public opinion data recorded by the cultural tourism large model application platform to construct a response effect evaluation model for the target cultural tourism interaction characteristics in each interaction scenario;

[0055] Step S5: Based on the response effect evaluation model, determine whether to adjust the target cultural tourism interaction characteristics in the model, and re-evaluate the response effect based on the adjusted interaction characteristics until the target cultural tourism interaction characteristics in each interaction scenario meet the preset evaluation value.

[0056] Furthermore, determining the interactive display feature set corresponding to the cultural tourism theme includes the following process:

[0057] Interactive display features refer to the use of cultural tourism models to scan and display the display features recorded in the AR interactive screen through the user's terminal device; such as sound, text and images;

[0058] Traverse all AR interaction scenes under the same cultural tourism theme, and take the same interactive display features in each AR interaction scene as a category of interactive display features; take the union to obtain an interactive display feature set of the cultural tourism theme containing several categories of interactive display features.

[0059] Furthermore, step S2 includes the following steps:

[0060] Step S21: When the interactive display features recorded in the same AR interaction scenario contain only one category, the interactive display features are output as the target cultural tourism interactive features in the corresponding AR interaction scenario; when the interactive display features recorded in the same AR interaction scenario contain multiple categories, the action duration T and the number of display feature interactions D of each interactive display feature in the interaction event are extracted, where the number of display feature interactions refers to the number of times the display form of the display feature changes from the initial display to the end of the interaction period; the action duration T and the number of display feature interactions D are normalized to obtain T0 and D0 respectively; and the dynamic index C of each interactive display feature in each AR interaction scenario is calculated using the formula: C = k1 × T0 + k2 × D0;

[0061] Step S22: Calculate the average dynamic index C0 of all interactive display features recorded in each AR interaction scene, and select the interactive display feature corresponding to the dynamic index greater than the average dynamic index C0 as the target cultural tourism interaction feature;

[0062] If the number of target cultural and tourism interaction features in the same AR interaction scene is greater than or equal to two, they are combined to form the target cultural and tourism interaction features in the scene.

[0063] Furthermore, step S3 includes the following specific steps:

[0064] Step S31: The behavior data includes interactive behavior data, visual perception data, and user feedback data, and the behavior data affected by the target cultural tourism interactive features are marked as key behavior data;

[0065] For example, the interactive behavior data includes the number of triggers, interactive operation type, dwell time, and navigation behavior. If the target cultural tourism interactive feature is sound, then amplifying the volume in the interactive operation type is the corresponding key behavior data.

[0066] Step S32: Extract the behavior data of each user stored on the user device in different AR interaction scenarios, and use the AR interaction scenario as the retrieval item for each type of behavior data to generate the behavior status index of various behavior data under the retrieval item; use the behavior status index as the target display item under the retrieval item; the analysis of the behavior status index is specifically as follows:

[0067] Obtain the actual output value P of the corresponding user under each evaluation index of each type of behavior data record, and generate the status score U corresponding to the actual output value according to the scoring basis divided under the evaluation index P ; Sum the status scores to obtain the behavior status index ∑U of each type of behavior data P 。

[0068] For example, analyze the four evaluation indexes of the trigger times, interaction operation type, stay duration, and navigation behavior included in the interaction behavior data, and record the actual output values as follows: the trigger times are 6 times, the interaction operation types are zoom and volume adjustment; the stay duration is 2 min, navigation behavior: record the navigation path and stay location of the user in the virtual scene both include the interaction scene;

[0069] Correspondingly give the scoring basis: when the trigger times are 0 - 5 times, the scoring value is 1; when the trigger times are 6 - 10 times, the scoring value is 2 points; and so on, each evaluation index records the corresponding scoring basis and gives the scoring value.

[0070] Further, step S3 also includes the following specific steps:

[0071] Step S33: Traverse the retrieval items and target display items generated by all AR interaction scenarios of the same user; extract the target display items corresponding to the same type of behavior data under different retrieval items to form the data analysis set of the same user, and calculate the characterization fluctuation value Q of the user in each data analysis set;

[0072] Q = {∑[(∑U P ) - U0] 2 / w} 1 / 2 ;

[0073] where U0 represents the average value of the target display items corresponding to the same type of behavior data recorded under different retrieval items of the same user; w represents the number of retrieval items stored by the user device corresponding to the user

[0074] Step S34: Set the characterization fluctuation value threshold Q0, and extract the behavior data type corresponding to the data analysis set when Q > Q0 as the key interaction judgment data of the user; if there is no Q > Q0, extract the behavior data type corresponding to the data analysis set when Q is the maximum as the key interaction judgment data of the user.

[0075] Analyzing key interaction evaluation data can effectively reduce data errors caused by users' subjective evaluation methods. Extracting key interaction evaluation data can reflect the behavioral data of users with different feelings in different interaction scenarios, avoiding low data accuracy caused by the consistent perception of behavioral feedback in different interaction scenarios due to users' subjective randomness, and providing a more accurate selection and preparation direction for the user data basis of each type of scenario.

[0076] Further, step S4 includes the following specific processes:

[0077] Step S41: Extract the monitoring period corresponding to the situation where the public opinion data in the same AR interaction scenario is judged to be in a positive state and the number of visitors is greater than or equal to the annual average number of visitors as the positive monitoring period of the AR interaction scenario. Obtain the key interaction evaluation data of all users and the corresponding behavioral status index during the positive monitoring period, and generate the positive response interval E of the corresponding AR interaction scenario, E = [(∑U P ) min ,(∑U P ) max ;

[0078] Step S42: Traverse all AR interaction scenarios to generate all positive response intervals E, and calculate the intersection ratio F of the positive response intervals corresponding to each AR interaction scenario, F = L(E1∩E2∩...∩E w ) / L E , where E1∩E2∩...∩E w represents the positive response intervals corresponding to the 1st, 2nd,..., wth AR interaction scenarios, L(E1∩E2∩...∩E w ) represents the difference in behavioral status index corresponding to the intersection interval, and L E represents the difference in behavioral status index in the positive response intervals of each AR interaction scenario;

[0079] Step S43: Set the intersection ratio threshold F0, and construct the response effect evaluation model y of the target cultural and tourism interaction characteristics under each interaction scenario,

[0080]

[0081] When F > F0, output y = 1; when F ≤ F0, output y = 0.

[0082] Construct a response effect evaluation model by combining user AR interaction data with actual public opinion data, and analyze the differences in user interaction data when positive public opinion is evaluated in different interaction scenarios. Since the interaction methods are different in different interaction scenarios, this is used to evaluate whether the current interaction method provides a better user interaction experience. When the model output value is 0, it indicates that although the user's perception of the interaction method in this scenario is positive, compared with the interaction modes in other scenarios, the corresponding user behavior state index is not good, so further analysis is required.

[0083] Further, step S5 includes the following specific steps:

[0084] Step S51: Mark the AR interaction scenarios with a model output value of 0 in the response effect evaluation model as interaction scenarios to be adjusted; extract the target cultural and tourism interaction features recorded in the interaction scenarios to be adjusted, and based on the interaction display feature set, excluding the target cultural and tourism interaction features, construct a combined form of all the remaining interaction features as the adjusted interaction features;

[0085] Step S52: After applying each type of adjusted interaction feature to the AR interaction scenario and storing the user interaction data, return to step S4 to calculate the output value of the response effect evaluation model, and select the interaction feature combination row corresponding to when the output value is equal to the preset evaluation value of 1 as the required adjusted display feature for the interaction scenario to be adjusted;

[0086] If the number of required adjusted display features is greater than one, obtain the intersection ratio required for calculating the response effect evaluation model, and select the interaction feature combination corresponding to the maximum intersection ratio in the AR interaction scenario as the required adjusted display feature for the interaction scenario to be adjusted.

[0087] As shown in the embodiment: The interaction display feature set records sound, image, and text; the target cultural and tourism interaction features in interaction scenario 1 are image and text;

[0088] Then, excluding the target cultural and tourism interaction features, construct a combined form of all the remaining interaction features as the adjusted interaction features; including (image, sound), (sound), (image), (text), (text, sound) these five types.

[0089] A large model fine-tuning training system for the cultural and tourism vertical domain, the system includes an interaction display feature set determination module, a target cultural and tourism interaction feature output module, a key interaction judgment data analysis module, a response effect evaluation model construction module, and a model adjustment module;

[0090] The interaction display feature set determination module is used to mark each AR interaction scenario applying the cultural and tourism large model under the same cultural and tourism theme, retrieve the interaction display features in each AR interaction scenario, and determine the interaction display feature set corresponding to the cultural and tourism theme;

[0091] The target cultural and tourism interaction feature output module is used to analyze and output the target cultural and tourism interaction features of each AR interaction scenario;

[0092] The key interaction evaluation data analysis module is used to analyze the key interaction evaluation data of each interaction user in the same interaction scenario;

[0093] The response effect evaluation model construction module is used to construct the response effect evaluation model of the target cultural and tourism interaction features in each interaction scenario;

[0094] The model adjustment module is used to judge whether to adjust the target cultural and tourism interaction features in the model based on the response effect evaluation model, and re-evaluate the response effect based on the adjusted interaction features until the target cultural and tourism interaction features in each interaction scenario meet the preset evaluation value.

[0095] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fine-tuning training method for a large model in the vertical field of culture and tourism, characterized in that: The method includes: Step S1: Mark each AR interaction scenario that applies the cultural and tourism large model under the same cultural and tourism theme, retrieve the interaction display features in each AR interaction scenario, and determine the interaction display feature set corresponding to the cultural and tourism theme; Step S2: Based on the interaction display feature set and the interaction events recorded in each AR interaction scenario, analyze and output the target cultural and tourism interaction features of each AR interaction scenario; Step S3: Based on the target cultural and tourism interaction features, extract the behavior data storing the user's response to the interaction event under the corresponding AR interaction scenario; analyze the key interaction evaluation data of each interacting user in the same interaction scenario; Step S4: Use the key interaction evaluation data as the target data, and combine it with the public opinion data recorded by the cultural and tourism large model application platform to construct a response effect evaluation model for the target cultural and tourism interaction features in each interaction scenario; Step S5: Based on the response effect evaluation model, determine whether to adjust the target cultural and tourism interaction features in the model, and re-evaluate the response effect based on the adjusted interaction features until the target cultural and tourism interaction features in each interaction scenario meet the preset evaluation value.

2. The fine-tuning training method for the large model in the vertical cultural and tourism field according to claim 1, wherein: The determination of the interaction display feature set corresponding to the cultural and tourism theme includes the following process: The interaction display feature refers to the display features recorded in the AR interaction screen scanned by the user terminal device using the cultural and tourism large model; such as sound, text, and images; Traverse all AR interaction scenarios under the same cultural and tourism theme, and regard the same interaction display features in each AR interaction scenario as a type of interaction display feature; take the union to obtain the interaction display feature set of the cultural and tourism theme containing several categories of interaction display features.

3. The fine-tuning training method for the large model in the vertical cultural and tourism field according to claim 1, wherein: Step S2 includes the following steps: Step S21: When the interaction display features recorded in the same AR interaction scenario only contain one category, output the interaction display feature as the target cultural and tourism interaction feature of the corresponding AR interaction scenario; when the interaction display features recorded in the same AR interaction scenario contain multiple categories, extract the action duration T and the display feature interaction times D of each interaction display feature in the interaction event. The display feature interaction times refer to the number of display form changes of the display feature during the interaction period from the initial display to the end; normalize the action duration T and the display feature interaction times D to obtain T0 and D0; use the formula: C = k1×T0 + k2×D0 to calculate the dynamic index C of each interaction display feature in each AR interaction scenario; Step S22: Calculate the average dynamic index C0 of all the interaction display features recorded in each AR interaction scenario, and select the interaction display features corresponding to the dynamic index greater than the average dynamic index C0 as the target cultural and tourism interaction features; If the number of target cultural and tourism interaction features in the same AR interaction scenario is greater than or equal to two, combine them to form the target cultural and tourism interaction features of this scenario.

4. The fine-tuning training method for the large model in the vertical cultural and tourism field according to claim 1, wherein: Step S3 includes the following specific steps: Step S31: The behavior data includes interaction behavior data, visual perception data, and user feedback data. Mark the behavior data affected by the target cultural and tourism interaction features as key behavior data; Step S32: Extract the behavior data of each user stored on the user device in different AR interaction scenarios, and use the AR interaction scenario as the retrieval item for each type of behavior data to generate the behavior status index of various behavior data under the retrieval item; use the behavior status index as the target display item under the retrieval item; the analysis of the behavior status index is specifically as follows: Obtain the actual output value P of the corresponding user under each type of behavior data record evaluation index, and generate the status score U corresponding to the actual output value according to the scoring basis divided under the evaluation index P ; Sum the status scores to obtain the behavior status index ∑U of each type of behavior data P .

5. The fine-tuning training method for the large model in the vertical cultural and tourism field according to claim 4, characterized in that: Step S3 also includes the following specific steps: Step S33: Traverse the retrieval items and target display items generated by all AR interaction scenarios of the same user; extract the target display items corresponding to the same type of behavior data under different retrieval items to form the data analysis set of the same user, and calculate the characterization fluctuation value Q of the user in each data analysis set; Q = {∑[(∑U P ) - U0] 2 / w} 1 / 2 ; Where U0 represents the average value of the target display items corresponding to the same type of behavior data recorded by different retrieval items of the same user; w represents the number of retrieval items stored by the user device corresponding to the user; Step S34: Set the characterization fluctuation value threshold Q0, and extract the behavior data type of the data analysis set corresponding to Q>Q0 as the key interaction evaluation data of the user; if there is no Q>Q0, extract the behavior data type of the data analysis set corresponding to the maximum value of Q as the key interaction evaluation data of the user.

6. The fine-tuning training method for the large model in the vertical field of culture and tourism according to claim 5, characterized in that: Step S4 includes the following specific processes: Step S41: Extract the monitoring period corresponding to the public opinion data in the same AR interaction scenario that is judged to be in a positive state and the number of visitors is greater than or equal to the annual average number of visitors, as the positive monitoring period of the AR interaction scenario. Obtain the key interaction evaluation data and the corresponding behavior status indices of all users during the positive monitoring period, and generate the positive response interval E of the corresponding AR interaction scenario, E = [(∑U P ) min ,(∑U P ) max ; Step S42: Traverse all AR interaction scenarios to generate all positive response intervals E, and calculate the intersection ratio F of the positive response intervals corresponding to each AR interaction scenario, F = L(E1 ∩ E2 ∩... ∩ E w ) / L E , where E1 ∩ E2 ∩... ∩ E w represents the positive response intervals corresponding to the 1st, 2nd,..., w-th AR interaction scenarios, and L(E1 ∩ E2 ∩... ∩ E w ) represents the difference in the behavior state indices corresponding to the intersection interval, and L E represents the difference in the behavior state indices in the positive response intervals of each AR interaction scenario; Step S43: Set the intersection ratio threshold F0, and construct the response effect evaluation model y of the target cultural and tourism interaction feature in each interaction scenario, When F>F0, output y = 1; when F≤F0, output y = 0.

7. The fine-tuning training method for the large model in the vertical cultural and tourism field according to claim 5, wherein: Step S5 includes the following specific steps: Step S51: Mark the AR interaction scenario with the output value of 0 of the response effect evaluation model as the interaction scenario to be adjusted; extract the target cultural and tourism interaction feature recorded in the interaction scenario to be adjusted, and based on the interaction display feature set, remove the target cultural and tourism interaction feature, and construct the combination form of all the remaining interaction features as the adjusted interaction feature; Step S52: After applying each type of adjusted interaction feature to the AR interaction scenario and storing the user interaction data, return to Step S4 to calculate the output value of the response effect evaluation model, and select the interaction feature combination corresponding to the output value equal to the preset evaluation value of 1 as the required adjustment display feature of the interaction scenario to be adjusted; If the number of required adjustment display features is greater than one, obtain the intersection ratio required for calculating the response effect evaluation model, and select the interaction feature combination of the AR interaction scenario corresponding to the maximum intersection ratio as the required adjustment display feature of the interaction scenario to be adjusted.

8. A fine-tuning training system for a large model in the cultural and tourism vertical field applying the fine-tuning training method for a large model in the cultural and tourism vertical field according to any one of claims 1-7, characterized in that, The system includes an interaction display feature set determination module, a target cultural and tourism interaction feature output module, a key interaction evaluation data analysis module, a response effect evaluation model construction module, and a model adjustment module; The interaction display feature set determination module is used to mark each AR interaction scenario applying the cultural and tourism large model under the same cultural and tourism theme, retrieve the interaction display features in each AR interaction scenario, and determine the interaction display feature set corresponding to the cultural and tourism theme; The target cultural and tourism interaction feature output module is used to analyze and output the target cultural and tourism interaction feature of each AR interaction scenario; The key interaction evaluation data analysis module is used to analyze the key interaction evaluation data of each interacting user in the same interaction scenario; The response effect evaluation model construction module is used to construct a response effect evaluation model for the target cultural and tourism interaction features in each interaction scenario; The model adjustment module is used to determine whether to adjust the target cultural and tourism interaction features in the model based on the response effect evaluation model, and re-evaluate the response effect based on the adjusted interaction features until the target cultural and tourism interaction features in each interaction scenario meet the preset evaluation value.

Images