System and method for generating mixed beverage formula

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a mixed beverage formula generation system and method.The system comprises an acquisition module, an extraction module and a generation module.The acquisition module is used for acquiring current multi-modal data, current environment data and current limiting conditions of a user; the extraction module is used for performing feature extraction on the current multi-mode data and the current environment data to obtain current situation information; and the generation module is used for inputting the plurality of situation characteristics and the current limiting condition into a preset mixed beverage formula generation model to obtain the weights of a plurality of formula elements in the mixed beverage matched with the current situation information, the name of the mixed beverage and the formula description of the mixed beverage, and generating a mixed beverage formula. Therefore, the problems of lack of literature expression, poor situational adaptability and the like in the background technology are solved, and personalized and emotional mixed beverage formula generation meeting industrial standards is realized by combining artificial intelligence and multi-dimensional situational information.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a system and method for generating a mixed drink formula. Background Art

[0002] A mixed drink (such as a cocktail) is a beverage made from a mixture of various alcoholic beverages, fruit juices, syrups, spices, etc. It can also be available as a non-alcoholic beverage. Its formula design is a complex process that is highly dependent on the designer's subjective perception of the humanistic environment and his or her literary literacy.

[0003] The design of mixed drink recipes has gone through a long development process, from the initial simple mixing to today's personalized customization, and has evolved from traditional methods to modern methods. The traditional manual design method is highly dependent on the experience, knowledge and creativity of the bartender. Bartenders need to consider many factors, such as the unique flavors and characteristics of different base liquors, the enrichment of the taste and layers of auxiliary materials (juice, syrup, spices, etc.), and the influence of the proportion of raw materials on the overall balance and taste of mixed drinks. Bartenders need to make precise blends based on their experience and understanding of taste. In addition, a good mixed drink name and description can enhance the attractiveness and cultural connotation of the mixed drink.

[0004] The related technologies mainly rely on data retrieval and keyword replacement technology (for example, based on the huge cocktail recipe database provided by relevant websites, users can query by entering conditions such as target taste, ingredients or alcohol concentration, and the system will filter and match appropriate recipes according to the rules), or automatically generate mixed drink recipes through machine learning algorithms (for example, using a multi-layer perceptron to analyze existing cocktail recipe data, learn the co-occurrence relationship between ingredients, and generate new recipes based on the ingredients or taste preferences entered by the user).

[0005] However, despite the continuous development of mixed drink formula design methods, the methods in related technologies have the following limitations: (1) Lack of literary expression dimension. Traditional mixed drink formulas often contain rich literary and artistic connotations, and existing generation methods cannot effectively construct formula descriptions with literary artistic conception; (2) The types of mixed drinks are limited. When relying on database retrieval and keyword replacement technology, only the types of drinks already in the database can be generated, and it is impossible to create new beverage categories or adjust the formula according to new raw materials and technological trends. In addition, traditional mixed drink formula databases usually focus on classic and popular mixed drinks, and lack the collection of regional characteristics or niche beverage types; (3) Poor situational adaptability. Existing methods find it difficult to dynamically generate beverage formulas that meet specific situations based on real-time situational information (such as geographic location, activity type, user stories and experiences, atmosphere, etc.); (4) Insufficient application of professional knowledge. Although deep learning-based methods can generate mixed drink formulas, they do not fully understand and apply professional knowledge related to mixed drink preparation, such as raw material ratios and taste matching, resulting in the generated formulas not meeting industry standards or practical feasibility, which urgently needs to be solved. Summary of the invention

[0006] The present invention provides a system and method for generating a mixed drink formula to solve the problems of lack of literary expression and poor situational adaptability of the background technology. By combining artificial intelligence and multi-dimensional situational information, personalized, emotional and industry-standard mixed drink formula generation is achieved.

[0007] A first aspect of the present invention provides a system for generating a mixed drink formula, comprising: A collection module, used to collect the user's current multimodal data, current environmental data and current constraints when there is a need for contextual beverage recipe generation; An extraction module, configured to extract features from the current multimodal data and the current environment data to obtain current context information, wherein the current context information includes a plurality of context features; A generation module is used to input the multiple situational features and the current constraints into a preset mixed drink formula generation model to obtain weights of multiple formula elements in the mixed drink that match the current situational information, the name of the mixed drink and the formula description of the mixed drink, so as to generate a mixed drink formula according to the weights of multiple formula elements in the mixed drink, the name of the mixed drink and the formula description of the mixed drink, wherein the preset mixed drink formula generation model is obtained by fine-tuning the original base model using a preset data set.

[0008] According to one embodiment of the present invention, the current multimodal data includes at least one of the current voice data of the user, the current visual feature data of the user, and the current interaction data of the user.

[0009] According to one embodiment of the present invention, the acquisition module includes: A voice collection component, used to collect the current voice data of the user; A video acquisition component, used to acquire current visual feature data of the user; The interaction component is used to collect current interaction data with the user.

[0010] According to one embodiment of the present invention, the system for generating a mixed drink formula further includes: a training module, wherein the training module is specifically used for: Based on the preset data structure, construct the initial training data set; Using the initial training data set to perform fine-tuning training on the original base model to obtain an initial pre-tuned model; Inputting the data in the preset test data set into the initial pre-adjusted model to generate new sample data, and evaluating and screening the new sample data according to a preset quantitative scoring formula to obtain data that meets the preset evaluation conditions, and adding the data that meets the preset evaluation conditions to the initial training data set to obtain a new training data set; The initial pre-adjustment model is iteratively trained using the new training data set, and the iteratively trained initial pre-adjustment model is verified until the iteratively trained initial pre-adjustment model meets the preset verification conditions, and the iteratively trained initial pre-adjustment model is used as the preset mixed beverage formula generation model.

[0011] According to one embodiment of the present invention, the preset quantitative scoring formula is:

[0012] in, is the quantitative scoring result of the current sample data, is the weight of task completion. Rate the task completion. The weight of the raw material accuracy, Score the accuracy of ingredients, The weight of literary color, Rate the literary color, The weight of diversity, Score diversity, is the weight of compliance with industry standards, Score compliance with industry standards.

[0013] According to one embodiment of the present invention, the system for generating a mixed beverage formula further includes: A feedback module, used to obtain user satisfaction and feedback information on the mixed beverage formula; The optimization module is used to adjust and optimize the model parameters of the preset mixed drink formula generation model based on the user's satisfaction and feedback information on the mixed drink formula.

[0014] According to the generation system of the mixed drink formula of the embodiment of the present invention, the collection module is used to collect the user's current multimodal data, current environmental data and current constraints, and the extraction module is used to extract features of the current multimodal data and current environmental data to obtain the current context information; the generation module is used to input multiple context features and current constraints into the preset mixed drink formula generation model, and the weights of multiple formula elements in the mixed drink that match the current context information, the name of the mixed drink and the formula description of the mixed drink are obtained, and the mixed drink formula is generated. In this way, the problems of lack of literary expression and poor context adaptability of the background technology are solved, and personalized, emotional and industry-standard mixed drink formula generation is achieved by combining artificial intelligence and multi-dimensional context information.

[0015] A second aspect of the present invention provides a method for generating a mixed drink formula, using the above-mentioned system for generating a mixed drink formula, wherein the method comprises the following steps: When there is a need to generate a contextualized beverage recipe, the acquisition module is used to acquire the user's current multimodal data, current environmental data, and current constraints; Using the extraction module to extract features from the current multimodal data and the current environment data to obtain current context information, wherein the current context information includes a plurality of context features; The generation module is used to input the multiple situational features and the current constraints into a preset mixed drink formula generation model to obtain weights of multiple formula elements in the mixed drink that match the current situational information, the name of the mixed drink and the formula description of the mixed drink, so as to generate a mixed drink formula according to the weights of multiple formula elements in the mixed drink, the name of the mixed drink and the formula description of the mixed drink, wherein the preset mixed drink formula generation model is obtained by fine-tuning the original base model using a preset data set.

[0016] According to one embodiment of the present invention, the current multimodal data includes at least one of the current voice data of the user, the current visual feature data of the user, and the current interaction data of the user.

[0017] According to one embodiment of the present invention, before using the generation module to input the plurality of situational features and the current restriction condition into a preset mixed drink formula generation model, the method further includes: Based on the preset data structure, construct the initial training data set; Using the initial training data set to perform fine-tuning training on the original base model to obtain an initial pre-tuned model; Inputting the data in the preset test data set into the initial pre-adjusted model to generate new sample data, and evaluating and screening the new sample data according to a preset quantitative scoring formula to obtain data that meets the preset evaluation conditions, and adding the data that meets the preset evaluation conditions to the initial training data set to obtain a new training data set; The initial pre-adjustment model is iteratively trained using the new training data set, and the iteratively trained initial pre-adjustment model is verified until the iteratively trained initial pre-adjustment model meets the preset verification conditions, and the iteratively trained initial pre-adjustment model is used as the preset mixed beverage formula generation model.

[0018] According to one embodiment of the present invention, after generating a mixed drink formula according to the weights of a plurality of formula elements in the mixed drink, the name of the mixed drink and the formula description of the mixed drink, the method further includes: Using a feedback module to obtain user satisfaction and feedback information on the mixed beverage formula; The optimization module is used to adjust and optimize the model parameters of the preset mixed drink formula generation model based on the user's satisfaction and feedback information on the mixed drink formula.

[0019] According to the method for generating a mixed drink formula in an embodiment of the present invention, a collection module is used to collect the user's current multimodal data, current environmental data and current constraints, and an extraction module is used to extract features from the current multimodal data and current environmental data to obtain current context information; a generation module is used to input multiple context features and current constraints into a preset mixed drink formula generation model to obtain the weights of multiple formula elements in the mixed drink that match the current context information, the name of the mixed drink and the formula description of the mixed drink, and generate a mixed drink formula. Thus, the problems of lack of literary expression and poor context adaptability of the background technology are solved, and personalized, emotional and industry-standard mixed drink formula generation is achieved by combining artificial intelligence and multi-dimensional context information.

[0020] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1A schematic diagram of a method for generating a mixed beverage formula in the prior art; Figure 2 A schematic diagram of an AI training data structure design according to an embodiment of the present invention; Figure 3 A schematic diagram of a data iteration and model training process according to an embodiment of the present invention; Figure 4 A schematic diagram of a flow chart of a mixed beverage formula AI generation model training method according to an embodiment of the present invention; Figure 5 A block diagram of a system for generating a mixed drink recipe according to an embodiment of the present invention; Figure 6 A flow chart of a method for generating a mixed drink formula according to an embodiment of the present invention; Figure 7 A schematic diagram of a method for generating a mixed drink formula based on AI according to an embodiment of the present invention; Figure 8 is a flow chart of a method for generating a mixed drink formula according to one embodiment of the present invention; Figure numerals: 21 - multimodal data, 22 - environmental data, 23 - contextual information, 24 - restrictive conditions, 25 - beverage formula; 10 - generation system of mixed beverage formula, 100 - acquisition module, 200 - extraction module, 300 - generation module. DETAILED DESCRIPTION

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The system and method for generating a mixed drink recipe according to an embodiment of the present invention will be described below with reference to the accompanying drawings.

[0024] Before introducing the mixed drink formula generation system of the embodiment of the present invention, the mixed drink formula generation method in the related art and the training method of the preset mixed drink formula generation model involved in the mixed drink formula generation system of the present invention are briefly introduced.

[0025] In the field of mixed beverage recipe design, existing generation technologies mainly use rule-based and database retrieval systems. Such systems usually require users to input required parameters, such as target taste, ingredients or alcohol concentration, and query existing recipes in the database, use certain rules to screen and match, and finally output a suitable beverage recipe. Figure 1A method for generating a mixed drink recipe in the related art is shown. The user first inputs his or her requirements for the mixed drink, which may include the target taste, ingredient requirements, alcohol concentration, or other specific preferences. The system will query the existing beverage recipe database according to the requirements input by the user. After retrieving multiple related recipes, the system will filter and match them based on certain rules. The filtering rules may include the combination of ingredients, the balance of tastes, the matching degree of user preferences, etc., and finally output one or more mixed drink recipes that best meet the user's needs.

[0026] However, existing technologies have many shortcomings in practical applications. First, although general large language models have powerful natural language processing capabilities, they show obvious limitations in professional fields such as beverage recipe generation. These models find it difficult to accurately understand and apply professional knowledge related to beverage preparation, such as raw material ratios, taste matching, production processes, etc., resulting in the generated recipes that may not meet industry standards or practical feasibility. Secondly, existing systems have poor understanding and analysis capabilities of situational information, and cannot accurately capture users' personalized needs, nor can they convert abstract user intentions (such as specific emotions or scenarios) into specific beverage characteristics. The current technical solutions also have the problem of poor situational adaptability, and it is difficult to dynamically adjust the generation strategy according to real-time environmental conditions (such as location, weather, activity type, etc.).

[0027] In order to solve the problems existing in the prior art, the present invention proposes a generation system for mixed drink formulas based on artificial intelligence and incorporating multi-dimensional contextual information. The system aims to capture the user's emotional state and cultural background by analyzing the user's conversation content, multimodal data, environmental data and personal stories, and generate a more literary and emotionally resonant recipe description; at the same time, it breaks through the limitations of traditional retrieval and keyword replacement technology, and constructs an intelligent generation solution that adapts to new raw materials, new blending styles and industry trends to meet the continuous innovation and development needs of mixed drink formulas. In addition, this system focuses on improving situational adaptability, and can dynamically generate beverage formulas that meet specific scenarios based on real-time situational information (such as geographic location, activity type, user experience and atmosphere, etc.), and deeply integrates professional bartending knowledge with AI technology, such as the understanding and application of raw material characteristics, taste matching, production process, etc., to ensure that the generated formula is both in line with industry standards and creative.

[0028] Furthermore, in order to train an AI model that can generate mixed drink recipes and visual presentations that incorporate situational features and literary overtones, the present invention proposes a training method for a preset mixed drink recipe generation model, which aims to solve the problems of existing AI models in the field of drink recipe generation, such as lack of expertise, poor situational adaptability, and insufficient data.

[0029] The training method of the preset mixed drink formula generation model involved in the generation system of the mixed drink formula of the present invention constructs a new data structure to map the user's dialogue information and multimodal environmental information to context information, and uses it as input for generating a drink formula with literary flavor and personalized instructions. A data set construction and expansion method based on AI using a small sample data set, literary data and industry expert review is adopted to form a high-quality incremental data set through multiple iterations, and the data set is used to fine-tune the model so that it can understand and generate high-quality formulas that match the input data, and quantitatively score the formula generated by the model to evaluate its accuracy, personalized customization effect, literary expression and compliance with industry specifications, and add high-quality data to the training data set to continuously improve the generalization and stability of the model.

[0030] Specifically, if Figure 2 As shown, in the preset mixed beverage recipe generation model training process, the present invention constructs a new data structure consisting of three parts, which maps the user's conversation information, multimodal data 21 and environmental data 22 information to contextual information 23 to support the needs of contextualized beverage recipe 25 generation, and then generates a beverage recipe with literary color and personalized description based on the contextual information.

[0031] Furthermore, the first part of the data structure includes multimodal data 21 and environmental data 22. The multimodal data 21 and environmental data 22 of the embodiment of the present invention specifically refer to raw data used to generate user situation descriptions after being standardized through different types of data sources such as voice, image, and text. The situation information 23 extracted after analysis includes the user's current emotional state, preferences, geographic location, climate characteristics, etc., which is further used as input for generating a beverage recipe 25.

[0032] The second part of the data structure is contextual information 23 and restrictive conditions 24, which include user information, experience stories, current geographic location, climate characteristics, beverage flavors, preferences, etc.; restrictive conditions 24 include specific requirements or restrictions that need to be followed when generating beverage recipes 25, such as the types of available raw materials, the taste of the beverage, health requirements, etc.

[0033] The third part of the data structure is the beverage formula 25, which contains text information such as descriptive information, beverage proportions, and beverage explanations. Through the association rule constraints with the contextual information 23 and restrictive conditions 24, the mixed beverage formula 25 with reasonable ingredients in the formula is creative, literary, meets user needs, and can generate emotional resonance.

[0034] Among them, there is a data conversion and analysis relationship between the first part and the second part of the data structure, mapping the multimodal data 21 (such as original data in various formats such as voice and images, and recordings of conversations with users) and environmental data 22 into contextual information 23 to support the needs of generating contextualized beverage recipes 25.

[0035] The second and third parts of the data structure have an input-output relationship, taking the context information 23 as input and outputting the mixed drink formula design, including raw materials, dosage, description, etc. The data structure mainly serves to combine the input context information 23 with the constraints 24 to generate a reasonable and creative mixed drink formula 25.

[0036] Furthermore, the input data is multimodal data 21, environmental data 22 and user conversation text. The multimodal data 21 and environmental data 22 have been standardized into textual data through a large language model and spliced.

[0037] Furthermore, the output data is context information 23, including user information, experienced stories, current geographical location, climate characteristics, beverage flavors, preferences, and other information.

[0038] Furthermore, the input data is context information 23 and restrictive conditions 24, where the restrictive conditions 24 refer to specific requirements or restrictions that need to be followed when generating a beverage recipe 25, such as the type of raw materials, the taste of the beverage, health requirements, etc.

[0039] Furthermore, the output data is a mixed drink recipe 25, and the structure includes description information. Through the constraints of the data structure, the recipe part not only meets the input raw material type restrictions, but also meets the name and description of the output drink. The output data is displayed in a strictly formatted JSON format, and each key-value pair clearly defines the response result of each variable in the input. For example, the output data may contain detailed information and proportions of the raw materials, step instructions, etc., to ensure the integrity and standardization of the data structure, which is convenient for subsequent system integration and automated processing.

[0040] Optionally, in a specific embodiment, the input data may include ingredient information: "rum", "gin", "Baileys", "strawberry puree", "peach puree"; the story scenario is: "Today is the National Day holiday, and I feel very happy in this small shop"; the preference information is: "No alcohol, sour drinks".

[0041] Based on the input information, the system generates the output data as follows: {"ingredients":[{"name":"Strawberry puree","proportion":0.3},{"name":"Peach puree","proportion":0.2},{"name":"Sparkling water","proportion":0.5}],"title":"Fruit sparkling drink","description":"A refreshing fruit sparkling drink, suitable for enjoying in the cheerful atmosphere of National Day. The sweetness of the fruit and the joy of the bubbles allow you to enjoy the wonderful time of the holiday.","type":"nonalcoholic"}.

[0042] This data structure provides a complete training sample framework for the AI ​​model by establishing the association between contextual information, recipe elements, and literary expression. Each dimension contains multiple subdivision fields to capture feature information at different levels, ensuring that the generated recipes meet professional standards while reflecting personalized emotional expression.

[0043] Furthermore, based on the data structure proposed above, the data set is constructed and the model is trained.

[0044] like Figure 3 As shown in the figure, we first artificially construct a high-quality small sample data set, and use it to expand the data by combining beverage design cases in literary works. Then we use this data set to fine-tune the AI ​​model so that it can initially generate expected results in the target field. On this basis, we use the pre-adjusted model to continuously generate new data samples, and the expert team evaluates and screens the generated results, adding high-quality data to the training data set. Through multiple iterations, we continuously improve the generalization ability and generation effect of the model, and finally enable the AI ​​model to have the ability to perform situational analysis and beverage generation.

[0045] Figure 4 A mixed drink recipe AI generation model training method proposed by the present invention is shown.

[0046] In step S401, a small sample data set is first manually written to construct an initial training data set. Each piece of data should contain data mapped from dialogue information, multimodal data and environmental data to context information, and data output as mixed drink recipe information based on context information input (including user information, beverage taste, preference, etc.) and output (including raw materials, dosage, description, etc.).

[0047] Furthermore, the expert review team will analyze the contextual information in each piece of data in detail and design beverage recipes based on these analysis results. Each design must meet the following requirements: safety of material use, rationality of ingredient combination, and literary expression. In addition, experts will ensure that the beverage description meets the needs of the context and is emotionally resonant, making the recipe more artistic and attractive.

[0048] Given a multimodal information and user input information, expert reviewers conduct data writing, context analysis, and context-based beverage recipe design in sequence. The model's capture of literary color, emotional value, and deep meaning mainly relies on the support of professional knowledge.

[0049] For "literary color", the model will learn an expression with emotional extension, which can incorporate a sense of scene into the description of beverages, such as "slightly sweet and sour lemon sparkling water brings a refreshing feeling of spring", and other descriptions with literary flavor, creating an atmosphere while meeting users' needs for personalized experience.

[0050] Emotional value is reflected in the fusion of the taste of the beverage and the scene. For example, in a leisure and relaxation scene, the system tends to recommend refreshing, fruity recipes, while in a quiet and solitary occasion, the recipe tends to be milder and mellower in taste.

[0051] Industry practices and safety guide the model to follow traditional blending methods, process requirements, and material taboos when generating recipes, ensuring that beverages meet industry standards while expressing creativity and are easy to use in practice.

[0052] For multimodal data and user input information, as well as context-based beverage design, reference is also made to beverage designs in existing literature. Specifically, the system collects beverage descriptions, recipe ingredients, and other content from professional bartender websites and literature related to food recipes, and pre-processes the data to remove noise information in unstructured text. The data is formatted and standardized in the post-crawling processing stage, and is processed and rewritten by review experts to ensure that the information is consistent with the existing data structure and the design method is the same.

[0053] In the process of constructing the initial training data set, according to different task requirements, the specific content of the input is changed, and the output content is rewritten so that the model data set responds to different variables. For example, for the "user preference" variable, multiple output data with different preferences can be written, and the specific requirements of the task can be adjusted to generate output results of different styles. This can cover diverse user needs and improve the generalization ability of the model.

[0054] Optionally, in a specific embodiment, the raw materials are "rum", "gin", "Baileys", "strawberry puree", "peach puree"; the preference is "no alcohol, sour drink". The task requirements include generating different beverage versions according to the preferences, such as two types of alcoholic and non-alcoholic. By adjusting the "preference" variable, for example, changing it to "alcoholic", and writing a beverage recipe with alcohol, the model can respond to different variables.

[0055] These data with similar but slightly different input information are called a "data group". When evaluating and screening data, each piece of data needs to be evaluated in detail, and finally the best data in each group is selected, and the excellent data in all groups are merged to form a high-quality data set to ensure the overall quality of the model training data. This method can cover more possibilities, generate diverse beverage recipes, and ultimately optimize the generation effect of the model, so that the model has a strong generalization ability and can effectively respond to complex and diverse user needs.

[0056] In the process of manually writing data or iterating models, the input data can be adjusted according to the needs of specific tasks. For example, some unnecessary variables can be deleted, new input variables can be added, or the description of existing variables can be modified to ensure that the generated data meets the latest requirements and improve the adaptability and flexibility of the model.

[0057] Optionally, in a specific embodiment, other task requirements and variables are added, such as auxiliary materials, and auxiliary material variables: "perfume lemon", "mint leaves", "cinnamon", "cocoa crumbs", etc. When the output information structure changes due to changes in input information and task requirements, it is necessary to add new key-value pairs while maintaining the original data structure. The result is as follows: {"ingredients":[{"name":"Strawberry puree","proportion":0.3},{"name":"Peach puree","proportion":0.2},{"name":"Sparkling water","proportion":0.5}],"title":"Fruit sparkling drink","description":"A refreshing fruit sparkling drink, suitable for enjoying in the cheerful atmosphere of National Day. The sweetness of the fruit and the joy of the bubbles allow you to enjoy the wonderful time of the holiday.","topping":"Perfume lemon","type":"nonalcoholic"}.

[0058] Through the above iterative adjustment method, the generalization and stability of the model can be continuously improved, so that it can respond more accurately to actual needs and meet the ever-changing market and user needs. In order to focus limited computing resources on valuable data and improve training efficiency, k% of the data is selected as the initial training set, and the rest is used as the evaluation data set for subsequent model performance evaluation and improvement. This method can ensure that the model focuses on high-quality data in the initial stage and reduce the impact of noise during training.

[0059] In step S402, the model is initially fine-tuned using the selected data set, with the goal of enabling the model to generate expected output results in the target domain. During the fine-tuning process, the model parameters are adjusted for the specific task of beverage recipe generation to ensure that it can understand and generate high-quality recipes that match the input data. In this process, there are no specific restrictions on the model base, training cycle, number of steps, and training method, allowing for flexible settings based on actual conditions to achieve the best training effect and model performance.

[0060] During the fine-tuning process, the initial data set should contain at least 30 pieces of data, which should not only be of high quality in content but also cover sufficiently diverse scenarios to enhance the generalization ability of the model and avoid overfitting in practical applications. The embodiments of the present invention do not restrict the specific amount of data.

[0061] In a specific embodiment, the initial data set contains multiple different beverage generation tasks, which cover a variety of scenarios and needs, such as the generation of non-alcoholic beverages, preferences for specific flavors (such as sour, sweet, bitter, etc.), and beverage design in festival-themed scenarios. In each task, the input data includes the user's personalized needs, raw material preferences, and blending restrictions, and the output is structured recipe information. Through fine-tuning training on these data, the model can accurately match the input requirements when generating beverage recipes, and show strong generalization capabilities in different scenarios, ensuring flexible response to diverse needs in practical applications.

[0062] In step S403, the test data set and other input data are used to make predictions through the fine-tuned model and collect output data. This process evaluates the quality of the output data through quantitative scoring.

[0063] In a specific embodiment, the test data set and the new data generated by the internal test are used as input, and the fine-tuned model is used for prediction. In the output results, for each data group, data quality assessment is performed, focusing on whether the generated output meets the task requirements of the input information, the effect of personalized customization, the richness of literary color, whether it conforms to industry practices, whether the ingredient description is accurate, and the diversity of recipe styles, etc., to ensure that the generated recipes meet both personalized needs and industry standards.

[0064] In a specific embodiment, each piece of raw data is quantitatively scored using the following formula:

[0065] in, is the quantitative scoring result of the current sample data, is the weight of task completion. Rate the task completion. The weight of the raw material accuracy, Score the accuracy of ingredients, The weight of literary color, Rate the literary color, The weight of diversity, Score diversity, is the weight of compliance with industry standards, Score compliance with industry standards.

[0066] Furthermore, for each piece of raw data, its diversity score needs to be calculated. The diversity score of the embodiment of the present invention can use the Shannon's Diversity Index algorithm to measure the rarity of the raw materials in each formula, thereby reflecting the difference between the formula and other formulas. Quantitative scoring is performed using the following formula:

[0067] in, Score diversity, is the frequency (proportion) of a certain raw material in the data set, that is, . The operator returns the maximum value on either side of the comma, which is a maximum diversity score of 20. is the total number of ingredients in the current recipe, It is the index value of a raw material in the current recipe.

[0068] Furthermore, the task completion score, literary color score, raw material accuracy score and industry standard compliance score are graded in a step-by-step manner according to the manual review criteria. These scores can be divided according to the criteria shown in Table 1: Table 1

[0069] In step S404, all data in the data group are sorted, and several data with the highest scores are selected as components of the final data set. The high-quality data results are integrated into the incremental data set, and the large language model base is fine-tuned again, so that the model can gradually improve the accuracy and diversity of generated beverage recipes.

[0070] In step S405, during the formal application of the model, each time a beverage recipe is generated, user satisfaction is collected, and immediate feedback from users on the generated recipe is obtained, including scores on taste, texture, whether personalized needs are met, and other subjective feelings of the user. All generated recipes and corresponding satisfaction scores will be automatically archived by the system and associated with the user's personal data. These archived data will serve as the basis for continuous improvement of the model. The system will analyze common problems or suggestions in user feedback through a feedback iteration mechanism, and adjust and optimize model parameters. Each periodic iteration will involve new data input to ensure that the model can continuously improve its accuracy and adaptability based on changes in market demand and upgrades in user personalized needs.

[0071] Therefore, after the above-mentioned preset mixed beverage formula generation model is trained, the preset mixed beverage formula generation model has the ability to perform situational analysis and beverage generation.

[0072] The following describes a mixed drink recipe generation system proposed by the present invention using the mixed drink recipe generation model.

[0073] Specifically, Figure 5 A block diagram of a system for generating a mixed beverage formula provided by an embodiment of the present invention.

[0074] like Figure 5 As shown, the mixed beverage formula generation system 10 includes: a collection module 100, an extraction module 200 and a generation module 300.

[0075] Among them, the collection module 100 is used to collect the user's current multimodal data, current environmental data and current constraints when there is a need to generate a contextualized beverage recipe; the extraction module 200 is used to extract features from the current multimodal data and current environmental data to obtain current context information, wherein the current context information includes multiple context features; the generation module 300 is used to input multiple context features and current constraints into a preset mixed beverage recipe generation model to obtain the weights of multiple recipe elements in the mixed beverage that match the current context information, the name of the mixed beverage and the recipe description of the mixed beverage, so as to generate a mixed beverage recipe based on the weights of multiple recipe elements in the mixed beverage, the name of the mixed beverage and the recipe description of the mixed beverage, wherein the preset mixed beverage recipe generation model is obtained by fine-tuning the original base model using a preset data set.

[0076] In some embodiments, the current multimodal data includes at least one of the user's current voice data, the user's current visual feature data, and the user's current interaction data.

[0077] Specifically, the mixed drink recipe generation system 10 of the embodiment of the present invention generates a drink style and recipe based on the contextual drink design principle and the user's current context information (including characters, events, scenes, environment, etc.). In detail, the physical entity of the mixed drink shows a unique flavor and texture combination through the blending of different raw materials; the text description of the mixed drink creates an emotional resonance through the close combination of name, description and context. For example, the drink description can be combined with rich text and context, making the drink a continuation and extension of the current context, allowing users to feel the emotional connection and atmosphere gain brought by the drink.

[0078] The people, things and scenes in the situational information will directly affect the style and formula of the beverage. The entity of the mixed beverage can shape the emotional resonance of the beverage by blending the raw materials, matching different flavors and textures, or the textual description of the mixed beverage can shape the emotional resonance of the beverage by combining the explanation and title of the beverage with the situation, making the beverage a continuation and extension of the current situation.

[0079] For example, in active social scenes (such as parties or gatherings), bartenders usually choose colorful drinks with bubbles or citrus aromas to create a lively atmosphere. The stimulation of bubbles and fruity flavors can enhance the interactive experience and create a lively atmosphere in social occasions. In contrast, in quiet scenes (such as book bars or small salons), bartenders prefer recipes such as vanilla, cinnamon or espresso to create a warm and private atmosphere, so that the drinks match the mood of the environment.

[0080] For example, in a winter scene of solitude, a warm chocolate drink or a hot drink with a cinnamon aroma can bring emotional comfort to users. In hot summer outdoor activities, refreshing drinks can better meet the needs of users and bring them a cool and comfortable experience. This kind of drink adjusted according to the environmental situation can further enhance the emotional fit of users.

[0081] For example, the description of a drink designed for an outdoor spring outing might be: "This drink is like a spring breeze, and the light lemon scent brings a pleasant mood." This type of description not only meets the flavor requirements of the drink, but also combines the context, using the drink as a continuation of the spring outing experience, thereby enhancing the user's emotional connection.

[0082] Furthermore, the mixed drink recipe generation system 10 of the embodiment of the present invention captures multimodal data in real time through a multimodal interactive system, including the user's personal story, emotions, behaviors, etc. The multimodal interactive system first uses automatic speech recognition (ASR) technology to convert the user's voice input into text form in real time, and analyzes the user's emotions using the tone, rhythm and speech speed in the voice. Then, the system obtains visual features such as the user's facial expression, movements and clothing through image analysis technology. In addition, the embodiment of the present invention can also obtain the specific environmental conditions (such as current weather, temperature, humidity, etc.) of the user's location through external data sources such as GPS positioning and weather query, and combine multimodal data for subsequent analysis of context information.

[0083] Furthermore, in the process of collecting the user's current multimodal data and current environmental data, the system fully considers the user's personal needs and dynamic context. For example, if the user's voice and facial expression show a happy state, and the system detects that the user is in an outdoor social scene, the system will tend to recommend a refreshing and social beverage recipe for the user. When implementing this function, the context information collection module will cooperate with the interaction module to further clarify the user's needs through multiple rounds of dialogue with the user, such as "How are you feeling today?" or "Is there anything special you want to talk about?" to help the system obtain more detailed context information.

[0084] Furthermore, in some embodiments, the acquisition module 100 includes: a voice acquisition component for acquiring the user's current voice data; a video acquisition component for acquiring the user's current visual feature data; and an interaction component for acquiring the current interaction data with the user.

[0085] Specifically, the voice collection component of the embodiment of the present invention converts the voice data in the natural conversation between the user and the system into an accurate text transcription through automatic speech recognition (ASR) technology, ensuring that every detail of the conversation is captured and processed. At the same time, by analyzing the characteristics of the voice such as tone, rhythm, and speaking speed, the user's emotional information is extracted and converted into a text description of the emotion. In addition, the current acquisition of visual feature data includes the user's expression, action, clothing, and appearance in front of the device. After being analyzed and processed by the deep learning model, this information forms a detailed description text, which provides comprehensive emotional and visual feature input for the beverage generation model.

[0086] Furthermore, the voice collection component of the embodiment of the present invention uses automatic speech recognition (ASR) technology to convert the user's voice into text information, and extract parameters such as pitch, volume and rhythm to analyze the emotional characteristics in the voice. Text input is usually directly entered by the user in the interactive device, and the system will send it together with the voice text into the natural language processing (NLP) engine to conduct a comprehensive analysis of the emotional vocabulary, semantic logic and potential emotional expression in the text. Through this multimodal information collection method, the system can obtain the user's multi-dimensional emotional state and provide data support for subsequent emotional analysis.

[0087] In a specific embodiment, the multimodal interaction system recognizes the user's voice input as: "I have a great time today, and I want to celebrate with a drink with a sweet fruit flavor." The system also determines the user's emotions as excited and high-pitched through the tone and speed of the voice. In addition, through video analysis, the system determines that the user is about 25-35 years old, female, wearing casual clothes, and there are two female friends who seem to be the same age next to him. Combining these emotional and environmental information, the system infers that the user is currently in a pleasant social environment. Therefore, in the process of generating beverage recipes, it is more inclined to choose non-alcoholic beverages with sweet, tropical fruit flavors to further amplify their current emotions and celebratory atmosphere.

[0088] Furthermore, the acquisition module 100 of the embodiment of the present invention collects the physical environment information of the user through built-in sensors and external data sources. The system can obtain the geographical information of the user's current location and the current weather conditions, including temperature, humidity, sunny or rainy, etc., through GPS positioning, weather query, etc. The system integrates this environmental information into the input variables of the beverage generation model to ensure that the generated formula can adapt to the actual environmental conditions of the user.

[0089] Furthermore, the mixed drink recipe generation system 10 of the embodiment of the present invention incorporates special events or festival themes that may occur at the current location into the input information of the generation model through manually written descriptions. For example, if the user is at a summer music festival, the manually written description will include the lively atmosphere of the music festival, typical drink preferences, and active characteristics of the crowd. These descriptions will be used to generate a drink recipe that is more suitable for the current scene, thereby further enhancing the user's situational experience.

[0090] Optionally, in a specific embodiment, the mixed drink recipe generation system 10 reads that the user is at a summer music festival, and the weather is sunny and hot in the location. The system learns through voice input that the user wants a refreshing but non-alcoholic drink. Based on these inputs, the generation module designs a refreshing drink containing "lemon juice", "mint leaves" and "sparkling water". The recipe is based on the theme of freshness and health, which meets the needs of cooling off and non-alcoholic drinks.

[0091] Furthermore, the mixed drink recipe generation system 10 ensures that the generated recipe is feasible in the actual preparation process by obtaining possible restrictive conditions in the beverage preparation process, including available raw materials, equipment restrictions, preparation time, etc. These restrictive conditions may come from user input, environmental conditions or system settings, and the system will automatically filter out recipe elements that do not meet the conditions to ensure that the generated beverage recipe meets the feasibility of actual operation.

[0092] In one embodiment, the user's input indicated that he was allergic to drinks containing caffeine, so the system excluded all caffeinated recipes when generating the beverage. Finally, a tropical flavored decaffeinated beverage with "mango juice" and "coconut water" as the main ingredients was generated, ensuring the safety and satisfaction of the user.

[0093] Thus, all the information acquired by the acquisition module 100, including the user's text conversation transcription, voice emotion description, expression and action description in the video, as well as the environment's weather, location information, system restrictive conditions, etc., are comprehensively converted into input data for guiding the generation of beverage recipes.

[0094] Furthermore, based on the multimodal information collected above, the extraction module 200 of the embodiment of the present invention constructs a mapping to generate a context feature vector by analyzing the multimodal data and environmental data to describe the user's personalized needs, emotional state, activity environment, etc.

[0095] Specifically, after the multimodal data (voice, image, text, etc.) and environmental data are standardized and structured, specific situational features (such as emotional state, preference, location, environment, etc.) are first extracted, and then these features are fused to generate a higher-dimensional situational feature set to ensure that each situational feature can be effectively applied when generating beverage recipes, providing users with personalized and emotional beverage designs.

[0096] Furthermore, the key to converting multimodal data and environmental data into contextual features lies in the process of standardization and structuring. Through the multimodal large language model (LLM), user voice, text, video and environmental data are standardized into a consistent text data format. This process uses natural language processing (NLP) technology to identify and extract implicit emotional features and semantic content in the text, as well as visual features in the video. After integration, this information forms data input in text format to ensure that each contextual information can be effectively converted into a unified contextual feature.

[0097] The feature fusion method of the embodiment of the present invention uses a specific data structure to combine standardized contextual information (including user needs, current environmental characteristics, emotional state, etc.) in order of priority. This data structure can reasonably allocate and parse contextual data from different sources, so that the system can accurately apply the weight of each contextual information when generating beverage recipes, thereby outputting the best beverage recommendation that meets the specific situation.

[0098] Furthermore, by establishing an association between the contextual features and the mixed beverage recipe results, the system can directly map the above-mentioned contextual features to the beverage recipe elements.

[0099] Specifically, a correlation model between situational features and mixed drink recipe results is established, and deep learning algorithms are used to directly map user situational features (such as emotions, environment, and taste preferences) to key elements of drink recipes. This model uses deep learning algorithms to dynamically analyze the impact of user characteristics on recipe elements (such as raw materials, mixing ratios, and presentation methods), thereby generating personalized drink recipes that are highly matched to the user's current situation. The mixed drink recipe generation system will adjust the recipe combination of the drink in real time based on situational features, making it more emotionally resonant and unique while meeting industry standards, and meeting the user's personalized situational needs.

[0100] Furthermore, after multiple situational features and current constraints are input into a preset mixed drink recipe generation model, an association between the situational features and the mixed drink recipe results is established to obtain weights of multiple recipe elements in the mixed drink that match the current situational information, the name of the mixed drink and the recipe description of the mixed drink, and a mixed drink recipe is generated based on the weights of multiple recipe elements in the mixed drink, the name of the mixed drink and the recipe description of the mixed drink.

[0101] Therefore, in the process of generating a mixed drink formula, the embodiment of the present invention not only takes into account industry professional standards, but also incorporates literary colors and cultural attributes, so that the generated formula has more emotional resonance and uniqueness.

[0102] Furthermore, in some embodiments, the mixed drink recipe generation system 10 also includes: a training module, wherein the training module is specifically used to: construct an initial training data set based on a preset data structure; use the initial training data set to fine-tune the original base model to obtain an initial pre-adjusted model; input the data in the preset test data set into the initial pre-adjusted model to generate new sample data, and evaluate and screen the new sample data according to a preset quantitative scoring formula to obtain data that meets preset evaluation conditions, and add the data that meets the preset evaluation conditions to the initial training data set to obtain a new training data set; use the new training data set to iteratively train the initial pre-adjusted model, and verify the iteratively trained initial pre-adjusted model until the iteratively trained initial pre-adjusted model meets the preset verification conditions, and use the iteratively trained initial pre-adjusted model as the preset mixed drink recipe generation model.

[0103] Furthermore, in some embodiments, the preset quantitative scoring formula is:

[0104] in, is the quantitative scoring result of the current sample data, is the weight of task completion. Rate the task completion. The weight of the raw material accuracy, Score the accuracy of ingredients, The weight of literary color, Rate the literary color, The weight of diversity, Score diversity, is the weight of compliance with industry standards, Score compliance with industry standards.

[0105] Among them, fine-tuning the original base model using the initial training data set refers to the process of further training the original base model using the initial training data set for a specific task, that is, while retaining most of the parameters of the original base model, the weights and parameters of the model are adjusted by performing a small amount of iterative training on the initial training data set to make it more focused on solving specific problems or adapting to specific scenarios.

[0106] It should be noted that the specific training process of the preset mixed beverage formula generation model in the embodiment of the present invention has been described in detail above, and will not be described in detail here to avoid redundancy.

[0107] Furthermore, in some embodiments, the mixed drink recipe generation system 10 also includes: a feedback module for obtaining user satisfaction and feedback information on the mixed drink recipe; and an optimization module for adjusting and optimizing model parameters of a preset mixed drink recipe generation model based on user satisfaction and feedback information on the mixed drink recipe.

[0108] Specifically, after the mixed drink recipe generation system 10 generates the mixed drink recipe, the feedback module will obtain the user's satisfaction score and feedback information on the generated mixed drink recipe, and the scoring dimensions may include taste, texture, degree of satisfaction of personalized needs, and overall experience, etc. The feedback information may include specific suggestions for the recipe (such as adjusting the proportion of a certain raw material, changing the description style, etc.) or expressing preferences for certain aspects (such as literary color or emotional resonance).

[0109] Furthermore, all satisfaction ratings and feedback information provided by users will be automatically recorded by the system and archived in association with the user's personal data (such as contextual information, preference settings, etc.) to form a continuously growing feedback database.

[0110] Furthermore, the optimization module will analyze user feedback data to identify common problems or directions for improvement. Based on the analysis results, the optimization module will dynamically adjust the key parameters of the model, including model parameter adjustment, raw material selection weights, ratio rules, or literary description styles, etc., which are not specifically limited here. For example, if multiple users report that a certain type of recipe tastes sweet, the system will adjust the ratio parameters of the relevant raw materials.

[0111] Furthermore, the optimization module will reapply the adjusted parameters to the model and use the newly added feedback data for fine-tuning training to further improve the generalization ability and generation quality of the model. After each optimization, the system will evaluate the performance of the model to verify whether the improvement effect has achieved the expected goal. If the expected goal is not achieved, the parameters will continue to be adjusted and the above process will be repeated to continuously improve the generation ability and generalization of the model.

[0112] According to the generation system of the mixed drink formula of the embodiment of the present invention, the collection module is used to collect the user's current multimodal data, current environmental data and current constraints, and the extraction module is used to extract features of the current multimodal data and current environmental data to obtain the current context information; the generation module is used to input multiple context features and current constraints into the preset mixed drink formula generation model, and the weights of multiple formula elements in the mixed drink that match the current context information, the name of the mixed drink and the formula description of the mixed drink are obtained, and the mixed drink formula is generated. In this way, the problems of lack of literary expression and poor context adaptability of the background technology are solved, and personalized, emotional and industry-standard mixed drink formula generation is achieved by combining artificial intelligence and multi-dimensional context information.

[0113] The following describes a method for generating a mixed drink recipe according to an embodiment of the present invention with reference to the accompanying drawings. The method adopts the above-mentioned system for generating a mixed drink recipe.

[0114] like Figure 6 As shown, the method for generating the mixed beverage formula comprises the following steps: In step S601, when there is a demand for generating a contextualized beverage recipe, a collection module is used to collect the user's current multimodal data, current environmental data, and current restriction conditions.

[0115] In step S602, an extraction module is used to perform feature extraction on current multimodal data and current environmental data to obtain current context information, wherein the current context information includes a plurality of context features.

[0116] In step S603, a generation module is used to input multiple situational features and current constraints into a preset mixed drink recipe generation model to obtain weights of multiple recipe elements in the mixed drink that match the current situational information, the name of the mixed drink and the recipe description of the mixed drink, so as to generate a mixed drink recipe based on the weights of multiple recipe elements in the mixed drink, the name of the mixed drink and the recipe description of the mixed drink, wherein the preset mixed drink recipe generation model is obtained by fine-tuning the original base model using a preset data set.

[0117] Specifically, if Figure 7 As shown, an AI-based mixed beverage formula generation method of an embodiment of the present invention aims to establish a multi-dimensional context information collection and analysis mechanism, combine a multimodal feature mapping system and a context feature association model, analyze context information, including the user's personal emotions, preferences, identity characteristics, past stories, emotional states, described spaces, scenery and places, and other dimensions, and generate a beverage formula design based on this. The generated result not only meets the industry professional standards, but also meets the user's emotional expression and personalized experience needs.

[0118] Further, in some embodiments, the current multimodal data includes at least one of the user's current voice data, the user's current visual feature data, and the user's current interaction data.

[0119] Furthermore, in some embodiments, before using the generation module to input multiple situational features and current constraints into a preset mixed drink formula generation model, it also includes: constructing an initial training data set based on a preset data structure; using the initial training data set to fine-tune the original base model to obtain an initial pre-adjusted model; inputting data in a preset test data set into the initial pre-adjusted model to generate new sample data, and evaluating and screening the new sample data according to a preset quantitative scoring formula to obtain data that meets preset evaluation conditions, and adding the data that meets the preset evaluation conditions to the initial training data set to obtain a new training data set; using the new training data set to iteratively train the initial pre-adjusted model, and verifying the iteratively trained initial pre-adjusted model until the iteratively trained initial pre-adjusted model meets the preset verification conditions, and using the iteratively trained initial pre-adjusted model as the preset mixed drink formula generation model.

[0120] Furthermore, in some embodiments, after generating a mixed drink formula based on the weights of multiple formula elements in the mixed drink, the name of the mixed drink and the formula description of the mixed drink, it also includes: using a feedback module to obtain user satisfaction and feedback information on the mixed drink formula; using an optimization module based on user satisfaction and feedback information on the mixed drink formula, adjusting and optimizing model parameters of a preset mixed drink formula generation model.

[0121] In order to facilitate those skilled in the art to understand the method for generating the mixed drink formula proposed by the present invention more clearly and intuitively, the following is a detailed description of the method. Figure 8 Provide detailed explanation.

[0122] Specifically, if Figure 8 As shown, the method for generating the mixed beverage formula comprises the following steps: S801, obtain multimodal data such as video and audio through a multimodal system, and obtain environmental data such as weather and sensor information through sensors.

[0123] S802, the AI ​​model analyzes multimodal data and environmental data to generate situational analysis results.

[0124] S803, combining the situation analysis results with the restrictive conditions, generating beverage production tasks and requirements.

[0125] S804, the AI ​​model generates a structured recipe based on demand.

[0126] It should be noted that the above explanations of the embodiment of the system for generating a mixed drink formula are also applicable to the method for generating a mixed drink formula of this embodiment, and will not be repeated here.

[0127] According to the method for generating a mixed drink formula in an embodiment of the present invention, a collection module is used to collect the user's current multimodal data, current environmental data and current constraints, and an extraction module is used to extract features from the current multimodal data and current environmental data to obtain current context information; a generation module is used to input multiple context features and current constraints into a preset mixed drink formula generation model to obtain the weights of multiple formula elements in the mixed drink that match the current context information, the name of the mixed drink and the formula description of the mixed drink, and generate a mixed drink formula. Thus, the problems of lack of literary expression and poor context adaptability of the background technology are solved, and personalized, emotional and industry-standard mixed drink formula generation is achieved by combining artificial intelligence and multi-dimensional context information.

[0128] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0129] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0130] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A system for generating a mixed drink recipe, characterized in that: include: A collection module, used to collect the user's current multimodal data, current environmental data and current constraints when there is a need for contextual beverage recipe generation; An extraction module, configured to extract features from the current multimodal data and the current environment data to obtain current context information, wherein the current context information includes a plurality of context features; A generation module is used to input the multiple situational features and the current constraints into a preset mixed drink formula generation model to obtain weights of multiple formula elements in the mixed drink that match the current situational information, the name of the mixed drink and the formula description of the mixed drink, so as to generate a mixed drink formula according to the weights of multiple formula elements in the mixed drink, the name of the mixed drink and the formula description of the mixed drink, wherein the preset mixed drink formula generation model is obtained by fine-tuning the original base model using a preset data set.

2. The system for generating a mixed drink recipe according to claim 1, characterized in that: The current multimodal data includes at least one of the current voice data of the user, the current visual feature data of the user, and the current interaction data of the user.

3. The system for generating a mixed drink recipe according to claim 1 or 2, characterized in that: The acquisition module comprises: A voice collection component, used to collect the current voice data of the user; A video acquisition component, used to acquire current visual feature data of the user; The interaction component is used to collect current interaction data with the user.

4. The system for generating a mixed drink recipe according to claim 1, characterized in that: Also includes: A training module, wherein the training module is specifically used for: Based on the preset data structure, construct the initial training data set; Using the initial training data set to perform fine-tuning training on the original base model to obtain an initial pre-tuned model; Inputting the data in the preset test data set into the initial pre-adjusted model to generate new sample data, and evaluating and screening the new sample data according to a preset quantitative scoring formula to obtain data that meets the preset evaluation conditions, and adding the data that meets the preset evaluation conditions to the initial training data set to obtain a new training data set; The initial pre-adjustment model is iteratively trained using the new training data set, and the iteratively trained initial pre-adjustment model is verified until the iteratively trained initial pre-adjustment model meets the preset verification conditions, and the iteratively trained initial pre-adjustment model is used as the preset mixed beverage formula generation model.

5. The system for generating a mixed drink recipe according to claim 4, characterized in that: The preset quantitative scoring formula is: in, is the quantitative scoring result of the current sample data, is the weight of task completion. Rate the task completion. The weight of the raw material accuracy, Score the accuracy of ingredients, The weight of literary color, Rate the literary color, The weight of diversity, Score diversity, is the weight of compliance with industry standards, Score compliance with industry standards.

6. The system for generating a mixed drink recipe according to claim 1, characterized in that: Also includes: A feedback module, used to obtain user satisfaction and feedback information on the mixed beverage formula; The optimization module is used to adjust and optimize the model parameters of the preset mixed drink formula generation model based on the user's satisfaction and feedback information on the mixed drink formula.

7. A method for generating a mixed drink formula, characterized in that: A system for generating a mixed drink recipe according to any one of claims 1 to 6, wherein the method comprises the following steps: When there is a need to generate a contextualized beverage recipe, the acquisition module is used to acquire the user's current multimodal data, current environmental data, and current constraints; Using the extraction module to extract features from the current multimodal data and the current environment data to obtain current context information, wherein the current context information includes a plurality of context features; The generation module is used to input the multiple situational features and the current constraints into a preset mixed drink formula generation model to obtain weights of multiple formula elements in the mixed drink that match the current situational information, the name of the mixed drink and the formula description of the mixed drink, so as to generate a mixed drink formula according to the weights of multiple formula elements in the mixed drink, the name of the mixed drink and the formula description of the mixed drink, wherein the preset mixed drink formula generation model is obtained by fine-tuning the original base model using a preset data set.

8. The method for generating a mixed drink recipe according to claim 7, characterized in that: The current multimodal data includes at least one of the current voice data of the user, the current visual feature data of the user, and the current interaction data of the user.

9. The method for generating a mixed drink recipe according to claim 7, characterized in that: Before using the generation module to input the plurality of situational features and the current restriction conditions into a preset mixed drink formula generation model, the method further includes: Based on the preset data structure, construct the initial training data set; Using the initial training data set to perform fine-tuning training on the original base model to obtain an initial pre-tuned model; Inputting the data in the preset test data set into the initial pre-adjusted model to generate new sample data, and evaluating and screening the new sample data according to a preset quantitative scoring formula to obtain data that meets the preset evaluation conditions, and adding the data that meets the preset evaluation conditions to the initial training data set to obtain a new training data set; The initial pre-adjustment model is iteratively trained using the new training data set, and the iteratively trained initial pre-adjustment model is verified until the iteratively trained initial pre-adjustment model meets the preset verification conditions, and the iteratively trained initial pre-adjustment model is used as the preset mixed beverage formula generation model.

10. The method for generating a mixed drink recipe according to claim 7, characterized in that: After generating a mixed drink formula according to the weights of the plurality of formula elements in the mixed drink, the name of the mixed drink and the formula description of the mixed drink, the method further includes: Using a feedback module to obtain user satisfaction and feedback information on the mixed beverage formula; The optimization module is used to adjust and optimize the model parameters of the preset mixed drink formula generation model based on the user's satisfaction and feedback information on the mixed drink formula.

Images