A voice-interactive cooking parameter generation method and system

Abstract

This invention relates to the field of automated cooking control technology, and discloses a voice-interactive cooking parameter generation method and system. By recognizing and semantically parsing the user's voice cooking commands, structured information containing core ingredient entities and at least one differentiating feature is obtained. The differentiating feature indicates personalized requirements for the cooking process and / or cooking result of the core ingredient entity. Based on this, benchmark cooking parameters corresponding to the core ingredient entity are retrieved from a pre-set cooking parameter knowledge base. The benchmark cooking parameters and the differentiating feature are input into a large language model trained with cooking domain data. The large language model outputs correction information to generate personalized cooking parameters. Finally, the parameters undergo security verification; if successful, a control command is generated and issued to the cooking appliance for execution. Therefore, implementing this invention can improve the accuracy and comprehensiveness of cooking appliances in understanding the cooking needs in user voice commands, thereby improving the accuracy of cooking parameter generation.

Description

Technical Field

[0001] This invention relates to the field of cooking automation control technology, and in particular to a voice-interactive cooking parameter generation method and system. Background Technology

[0002] With the development of artificial intelligence technology, smart cooking appliances that support voice control are becoming increasingly popular. Users can conveniently start the corresponding standardized cooking programs preset in the appliance by saying commands such as "steam fish" or "cook porridge".

[0003] However, in practice, it has been found that this control method has limited adaptability. When users give more complex voice commands, such as "cook frozen egg tarts," "process steaks just taken out of the freezer," "want them a little charred," or "for the baby," existing cooking appliances often cannot fully and accurately understand the complete requirements. This manifests as either displaying an error message and being unable to execute the command, or selectively executing its internally preset default program. This often leads to a significant deviation between the cooking result and the user's actual expectations, such as frozen ingredients being undercooked or requiring a crispy texture but not being cooked enough.

[0004] Therefore, it is particularly important to propose a technical solution that improves the accuracy and comprehensiveness of cooking appliances in understanding the cooking needs in users' voice commands, thereby improving the accuracy of cooking parameter generation. Summary of the Invention

[0005] This invention provides a voice-interactive cooking parameter generation method and system, which can improve the accuracy and comprehensiveness of cooking appliances in understanding the cooking needs in user voice commands, thereby improving the accuracy of cooking parameter generation.

[0006] To address the aforementioned technical problems, the first aspect of this invention discloses a voice-interactive cooking parameter generation method, the method comprising: Receive user's voice cooking instructions; The voice cooking instructions are recognized and semantically parsed to obtain corresponding structured information containing core ingredient entities and at least one differentiating feature. The differentiating feature is used to indicate personalized requirements for the cooking process and / or cooking results of the core ingredient entities. Based on the core ingredient entity, the benchmark cooking parameters corresponding to the core ingredient entity are retrieved from the preset cooking parameter knowledge base; The baseline cooking parameters and at least one of the differential features are input into a large language model trained with cooking domain data. The large language model outputs correction information for the baseline cooking parameters and generates personalized cooking parameters based on the correction information. The personalized cooking parameters are subject to security verification. If the verification passes, control instructions are generated based on the personalized cooking parameters and sent to the cooking appliances for execution.

[0007] As an optional implementation, in the first aspect of the present invention, the step of recognizing and semantically parsing the voice cooking command to obtain corresponding structured information containing core ingredient entities and at least one differentiated feature includes: Semantic parsing is performed on the text instructions obtained from the voice cooking instructions to identify the core ingredient entities from the text instructions, and at least one modifier entity is identified to define the cooking process or cooking result of the core ingredient entities. Semantic classification is performed on at least one of the said modifier entities to classify each of the said modifier entities into a first type of feature corresponding to pre-cooking conditions or a second type of feature corresponding to post-cooking expectations, wherein the first type of feature is used to represent physical conditions applied to the core ingredient entity that exist before cooking begins, and the second type of feature is used to represent requirements for the sensory and / or physical properties that the core ingredient entity is expected to present after cooking. Based on the core ingredient entity and the first and / or second type features obtained after classification, structured information is generated, wherein the first and second type features constitute at least one of the differentiated features.

[0008] As an optional implementation, in a first aspect of the invention, the semantic classification of at least one of the modifying entities to classify each of the modifying entities into a first type of feature corresponding to pre-cooking conditions or a second type of feature corresponding to post-cooking expectations includes: Based on the complete semantics of the text instructions, a contextual semantic field is constructed to describe the current cooking task; Obtain typical cooking operation sequences associated with the core ingredient entity from a pre-set cooking process knowledge base; The semantics of the modifier entity are matched and parsed with the context semantic field and the typical cooking operation sequence to determine the attribute content described by the modifier entity and the associated cooking stage. In response to determining that the attribute content describes the physical properties of the core ingredient entity that can be directly observed or measured before the cooking operation begins, the decorative entity is classified as the first type of feature. In response to the determination that the attribute content describes the attributes presented by the core ingredient entity that can only be determined through sensory evaluation or physical measurement after the cooking operation is completed, the modifying entity is classified as the second type of feature.

[0009] As an optional implementation, in the first aspect of the present invention, retrieving the benchmark cooking parameters corresponding to the core ingredient entity from a preset cooking parameter knowledge base based on the core ingredient entity includes: Based on the type of the differentiated features contained in the structured information, a parameter retrieval strategy is determined for the core ingredient entity; According to the parameter retrieval strategy, at least one set of candidate benchmark parameters associated with the core ingredient entity is retrieved from the preset cooking parameter knowledge base. Based on the differentiated features in the structured information, at least one set of candidate benchmark parameters is subjected to context-adaptive filtering to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one set of candidate benchmark parameters.

[0010] As an optional implementation, in the first aspect of the present invention, the step of performing context-adaptive filtering on at least one set of candidate benchmark parameters based on the differential features in the structured information to determine benchmark cooking parameters corresponding to the core ingredient entity from at least one set of candidate benchmark parameters includes: For each of the differentiated features in the structured information, determine the cooking effect influencing factor that the differentiated feature points to; For each candidate benchmark parameter in at least one of the candidate benchmark parameter sets, a scenario fit score between the candidate benchmark parameter and the differentiated feature is calculated based on the cooking effect influence factor. Based on the logical dependencies between the different differentiated features in the structured information, a priority integration strategy for integrating multiple scene adaptability scores is determined; Based on the priority integration strategy, the multiple scene adaptability scores corresponding to each candidate benchmark parameter are integrated, and the benchmark cooking parameters corresponding to the core ingredient entity are determined from the candidate benchmark parameter set according to the integration result.

[0011] As an optional implementation, in a first aspect of the invention, the step of inputting the benchmark cooking parameters and at least one of the differential features into a large language model trained on cooking domain data, and having the large language model output correction information for the benchmark cooking parameters, includes: Construct an inference input context that includes the benchmark cooking parameters, at least one of the differentiated features, and preset safety constraints; The reasoning input context is input into a large language model trained with data from the culinary domain, triggering the large language model to perform multi-stage chained reasoning that includes task decomposition, constraint injection, and corrective derivation. Receive the intermediate structured output generated by the large language model based on the multi-stage chained reasoning, which includes at least one cooking parameter correction item and its correction basis; Based on the at least one cooking parameter correction item in the intermediate structured output, the correction information for the baseline cooking parameters is generated.

[0012] As an optional implementation, in the first aspect of the present invention, the security verification of the personalized cooking parameters includes: Based on the personalized cooking parameters and the cooking parameter knowledge base, at least one security verification dimension associated with the current cooking task is determined; For each of the security verification dimensions, the verification benchmark and tolerance range corresponding to the security verification dimension are obtained. The verification benchmark and tolerance range are obtained based on the security rules in the cooking parameter knowledge base and the expected cooking process state derived from the personalized cooking parameters. In at least one of the aforementioned safety verification dimensions, the personalized cooking parameters or the process state parameters derived therefrom are compared with the corresponding verification benchmarks and tolerance ranges to generate compliance judgment results for each dimension. Based on the risk level predefined in the cooking parameter knowledge base and the structured information of the safety verification dimension, an importance weight for each safety verification dimension to cooking safety is generated; Based on the aforementioned importance weights, the compliance judgment results for each of the aforementioned security verification dimensions are weighted and calculated to generate a weighted security assessment score. The weighted safety assessment score is compared with a preset safety threshold, and the result of the comparison is used to output a judgment result on whether the personalized cooking parameters have passed the safety verification.

[0013] A second aspect of the present invention discloses a voice-interactive cooking parameter generation system, the system comprising: The receiving module is used to receive the user's voice cooking instructions; The recognition and parsing module is used to recognize and semantically parse the voice cooking instructions to obtain corresponding structured information containing core ingredient entities and at least one differentiated feature. The differentiated feature is used to indicate personalized requirements for the cooking process and / or cooking results of the core ingredient entities. The cooking retrieval module is used to retrieve the benchmark cooking parameters corresponding to the core ingredient entity from a preset cooking parameter knowledge base based on the core ingredient entity. The inference correction module is used to input the benchmark cooking parameters and at least one of the differential features into a large language model trained with cooking domain data, and the large language model outputs correction information for the benchmark cooking parameters, and generates personalized cooking parameters based on the correction information. A security verification module is used to perform security verification on the personalized cooking parameters. The execution control module is used to generate control commands based on the personalized cooking parameters and send them to the cooking appliances for execution if the security verification module passes the verification.

[0014] As an optional implementation, in the second aspect of the present invention, the specific method by which the recognition and parsing module recognizes and semantically parses the voice cooking instructions to obtain corresponding structured information containing core ingredient entities and at least one differentiated feature includes: Semantic parsing is performed on the text instructions obtained from the voice cooking instructions to identify the core ingredient entities from the text instructions, and at least one modifier entity is identified to define the cooking process or cooking result of the core ingredient entities. Semantic classification is performed on at least one of the said modifier entities to classify each of the said modifier entities into a first type of feature corresponding to pre-cooking conditions or a second type of feature corresponding to post-cooking expectations, wherein the first type of feature is used to represent physical conditions applied to the core ingredient entity that exist before cooking begins, and the second type of feature is used to represent requirements for the sensory and / or physical properties that the core ingredient entity is expected to present after cooking. Based on the core ingredient entity and the first and / or second type features obtained after classification, structured information is generated, wherein the first and second type features constitute at least one of the differentiated features.

[0015] As an optional implementation, in a second aspect of the invention, the specific manner in which the identification and parsing module performs semantic classification on at least one of the modifying entities to classify each of the modifying entities into a first type of feature corresponding to the pre-cooking conditions or a second type of feature corresponding to the post-cooking expectations includes: Based on the complete semantics of the text instructions, a contextual semantic field is constructed to describe the current cooking task; Obtain typical cooking operation sequences associated with the core ingredient entity from a pre-set cooking process knowledge base; The semantics of the modifier entity are matched and parsed with the context semantic field and the typical cooking operation sequence to determine the attribute content described by the modifier entity and the associated cooking stage. In response to determining that the attribute content describes the physical properties of the core ingredient entity that can be directly observed or measured before the cooking operation begins, the decorative entity is classified as the first type of feature. In response to the determination that the attribute content describes the attributes presented by the core ingredient entity that can only be determined through sensory evaluation or physical measurement after the cooking operation is completed, the modifying entity is classified as the second type of feature.

[0016] As an optional implementation, in a second aspect of the present invention, the cooking retrieval module retrieves the benchmark cooking parameters corresponding to the core ingredient entity from a preset cooking parameter knowledge base based on the core ingredient entity in the following specific ways: Based on the type of the differentiated features contained in the structured information, a parameter retrieval strategy is determined for the core ingredient entity; According to the parameter retrieval strategy, at least one set of candidate benchmark parameters associated with the core ingredient entity is retrieved from the preset cooking parameter knowledge base. Based on the differentiated features in the structured information, at least one set of candidate benchmark parameters is subjected to context-adaptive filtering to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one set of candidate benchmark parameters.

[0017] As an optional implementation, in a second aspect of the invention, the cooking retrieval module performs context-adaptive filtering on at least one set of candidate benchmark parameters based on the differential features in the structured information to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one set of candidate benchmark parameters. The specific method includes: For each of the differentiated features in the structured information, determine the cooking effect influencing factor that the differentiated feature points to; For each candidate benchmark parameter in at least one of the candidate benchmark parameter sets, a scenario fit score between the candidate benchmark parameter and the differentiated feature is calculated based on the cooking effect influence factor. Based on the logical dependencies between the different differentiated features in the structured information, a priority integration strategy for integrating multiple scene adaptability scores is determined; Based on the priority integration strategy, the multiple scene adaptability scores corresponding to each candidate benchmark parameter are integrated, and the benchmark cooking parameters corresponding to the core ingredient entity are determined from the candidate benchmark parameter set according to the integration result.

[0018] As an optional implementation, in a second aspect of the invention, the inference correction module inputs the baseline cooking parameters and at least one of the differential features into a large language model trained with cooking domain data, and the large language model outputs correction information for the baseline cooking parameters in the following specific manner: Construct an inference input context that includes the benchmark cooking parameters, at least one of the differentiated features, and preset safety constraints; The reasoning input context is input into a large language model trained with data from the culinary domain, triggering the large language model to perform multi-stage chained reasoning that includes task decomposition, constraint injection, and corrective derivation. Receive the intermediate structured output generated by the large language model based on the multi-stage chained reasoning, which includes at least one cooking parameter correction item and its correction basis; Based on the at least one cooking parameter correction item in the intermediate structured output, the correction information for the baseline cooking parameters is generated.

[0019] As an optional implementation, in the second aspect of the present invention, the specific method by which the security verification module performs security verification on the personalized cooking parameters includes: Based on the personalized cooking parameters and the cooking parameter knowledge base, at least one security verification dimension associated with the current cooking task is determined; For each of the security verification dimensions, the verification benchmark and tolerance range corresponding to the security verification dimension are obtained. The verification benchmark and tolerance range are obtained based on the security rules in the cooking parameter knowledge base and the expected cooking process state derived from the personalized cooking parameters. In at least one of the aforementioned safety verification dimensions, the personalized cooking parameters or the process state parameters derived therefrom are compared with the corresponding verification benchmarks and tolerance ranges to generate compliance judgment results for each dimension. Based on the risk level predefined in the cooking parameter knowledge base and the structured information of the safety verification dimension, an importance weight for each safety verification dimension to cooking safety is generated; Based on the aforementioned importance weights, the compliance judgment results for each of the aforementioned security verification dimensions are weighted and calculated to generate a weighted security assessment score. The weighted safety assessment score is compared with a preset safety threshold, and the result of the comparison is used to output a judgment result on whether the personalized cooking parameters have passed the safety verification.

[0020] A third aspect of the present invention discloses another voice-interactive cooking parameter generation system, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the voice-interactive cooking parameter generation method disclosed in the first aspect of the present invention.

[0021] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute the voice-interactive cooking parameter generation method disclosed in the first aspect of the present invention.

[0022] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: In this embodiment of the invention, a user's voice cooking command is received; the voice cooking command is recognized and semantically parsed to obtain corresponding structured information containing core ingredient entities and at least one differentiating feature, the differentiating feature being used to indicate personalized requirements for the cooking process and / or cooking result of the core ingredient entity; based on the core ingredient entity, benchmark cooking parameters corresponding to the core ingredient entity are retrieved from a pre-set cooking parameter knowledge base; the benchmark cooking parameters and at least one differentiating feature are input into a large language model trained with cooking domain data, the large language model outputs correction information for the benchmark cooking parameters, and personalized cooking parameters are generated based on the correction information; the personalized cooking parameters are subjected to security verification; if the verification passes, control commands are generated based on the personalized cooking parameters and sent to the cooking appliance for execution. It is evident that implementing this invention can improve the depth and structure of understanding user natural language commands by "recognizing and semantically parsing voice cooking commands to obtain structured information containing core ingredient entities and differentiating features," thereby facilitating the accurate extraction of the user's personalized cooking intent from ambiguous voice commands. The system can improve the reliability and scientific rigor of cooking parameter generation by retrieving baseline cooking parameters from a pre-built cooking parameter knowledge base based on core ingredient entities. This ensures that subsequent personalized adjustments are based on a safe and compliant baseline that conforms to culinary common sense. Furthermore, it can enhance the intelligence and adaptability of cooking parameter generation by inputting baseline cooking parameters and differentiated features into a large language model trained on culinary domain data. This allows the model to output correction information and generate personalized cooking parameters, transforming users' personalized, non-quantifiable descriptions (such as "crispy on the outside and tender on the inside") into concrete and actionable cooking parameter adjustments. Finally, it can improve the safety and reliability of the final output parameters by performing safety checks on personalized cooking parameters. This ensures that the cooking process complies with safety standards while meeting personalized needs, preventing safety risks caused by inappropriate parameters. Through the complete process described above, end-to-end automatic generation of cooking control commands, from fuzzy voice commands to safe, personalized, and executable commands, can be achieved. This improves the intelligence level of cooking control and user experience, allowing users to complete personalized cooking by driving kitchen appliances with natural voice commands without needing professional cooking knowledge. It also improves the accuracy and comprehensiveness of cooking appliances in understanding the cooking needs in user voice commands, thereby improving the accuracy of cooking parameter generation. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a voice-interactive cooking parameter generation method disclosed in an embodiment of the present invention; Figure 2 This is a flowchart illustrating another voice-interactive cooking parameter generation method disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a voice-interactive cooking parameter generation system disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of another voice-interactive cooking parameter generation system disclosed in an embodiment of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] This invention discloses a voice-interactive cooking parameter generation method and system. It improves the depth and structure of understanding user natural language commands by "recognizing and semantically parsing voice cooking instructions to obtain structured information containing core ingredient entities and differentiated features," thereby facilitating the accurate extraction of users' personalized cooking intentions from ambiguous voice commands. It enhances the reliability and scientific rigor of the starting point for cooking parameter generation by "retrieving benchmark cooking parameters from a pre-set cooking parameter knowledge base based on core ingredient entities," ensuring that subsequent personalized adjustments are based on a safe benchmark consistent with culinary common sense. It improves the intelligence and adaptability of cooking parameter generation by "inputting benchmark cooking parameters and differentiated features into a large language model trained with culinary domain data, from which the model outputs correction information and generates personalized cooking parameters," thus facilitating the transformation of users' personalized, non-quantitative descriptions (such as "crispy on the outside and tender on the inside") into specific, executable cooking parameter adjustments. Finally, it enhances the safety and reliability of the final output parameters by "performing safety checks on personalized cooking parameters," ensuring that the cooking process complies with safety standards while meeting personalized needs, preventing safety risks caused by inappropriate parameters. Through the complete process described above, end-to-end automatic generation of safe, personalized, and executable cooking control commands, from fuzzy voice commands, can be achieved. This improves the intelligence level of cooking control and user experience, allowing users to complete personalized cooking by driving kitchen appliances with natural voice commands without needing professional cooking knowledge. It also improves the accuracy and comprehensiveness of cooking appliances in understanding the cooking needs in user voice commands, thereby enhancing the accuracy of cooking parameter generation. These will be explained in detail below.

[0029] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a voice-interactive cooking parameter generation method disclosed in an embodiment of the present invention. Figure 1The described voice-interactive cooking parameter generation method can be applied to cooking appliances such as smart ovens, steam ovens, rice cookers, and pressure cookers—that is, devices involving "standard programs + personalized fine-tuning." It can also be applied to smart devices associated with the aforementioned cooking appliances, including but not limited to one or more of smart home devices, cloud devices, edge computing devices, relay devices, base station devices, urban management devices, and smart connected devices. This invention does not limit the scope of these devices. Furthermore, the execution of this invention can be uniformly scheduled by an agent module. This agent is responsible for receiving the speech recognition results, sequentially calling sub-modules such as receiving, recognition and parsing, cooking retrieval, inference correction, security verification, and execution control, and finally generating control commands. For example, after obtaining structured information, the agent first accesses the cooking parameter knowledge base to obtain baseline parameters, then combines the baseline parameters with differentiated features and submits them to the large language model. After receiving correction information and generating personalized parameters, it triggers the security verification process. Figure 1 As shown, the voice-interactive cooking parameter generation method may include the following operations: 101. Receive user's voice cooking instructions; In this embodiment of the invention, optionally, the user issues cooking commands in natural language via a voice acquisition module mounted on a smart cooking device (such as a smart stove, smart oven, or cooking robot) or a connected terminal (such as a mobile app or smart speaker). For example, commands could include: "Fry this frozen steak until it's crispy on the outside and tender on the inside," or "Cook a pot of slightly firm mixed grain rice." The voice acquisition module (microphone array) captures the audio signal, performs noise reduction and enhancement processing, converts it into a digital audio stream, and transmits it to the subsequent processing module.

[0030] 102. Recognize and semantically parse the voice cooking instructions to obtain the corresponding structured information containing the core ingredient entity and at least one differentiated feature. The differentiated feature is used to indicate the personalized requirements for the cooking process and / or cooking result of the core ingredient entity. In this embodiment of the invention, optionally, this step aims to convert ambiguous natural language instructions into structured data that can be accurately understood and processed by machines.

[0031] Speech recognition: Utilizes an automatic speech recognition engine (such as an end-to-end ASR model) to convert digital audio streams into text instructions. For example, converting audio to text: "Fry this frozen steak until it's crispy on the outside and tender on the inside."

[0032] Semantic parsing: Performing natural language understanding on the above text instructions, specifically including: Named Entity Recognition: Using a NER (Named Entity Recognition) model fine-tuned on a corpus in the culinary domain, entities are identified from the text. The core ingredient entity "steak" is identified.

[0033] For the NER model fine-tuned on the corpus of the cooking domain, its training data can contain approximately 500 to 1000 labeled cooking instruction texts, covering common ingredients, states, tastes, and taboos. Training can be conducted using supervised learning, fine-tuning the model based on lightweight pre-trained models such as BERT or RoBERTa to achieve an accuracy rate of over 99% in recognizing entities in the cooking domain.

[0034] Relationship and Attribute Extraction: Through dependency parsing, semantic role labeling, or sequence labeling models, descriptive phrases modifying core food entities and their relationships with these entities are identified. For example, the phrases "frozen" and "cooked to a crispy exterior and tender interior" modifying "steak" are identified.

[0035] Structured Information Generation: The parsed results are organized into structured information, such as a JSON object: {"core_ingredient": "steak", "features": [{"type": "pre_condition", "value": "frozen"},{"type": "post_expectation", "value": "crispy on the outside and tender on the inside"}]}. Here, "core_ingredient" represents the core ingredient entity, and each item in the "features" list is a differentiating feature used to express the user's personalized requirements for the cooking process or result. Each object in the "features" list contains two fields: "type" (feature type) and "value" (feature value). The value of "type," "pre_condition," represents the pre-cooking conditions (such as the initial state of the ingredients), and "post_expectation" represents the post-cooking expectation (such as sensory requirements for the finished product). Each object corresponds to a specific differentiating feature.

[0036] Furthermore, the aforementioned corpus of culinary domain information can include publicly available recipe databases, food community texts, and food science literature. Fine-tuning models such as NER aims to enable them to accurately identify specific entities in the culinary domain, such as "sirloin steak," "medium-rare," and "high heat." In addition, a culinary domain ontology or knowledge graph can be constructed to define entity types and their relationships, such as core ingredients, cooking methods, kitchen utensils, and sensory attributes, providing structured knowledge constraints for semantic parsing and further improving parsing accuracy.

[0037] In this embodiment of the invention, as an optional implementation, the above-mentioned recognition and semantic parsing of voice cooking commands to obtain corresponding structured information containing core ingredient entities and at least one differentiated feature includes: Semantic parsing is performed on the text instructions obtained from the recognized voice cooking instructions to identify the core ingredient entities from the text instructions, and at least one modifier entity used to define the cooking process or cooking result of the core ingredient entities. Semantic classification is performed on at least one modifier entity to classify each modifier entity into a first type of feature corresponding to pre-cooking conditions or a second type of feature corresponding to post-cooking expectations, wherein the first type of feature is used to represent the physical conditions applied to the core ingredient entity that exist before cooking begins, and the second type of feature is used to represent the requirements for the sensory and / or physical properties that the core ingredient entity is expected to present after cooking. Based on the core ingredient entities and the first and / or second types of features obtained after classification, structured information is generated, wherein the first and second types of features constitute at least one differentiating feature.

[0038] In this embodiment of the invention, optionally, for the identification of core ingredient entities: core ingredient entities are the main objects of the cooking operation, usually specific ingredient names, such as "chicken wings," "vegetables," and "noodles." Identification is performed through domain dictionary matching or a NER model.

[0039] Further optionally, for the identification of modifying entities: Modifying entities are words or phrases that describe the state of the core ingredient entity or the user's expectations of the process or result. They can be divided into two categories: 1) Attribute modifications: directly describing the current physical state of the ingredient, such as "frozen," "thick," or "fresh." 2) Expectation modifications: including descriptions of the user's sensory or physical attribute requirements for the cooking result (finished product), such as texture modifications: "soft and tender," "golden," or "crispy on the outside and tender on the inside." Furthermore, they can also include plating features, garnish features, timeliness features, and serving size features that describe the expected quality of the dish. Dependency parsing is used to identify words or phrases that have modifying relationships such as modifiers, adverbs, and complements with the core ingredient entity, and these are then combined with a domain dictionary for identification.

[0040] Further optional, for the classification criteria: binary classification based on the attribute "observable time point" described by the descriptive entity.

[0041] The first type of characteristic corresponds to pre-cooking conditions. These are conditions that can be determined by direct observation or simple physical measurement of the ingredients before cooking begins. For example, "frozen" (temperature state), "cut into pieces" (physical form), "300 grams" (mass). These characteristics are the initial input conditions for the cooking process.

[0042] The second category of characteristics corresponds to post-cooking expectations. These are attributes that can only be determined after the cooking process is complete, through sensory evaluation (sight, taste, smell) or physical measurement (such as inserting a thermometer). Examples include "soft and tender" (texture), "golden yellow" (color), and "well-done" (doneness). These characteristics are the output goals of the cooking process.

[0043] Optionally, the core ingredient entities, along with the first type of features (pre-cooking conditions) and the second type of features (post-cooking expectations) obtained from the classification, can be organized according to a predefined structure (such as key-value pairs or lists) to form machine-readable structured information. For example: {"ingredient": "steak", "pre_conditions": ["frozen"], "post_expectations": ["crispy on the outside and tender on the inside"]}. Here, "ingredient" represents the core ingredient entity, "pre_conditions" represents the pre-cooking conditions (each element corresponds to a first type of feature), and "post_expectations" represents the post-cooking expectations (each element corresponds to a second type of feature). Here, "pre_conditions" and "post_expectations" together constitute the aforementioned "at least one differentiated feature".

[0044] Furthermore, differentiating features can also include dietary restrictions and taboos, which are unified with the above-mentioned post-cooking expectations (the second type of feature).

[0045] As can be seen, implementing this optional embodiment can improve the completeness of extracting key information elements from text instructions by "identifying core ingredient entities and decorative entities," thus ensuring that no user descriptions affecting the cooking process are overlooked. It can also improve the refined and structured understanding of user intent descriptions by "semantically classifying decorative entities into pre-cooking conditions (first-class feature) or post-cooking expectations (second-class feature)," thereby facilitating a clear distinction between the objective initial conditions and subjective expected results described by the user. Furthermore, this classification provides clear and well-defined differentiated features for subsequent steps, enabling more targeted processing strategies in parameter retrieval and model inference. For example, adjusting thermodynamic processes for "pre-cooking conditions" and optimizing flavor and texture parameters for "post-cooking expectations."

[0046] In this optional embodiment, as an optional implementation, the above-described semantic classification of at least one modifying entity to classify each modifying entity into a first type of feature corresponding to the pre-cooking condition or a second type of feature corresponding to the post-cooking expectation includes: Based on the complete semantics of the text instructions, a contextual semantic field is constructed to describe the current cooking task; Obtain typical cooking operation sequences associated with core ingredient entities from a pre-built cooking process knowledge base; The semantics of the modifier entity are matched with the context semantic field and the typical cooking operation sequence to determine the attribute content described by the modifier entity and the associated cooking stage. In response to the fact that the attribute content describes the physical properties of the core ingredient entity that can be directly observed or measured before the cooking operation begins, the decorative entity is classified as a first-class feature. In response to the requirement that the attribute content description is an attribute of the core ingredient entity that can only be determined through sensory evaluation or physical measurement after the cooking operation is completed, the decorative entity is classified as a second type of feature.

[0047] In this embodiment of the invention, optionally, the context semantic field is an abstract representation of the cooking task scenario implied by the current instruction. A scenario framework is constructed by analyzing verbs (such as "fry", "roast", "stew"), kitchen utensils (if mentioned), and common sense in the instruction. For example, for the instruction "fry frozen steak", its semantic field includes: [cooking method: frying (heat conduction, short-term high temperature)], [common kitchen utensils: pan], [common process: preheating, putting in the pan, flipping, taking out the pan and letting it stand].

[0048] Optionally, the typical cooking operation sequence associated with the core ingredient entity can be retrieved from a pre-built cooking process knowledge base. The cooking process knowledge base stores the standard cooking steps for common ingredients. For example, the typical operation sequence for "steak" could be: [Thaw (optional)] -> Pre-treatment (wiping, marinating) -> Preheating the pan -> High-temperature searing -> Flipping -> Resting.

[0049] Further, optionally, the semantics of the modifying entity can be matched with the contextual semantic field and the typical cooking operation sequence for relationship parsing.

[0050] Matching with the contextual semantic field: Analyze the state described by the modifying entity to determine whether it leans more towards a "cause" or a "result". For example, "frozen" is a "cause" that leads to the need to adjust cooking time and temperature, and it exists before cooking begins.

[0051] Matching with typical operation sequences: Determine at which stage of a typical operation sequence the state described by the descriptive entity is typically changed or ultimately achieved. For example, the state of "frozen" is changed in the "thawing" stage, while "crispy on the outside and tender on the inside" is ultimately achieved in the "high-temperature frying" stage.

[0052] Based on the above matching, the "attribute content" and "associated cooking stage" of the descriptive entity are determined.

[0053] Further, alternatively, classification can be performed based on the matching and parsing results.

[0054] If the analysis determines that the attribute describes a physical property (such as temperature, shape, or mass) that exists and can be directly observed before the start of the cooking operation (i.e., before the first typical operation step, such as "frying in a pan"), then it is classified as a first-class feature.

[0055] If the analysis determines that the attribute describes an attribute that is presented as a result after the cooking operation is completed and needs to be evaluated by senses or tools (such as taste, color, internal temperature), then it is classified as a second type of feature.

[0056] As can be seen, implementing this optional embodiment can improve the overall contextual understanding of instructions by "constructing a contextual semantic field of the current cooking task," thereby facilitating a more accurate grasp of the role of decorative entities in a specific cooking task. It can introduce domain knowledge as a classification reference by "obtaining typical cooking operation sequences associated with core ingredient entities," thus enabling the understanding of user descriptions within a standard cooking process framework. It can improve the accuracy and contextual relevance of feature classification by "matching and resolving the relationship between the semantics of decorative entities and the contextual semantic field and typical operation sequences," thereby helping to solve classification problems that are semantically ambiguous or dependent on context (e.g., whether "tender" refers to the initial meat quality or the desired texture). By clearly defining the "time point of attribute content description (before or after cooking)" as the classification basis, feature classification has clear and objective judgment criteria, thereby improving the consistency and interpretability of classification results.

[0057] 103. Based on the core ingredient entity, retrieve the benchmark cooking parameters corresponding to the core ingredient entity from the pre-set cooking parameter knowledge base; In this embodiment of the invention, optionally, the cooking parameter knowledge base is a structured database or atlas that stores a large number of mapping relationships between ingredients, dishes, and standard cooking parameters. Parameters typically include, but are not limited to: heating temperature, heating time, heating mode (such as frying, baking, steaming), power level, whether preheating is required, and recommended cookware.

[0058] Furthermore, this knowledge base can be constructed in the form of a graph, where nodes include ingredients, dishes, kitchen utensils, cooking parameters, etc., and edges represent the relationships between them (such as "applicable to", "reference temperature", "achieved effect"). For example, the "steak" node can be associated with the "pan-frying" method node, and the edge can be attached with attributes such as "baseline temperature: 200℃" and "baseline time: 2.5 minutes / side". This structure facilitates complex multi-hop association retrieval and contextual reasoning.

[0059] Search: Using the core ingredient entity "steak" as the search key, the system searches the knowledge base. Multiple baseline records can be returned, for example: "Sirloin Steak - Baseline Parameters: Preheat pan to high heat, sear each side for 2 minutes, rest for 3 minutes"; "Filet Mignon - Baseline Parameters: Medium heat, sear each side for 3 minutes, rest for 5 minutes." The system can further filter based on default rules (such as the most common category) or subsequent steps, initially selecting one, such as "General Baseline Parameters for Steak: High heat, sear each side for 2.5 minutes."

[0060] In this embodiment of the invention, as another optional implementation, the above-mentioned method of retrieving benchmark cooking parameters corresponding to the core ingredient entity from a preset cooking parameter knowledge base based on the core ingredient entity includes: Based on the types of differentiated features contained in the structured information, a parameter retrieval strategy is determined for core food entities; Based on the parameter retrieval strategy, at least one set of candidate benchmark parameters associated with the core ingredient entity is retrieved from the pre-set cooking parameter knowledge base. Based on the differentiated features in the structured information, at least one candidate benchmark parameter set is subjected to context-adaptive screening in order to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one candidate benchmark parameter set.

[0061] In this embodiment of the invention, optionally, a parameter retrieval strategy for core ingredient entities is determined based on the type of differentiated features contained in the structured information.

[0062] The parameter retrieval strategy determines the breadth and focus of the retrieval.

[0063] Strategy types can include: Precise matching priority strategy: When the structured information contains both Type I and Type II features, and the features are specific, priority is given to retrieving fine parameters that can match both the core ingredient and some key features. For example, for "thick-cut frozen steak", parameters for "thick-cut steak" and / or "frozen steak" are retrieved first.

[0064] Core matching expansion strategy: When there are few or ambiguous features, the core ingredient entity is used as the primary key for retrieval, and all relevant benchmark parameters are returned for subsequent filtering.

[0065] Results-oriented strategy: When the second type of feature (expected result after cooking) is very prominent (such as "soft and tender"), the search will also query the cooking parameters of other ingredients that can achieve similar results as a reference.

[0066] Further optionally, based on a parameter retrieval strategy, at least one set of candidate benchmark parameters associated with the core ingredient entity can be retrieved from a pre-set cooking parameter knowledge base.

[0067] Execute the search strategy. For example, using the "core matching expansion strategy" with "steak" as the keyword, multiple candidate sets can be retrieved, such as "sirloin steak benchmark parameter set", "filet mignon benchmark parameter set", and "general steak benchmark parameter set".

[0068] Further, optionally, for differential features based on structured information, context-adaptive screening is performed on at least one set of candidate benchmark parameters.

[0069] Feature and parameter association analysis: Similarity calculation or rule matching is performed between differentiated features and metadata (such as applicable scenarios, applicable ingredient status, and expected effect description) of candidate benchmark parameter sets.

[0070] Filtering logic: For the first type of feature (such as "frozen"), filter the set of parameters that explicitly include "frozen" or "low temperature" in "applicable state" or have a longer base cooking time.

[0071] For the second type of feature (such as "crispy on the outside and tender on the inside"), select the set of parameters that contain similar semantics such as "crispy on the outside" and "juicy on the inside" in the description of "expected effect".

[0072] Determine the final baseline cooking parameters: Through the above screening, select the set of baseline cooking parameters that best fits the current context (core ingredients + differentiating features) from the candidate set, and use the standard parameters in this set as the baseline for subsequent adjustments. For example, from multiple steak parameter sets, select "benchmark parameters suitable for frozen prepared steaks" as the baseline.

[0073] As can be seen, implementing this optional embodiment can improve the strategic nature and flexibility of benchmark parameter retrieval by "determining parameter retrieval strategies based on the type of differentiated features," thereby facilitating dynamic adjustment of the retrieval method according to the user's emphasis (whether it is a special condition or a special expectation). It can improve the richness and coverage of retrieval results by "retrieving at least one set of candidate benchmark parameters," thus providing selection space to avoid errors caused by a single benchmark mismatch. Furthermore, it can improve the scenario adaptability and personalized relevance of the finally selected benchmark parameters by "context-adaptive filtering of the candidate benchmark parameter set based on differentiated features," thereby providing a higher-quality starting point for subsequent large language model correction that is closer to the user's current needs.

[0074] In this optional embodiment, as an optional implementation, the above-mentioned context-adaptive screening of at least one candidate benchmark parameter set based on differential features in structured information to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one candidate benchmark parameter set includes: For each differential feature in the structured information, determine the cooking effect influencing factor that the differential feature points to; For each candidate benchmark parameter in at least one set of candidate benchmark parameters, a scenario fit score between the candidate benchmark parameter and the differentiated feature is calculated based on the cooking effect influence factor. Based on the logical dependencies between different differentiated features in the structured information, a priority integration strategy is determined for integrating multiple scene adaptability scores. Based on the priority integration strategy, the multiple scene adaptability scores corresponding to each candidate benchmark parameter are combined, and the benchmark cooking parameters corresponding to the core ingredient entity are determined from the candidate benchmark parameter set according to the combined results.

[0075] In this embodiment of the invention, optionally, for each differentiated feature in the structured information, the cooking effect influencing factor pointed to by the differentiated feature is determined.

[0076] Each differentiating feature corresponds to one or more key physical or chemical factors that affect the final cooking result.

[0077] For example, factors influencing the "frozen" characteristic may include: initial temperature (requiring additional heat to thaw) and heat conduction rate (slow internal heating).

[0078] Factors influencing the characteristic of "crispy on the outside and tender on the inside" may include: surface Maillard reaction intensity (crispy), internal moisture retention rate (tender), and internal and external temperature gradient.

[0079] Further optionally, for each candidate benchmark parameter in at least one set of candidate benchmark parameters, a scenario fit score is calculated based on the cooking effect influence factor.

[0080] For each candidate parameter, evaluate how well its default settings (such as temperature and time) satisfy the influencing factors corresponding to the current feature.

[0081] Calculation method: A scoring function can be established. For example, for a candidate parameter, check whether its "remarks" or "applicability instructions" mention "applicable to frozen foods". If yes, it gets a high score on that feature dimension; otherwise, it gets a low score. Alternatively, more precisely, the score can be calculated by measuring the degree of match between the candidate parameter's suggested "total heating amount" and the "theoretical heat required to heat frozen foods to the target temperature".

[0082] Further, optionally, a priority integration strategy can be determined based on the logical dependencies between differentiated features.

[0083] Identifying dependencies: For example, achieving the desired result of "crispy on the outside and tender on the inside" heavily depends on how the initial condition of "freezing" is handled. There is a strong dependency between the two.

[0084] Develop a strategy: Weighted average strategy: If the features are relatively independent, the fit scores for each feature dimension are summed using weighted averages. The weights can be preset based on the feature type (e.g., the first type of feature usually has a higher weight because it directly affects the thermodynamic process).

[0085] Key factor veto strategy: If there is a strong dependency, the score of the dependent feature (such as "frozen") dimension has higher priority, and its low score can directly lead to the exclusion of the candidate parameter.

[0086] Sequential decision-making strategy: First, select a batch of parameters based on the first type of features (conditions), and then sort these parameters based on the second type of features (expectations).

[0087] Alternatively, based on a priority integration strategy, the multiple scenario fit scores corresponding to each candidate benchmark parameter are integrated, and the benchmark cooking parameters are determined based on the integrated results.

[0088] Execute the selected strategy. For example, using a weighted average strategy, calculate the final comprehensive score for each candidate parameter. Select the candidate parameter with the highest comprehensive score as the benchmark cooking parameter that best corresponds to the current core ingredient entity and differentiating features. If the highest scores are close, multiple parameters can be selected for subsequent large language model correction, with the model making the final decision.

[0089] As can be seen, implementing this optional embodiment can improve the analytical depth of the physical / chemical driving factors behind user needs by "determining the influencing factors of cooking effects pointed to by each differentiated feature," thereby facilitating the linking of qualitative descriptions to quantifiable culinary science principles. It can transform the suitability assessment from qualitative judgment to quantitative comparison by "calculating the scenario suitability score between candidate benchmark parameters and differentiated features," thereby improving the objectivity and accuracy of the screening process. It can improve the logical rationality and strategic nature of multi-feature comprehensive decision-making by "determining priority integration strategies based on the logical dependencies between differentiated features," thereby facilitating the correct handling of potential conflicts or hierarchical relationships between features (e.g., the processing priority of "frozen" state is higher than that of "crispy" state). It can achieve the selection of optimal benchmark parameters based on quantitative assessment and logical strategies by "combining multiple scenario suitability scores according to the priority integration strategy," thereby improving the overall matching degree between the selected benchmark parameters and the current complex personalized needs.

[0090] 104. Input the baseline cooking parameters and at least one differential feature into a large language model trained with cooking domain data. The large language model outputs correction information for the baseline cooking parameters and generates personalized cooking parameters based on the correction information. In this embodiment of the invention, optionally, this is the core of achieving intelligent parameter tuning. The large language model in the culinary field is a model obtained through pre-training and fine-tuning using massive corpora of recipes, food science literature, and culinary experimental data.

[0091] Model Input Construction: The baseline parameters and differentiated features from structured information, along with predefined safety constraints (such as maximum temperature limits), are organized into a clear prompt and input into the large language model. For example: "Baseline cooking parameters: For steak, use high heat, sear for 2.5 minutes per side. User special requirements: 1. Ingredients are initially frozen. 2. Desired cooking result is crispy on the outside and tender on the inside. Based on cooking knowledge, please provide suggestions for correcting the baseline parameters while ensuring safety (e.g., core temperature must meet safety standards, avoid excessive production of harmful substances)." Model Inference and Output: The large language model infers based on its internal knowledge and outputs corrective information. For example: "Correction Suggestion: Since the ingredients are frozen, it is recommended to: 1. Use medium-low heat (compared to high heat) to extend the frying time on one side by about 1.5 times to ensure the inside is thawed and cooked through, avoiding a burnt outside and raw inside. 2. Switch to high heat in the final stage and add 30 seconds to each side to achieve a 'burnt outside' effect. 3. The total cooking time needs to be increased accordingly." Parameter generation: The system analyzes the model's output correction suggestions and converts them into quantifiable parameter adjustments. For example, it generates personalized parameters: "Stage 1: Medium heat, 3 minutes and 45 seconds per side (2.5 minutes * 1.5); Stage 2: High heat, 30 seconds per side. Total 4 minutes and 15 seconds per side." Furthermore, the aforementioned large-scale language model can be based on a small-scale model in a vertical domain, using approximately 1500 to 3000 sample pairs of "baseline parameters - user needs - output parameters" covering various ingredients, states, and textures for low-resource fine-tuning (such as LoRA). The training objective is to enable the model to learn fixed parameter correction logic, such as recognizing "frozen" state corresponding to increasing cooking time, and recognizing "crispy" texture corresponding to extending high-temperature time or increasing the final temperature.

[0092] In another optional implementation of this invention, the above-mentioned inputting the baseline cooking parameters and at least one differential feature into a large language model trained with cooking domain data, and having the large language model output correction information for the baseline cooking parameters, includes: Construct an inference input context that includes baseline cooking parameters, at least one differentiated feature, and pre-defined safety constraints; The reasoning input context is fed into a large language model trained on data from the culinary domain, triggering the large language model to perform multi-stage chained reasoning, which includes task decomposition, constraint injection, and corrective derivation. Receives an intermediate structured output generated by a large language model based on multi-stage chained reasoning, containing at least one cooking parameter correction term and its correction basis; Based on at least one cooking parameter correction item in the intermediate structured output, correction information for the baseline cooking parameters is generated.

[0093] In this embodiment of the invention, optionally, an inference input context is constructed that includes baseline cooking parameters, at least one differentiated feature, and preset safety constraints.

[0094] Organize the above information into a structured prompt engineering template and input it into the large language model. For example: [Task] Please adjust the baseline cooking parameters according to user needs.

[0095] [Baseline Parameters] Ingredient: Steak. Steps: Heat a pan over high heat and sear each side for 2.5 minutes.

[0096]

User Requirements

[0097] [Safety Constraints] During cooking, the core temperature must reach a safe standard (e.g., above X degrees Celsius); the surface temperature should not exceed Y degrees Celsius for an extended period to prevent excessive carbonization.

[0098]

Output Requirements

[0099] Further optional, multi-stage chained reasoning can be performed to trigger a large language model.

[0100] Through carefully designed prompts, the model's chain-of-thought ability is stimulated. The model's internal reasoning process is guided as follows: Task breakdown phase: Understanding "crispy on the outside and tender on the inside" requires two sub-goals: "high temperature for a short time to form a crispy crust" and "slow internal heating to retain juice".

[0101] Constraint injection phase: Identify the conflict between the "frozen" state and the "slow internal heating" sub-target, which requires extending the heating time; Identify the "safety constraint" requirement that the center temperature must meet the standard.

[0102] Correction and derivation stage: Regarding "freezing" and "internal compliance": it is deduced that "the total heating time should be extended or the initial heat should be reduced to allow the heat to penetrate deeper".

[0103] Regarding "external charring" and "safety": it is deduced that "the firepower should be increased in the final stage, but the time should be controlled to prevent excessive surface carbonization."

[0104] Comprehensive balance: A two-stage correction scheme was derived, which is to first heat at medium-low heat for a longer time to ensure the inside is thoroughly cooked, and then switch to high heat for a short time to form a charred crust.

[0105] Further optional, for receiving intermediate structured output generated by a large language model.

[0106] The model output includes not only the conclusion but also the intermediate steps of the reasoning, presented in a semi-structured format. For example: Reasoning steps: (1) The freezing state leads to an increase in the initial heat load demand, requiring an extension of the total time or adjustment of the firepower.

[0107] (2) "Tender inside" requires that the internal temperature should not rise too quickly. It is recommended to use medium-low heat initially.

[0108] (3) "Outer char" requires a high final surface temperature, so it is recommended to use high heat in the later stages.

[0109] (4) In accordance with safety constraints, it is necessary to ensure that the extended time is sufficient to reach the core temperature standard, and that the fire time does not cause surface carbonization.

[0110] Suggested revisions: Phase 1 firepower: Adjusted to "medium firepower".

[0111] Phase 1 time: One side extended to approximately 3 minutes and 45 seconds.

[0112] Phase 2 firepower: Adjusted to "Large Fire".

[0113] Phase 2 time: 30 seconds more per side.

[0114] Further, optionally, correction information can be generated based on the intermediate structured output.

[0115] The system analyzes the model's output, extracts specific "correction items" (such as "firepower: medium fire -> high fire", "time: +30 seconds") and "correction amounts", and formats them into a list of correction instructions that the system can process internally, i.e., correction information.

[0116] Furthermore, to stably guide the large language model to output structured content, a few-shot suggestion method can be used, providing 1-3 formatted input-output examples in the system prompts. Model output can be constrained to a specific JSON or YAML format for automatic system parsing. For example, the model can be required to output in the format {"reasoning": ["reasoning step 1", "reasoning step 2"], "adjustments": [{"parameter": "heat", "change": "medium heat", "duration": "3 minutes 45 seconds"}, ...]}. Here, "reasoning" is a list of reasoning steps (recording the model's logical deduction process), and "adjustments" is a list of parameter adjustments (each object contains fields such as "parameter" (the type of cooking parameter to be adjusted), "change" (the adjusted parameter value), and "duration" (the duration the parameter is effective), ensuring the machine readability and parsability of the output content.

[0117] As can be seen, implementing this optional embodiment can provide the large language model with a reasoning task that is complete in information and has clear boundaries by "constructing a reasoning input context that includes baseline cooking parameters, differentiated features, and pre-set safety constraints," thereby improving the relevance and safety awareness of the model output. It can guide the model to perform structured and in-depth thinking by "triggering the large language model to execute multi-stage chained reasoning that includes task decomposition, constraint injection, and correction derivation," thereby improving the logic, rationality, and interpretability of parameter correction suggestions and avoiding "black box" leaps in reasoning. It can obtain a transparent and traceable reasoning process and results by "receiving the intermediate structured output generated by the large language model that includes correction items and correction basis," thereby improving the credibility of the entire parameter generation system and providing a basis for possible verification or manual review. Through the above-mentioned guided reasoning and structured output, it can effectively leverage the advantages of the large language model in knowledge fusion and logical reasoning, while constraining its output format and content scope, thereby improving the reliability, controllability, and practicality of using the large model for cooking parameter correction.

[0118] 105. Perform safety verification on personalized cooking parameters; In this embodiment of the invention, optionally, to ensure security, the generated personalized parameters are verified in multiple dimensions.

[0119] Verification dimensions include, but are not limited to: temperature safety (e.g., whether the cooking temperature is within the upper limit of the cookware's safety, and whether it may generate excessive oil fumes instantly), time safety (e.g., whether the total cooking time will lead to excessive carbonization of the food), combination safety (e.g., whether cooking at a specific temperature for a specific time may produce harmful substances), and device compatibility (e.g., whether the parameters exceed the capability range of the currently connected cooking appliance).

[0120] Verification method: Compare the personalized parameters with the safety rule base stored in the cooking parameter knowledge base. For example, the rule base may stipulate that "when pan-frying steak, the center temperature should reach above X degrees to ensure safety" and "the smoke point of cooking oil should generally not be higher than Y degrees".

[0121] Verification Execution: Based on personalized parameters and ingredients, the expected core temperature, surface temperature, etc., are calculated and logically compared with safety rules. If all dimensions meet the safety rules, the verification passes; if any dimension exceeds the limit, the verification fails.

[0122] 106. If the verification passes, control instructions will be generated based on the personalized cooking parameters and sent to the cooking appliances for execution.

[0123] In this embodiment of the invention, optionally, the verified personalized cooking parameters are converted into a control instruction set recognizable by the cooking appliance. For example, for a smart stove, the instruction set may include: "Set power level P1 (corresponding to medium heat), duration t1; then set power level P2 (corresponding to high heat), duration t2". The instruction set is sent to the corresponding cooking appliance via an Internet of Things communication protocol (such as Wi-Fi, Bluetooth), driving it to execute sequentially, thereby automatically completing the cooking process that meets the user's personalized needs.

[0124] As an alternative implementation, the aforementioned cooking parameter knowledge base can also be deployed locally on the cooking appliance. In this approach, the retrieval of baseline parameters is completed locally, with only the differentiated features uploaded to the cloud-based large language model for correction calculations, thereby reducing reliance on cloud storage and the amount of data transmitted over the network.

[0125] As an alternative extension, to more accurately obtain the status of ingredients, the system can integrate an image sensor. The agent can call an image recognition model to analyze the image of the ingredients (such as whether there is frost), and use the identified status information such as "frozen" and "room temperature" as the first type of feature. This information, together with the results of voice command parsing, constitutes structured information, thereby reducing reliance on user language descriptions.

[0126] As can be seen, implementing the embodiments of the present invention can improve the depth and structuring of understanding user natural language commands by "recognizing and semantically parsing voice cooking commands to obtain structured information containing core ingredient entities and differentiated features," thereby facilitating the accurate extraction of users' personalized cooking intentions from ambiguous voice commands. It can also improve the reliability and scientific rigor of the starting point for generating cooking parameters by "retrieving benchmark cooking parameters from a pre-set cooking parameter knowledge base based on core ingredient entities," thus ensuring that subsequent personalized adjustments are based on a safe benchmark that conforms to common cooking knowledge. Furthermore, it can enhance the intelligence and adaptability of cooking parameter generation by "inputting benchmark cooking parameters and differentiated features into a large language model trained with data from the cooking domain, and having the large language model output correction information and generate personalized cooking parameters," thereby facilitating the transformation of users' personalized, non-quantitative descriptions (such as "crispy on the outside and tender on the inside") into specific, executable cooking parameter adjustments. Finally, it can improve the safety and reliability of the final output parameters by "performing safety verification on personalized cooking parameters," thus ensuring that the cooking process complies with safety standards while meeting personalized needs, preventing safety risks caused by inappropriate parameters. Through the complete process described above, end-to-end automatic generation of cooking control commands, from fuzzy voice commands to safe, personalized, and executable commands, can be achieved. This improves the intelligence level of cooking control and user experience, allowing users to complete personalized cooking by driving kitchen appliances with natural voice commands without needing professional cooking knowledge. It also improves the accuracy and comprehensiveness of cooking appliances in understanding the cooking needs in user voice commands, thereby improving the accuracy of cooking parameter generation.

[0127] Example 2 Please see Figure 2 , Figure 2 This is a flowchart illustrating another voice-interactive cooking parameter generation method disclosed in an embodiment of the present invention. Figure 2 The described voice-interactive cooking parameter generation method can be applied to cooking appliances such as smart ovens, steam ovens, rice cookers, and pressure cookers—that is, devices involving "standard programs + personalized fine-tuning." It can also be applied to smart devices associated with the aforementioned cooking appliances, including but not limited to one or more of smart home devices, cloud devices, edge computing devices, relay devices, base station devices, urban management devices, and smart connected devices. This invention does not limit the scope of these devices. Figure 2 As shown, the voice-interactive cooking parameter generation method may include the following operations: 201. Receive user's voice cooking instructions; 202. Recognize and semantically parse the voice cooking instructions to obtain the corresponding structured information containing the core ingredient entity and at least one differentiated feature. The differentiated feature is used to indicate the personalized requirements for the cooking process and / or cooking result of the core ingredient entity. 203. Based on the core ingredient entity, retrieve the benchmark cooking parameters corresponding to the core ingredient entity from the pre-set cooking parameter knowledge base; 204. Input the baseline cooking parameters and at least one differential feature into a large language model trained with cooking domain data. The large language model outputs correction information for the baseline cooking parameters and generates personalized cooking parameters based on the correction information. 205. Based on personalized cooking parameters and a cooking parameter knowledge base, determine at least one security verification dimension associated with the current cooking task; 206. For each safety verification dimension, obtain the verification benchmark and tolerance range corresponding to that safety verification dimension. The verification benchmark and tolerance range are obtained based on the expected cooking process state derived from the safety rules and personalized cooking parameters in the cooking parameter knowledge base. 207. On at least one safety verification dimension, compare the personalized cooking parameters or the process state parameters derived from them with the corresponding verification benchmarks and tolerance ranges to generate compliance judgment results for each dimension. 208. Based on the risk levels and structured information predefined in the cooking parameter knowledge base for safety verification dimensions, generate the importance weight of each safety verification dimension for cooking safety; 209. Based on the importance weight, the compliance judgment results of each security verification dimension are weighted and calculated to generate a weighted security assessment score; 210. Compare the weighted safety assessment score with the preset safety threshold, and output the judgment result of whether the personalized cooking parameters pass the safety verification based on the comparison result; 211. If the verification passes, control instructions will be generated based on the personalized cooking parameters and sent to the cooking appliances for execution.

[0128] In this embodiment of the invention, optionally, at least one security verification dimension associated with the current cooking task is determined.

[0129] Based on personalized cooking parameters (such as temperature, time, and power curves) and a safety rule base in the cooking parameter knowledge base, the dimensions that need to be verified are dynamically determined. For example, for the "pan-frying steak" task, the following dimensions may be involved: core temperature safety dimension, surface overheating / carbonization dimension, oil fume generation dimension, oil peroxidation dimension, and equipment power load dimension.

[0130] Alternatively, for each security verification dimension, the corresponding verification benchmark and tolerance range can be obtained.

[0131] Verification benchmarks: Obtained from safety rules in the knowledge base. For example, the benchmark for the core temperature safety dimension is "the lower limit of the safe core temperature of beef (e.g., X℃)"; the benchmark for the surface carbonization dimension could be "the duration for which the food surface temperature does not exceed Y℃".

[0132] Tolerance range: The allowable deviation range. For example, the target center temperature can be set with a tolerance of +Z℃, and the equipment power has a tolerance range of the rated value.

[0133] Further, optionally, for at least one safety verification dimension, the personalized cooking parameters or the process state parameters derived therefrom are compared with the corresponding verification benchmark and tolerance range.

[0134] Derivation of process state parameters: Using thermodynamic models or empirical formulas, process state parameters are calculated based on individual parameters (heat, time, initial state of ingredients). For example, the estimated core temperature, maximum surface temperature and duration, and amount of smoke generated at the end of cooking are calculated.

[0135] Comparison: Compare the calculated state parameters with the baseline and tolerance of the corresponding dimension. For example, determine whether the "estimated center temperature" is greater than the "safe lower limit"; and whether the "surface high temperature duration" exceeds the "carbonization risk threshold time".

[0136] Further optional, for generating the importance weight of each security verification dimension to cooking safety.

[0137] The weights are not fixed, but dynamically generated: Based on risk level: The knowledge base predefines a basic risk level for each safety dimension (e.g., "insufficient center temperature" is "high" risk level, "slightly heavy oil fumes" is "medium" risk level).

[0138] Based on the current context: Adjust weights by incorporating structured information. For example, if the user is a "pregnant woman" or a "child," the weight of "core temperature safety" will be increased; if the cooking ingredient is "deep-sea fish," the weight of the "oxidation / nutrient loss" dimension can be increased.

[0139] Furthermore, the dynamic generation of importance weights can be achieved through a weight calculation function. The inputs to this function include: the basic risk level of the safety dimension, user or food characteristics extracted from structured information (such as "user type: child", "food category: poultry"), and current environmental parameters (such as "equipment model: XX oven"). The function can internally preset a set of rules or a small neural network to map the final weight values ​​based on the input combinations. For example, the rule could be: IF (user includes "child") AND (safety dimension == "core temperature") THEN Weight = basic weight * 1.5.

[0140] Optionally, a weighted calculation can be performed on the compliance determination results for each security verification dimension to generate a weighted security assessment score.

[0141] Single-dimensional judgment: Each dimension gives a qualitative judgment such as "compliant", "minor violation", "serious violation" or "serious violation", or quantifies it into a score (e.g., compliant = 1.0, minor violation = 0.5, serious violation = 0).

[0142] Weighted summation: Multiply the quantitative score of each dimension by its importance weight, and then sum them to obtain the total weighted security assessment score. For example, total score = Σ(dimensional i score * weight i).

[0143] Further, optionally, the weighted security assessment score is compared with a preset security threshold, and the final judgment result is output.

[0144] Set a passing threshold (e.g., 0.85). If the weighted total score is greater than or equal to the threshold, it is considered "verification passed". If it is lower than the threshold, it is considered "verification failed", and feedback can be provided on which high-risk dimension(s) caused the failure. It may even trigger the correction process in step 204 above to readjust the parameters, or directly prompt the user to intervene.

[0145] As can be seen, implementing this optional embodiment can improve the comprehensiveness and relevance of safety verification by "determining at least one safety verification dimension associated with the current cooking task," thereby facilitating dynamic coverage of various potential risk points related to the current cooking, rather than performing fixed checks. It can also improve the precision and scientific rigor of verification by "obtaining verification benchmarks and tolerance ranges derived from safety rules and expected process states," allowing safety standards to be dynamically adapted to specific personalized cooking solutions. Furthermore, it can achieve multi-dimensional and in-depth safety status assessment by "compliance comparison across multiple safety verification dimensions," facilitating cross-verification of parameter safety from different perspectives. Finally, it can achieve risk-aware, context-sensitive dynamic weighting by "generating importance weights for each safety verification dimension based on predefined risk levels and structured information," thereby improving the sensitivity of safety assessments to the actual risk level of the current cooking task (e.g., when cooking for children, the risk weight for insufficient core temperature is automatically increased). By "weighting the compliance assessment results to generate a weighted security assessment score," multi-dimensional security assessments with different weights can be integrated into a comprehensive quantitative security indicator, facilitating clear and unified decisions on whether security passes or fails. Through this multi-dimensional, weighted security verification process, while encouraging personalized cooking parameters, a dynamic, intelligent, and reliable security defense line is constructed, fundamentally improving the overall security and robustness of voice-interactive cooking.

[0146] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of a voice-interactive cooking parameter generation system disclosed in an embodiment of the present invention. This voice-interactive cooking parameter generation system can be applied to cooking appliances such as smart ovens, steam ovens, rice cookers, pressure cookers, etc., i.e., devices involving "standard programs + personalized fine-tuning." It can also be applied to smart devices associated with the aforementioned cooking appliances, including but not limited to one or more of smart home devices, cloud devices, edge computing devices, relay devices, base station devices, urban management devices, and smart connected devices. The embodiments of the present invention do not limit this. Figure 3 As shown, the voice-interactive cooking parameter generation system may include: Receiver module 301 is used to receive the user's voice cooking instructions; The recognition and parsing module 302 is used to recognize and semantically parse the voice cooking instructions to obtain the corresponding structured information containing the core ingredient entity and at least one differentiated feature. The differentiated feature is used to indicate the personalized requirements for the cooking process and / or cooking result of the core ingredient entity. The cooking retrieval module 303 is used to retrieve the benchmark cooking parameters corresponding to the core ingredient entity from a pre-set cooking parameter knowledge base based on the core ingredient entity. The inference correction module 304 is used to input the baseline cooking parameters and at least one differential feature into a large language model trained with cooking domain data, and the large language model outputs correction information for the baseline cooking parameters and generates personalized cooking parameters based on the correction information. The safety verification module 305 is used to verify the safety of personalized cooking parameters; The execution control module 306 is used to generate control commands based on personalized cooking parameters and send them to the cooking appliances for execution if the safety verification module passes the verification.

[0147] As can be seen, implementing the embodiments of the present invention can improve the depth and structuring of understanding user natural language commands by "recognizing and semantically parsing voice cooking commands to obtain structured information containing core ingredient entities and differentiated features," thereby facilitating the accurate extraction of users' personalized cooking intentions from ambiguous voice commands. It can also improve the reliability and scientific rigor of the starting point for generating cooking parameters by "retrieving benchmark cooking parameters from a pre-set cooking parameter knowledge base based on core ingredient entities," thus ensuring that subsequent personalized adjustments are based on a safe benchmark that conforms to common cooking knowledge. Furthermore, it can enhance the intelligence and adaptability of cooking parameter generation by "inputting benchmark cooking parameters and differentiated features into a large language model trained with data from the cooking domain, and having the large language model output correction information and generate personalized cooking parameters," thereby facilitating the transformation of users' personalized, non-quantitative descriptions (such as "crispy on the outside and tender on the inside") into specific, executable cooking parameter adjustments. Finally, it can improve the safety and reliability of the final output parameters by "performing safety verification on personalized cooking parameters," thus ensuring that the cooking process complies with safety standards while meeting personalized needs, preventing safety risks caused by inappropriate parameters. Through the complete process described above, end-to-end automatic generation of cooking control commands, from fuzzy voice commands to safe, personalized, and executable commands, can be achieved. This improves the intelligence level of cooking control and user experience, allowing users to complete personalized cooking by driving kitchen appliances with natural voice commands without needing professional cooking knowledge. It also improves the accuracy and comprehensiveness of cooking appliances in understanding the cooking needs in user voice commands, thereby improving the accuracy of cooking parameter generation.

[0148] In this embodiment of the invention, as an optional implementation, the specific method by which the recognition and parsing module 302 recognizes and semantically parses the voice cooking instructions to obtain the corresponding structured information containing core ingredient entities and at least one differentiated feature includes: Semantic parsing is performed on the text instructions obtained from the recognized voice cooking instructions to identify the core ingredient entities from the text instructions, and at least one modifier entity used to define the cooking process or cooking result of the core ingredient entities. Semantic classification is performed on at least one modifier entity to classify each modifier entity into a first type of feature corresponding to pre-cooking conditions or a second type of feature corresponding to post-cooking expectations, wherein the first type of feature is used to represent the physical conditions applied to the core ingredient entity that exist before cooking begins, and the second type of feature is used to represent the requirements for the sensory and / or physical properties that the core ingredient entity is expected to present after cooking. Based on the core ingredient entities and the first and / or second types of features obtained after classification, structured information is generated, wherein the first and second types of features constitute at least one differentiating feature.

[0149] As can be seen, implementing this optional embodiment can improve the completeness of extracting key information elements from text instructions by "identifying core ingredient entities and decorative entities," thus ensuring that no user descriptions affecting the cooking process are overlooked. It can also improve the refined and structured understanding of user intent descriptions by "semantically classifying decorative entities into pre-cooking conditions (first-class feature) or post-cooking expectations (second-class feature)," thereby facilitating a clear distinction between the objective initial conditions and subjective expected results described by the user. Furthermore, this classification provides clear and well-defined differentiated features for subsequent steps, enabling more targeted processing strategies in parameter retrieval and model inference. For example, adjusting thermodynamic processes for "pre-cooking conditions" and optimizing flavor and texture parameters for "post-cooking expectations."

[0150] In this optional embodiment, as an optional implementation, the identification and parsing module 302 performs semantic classification on at least one modifying entity to classify each modifying entity into a first type of feature corresponding to the pre-cooking condition or a second type of feature corresponding to the post-cooking expectation. The specific methods include: Based on the complete semantics of the text instructions, a contextual semantic field is constructed to describe the current cooking task; Obtain typical cooking operation sequences associated with core ingredient entities from a pre-built cooking process knowledge base; The semantics of the modifier entity are matched with the context semantic field and the typical cooking operation sequence to determine the attribute content described by the modifier entity and the associated cooking stage. In response to the fact that the attribute content describes the physical properties of the core ingredient entity that can be directly observed or measured before the cooking operation begins, the decorative entity is classified as a first-class feature. In response to the requirement that the attribute content description is an attribute of the core ingredient entity that can only be determined through sensory evaluation or physical measurement after the cooking operation is completed, the decorative entity is classified as a second type of feature.

[0151] As can be seen, implementing this optional embodiment can improve the overall contextual understanding of instructions by "constructing a contextual semantic field of the current cooking task," thereby facilitating a more accurate grasp of the role of decorative entities in a specific cooking task. It can introduce domain knowledge as a classification reference by "obtaining typical cooking operation sequences associated with core ingredient entities," thus enabling the understanding of user descriptions within a standard cooking process framework. It can improve the accuracy and contextual relevance of feature classification by "matching and resolving the relationship between the semantics of decorative entities and the contextual semantic field and typical operation sequences," thereby helping to solve classification problems that are semantically ambiguous or dependent on context (e.g., whether "tender" refers to the initial meat quality or the desired texture). By clearly defining the "time point of attribute content description (before or after cooking)" as the classification basis, feature classification has clear and objective judgment criteria, thereby improving the consistency and interpretability of classification results.

[0152] In this embodiment of the invention, as another optional implementation, the cooking retrieval module 303 retrieves the benchmark cooking parameters corresponding to the core ingredient entity from a preset cooking parameter knowledge base based on the core ingredient entity in the following specific ways: Based on the types of differentiated features contained in the structured information, a parameter retrieval strategy is determined for core food entities; Based on the parameter retrieval strategy, at least one set of candidate benchmark parameters associated with the core ingredient entity is retrieved from the pre-set cooking parameter knowledge base. Based on the differentiated features in the structured information, at least one candidate benchmark parameter set is subjected to context-adaptive screening in order to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one candidate benchmark parameter set.

[0153] As can be seen, implementing this optional embodiment can improve the strategic nature and flexibility of benchmark parameter retrieval by "determining parameter retrieval strategies based on the type of differentiated features," thereby facilitating dynamic adjustment of the retrieval method according to the user's emphasis (whether it is a special condition or a special expectation). It can improve the richness and coverage of retrieval results by "retrieving at least one set of candidate benchmark parameters," thus providing selection space to avoid errors caused by a single benchmark mismatch. Furthermore, it can improve the scenario adaptability and personalized relevance of the finally selected benchmark parameters by "context-adaptive filtering of the candidate benchmark parameter set based on differentiated features," thereby providing a higher-quality starting point for subsequent large language model correction that is closer to the user's current needs.

[0154] In this optional embodiment, as an optional implementation, the cooking retrieval module 303 performs context-adaptive filtering on at least one candidate benchmark parameter set based on the differential features in the structured information to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one candidate benchmark parameter set. The specific methods include: For each differential feature in the structured information, determine the cooking effect influencing factor that the differential feature points to; For each candidate benchmark parameter in at least one set of candidate benchmark parameters, a scenario fit score between the candidate benchmark parameter and the differentiated feature is calculated based on the cooking effect influence factor. Based on the logical dependencies between different differentiated features in the structured information, a priority integration strategy is determined for integrating multiple scene adaptability scores. Based on the priority integration strategy, the multiple scene adaptability scores corresponding to each candidate benchmark parameter are combined, and the benchmark cooking parameters corresponding to the core ingredient entity are determined from the candidate benchmark parameter set according to the combined results.

[0155] As can be seen, implementing this optional embodiment can improve the analytical depth of the physical / chemical driving factors behind user needs by "determining the influencing factors of cooking effects pointed to by each differentiated feature," thereby facilitating the linking of qualitative descriptions to quantifiable culinary science principles. It can transform the suitability assessment from qualitative judgment to quantitative comparison by "calculating the scenario suitability score between candidate benchmark parameters and differentiated features," thereby improving the objectivity and accuracy of the screening process. It can improve the logical rationality and strategic nature of multi-feature comprehensive decision-making by "determining priority integration strategies based on the logical dependencies between differentiated features," thereby facilitating the correct handling of potential conflicts or hierarchical relationships between features (e.g., the processing priority of "frozen" state is higher than that of "crispy" state). It can achieve the selection of optimal benchmark parameters based on quantitative assessment and logical strategies by "combining multiple scenario suitability scores according to the priority integration strategy," thereby improving the overall matching degree between the selected benchmark parameters and the current complex personalized needs.

[0156] In another optional implementation of this invention, the inference correction module 304 inputs the baseline cooking parameters and at least one differential feature into a large language model trained with cooking domain data. The specific method by which the large language model outputs correction information for the baseline cooking parameters includes: Construct an inference input context that includes baseline cooking parameters, at least one differentiated feature, and pre-defined safety constraints; The reasoning input context is fed into a large language model trained on data from the culinary domain, triggering the large language model to perform multi-stage chained reasoning, which includes task decomposition, constraint injection, and corrective derivation. Receives an intermediate structured output generated by a large language model based on multi-stage chained reasoning, containing at least one cooking parameter correction term and its correction basis; Based on at least one cooking parameter correction item in the intermediate structured output, correction information for the baseline cooking parameters is generated.

[0157] As can be seen, implementing this optional embodiment can provide the large language model with a reasoning task that is complete in information and has clear boundaries by "constructing a reasoning input context that includes baseline cooking parameters, differentiated features, and pre-set safety constraints," thereby improving the relevance and safety awareness of the model output. It can guide the model to perform structured and in-depth thinking by "triggering the large language model to execute multi-stage chained reasoning that includes task decomposition, constraint injection, and correction derivation," thereby improving the logic, rationality, and interpretability of parameter correction suggestions and avoiding "black box" leaps in reasoning. It can obtain a transparent and traceable reasoning process and results by "receiving the intermediate structured output generated by the large language model that includes correction items and correction basis," thereby improving the credibility of the entire parameter generation system and providing a basis for possible verification or manual review. Through the above-mentioned guided reasoning and structured output, it can effectively leverage the advantages of the large language model in knowledge fusion and logical reasoning, while constraining its output format and content scope, thereby improving the reliability, controllability, and practicality of using the large model for cooking parameter correction.

[0158] In an optional embodiment, the specific method by which the security verification module 305 performs security verification on personalized cooking parameters includes: Based on personalized cooking parameters and a cooking parameter knowledge base, at least one security verification dimension associated with the current cooking task is determined. For each security verification dimension, the verification benchmark and tolerance range corresponding to that security verification dimension are obtained. The verification benchmark and tolerance range are derived from the expected cooking process state derived from the safety rules and personalized cooking parameters in the cooking parameter knowledge base. In at least one safety verification dimension, personalized cooking parameters or process state parameters derived from them are compared with the corresponding verification benchmarks and tolerance ranges to generate compliance judgment results for each dimension. Based on the risk levels and structured information predefined in the cooking parameter knowledge base for safety verification dimensions, the importance weight of each safety verification dimension to cooking safety is generated. Based on the importance weight, the compliance judgment results of each security verification dimension are weighted and calculated to generate a weighted security assessment score; The weighted safety assessment score is compared with a preset safety threshold, and the result of the comparison is used to determine whether the personalized cooking parameters pass the safety verification.

[0159] As can be seen, implementing this optional embodiment can improve the comprehensiveness and relevance of safety verification by "determining at least one safety verification dimension associated with the current cooking task," thereby facilitating dynamic coverage of various potential risk points related to the current cooking, rather than performing fixed checks. It can also improve the precision and scientific rigor of verification by "obtaining verification benchmarks and tolerance ranges derived from safety rules and expected process states," allowing safety standards to be dynamically adapted to specific personalized cooking solutions. Furthermore, it can achieve multi-dimensional and in-depth safety status assessment by "compliance comparison across multiple safety verification dimensions," facilitating cross-verification of parameter safety from different perspectives. Finally, it can achieve risk-aware, context-sensitive dynamic weighting by "generating importance weights for each safety verification dimension based on predefined risk levels and structured information," thereby improving the sensitivity of safety assessments to the actual risk level of the current cooking task (e.g., when cooking for children, the risk weight for insufficient core temperature is automatically increased). By "weighting the compliance assessment results to generate a weighted security assessment score," multi-dimensional security assessments with different weights can be integrated into a comprehensive quantitative security indicator, facilitating clear and unified decisions on whether security passes or fails. Through this multi-dimensional, weighted security verification process, while encouraging personalized cooking parameters, a dynamic, intelligent, and reliable security defense line is constructed, fundamentally improving the overall security and robustness of voice-interactive cooking.

[0160] Example 4 Please see Figure 4 , Figure 4 This is a schematic diagram of another voice-interactive cooking parameter generation system disclosed in an embodiment of the present invention. This voice-interactive cooking parameter generation system can be applied to cooking appliances such as smart ovens, steam ovens, rice cookers, pressure cookers, etc., i.e., devices involving "standard programs + personalized fine-tuning." It can also be applied to smart devices associated with the aforementioned cooking appliances. These smart devices include, but are not limited to, one or more of smart home devices, cloud devices, edge computing devices, relay devices, base station devices, urban management devices, and smart connected devices; the embodiments of the present invention do not impose limitations. Figure 4 As shown, the voice-interactive cooking parameter generation system may include: Memory 401 that stores executable program code.

[0161] Processor 402 coupled to memory 401.

[0162] The processor 402 calls the executable program code stored in the memory 401 to execute the steps in the voice-interactive cooking parameter generation method described in Embodiment 1 or Embodiment 2 of the present invention.

[0163] Example 5 This invention discloses a computer storage medium storing computer instructions. When these computer instructions are invoked, they are used to execute the steps in the voice-interactive cooking parameter generation method described in Embodiment 1 or Embodiment 2 of this invention.

[0164] Example 6 This invention discloses a computer program product, which includes a non-transitory computer storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the voice-interactive cooking parameter generation method described in Embodiment 1 or Embodiment 2.

[0165] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0166] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0167] Finally, it should be noted that the voice-interactive cooking parameter generation method and system disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating cooking parameters through voice interaction, characterized in that, The method includes: Receive user's voice cooking instructions; The voice cooking instructions are recognized and semantically parsed to obtain corresponding structured information containing core ingredient entities and at least one differentiating feature. The differentiating feature is used to indicate personalized requirements for the cooking process and / or cooking results of the core ingredient entities. Based on the core ingredient entity, the benchmark cooking parameters corresponding to the core ingredient entity are retrieved from the preset cooking parameter knowledge base; The baseline cooking parameters and at least one of the differential features are input into a large language model trained with cooking domain data. The large language model outputs correction information for the baseline cooking parameters and generates personalized cooking parameters based on the correction information. The personalized cooking parameters are subject to security verification. If the verification passes, control instructions are generated based on the personalized cooking parameters and sent to the cooking appliances for execution.

2. The voice-interactive cooking parameter generation method according to claim 1, characterized in that, The process of recognizing and semantically parsing the voice cooking instructions to obtain corresponding structured information containing core ingredient entities and at least one differentiating feature includes: Semantic parsing is performed on the text instructions obtained from the voice cooking instructions to identify the core ingredient entities from the text instructions, and at least one modifier entity is identified to define the cooking process or cooking result of the core ingredient entities. Semantic classification is performed on at least one of the said modifier entities to classify each of the said modifier entities into a first type of feature corresponding to pre-cooking conditions or a second type of feature corresponding to post-cooking expectations, wherein the first type of feature is used to represent physical conditions applied to the core ingredient entity that exist before cooking begins, and the second type of feature is used to represent requirements for the sensory and / or physical properties that the core ingredient entity is expected to present after cooking. Based on the core ingredient entity and the first and / or second type features obtained after classification, structured information is generated, wherein the first and second type features constitute at least one of the differentiated features.

3. The voice-interactive cooking parameter generation method according to claim 2, characterized in that, The step of semantically classifying at least one of the modifying entities to classify each of the modifying entities into a first type of feature corresponding to the pre-cooking conditions or a second type of feature corresponding to the post-cooking expectations includes: Based on the complete semantics of the text instructions, a contextual semantic field is constructed to describe the current cooking task; Obtain typical cooking operation sequences associated with the core ingredient entity from a pre-set cooking process knowledge base; The semantics of the modifier entity are matched and parsed with the context semantic field and the typical cooking operation sequence to determine the attribute content described by the modifier entity and the associated cooking stage. In response to determining that the attribute content describes the physical properties of the core ingredient entity that can be directly observed or measured before the cooking operation begins, the decorative entity is classified as the first type of feature. In response to the determination that the attribute content describes the attributes presented by the core ingredient entity that can only be determined through sensory evaluation or physical measurement after the cooking operation is completed, the modifying entity is classified as the second type of feature.

4. The voice-interactive cooking parameter generation method according to claim 2, characterized in that, The step of retrieving the benchmark cooking parameters corresponding to the core ingredient entity from a pre-set cooking parameter knowledge base based on the core ingredient entity includes: Based on the type of the differentiated features contained in the structured information, a parameter retrieval strategy is determined for the core ingredient entity; According to the parameter retrieval strategy, at least one set of candidate benchmark parameters associated with the core ingredient entity is retrieved from the preset cooking parameter knowledge base. Based on the differentiated features in the structured information, at least one set of candidate benchmark parameters is subjected to context-adaptive filtering to determine the benchmark cooking parameters corresponding to the core ingredient entity from at least one set of candidate benchmark parameters.

5. The voice-interactive cooking parameter generation method according to claim 4, characterized in that, The step of performing context-adaptive filtering on at least one set of candidate benchmark parameters based on the differentiated features in the structured information to determine benchmark cooking parameters corresponding to the core ingredient entity from at least one set of candidate benchmark parameters includes: For each of the differentiated features in the structured information, determine the cooking effect influencing factor that the differentiated feature points to; For each candidate benchmark parameter in at least one of the candidate benchmark parameter sets, a scenario fit score between the candidate benchmark parameter and the differentiated feature is calculated based on the cooking effect influence factor. Based on the logical dependencies between the different differentiated features in the structured information, a priority integration strategy for integrating multiple scene adaptability scores is determined; Based on the priority integration strategy, the multiple scene adaptability scores corresponding to each candidate benchmark parameter are integrated, and the benchmark cooking parameters corresponding to the core ingredient entity are determined from the candidate benchmark parameter set according to the integration result.

6. The voice-interactive cooking parameter generation method according to any one of claims 1-5, characterized in that, The step of inputting the benchmark cooking parameters and at least one of the differential features into a large language model trained on cooking domain data, and having the large language model output correction information for the benchmark cooking parameters, includes: Construct an inference input context that includes the benchmark cooking parameters, at least one of the differentiated features, and preset safety constraints; The reasoning input context is input into a large language model trained with data from the culinary domain, triggering the large language model to perform multi-stage chained reasoning that includes task decomposition, constraint injection, and corrective derivation. Receive the intermediate structured output generated by the large language model based on the multi-stage chained reasoning, which includes at least one cooking parameter correction item and its correction basis; Based on the at least one cooking parameter correction item in the intermediate structured output, the correction information for the baseline cooking parameters is generated.

7. The voice-interactive cooking parameter generation method according to any one of claims 1-5, characterized in that, The security verification of the personalized cooking parameters includes: Based on the personalized cooking parameters and the cooking parameter knowledge base, at least one security verification dimension associated with the current cooking task is determined; For each of the security verification dimensions, the verification benchmark and tolerance range corresponding to the security verification dimension are obtained. The verification benchmark and tolerance range are obtained based on the security rules in the cooking parameter knowledge base and the expected cooking process state derived from the personalized cooking parameters. In at least one of the aforementioned safety verification dimensions, the personalized cooking parameters or the process state parameters derived therefrom are compared with the corresponding verification benchmarks and tolerance ranges to generate compliance judgment results for each dimension. Based on the risk level predefined in the cooking parameter knowledge base and the structured information of the safety verification dimension, an importance weight for each safety verification dimension to cooking safety is generated; Based on the aforementioned importance weights, the compliance judgment results for each of the aforementioned security verification dimensions are weighted and calculated to generate a weighted security assessment score. The weighted safety assessment score is compared with a preset safety threshold, and the result of the comparison is used to output a judgment result on whether the personalized cooking parameters have passed the safety verification.

8. A voice-interactive cooking parameter generation system, characterized in that, The system includes: The receiving module is used to receive the user's voice cooking instructions; The recognition and parsing module is used to recognize and semantically parse the voice cooking instructions to obtain corresponding structured information containing core ingredient entities and at least one differentiated feature. The differentiated feature is used to indicate personalized requirements for the cooking process and / or cooking results of the core ingredient entities. The cooking retrieval module is used to retrieve the benchmark cooking parameters corresponding to the core ingredient entity from a preset cooking parameter knowledge base based on the core ingredient entity. The inference correction module is used to input the benchmark cooking parameters and at least one of the differential features into a large language model trained with cooking domain data, and the large language model outputs correction information for the benchmark cooking parameters, and generates personalized cooking parameters based on the correction information. A security verification module is used to perform security verification on the personalized cooking parameters. The execution control module is used to generate control commands based on the personalized cooking parameters and send them to the cooking appliances for execution if the security verification module passes the verification.

9. A voice-interactive cooking parameter generation system, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the voice-interactive cooking parameter generation method as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked, are used to execute the voice-interactive cooking parameter generation method as described in any one of claims 1-7.

Images