A method of flavour wheel construction

Abstract

The present application relates to a kind of flavor wheel construction method, belong to the cross technical field of chemical industry and computer.The present application first obtains the sensory evaluation data of target product, then extracts sensory description language sequence through sensory evaluation data, again digitizes the description language sequence extracted, then constructs the hierarchical system between flavor using clustering method, then constructs each layer flavor transition system respectively, the node order corresponding to the Hamilton circuit of shortest length is used as the transition system of each layer flavor of flavor wheel, finally each layer flavor transition system is matched and associated according to the hierarchical system of flavor wheel, in the form of ring chart is displayed, constitutes the flavor wheel of target product.The present application gives consideration to practicability and scientificity, can efficiently, intelligently construct the flavor wheel system for a certain sub-product field under the condition of avoiding excessive dependence on artificial experience, so that evaluation personnel can better carry out standardized, standardized, efficient product sensory evaluation work.

Description

Technical Field

[0001] This invention relates to a method for constructing flavor wheels, belonging to the interdisciplinary technical field of chemical engineering and computer science. Background Technology

[0002] The flavor wheel is an internationally recognized, scientifically effective, hierarchical sensory evaluation system widely used in the development, design, and evaluation of flavored products. It has numerous applications in the food and flavoring industries. Internationally, beverages such as wine, whiskey, and beer have long had their own dedicated flavor wheel systems. There are also research reports on flavor wheels for products such as tea, confectionery, cigarettes, cigars, and coffee. A flavor wheel typically contains 2-3 layers of flavor descriptions, progressively classifying and refining flavors from abstract to concrete. Adjacent flavor descriptions share similar flavor characteristics, and by establishing similarity and hierarchy relationships between descriptions, a cohesive whole with a transitional structure is formed. The hierarchy and transitional relationships within the flavor wheel allow evaluators to better understand the flavor characteristics of products, enabling standardized, regulated, and efficient sensory evaluation. Therefore, the efficient and intelligent construction of flavor wheel systems for specific product segments is of great significance for sensory evaluation work.

[0003] Currently, flavor wheels are often constructed by professionals relying on extensive product design experience and tasting experiments. For example, the comparative literature "Sensory profile of rooibos originating from the Western and Northern Cape governed by production year and development of rooibos aromawheel" uses an expert panel to evaluate samples of rooibos tea from different origins to ultimately form the product's flavor wheel. This approach makes the construction process challenging and introduces a degree of subjectivity into the design of flavor wheels, affecting their accuracy and universality. To address this, some have proposed intelligent methods for constructing flavor wheels. For instance, Chinese invention patent CN112418919B discloses a quality evaluation method that uses online review data to extract sensory attribute words to build a lexicon, constructing a flavor wheel from high-frequency words. However, the constructed flavor wheels often lack transitional relationships between levels and exhibit low similarity between levels, which also impacts the accuracy and applicability of the flavor wheels. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing flavor wheels to solve the problems of high difficulty, low accuracy and low universality in the current construction process.

[0005] To solve the above-mentioned technical problems, the present invention provides a method for constructing flavor wheels, the method comprising the following steps:

[0006] 1) Select the target product and obtain sensory description data of the target product;

[0007] 2) Extract the sensory descriptive sequence of the target product from the acquired sensory evaluation data;

[0008] 3) Digitize the sensory descriptions in the sensory description sequence;

[0009] 4) Cluster the digitized sensory descriptions and treat each sensory description as a specific flavor to construct a flavor wheel hierarchy system for the target product. This hierarchy system includes at least two layers. The flavors with broader meanings in each category are taken as the first layer flavors. The remaining flavors in each category are used to determine whether to continue dividing. If no further division is needed, the remaining flavors in each category are taken as the second layer flavors. If further division is needed, some relatively broad flavors are selected from the remaining flavors in each category as the second layer flavors. The remaining flavors are then further divided until no further division is possible.

[0010] 5) Take the first layer flavor as the upper layer flavor, take the upper layer flavor in the flavor wheel hierarchy as nodes, construct a topological undirected graph composed of upper layer flavors, calculate the length of all Hamiltonian circuits in the undirected graph based on the distance between each node, and take the node order corresponding to the shortest Hamiltonian circuit as the transition system of upper layer flavor in the flavor wheel.

[0011] 6) Take the second layer flavor as the lower layer flavor, and take each flavor in the lower layer adjacent to the upper layer as a node. Calculate the distance between all the lower layer flavors and the adjacent flavors in the upper layer. Select the shortest distance as the upper boundary lower layer flavor and the lower boundary lower layer flavor respectively. Construct a topological undirected graph composed of each lower layer flavor. Calculate the Hamiltonian path length of the topological undirected graph based on the distance between each node of the lower layer flavor. Take the node order corresponding to the shortest Hamiltonian path as the transition system of the lower layer flavors inside the upper layer flavor.

[0012] 7) After completing the second layer of flavor transition system, determine whether there is a third layer of flavor. If so, take the second layer of flavor as the upper layer flavor and the third layer flavor as the lower layer flavor, and repeat step 6) until the transition system of all layers of flavor is established.

[0013] 8) Sort the flavors of each level in the transition system of each layer according to their specific flavors, and display them in the form of a two- or multi-layered circular diagram to form the flavor wheel of the target product.

[0014] This invention balances practicality and scientific rigor. It transforms large amounts of product sensory description data into computer-processable digital vectors using digital technology. A hierarchical system is established through clustering algorithms, with flavor profiles at each level serving as nodes. An undirected graph is constructed for each level, and this undirected graph technique is used to determine the transition systems between levels. Based on these transition systems, the entire flavor wheel is built. This invention, while avoiding excessive reliance on human experience, enables efficient and intelligent construction of flavor wheel systems for specific product segments, allowing evaluators to conduct standardized, regulated, and efficient product sensory evaluations.

[0015] Furthermore, step 1) employs a method of collecting existing publicly available data or collecting sensory evaluation data.

[0016] This invention can obtain sensory description data of a target product by using publicly available data or by collecting sensory evaluation data through surveys and questionnaires, and the acquisition methods are flexible and diverse.

[0017] Further, step 2) of extracting the sensory description sequence of the target product is as follows: candidate strings are extracted from text data to construct a sensory description word library; the sensory evaluation data is segmented using a word segmentation model, and the segmentation results are compared with the sensory description word library to extract sensory description words, thereby obtaining the corresponding sensory description word sequence.

[0018] This invention improves the accuracy of the obtained sensory description sequences by comparing the segmentation results with a sensory description lexicon when extracting sensory description sequences of target products using a word segmentation model, and then using the lexicon to filter the segmentation results.

[0019] Furthermore, the construction process of the lexicon is as follows:

[0020] A. Collect all the text data from the sensory description data into a corpus S;

[0021] B. Extract candidate strings by taking all string fragments with a length between 1 and L that appear in the corpus as the candidate string set;

[0022] C. Calculate the word frequency of the candidate strings;

[0023] D. Calculate the aggregation degree of the candidate strings and select the mutual information value with the lowest value as the aggregation degree of the candidate string;

[0024] E. Calculate the adjacency entropy of the candidate string, and take the lower value between the left and right adjacency entropies as the adjacency entropy of the candidate string;

[0025] F. By setting thresholds for three indicators—word frequency, aggregation degree, and adjacency entropy—or any combination of these three indicators, candidate strings that do not meet the conditions for word formation are removed, while candidate strings that do meet the conditions for word formation are retained, ultimately forming a set of descriptive words for the corpus.

[0026] G. The set of descriptive words is filtered to form a sensory descriptive word library.

[0027] The word extraction method based on word frequency, aggregation degree, and adjacency entropy in this invention is more flexible than conventional word segmentation tools when constructing a lexicon. It has higher extraction accuracy for new words and professional terms and can better construct a sensory description lexicon.

[0028] Furthermore, the digitization process in step 3) is as follows:

[0029] A. For each sensory descriptor sequence seq i =[w i 1 ,w i 2 ,w i 3 ,...,w i n Training data is generated by sequentially using a single descriptive word as the central word and the remaining words as context words.

[0030] B. Construct a vocabulary dictionary using a sensory description lexicon as the set, and build an initial center word vector matrix and an initial context word vector matrix with dimension N, where N is the length of the set flavor digital representation vector.

[0031] C. Construct a word vector training model, input the training dataset into the model, and obtain the output results.

[0032] Each sensory description represents a specific flavor. Without incorporating human expert experience, each specific flavor is represented in digital form, making similar flavors in the representation system closer together, and giving the entire representation system a certain degree of inference.

[0033] Furthermore, step 4) uses a hierarchical clustering method to construct a flavor wheel hierarchy.

[0034] Furthermore, the distance between the nodes in step 5) is a cosine distance.

[0035] Further, step 6) involves constructing the flavor transition system in the lower layer of the flavor wheel as follows:

[0036] A. For each upper flavor layer, calculate the average vector of all upper and lower flavor layers;

[0037] B. For each upper-layer flavor, calculate the distance between all its lower-layer vectors and the average vector of its upper-adjacent upper-layer flavors, sort the calculation results, and select the one with the shortest distance as the upper boundary lower-layer flavor;

[0038] C. Calculate the distance between all lower-level vectors of the upper-level flavor and the average vector of the lower adjacent upper-level flavor, sort the calculation results, and select the one with the shortest distance as the lower boundary lower-level flavor;

[0039] D. Construct a topological undirected graph consisting of all lower-level flavors under the upper-level flavor, calculate the distance between each lower-level flavor node and the Hamiltonian path length, and take the node order corresponding to the shortest Hamiltonian path as the transition system of lower-level flavors within the upper-level flavor.

[0040] Hamiltonian circuits are circuits that pass through each vertex in an undirected graph exactly once. Selecting the node sequence corresponding to the shortest circuit as the flavor transition system can make the flavor wheel better reflect the transition and approximation relationships between flavors. Attached Figure Description

[0041] Figure 1 This is a diagram of the flavor wheel construction method of the present invention;

[0042] Figure 2 This is a diagram of the word vector training model constructed in this embodiment of the invention;

[0043] Figure 3 This is the flavor wheel obtained in the embodiments of the present invention. Detailed Implementation

[0044] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0045] Examples of Flavor Wheel Construction Methods

[0046] This invention first sets up a target product and obtains its sensory evaluation data. Then, it extracts sensory descriptive sequences from the sensory evaluation data. Next, it constructs digital representation vectors from the extracted descriptive sequences. Then, it uses a hierarchical clustering method to construct a hierarchical system between flavors. The upper-level flavor transition system is constructed by calculating and sorting all Hamiltonian circuits in a topological undirected graph composed of upper-level flavors. The lower-level flavor transition system is constructed by calculating and sorting Hamiltonian paths in a topological undirected graph composed of each lower-level flavor, forming a lower-level flavor transition system between adjacent upper-level flavors. Finally, the lower-level and upper-level flavor transition systems are sorted according to the specific flavors of each level in the corresponding hierarchy of each transition system, and displayed in the form of a two-sided or multi-layered ring diagram, constituting the flavor wheel of the target product. The implementation process of this method is as follows: Figure 1As shown below, a specific example will be used for illustration.

[0047] The target product can be selected according to the actual production application requirements. For example, broad target products such as perfumes, gourmet foods, cigarettes, etc. can be selected, or specific target products such as men's perfumes, cigars, Southeast Asian cuisine, etc. can be set. Here, cigarettes will be taken as an example to elaborate on the implementation process of the present invention in detail.

[0048] 1. Select tobacco flavor raw materials as the target product and obtain the sensory description data of the flavor raw materials.

[0049] According to the target product range and the difficulty of data acquisition, the specific implementation methods include the following two types:

[0050] (1) Collect existing public data. Obtain information such as product introductions, sensory descriptions, user evaluations, etc. of the target product through existing literature or public materials. The information sources include websites, journals, books, data documents, etc.

[0051] (2) Collect sensory evaluation data. The sensory evaluation data of the target product can be collected through forms such as questionnaires and aroma sniffing experiments.

[0052] In this embodiment, information is collected from existing public data, and the sensory description text of tobacco flavor raw materials is obtained from the book "Tobacco Flavor Raw Materials". For example, the sensory description of the flavor raw material 2-methyl-3-ethylpyrazine is "colorless to slightly yellow liquid, with the fragrance of potato, strong, nutty, roasted grain, and earthy".

[0053] 2. Extract the sensory description word sequence of the flavor raw materials through the obtained sensory evaluation data.

[0054] This process requires processing the collected sensory description text, extracting the sensory description words to construct a word bank. Using the constructed word bank as the basis for word segmentation, the other evaluation data obtained is segmented. The segmentation results are compared with the sensory description word bank, and the sensory description words are extracted to form the sensory description word sequence corresponding to each sensory evaluation data. The construction process of the word bank includes the following steps:

[0055] (1) Take all the collected text data as a corpus set S. In this embodiment, the sensory description text collected from the book "Tobacco Flavor Raw Materials" is used as the corpus set S.

[0056] (2) Extract candidate strings. Take all the string segments with lengths between 1 and L in the corpus as the candidate string set. Here, L is the set maximum length of the words. In this embodiment, the value of L is taken as 5, and the strings with lengths from 1 to 5 in the corpus are taken as candidate strings respectively. Such as "soil", "potato", "potato's", "potato's,", "potato's, strong", etc.

[0057] (3) Calculate the word frequency of the candidate strings. By counting the corpus set S, obtain the word frequency of each candidate string in the corpus set S.

[0058] (4) Calculate the aggregation degree of the candidate strings. Perform all possible binary segmentations on the candidate strings. Taking the candidate string "of potatoes" as an example, its binary segmentation combinations include ("soil"; "of beans") and ("potatoes"; "of"), a total of 2 kinds. Calculate the mutual information values MI("soil"; "of beans") and MI("potatoes"; "of") of all binary segmentation combinations, and select the lowest MI value as the aggregation degree of the candidate string "of potatoes". Perform the above operations on all candidate strings to obtain the aggregation degrees of all candidate strings. The mutual information calculation method is as follows:

[0059]

[0060] (5) Calculate the adjacency entropy of the candidate strings. Count the left adjacent characters of the candidate strings in the corpus and their occurrence frequencies. For example, the left adjacent character of the candidate string "cereals" is "roasted", and its occurrence frequency in the example text is 1 time. After obtaining the distribution of the left adjacent characters, calculate the information entropy of this distribution as the left adjacent entropy of the candidate string. Take the smaller value of the left and right adjacent entropies as the adjacency entropy of the candidate string. Perform the above operations on all candidate strings to obtain the adjacency entropies of all candidate strings. The information entropy calculation method is as follows:

[0061]

[0062] where s i represents the discrete value of the random variable.

[0063] (6) By setting thresholds for the three indicators of word frequency, aggregation degree, and adjacency entropy or any combination relationship between the three indicators, eliminate the candidate strings that do not meet the word formation conditions, and retain the candidate strings that meet the word formation conditions, finally forming the description word set of this corpus. Among them, dynamic thresholds can be adopted according to the different lengths of the candidate strings. For example, for candidate strings with a length of 2, retain those with an occurrence frequency greater than 20 times, and for candidate strings with a length of 3, retain those with an occurrence frequency greater than 15 times.

[0064] (7) Screen the description word set and screen out the words related to sensory descriptions as the sensory description word library. Such as "potatoes", "strong", "nuts", "roasted cereals", "soil", etc.

[0065] After successful lexicon extraction, the Wirtschaft segmentation model provided by HanLP was used as the segmentation tool. The sensory description lexicon was introduced into the segmentation model as an external dictionary to form a segmentation model suitable for sensory description word extraction. This model was then used to segment each sensory evaluation data point. The segmentation results were compared with the sensory description lexicon to extract the sensory description words, resulting in a sequence of sensory description words corresponding to each sensory evaluation data point. Examples include [“potato”, “strong”, “nuts”, “roasted grains”, “soil”].

[0066] 3. Construct digital representation vectors from the extracted descriptive sequences.

[0067] Each sensory descriptor represents a specific aroma. Representing each specific aroma in digital form ensures that similar aromas within the representation system are close together, and the entire representation system possesses a certain degree of inferentiality. This invention constructs embedded expression vectors for words based on co-occurrence relationships, thereby achieving digital representation of aromas. The specific implementation includes the following steps:

[0068] (1) For each sensory descriptor sequence seq i =[w i 1 ,w i 2 ,w i 3 ,...,w i n Training data is generated by sequentially using a single descriptive word as the center word and the remaining words as context words. In this embodiment, the training data generated from the sensory descriptive word sequence ["potato", "intense", "nuts", "roasted grains", "soil"] is as follows:

[0069] The central word is "potato," and the contextual words are "strong," "nuts," "roasted grains," and "soil."

[0070] The central word is "intense," and the contextual words are ["potatoes," "nuts," "roasted grains," and "soil"].

[0071] The central word is "nut," and the contextual words are ["potato," "strong," "roasted grain," "soil"].

[0072] The central word is "roasted grains," and the contextual words are "potatoes," "strong," "nuts," and "soil."

[0073] The central word is "soil," and the contextual words are "potato," "strong," "nuts," and "roasted grains."

[0074] (2) A vocabulary dictionary is constructed using a sensory description lexicon as a set. An initialization center word vector matrix and an initialization context word vector matrix with dimension N are constructed, where N is the length of the set aroma digital representation vector. In this embodiment, an initialization center word vector matrix and an initialization context word vector matrix with dimension 80 are constructed, and the initialization method is a random distribution of (-1,1).

[0075] (3) Construct a word vector training model based on the Python programming language, such as Figure 2 As shown, the model consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer is a one-hot encoding of the center word. A fully connected layer connects the input layer and the first hidden layer, with weights equal to the center word vector matrix. A fully connected layer connects the first hidden layer and the second hidden layer, with weights equal to the context word vector matrix. The output layer is obtained by applying a softmax operation to the second hidden layer.

[0076] (4) Input the training dataset into the model in batches to obtain the output results. Train the parameters of the center word matrix and the context word matrix through backpropagation, providing positive feedback for context words and negative feedback for negative samples. After the model converges, use the linear combination of the center word vector matrix and the context word vector matrix as the final word vector matrix, i.e., the aroma digital representation matrix. For example, take the average of the center word vector matrix and the context word vector matrix.

[0077] 4. Use hierarchical clustering to construct a aroma wheel hierarchy.

[0078] In the flavor wheel system, aromas with broad meanings are selected as top-level aromas, such as floral, green, and fruity aromas, while other aromas are selected as bottom-level aromas, such as apple, grass, and lemon. These aromas have a certain hierarchical relationship. Furthermore, aromas can be divided into top, middle, and bottom layers, for example, fruity-citrus-lemon. In this embodiment, the fragrance ingredients are divided into top and bottom-level aromas. The specific process is as follows:

[0079] (1) Based on the digital representation vector of aroma, the number of cluster categories M is set, and the aroma is clustered using a hierarchical clustering method; the setting of parameter M can be determined according to the number of aroma wheel categories. In this embodiment, the number of categories is set to 14.

[0080] (2) Manually adjust the clustering results, including operations such as adjustment, merging, and splitting.

[0081] (3) Select aromas that represent the overall characteristics of the category as top-level aromas, such as floral, fruity, and wine-like aromas; the rest are selected as bottom-level aromas, such as floral aromas—lavender, rose, etc. The resulting hierarchy of fragrance ingredients is as follows:

[0082]

[0083]

[0084] 5. Construct an aroma transition system in the upper layer of the flavor wheel.

[0085] Flavor wheels need to reflect the transitions and similarities between aromas. The similarity between some aromas is mainly due to the common aroma-producing components often found in natural ingredients. The specific construction method is as follows:

[0086] (1) Construct a topological undirected graph consisting of the upper-layer aromas. Each upper-layer aroma is treated as a node, and the distance between nodes is the cosine of the distance between the two upper-layer aromas. The cosine formula is as follows:

[0087]

[0088] (2) Then calculate all Hamiltonian cycles of the undirected graph and calculate the length of the cycle based on the distance.

[0089] (3) Sort the lengths of the Hamiltonian circuits used and select the shortest circuit as the transition system for the upper aroma. The aroma transition system of the fragrance raw materials is: "Fruit aroma → Floral aroma → Grass aroma → Woody aroma → Spicy aroma → Mossy aroma → Smoky aroma → Roasted aroma → Nutty aroma → Milky aroma → Oily aroma → Animal aroma → Paste aroma → Wine aroma → Fruity aroma.

[0090] 6. Construct a lower aroma transition system in the flavor wheel.

[0091] Upper aroma V i The adjacent aromas are V i-1 With V i+1 Each upper layer of aroma V i They all possess their own lower-level aroma V i1 V i2 V i3 ,...,V in A lower aroma transition system needs to be constructed beneath the upper aroma wheel structure. The specific steps are as follows:

[0092] (1) For each upper-layer aroma, calculate the average vector of all upper-layer and lower-layer aromas. Taking the upper-layer aroma "fruity" as an example, its average vector is:

[0093]

[0094] Each aroma vector is the vector corresponding to the aroma in the aroma digital representation matrix.

[0095] (2) For each upper-layer aroma, calculate the distance between all its lower-layer vectors and the average vector of its adjacent upper-layer aroma. Taking "fruity aroma" as an example, calculate the distance between all its lower-layer vectors and the average vector of its adjacent upper-layer aroma, "wine aroma". That is, V 柑橘 V 浆果 V 树果 V 热带水果 Each vector in The distance is calculated. The results are sorted, and the one with the shortest distance is selected as the upper boundary lower layer aroma, i.e., V is selected. 树果 The lower aroma acts as the upper boundary of the upper aroma "fruity aroma".

[0096] (3) Calculate the distance between all lower-level vectors of the upper-level aroma and the average vector of the next lower-adjacent upper-level aroma. Taking "fruity aroma" as an example, calculate the distance between all lower-level vectors of the upper-level aroma "fruity aroma" and the average vector of the next lower-adjacent upper-level aroma "floral aroma". That is, V 柑橘 V 浆果 V 树果 V 热带水果 Each vector in The distance is calculated. The results are sorted, and the one with the shortest distance is selected as the lower boundary lower layer aroma, i.e., V is selected. 柑橘 The lower layer aroma serves as the lower boundary of the upper layer aroma, "fruity aroma".

[0097] (4) Construct a topological undirected graph consisting of all lower-level aromas under the upper-level aroma "fruity". Each lower-level aroma is represented as a node, and the distance between nodes is the cosine of the distance between the two lower-level aromas. The lower-level aroma V is defined by the upper boundary. 树果 The lower layer aroma V below the starting point boundary 柑橘 Calculate all Hamiltonian paths in the undirected graph with the endpoint as the endpoint, and calculate the path length based on the distance. Sort all the obtained Hamiltonian path lengths, and select the shortest path as the transition system for the lower aroma layers within the "fruity" aroma. That is, "tree fruit → berry → tropical fruit → citrus".

[0098] Repeat the above steps until a transitional system of lower aromas within all upper aromas is established.

[0099] 7. Construct the flavor wheel for the target product.

[0100] The flavor wheel of the product is constructed by sorting the specific aromas of each level within the transition system of each layer and displaying them in the form of a two-sided or multi-layered annular diagram. For example, in this embodiment, the lower layer aroma "tree fruit" at the upper boundary of "fruit aroma" is connected to the lower layer aroma "fruit wine" at the lower boundary of the adjacent upper layer aroma "wine aroma," and the lower layer aroma "citrus" at the lower boundary of "fruit aroma" is connected to the lower layer aroma "iris" at the upper boundary of the adjacent upper layer aroma "floral aroma," thus obtaining the lower layer aroma transition system between the upper layer aroma "fruit aroma" and the adjacent upper layer aromas. After establishing the transition system for all lower layer aromas, the flavor wheel of the product is constructed by sorting the specific flavors of each level within the transition system of each layer and displaying them in the form of an annular diagram. Figure 3 As shown.

[0101] In this embodiment, a two-layer flavor wheel is established. If necessary, three, four, or more layers of flavor wheels can be established. The transition system between layers is established according to step 5. For each additional layer, the transition system between each layer can be calculated and constructed according to step 6. Finally, step 7 is used to construct a multi-layer product flavor wheel.

[0102] This invention proposes a flavor wheel construction method based on a large amount of product sensory description data and natural language processing technology. Supported by training data, it can perform machine self-learning based on the characteristics of the data itself. Under the condition of avoiding excessive reliance on human experience, it can efficiently and intelligently construct a flavor wheel system for a specific product segment, enabling evaluators to better carry out standardized, regulated and efficient product sensory evaluation work.

Claims

1. A method for constructing a flavor wheel, characterized in that, The construction method includes the following steps: 1) Select the target product and obtain sensory description data of the target product; 2) Extract the sensory descriptive sequence of the target product from the acquired sensory evaluation data; 3) Digitize the sensory descriptions in the sensory description sequence; 4) Cluster the digitized sensory descriptions and treat each sensory description as a specific flavor to construct a flavor wheel hierarchy system for the target product. This hierarchy system includes at least two layers. The flavors with the broadest meaning in each category are taken as the first-layer flavors. The remaining flavors in each category are used to determine whether to continue dividing. If no further division is needed, the remaining flavors in each category are taken as the second-layer flavors. If further division is needed, some relatively broad flavors are selected from the remaining flavors in each category as the second-layer flavors. The remaining flavors are then further divided until no further division is possible. 5) Take the first layer flavor as the upper layer flavor, take the upper layer flavor in the flavor wheel hierarchy as nodes, construct a topological undirected graph composed of upper layer flavors, calculate the length of all Hamiltonian circuits in the undirected graph based on the distance between each node, and take the node order corresponding to the shortest Hamiltonian circuit as the transition system of upper layer flavor in the flavor wheel. 6) Taking the second layer flavor as the lower layer flavor, after the first layer transition system is established, take each flavor in the lower layer adjacent to the upper layer as a node, calculate the distance between all the lower layer flavors and the adjacent flavors in the upper layer, select the shortest distance as the upper boundary lower layer flavor and the lower boundary lower layer flavor respectively, construct a topological undirected graph composed of each lower layer flavor, calculate the Hamiltonian path length of the topological undirected graph based on the distance between each node of the lower layer flavor, and take the node order corresponding to the shortest Hamiltonian path as the transition system of the lower layer flavors inside the upper layer flavor; 7) After completing the second layer of flavor transition system, determine whether there is a third layer of flavor. If so, take the second layer of flavor as the upper layer flavor and the third layer flavor as the lower layer flavor, and repeat step 6) until the transition system of all layers of flavor is established. 8) Sort the flavors of each level in the transition system of each layer according to their specific flavors, and display them in the form of a two- or multi-layered circular diagram to form the flavor wheel of the target product.

2. The flavor wheel construction method according to claim 1, characterized in that, Step 1) involves collecting existing publicly available data or gathering sensory evaluation data.

3. The method for constructing a flavor wheel according to claim 1, characterized in that, The process of extracting the sensory description sequence of the target product in step 2) is as follows: candidate strings are extracted from text data to construct a sensory description lexicon; the sensory description lexicon is introduced into the word segmentation model as a dictionary; the sensory evaluation data is segmented using the word segmentation model; the word segmentation results are compared with the sensory description lexicon to extract sensory description words and obtain the corresponding sensory description word sequence.

4. The method for constructing a flavor wheel according to claim 3, characterized in that, The construction process of the sensory description lexicon is as follows: A. Collect all the text data from the sensory description data into a corpus S; B. Extract candidate strings, and take all string fragments with a length between 1 and L that appear in the corpus as the candidate string set; C. Calculate the term frequency of the candidate strings; D. Calculate the aggregation degree of the candidate strings and select the candidate string with the lowest mutual information value as the aggregation degree; E. Calculate the adjacency entropy of the candidate string, and take the lower value between the left and right adjacency entropies as the adjacency entropy of the candidate string; F. By setting thresholds for three indicators—word frequency, aggregation degree, and adjacency entropy—or any combination of these three indicators, candidate strings that do not meet the conditions for word formation are removed, while candidate strings that do meet the conditions for word formation are retained, ultimately forming a set of descriptive words for the corpus. G. Filter the set of descriptive words to form a sensory descriptive word library.

5. The method for constructing a flavor wheel according to claim 1, characterized in that, The digitization process in step 3) is as follows: A. For each sensory descriptor sequence seq i = [w i 1 , w i 2 , w i 3 , ..., w i n Training data is generated by sequentially using a single descriptive word as the center word and the remaining words as context words, where n is the length of the sensory descriptive word sequence; B. Construct a vocabulary dictionary using a sensory description lexicon as a set, and construct an initialization center word vector matrix and an initialization context word vector matrix with dimension N, where N is the length of the set flavor digital representation vector; C. Construct a word vector training model, input the training dataset into the model, and obtain the output results.

6. The method for constructing a flavor wheel according to claim 1, characterized in that, Step 4) uses a hierarchical clustering method to construct a flavor wheel hierarchy.

7. The method for constructing a flavor wheel according to claim 1, characterized in that, The distance between nodes in step 5) is the cosine distance.

8. The method for constructing a flavor wheel according to claim 1, characterized in that, Step 6) involves constructing the lower-level flavor transition system of the flavor wheel as follows: A. For each upper flavor layer, calculate the average vector of all upper and lower flavor layers; B. For each upper flavor, calculate the distance between all its lower vectors and the average vector of its upper neighboring upper flavor, sort the calculation results, and select the one with the shortest distance as the upper boundary lower flavor; C. Calculate the distance between all lower-level vectors of the upper-level flavor and the average vector of the lower adjacent upper-level flavor, sort the calculation results, and select the one with the shortest distance as the lower boundary lower-level flavor; D. Construct a topological undirected graph consisting of all lower-level flavors under the upper-level flavor, calculate the distance and Hamiltonian path length between each lower-level flavor node, and take the node order corresponding to the shortest Hamiltonian path as the transition system of lower-level flavors within the upper-level flavor.

Images