Information processing system, information processing method, and program

Abstract

To clarify scientific characteristics of products contributing to specific feeling which a consumer feels with respect to the products such as cosmetics.SOLUTION: An information processing system comprises an acquisition unit and a control unit. The acquisition unit acquires sentence data indicating a sentence freely described by a consumer about the feeling which the consumer feels with respect to a plurality of products which the consumer uses for the body or clothes, and measurement data indicating a measurement result related to predetermined scientific characteristics of the product. The control unit creates measurement conversion data obtained by converting the measurement data into a numerical value or an image for each product, and words and phrases conversion data obtained by converting words and phrases extracted from the sentence data into the numerical value for each product. In addition the control unit estimates a scientific characteristic of the product contributing to feeling, by creating a learning model which predicts the words and phrases conversion data from the measurement conversion data, and analyzing the data having a high contribution degree in prediction of the words and phrases conversion data based on the learning model, among the measurement conversion data.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] The present invention relates to an information processing system, an information processing method, and a program capable of executing information processing for estimating scientific properties that contribute to the sensation felt by consumers of products that consumers use on their bodies, such as cosmetics or quasi-drugs, or products that they use on their clothing. [Background technology]

[0002] Conventionally, systems for predicting the user's feeling when using a product such as cosmetics have been known. For example, the following Patent Document 1 discloses a system for predicting the feeling when a user uses a sensory evaluation object such as cosmetics or medicines (external skin preparations) applied to the skin, using a neural network to evaluate usability and the like of the sensory evaluation object from the base property information of the sensory evaluation object. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2012-123749 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, while the technology of Patent Document 1 above can infer consumer sensitivity when using a product such as a cosmetic product from information on the base physical properties of the product, it cannot infer the scientific characteristics of substances or physical properties that cause the user to feel that sensitivity.

[0005] An object of the present invention is to provide an information processing system, an information processing method, and a program capable of clarifying the scientific characteristics of a product, such as a cosmetic product, that contribute to a particular sensation felt by consumers toward the product. [Means for solving the problem]

[0006] An information processing system according to an embodiment of the present invention includes an acquisition unit and a control unit. The acquisition unit acquires text data indicating text written by a consumer about sensations felt by the consumer for a plurality of products used on the body or clothing, and measurement data indicating measurement results regarding predetermined scientific properties of the products. The control unit generates measurement conversion data by converting the measurement data into a numerical value or image for each product, and phrase conversion data by converting words extracted from the text data into numerical values ​​for each product. The control unit further generates a learning model that predicts the phrase conversion data from the measurement conversion data, and estimates the scientific properties of the product that contribute to the sensation by analyzing data from the measurement conversion data that has a high contribution rate in the prediction of the phrase conversion data by the learning model.

[0007] An information processing method according to another aspect of the present invention includes: Obtaining text data indicating free-form text written by a consumer regarding sensations felt by the consumer with respect to a plurality of products used by the consumer on the body or clothing, and measurement data indicating measurement results regarding predetermined scientific properties of the products; generating measurement conversion data by converting the measurement data into a numerical value or an image for each of the products, and generating word conversion data by converting the words extracted from the text data into a numerical value for each of the products; generating a learning model for predicting the phrase conversion data from the measurement conversion data; The method includes estimating scientific properties of the product that contribute to the sensation by analyzing data from the measurement conversion data that has a high contribution to the prediction of the phrase conversion data by the learning model.

[0008] A program according to another aspect of the present invention includes: A step of acquiring text data representing free text written by a consumer about the sensations the consumer felt with respect to a plurality of products used on the body or clothing by the consumer, and measurement data representing measurement results regarding predetermined scientific properties of the products; generating measurement conversion data by converting the measurement data into a numerical value or an image for each of the products, and generating word conversion data by converting words extracted from the text data into a numerical value for each of the products; generating a learning model for predicting the phrase conversion data from the measurement conversion data; The method also executes a step of estimating the scientific properties of the product that contribute to the sensation by analyzing data from the measurement conversion data that has a high contribution to the prediction of the phrase conversion data by the learning model. Effect of the Invention

[0009] The information processing system of the present invention makes it possible to clarify the scientific characteristics of a product, such as a cosmetic product, that contribute to a particular sensation felt by consumers toward the product. [Brief description of the drawings]

[0010] [Figure 1] 1 is a diagram showing a configuration of a cosmetics information providing system according to an embodiment of the present invention. [Diagram 2] 2 is a diagram showing a hardware configuration of a product characteristic analysis server according to an embodiment of the present invention. FIG. [Diagram 3] 2 is a diagram showing a configuration of a database included in a product characteristic analysis server according to an embodiment of the present invention. FIG. [Figure 4] 1 is a flowchart showing the flow of a process for identifying scientific properties that contribute to the sensation of a product, performed by a product property analysis server according to an embodiment of the present invention. [Diagram 5] 1 is a graph showing the relationship between the evaluation values ​​of multiple products and the appearance rate of a certain word in reviews in one embodiment of the present invention. [Figure 6] FIG. 4 is a diagram showing a flow of product measurement in one embodiment of the present invention. [Figure 7] FIG. 2 is a diagram conceptually illustrating a process for generating a machine learning model in one embodiment of the present invention. [Figure 8] 1 is a graph showing the contribution of explanatory variables in one embodiment of the present invention. [Figure 9] 1 is a graph showing explanatory variables with high contributions transferred onto an NMR spectrum in one embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0012] [System configuration] As shown in FIG. 1, the system includes a product characteristic analysis server 100 and a plurality of user terminals 200.

[0013] The product property analysis server 100 is connected to a plurality of user terminals 200, for example, via the Internet 50. The product property analysis server 100 provides analysis information on a plurality of products that consumers use on their bodies or clothes, such as cosmetics, quasi-drugs (including daily necessities such as oral care products), medicines (skin preparations, etc.), sweat wipes, cosmetic sheets, laundry detergents, and fabric softeners. Specifically, the product property analysis server 100 estimates the scientific properties of the product that contribute to the sensations felt by consumers regarding the product, using a learning model generated based on text data and measurement data related to the product.

[0014] The user terminals 200 (200A, 200B, 200C...) are terminals used by users (users of this system such as researchers at companies that manufacture the above-mentioned cosmetics, quasi-drugs, and other products), and are, for example, smartphones, mobile phones, tablet PCs (Personal Computers), notebook PCs, desktop PCs, etc. The user terminals 200 access the product characteristic analysis server 100, receive web pages and other information related to the above-mentioned learning models, scientific properties, etc., and display them on the screen using a browser, other applications, etc.

[0015] The learning model is generated based on words extracted from text data representing free-form writing by consumers about the sensations they had regarding the product, and measurement data representing measurement results regarding specified scientific properties of the product, and predicts the words from the measurement data.

[0016] Text data is, for example, word-of-mouth data, but it can also be comments on SNS, responses to surveys, etc. Scientific properties refer to various chemical and material properties such as the ingredients or combinations of ingredients contained in a product that contribute to a certain product function, the physical properties of the product (viscosity, elasticity, contact angle, etc.), the spatial distribution of substances in the product (dispersion state of inorganic particles, etc.), and the coating state when applied to the skin (compatible mode of polymers, etc.).

[0017] The sensations felt by consumers include, for example, the feel and comfort of the product or clothing using the product on the skin, as well as visual, auditory, taste, smell and other sensations (specifically, a thick feeling, a moist feeling, a stretchy feeling, a sense of protection, a lasting feeling, a moisturizing feeling, ease of application, a fluffy feeling, a smooth feeling, etc.).

[0018] The process of generating the learning model and the process of identifying scientific property information using the model will be described in detail later.

[0019] [Hardware configuration of product characteristic analysis server] As shown in FIG. 2, the product characteristic analysis server 100 includes a central processing unit (CPU) 11, a read only memory (ROM) 12, a random access memory (RAM) 13, an input / output interface 15, and a bus 14 connecting these to each other.

[0020] The CPU 11 appropriately accesses the RAM 13 etc. as necessary, and performs various arithmetic processing while controlling all the blocks of the product characteristic analysis server 100. The ROM 12 is a non-volatile memory in which the OS, programs, various parameters, and other firmware to be executed by the CPU 11 are fixedly stored. The RAM 13 is used as a working area for the CPU 11, and temporarily stores the OS, various applications being executed, and various data being processed.

[0021] The input / output interface 15 is connected to a display unit 16, an operation reception unit 17, a storage unit 18, a communication unit 19, and the like.

[0022] The display unit 16 is a display device using, for example, a Liquid Crystal Display (LCD), an Organic ElectroLuminescence Display (OLED), a Cathode Ray Tube (CRT), or the like.

[0023] The operation reception unit 17 is, for example, a pointing device such as a mouse, a keyboard, a touch panel, or other input device. When the operation reception unit 17 is a touch panel, the touch panel can be integrated with the display unit 16.

[0024] The storage unit 18 is, for example, a non-volatile memory such as a hard disk drive (HDD), a flash memory (SSD; Solid State Drive), or other solid-state memory. The storage unit 18 stores the OS, various applications, and various data.

[0025] As will be described later, particularly in this embodiment, the memory unit 18 has programs such as applications necessary for the generation process of the learning model described later and the process of identifying scientific characteristics using the same, as well as a consumer information database, a product text information database, a product measurement information database, and a product characteristic information database.

[0026] The communication unit 19 is, for example, a NIC (Network Interface Card) for Ethernet or various modules for wireless communication such as wireless LAN, and is responsible for communication processing with the user terminal 200.

[0027] Although not shown, the basic hardware configuration of the user terminal 200 is also substantially the same as the hardware configuration of the product characteristic analysis server 100 described above.

[0028] [Database configuration of product characteristic analysis server]

[0029] 3, the product characteristic analysis server 100 has, in the storage unit 18, a consumer information database 31, a product text information database 32, a product measurement information database 33, and a product characteristic information database 34. Note that each of these databases may be stored in a storage device or a server externally connected to the product characteristic analysis server 100, rather than in the storage unit 18.

[0030] The consumer information database 31 stores attribute information of the consumer (monitor) who provided the text data information on the product for each consumer. The consumer attribute information may include general information such as name, user ID for identifying the user, age (generation), occupation, address (residential area), sex, and email address, as well as preferences and purchase history of cosmetics and other products, evaluation values ​​given to products (e.g., number of stars or numerical value), and skin-related information such as skin type.

[0031] The product text information database 32 stores, for each product, text data (such as word-of-mouth data) indicating freely written text by consumers about the feelings they had about the product, words and phrases that indicate the feelings extracted from the text data, and data in which the extracted words and phrases are converted into numerical values ​​(word and phrase conversion data; described below).

[0032] The product measurement information database 33 stores, for each product, measurement data indicating the measurement results of the product sample by the measuring device and data obtained by converting the measurement data into numerical values ​​(measurement conversion data; described later). The measurement method is not particularly limited as long as it is a method capable of measuring the scientific properties of the product sample, and for example, the nuclear magnetic resonance method (NMR) is used.

[0033] The product characteristic information database 34 stores, for each product, information on the scientific characteristics of the product that contribute to consumer sensations, identified based on a learning model generated from the above-mentioned text data and measurement data.

[0034] These databases are mutually referenced and used as necessary in the learning model generation process and scientific property identification process by the product property analysis server 100, which will be described later.

[0035] [Product characteristic analysis server operation] Next, an operation of the product characteristic analysis server 100 configured as above will be described. The operation is executed by the hardware such as the CPU 11 and the communication unit 19 of the product characteristic analysis server 100 in cooperation with the software stored in the storage unit 18. For convenience, in the following description, the CPU 11 is the subject of the operation.

[0036] FIG. 4 is a flowchart showing the flow of operations of the product characteristic analysis server 100 according to this embodiment.

[0037] As shown in the figure, the CPU 11 of the product characteristic analysis server 100 first inputs word-of-mouth data as text data related to the target product, and stores it in the product text information database 32 (step 41). For example, if the product is a cosmetic product, the CPU 11 acquires text data of the word-of-mouth data from a cosmetic product review site. The data is stored in the product text information database 32.

[0038] The word-of-mouth on the cosmetics review site is a free description of impressions posted by users (consumers) of the site on the target product, and includes, for example, evaluation value data indicating evaluation values ​​given by consumers to the product on multiple levels (for example, 8 levels from 0 to 7; the higher the number, the higher the evaluation). In this embodiment, more than 60,000 word-of-mouth reviews were used.

[0039] The text data to be input is not limited to word-of-mouth data, but may be any text freely written by any consumer, such as a questionnaire or SNS.

[0040] In this embodiment, the above word-of-mouth data was input for about 100 commercially available UV preparations (cosmetics and quasi-drugs with SPF and PA indications), but the products may be other cosmetics and quasi-drugs. Also, the products do not have to be commercially available products, and may be any products for which opinions on the products can be obtained as free-form text, such as prototype products under development.

[0041] Next, the CPU 11 executes natural language processing on the input word-of-mouth data (step 42).

[0042] Specifically, the CPU 11 performs morphological analysis (word division) on the sentences of the word-of-mouth data using a natural language processing tool. As the natural language processing tool, various Japanese natural language processing tools such as GiNZA (a natural language processing library driven by the Python language), MeCab, Janome, etc. can be freely used. In addition, parameters such as parts of speech, analysis algorithms, and N-grams to be adopted in the analysis can be freely selected and set. In addition, the language to be processed in the natural language is not limited to Japanese, and can be, for example, English or Chinese, and natural language processing libraries such as spacy for English and jieba for Chinese can be used.

[0043] Next, the CPU 11 extracts words and phrases that indicate the consumer's sensibilities from the sentences that have been subjected to the natural language processing (step 43). Specifically, the words and phrases are extracted according to the following steps (1) to (5).

[0044] (1) From the parts of speech tagged to words by the above tool, extract nouns, verbs, adjectives, and adverbs that have a large influence on the meaning of the word-of-mouth review. Furthermore, convert the extracted parts of speech into original word forms based on the tool's registered dictionary (for example, "taka" is converted to "takai").

[0045] (2) For example, using a tool such as the Python library CountVectorizer, obtain frequency information for each word in the review text. Furthermore, generate a Bag-of-Words vector by removing words that appear in 95% or more and 5% or less of the products and become analysis noise. At this time, in order to obtain an expression that reflects the meaning of the review, word frequency information is obtained when two consecutive words are considered as one word (N-gram = 2). Note that the number of words obtained is not limited to two, and may be three or more.

[0046] (3) Based on the Bag-of-Words vector generated above, calculate the term frequency (tf, %) of each word that appears in the reviews of each product using the following formula. tf(t,d) = n t,d / Σ k n k,d where n t,d indicates the number of occurrences of a given word t in a document d (in this case, product reviews), and Σ k n k,d denotes the number of occurrences of all words detected in document d.

[0047] (4) On the other hand, calculate the average rating value (e.g., 0 to 7) assigned to each review for each product to determine the average rating value for each product. Reviews that do not have a rating value written on them are excluded from the calculation of the average rating value.

[0048] (5) Calculate the correlation coefficient of Pearson, Spearman, etc. between the evaluation value and tf in a brute force manner, and extract word pairs with high correlation coefficients with the evaluation value. The extracted words can be estimated as keywords related to the points on which consumers evaluated the product. From the keywords, select keywords related to consumers' feelings about the product and set them as specific words.

[0049] Next, the CPU 11 converts the extracted words and phrases into numerical data (word and phrase conversion data) (step 44). Specifically, the tf value of the specific word is obtained as numerical information. As shown in the graph of FIG. 5, in this embodiment, the tf value of the word pair "thick" which has the highest positive correlation coefficient with the evaluation value as an important expression representing the tactile function of the preparation, is used as the word and phrase conversion data. The right side of the figure shows a graph related to "moisturizing", which has the second highest positive correlation with the evaluation value after the word pair "thick".

[0050] Here, the method of setting the numerical data is not limited to the above, and can be set freely. For example, the tf value of any keyword desired by the analyst may be used as the index value of the product function as it is. In addition, any numerical information related to the keyword calculated by other language analysis or data analysis (for example, topic model analysis, dimensionality reduction, cluster analysis, etc.) may be set.

[0051] Next, the CPU 11 inputs the measurement data of the target product (step 45). The measurement data is stored in the product measurement information database 33. Here, the flow of the product measurement process will be described. Figure 6 is a diagram showing the flow of the measurement.

[0052] As shown in the figure, a product (UV formulation) sample is pretreated for measurement, and then measured using a measuring device (nuclear magnetic resonance device). Specifically, the following steps (1) to (3) are carried out.

[0053] (1) Weigh out approximately 0.2 g of a product sample and freeze-dry it overnight using a freeze-dryer.

[0054] (2) The sample is suspended in a specified amount of deuterated chloroform solution containing an internal standard (1,4-dinitrobenzene). An ultrasonic homogenizer is used for suspension. Other substances or solutions such as trimethylsilylpropanoic acid (TSP) as the internal standard and deuterated methanol solution as the solution may also be used.

[0055] (3) After filtering the solution using a PTFE filter, one-dimensional 1H-NMR measurement is performed to obtain measurement data (NMR spectrum).

[0056] The analytical method of the product sample using the measuring equipment is not limited to the nuclear magnetic resonance method (NMR) and can be freely set. Various measuring equipment can be used, such as infrared spectroscopy, Raman spectroscopy, gas chromatography, liquid chromatography, mass spectrometry, rheological analysis, elemental analysis, gel filtration chromatography, etc. The appropriate pretreatment method is also changed appropriately depending on the measuring equipment used.

[0057] Next, the CPU 11 converts the measurement data into numerical data (measurement numerical data) (step 46). Specifically, the following steps (1) and (2) are performed using software such as Delta (JEOL, ver. 5.3.1).

[0058] (1) The phase and chemical shift of the acquired NMR spectrum are adjusted based on an internal standard.

[0059] (2) As shown in the lower part of Figure 7(B), bucket integration of the adjusted NMR spectrum is performed in the range of 0 to 12 ppm. In bucket integration, standardization is performed so that the total integral value is constant. The NMR spectrum is divided along the horizontal axis at intervals of 0.005 ppm, and the integral value in each interval is obtained. This value is used as the measurement numerical data to be input into machine learning.

[0060] The method of digitizing the measurement data is not limited to Bucket Integration and can be freely set. For example, txt and csv output data for each point of the measurement data, or information obtained from the measurement data may be digitized using some algorithm. Also, image data such as image data of the measurement data may be used instead of the numerical data.

[0061] Next, the CPU 11 generates a machine learning model based on the generated phrase conversion data and measurement conversion data (step 47).

[0062] Specifically, as shown in FIG. 7(A), the products are divided into two groups based on the median tf value of the above-mentioned "nokkuri suru" as the phrase conversion data, and a high occurrence group and a low occurrence group are defined.

[0063] Then, as shown in FIG. 1B, a machine learning model is constructed by using the two groups defined above as the objective variables for machine learning (classification problem) and the measurement conversion data (bucket integral values ​​of NMR spectra) as the explanatory variables for machine learning.

[0064] Next, the CPU 11 calculates the prediction accuracy of the machine learning model generated as described above by cross-validation, and determines whether the prediction accuracy is sufficiently high (step 48). Specifically, the prediction accuracy is analyzed by the following steps (1) to (3).

[0065] (1) Appropriately select variables for the measurement conversion data. In this embodiment, up to 30 explanatory variables are selected using recursive feature elimination with a linear SVM as a learner. Implementation is performed using the Python language.

[0066] The measurement conversion data, which is a digitalized version of the NMR spectrum, has many variables (2467 variables), and an appropriate method for variable selection is required to build a machine learning model with high discrimination and prediction accuracy. In this embodiment, variable selection using various methods such as the Filter method, the Embedded method, and the Wrapper method was examined. As a result, the method of reducing the number of variables to 30 using Recursive Feature Elimination (RFE; a method of recursively selecting variables based on the regression coefficients and importance of any machine learning model) using a linear SVM as a machine learning device achieved the highest accuracy.

[0067] (2) A machine learning model is constructed using the selected explanatory variables and objective variables. In this embodiment, a multilayer perceptron is used. The model is implemented using the Python language.

[0068] In this embodiment, discrimination prediction using five types of nonlinear machine learning methods (k-nearest neighbor method, rbf kernel support vector machine, random forest, XGBoost, and multilayer perceptron) was examined. As a result, discrimination prediction using multilayer perceptron achieved the highest prediction accuracy. No hyperparameter tuning of the model was performed, and default parameters such as scikit-learn were used.

[0069] (3) The accuracy of class discrimination prediction by combining various variable selection methods and machine learning models is calculated by 5-fold stratified cross-validation. This method is capable of averaging out the bias of the objective variable during cross-validation, and the machine learning method with the highest average discrimination accuracy over five runs is adopted.

[0070] By calculating and comparing the discrimination accuracy of each method, we select and generate the optimal machine learning model that expresses the relationship between explanatory variables and target variables. The implementation is done using Python language.

[0071] As described above, various values ​​can be used as the objective variable. As described above, the tf value may be directly used for regression prediction, or a newly calculated value may be predicted based on the tf value. For example, as in this embodiment, a machine learning model may be constructed to classify products based on the tf value and distinguish the classes.

[0072] The variable selection (reduction) method, the type of machine learning algorithm, and the method of accuracy assessment in this embodiment are merely examples, and are not limited to those described above and can be freely selected and set. For example, a method of selection based on the correlation coefficient between explanatory variables, a method of variable selection using Akaike Information Criterion (AIC), a method of reducing the dimension of variables using principal component analysis, etc. can be freely selected.

[0073] In addition, various machine learning algorithms, such as logistic regression, Lasso regression, Ridge regression, decision tree, Gaussian process regression, gradient boosting tree, LightGBM, and deep neural network, can be freely selected according to the purpose.

[0074] The method of determining prediction accuracy can be freely selected, for example, cross-validation without stratification, leave-one-out cross-validation, etc. Furthermore, the discrimination index of the optimal machine learning model can be freely selected in addition to prediction accuracy (Accuracy). For example, error indices such as mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE), and accuracy indices in the confusion matrix such as sensitivity and specificity can be freely selected.

[0075] When it is determined that the prediction accuracy is sufficiently high through the above-mentioned verification (Yes in step 48), the CPU 11 executes an interpretation process of the machine learning model (step 49).

[0076] That is, the CPU 11 calculates the importance of variables in prediction by the machine learning model. The selected / generated machine learning model is interpreted by, for example, SHAP, and a SHAP value, which means the importance of an explanatory variable, is calculated.

[0077] Here, the Shapley value in SHAP is a value that quantifies each player's contribution, calculated based on cooperative game theory.

[0078] SHAP is a method of applying the Shapley value to a machine learning model and calculating the contribution of each explanatory variable to the predicted output value by considering each explanatory variable as a player in the game. The calculated Shapley value is an approximate value and is called the SHAP value. The SHAP value reflects both negative and positive contributions. Based on the ranking of these SHAP values, explanatory variables that are important (contributing) to discrimination prediction are extracted.

[0079] Note that the variable importance that can be used is not limited to the SHAP value. For example, feature importance and permutation importance in decision tree models, regression coefficients in linear models, integrated gradients, etc. can be freely selected.

[0080] Then, the CPU 11 estimates scientific property information that contributes to the consumer's sense of the product based on the extracted explanatory variables (step 50). Specifically, the CPU 11 executes the following steps (1) to (5).

[0081] (1) Rearrange the explanatory variables in descending order of the absolute value of the SHAP value (see Figure 8).

[0082] (2) Among explanatory variables with large absolute values ​​of the SHAP value, extract explanatory variables that are meaningful as measurement data (for example, excluding information such as noise).

[0083] (3) The extracted explanatory variables are compared with the measurement data to map important measurement regions in the acquired measurement data. In this embodiment, the mapping is performed using chemical shifts on the NMR spectrum as indices (see FIG. 9).

[0084] (4) Interpret the mapped information to obtain important scientific property information. In this embodiment, information on the ingredients contained in the product is obtained from the NMR chemical shifts. As a result of the analysis, ingredient information on two types of UV absorbers and silicones was obtained.

[0085] (5) Verify the validity of the identified scientific property information. In this embodiment, a machine learning model was constructed to predict the high and low of the above-mentioned phrase "Kokkuri suru" from the NMR spectrum, the model was interpreted by SHAP, and important ingredient information was extracted from the SHAP value information. As a result, two types of UV absorbers (ethylhexyl methoxycinnamate and diethylamino hydroxybenzoyl hexyl benzoate) and silicones were extracted. The blending tendency of the extracted ingredients was confirmed from the full ingredient display of the product, and differences in the blending tendency were confirmed depending on the tf value (occurrence frequency) of "Kokkuri suru."

[0086] It has long been known that silicones and UV absorbers contribute to the feel of UV preparations, so it is considered appropriate that the above scientific properties were identified from the above phrases.

[0087] In this way, machine learning analysis using language data and measurement data made it possible to mechanically, systematically, and efficiently derive previous knowledge about product functions and scientific properties with little reliance on human thought.

[0088] The CPU 11 stores the specified scientific property information in the product property information database 34 in association with the information indicating the product and the extracted word data, and outputs the information appropriately in response to a user request. This allows a user such as a product developer to easily understand the correspondence between the words extracted from the text data regarding the sensation of a certain product and the scientific properties that contribute to the sensation.

[0089] As described above, according to this embodiment, the product characteristic analysis server 100 can contribute to the product development of researchers and others by clarifying the scientific characteristics of a product, such as a cosmetic product, that contribute to a specific sensation felt by consumers toward the product.

[0090] [Variations] Although the embodiment of the present invention has been described above, the present invention is not limited to the above-described embodiment, and it goes without saying that various modifications can be made without departing from the spirit of the present invention.

[0091] In the above embodiment, a machine learning model was generated based on the word pair "noddy" extracted from word-of-mouth data on a specific product group (UV formulation group) and the measurement data of the product. However, the present inventor generated a machine learning model in the same manner as above using other word pairs extracted from word-of-mouth data on the same product (UV formulation), as shown below, and performed analysis.

[0092] The word pair "moisturizing" was extracted separately from the word-of-mouth data for the same product, and the measurement data was used to interpret the generated machine learning model using SHAP. Important ingredient information was extracted from the SHAP value information, and ingredients such as UV absorbents, ethers with long-chain alkyl groups, and alcohol were extracted.

[0093] It has long been known that long-chain alcohols and POE alkyl ethers are commonly used for emulsification, emulsion stabilization, and viscosity adjustment. Their use in emulsification control and formulation viscosity adjustment of oils such as UV absorbents is thought to contribute to improving the moist and hydrating sensation felt by users, so the analysis results are considered to be valid.

[0094] Similarly, a machine learning model was generated using the word pair "very stretchy" extracted separately from the word-of-mouth data of the same product and the measurement data, and the results showed that ingredients such as polyether (PE) modified silicone, emulsifiers, inorganic particle surface modifiers, and ester oil were extracted.

[0095] PE-modified silicone is commonly used as a feel improver (to improve slipperiness and provide a smooth, moist feel) and emulsifier, and ester oil has long been known to have an emollient effect, so it is thought that they contribute to the stretchiness felt by users (the feeling that it stretches well), and the analysis results are therefore considered to be valid.

[0096] Similarly, when a machine learning model was generated using the word pair “protect from UV rays” extracted separately from the word-of-mouth data of the same product and the measurement data, the components triester oil and PE-modified silicone were extracted.

[0097] It has long been known that triester oil has the effect of improving the solubility of UV absorbents, and PE-modified silicone has the effect of forming a film and improving the feel, and it is believed that these contribute to the sense of protection felt by users from the product, so the above analysis results are considered to be valid.

[0098] In this way, by generating and analyzing machine learning models in the same manner as described above for other words extracted from word-of-mouth data about the same product and the measurement results of those other products, it is possible to clarify the scientific characteristics of the product that contribute to the specific sensations that consumers feel toward the product.

[0099] In the above-described embodiment, UV preparations have been given as an example of the product that is the subject of the present invention, but the present invention may also cover other products that consumers use on their bodies or clothing, such as other cosmetics, quasi-drugs (including daily necessities such as oral care products), medicines (skin preparations, etc.), sweat wipes, cosmetic sheets, laundry detergents, and fabric softeners.

[0100] In the above embodiment, only one product characteristic analysis server 100 is shown, but the processes executed by the product characteristic analysis server 100 may be distributed and executed by a plurality of servers. For example, the natural language processing of text data such as word-of-mouth data and the generation process of a machine learning model may be executed by separate servers.

[0101] Among the inventions described in the claims of this application, the invention described as an "information processing method" is one in which each step is automatically performed by at least one device such as a computer through information processing by software, and is not performed by a human using a device such as a computer. In other words, the "information processing method" is an information processing method by computer software, and is not a method in which a human operates a calculation tool called a computer. [Explanation of symbols]

[0102] 11...CPU 18...Storage section 19…Communications Department 31…Consumer information database 32...Product document information database 33...Product measurement information database 34...Product characteristics information database 100…Product characteristic analysis server 200...User terminal

Claims

1. an acquisition unit that acquires text data representing free-form text written by a consumer about the sensations the consumer felt with respect to a plurality of products used on the consumer's body or clothing, and measurement data representing measurement results regarding predetermined scientific properties of the products; a control unit that generates measurement conversion data by converting the measurement data into numerical values ​​or images for each of the products, and phrase conversion data by converting words extracted from the text data into numerical values ​​for each of the products, generates a learning model that predicts the phrase conversion data from the measurement conversion data, and estimates scientific properties of the product that contribute to the sensation by analyzing data from the measurement conversion data that has a high contribution to the prediction of the phrase conversion data by the learning model; An information processing system comprising:

2. 2. The information processing system according to claim 1, the acquiring unit acquires evaluation value data indicating an evaluation value of the product corresponding to the text data; The control unit extracts, from the text data, words that have a high correlation with the evaluation value. Information processing system.

3. 2. The information processing system according to claim 1, The control unit extracts two or more consecutive words as the phrase from the text data. Information processing system.

4. 2. The information processing system according to claim 1, The control unit stores the information indicating the product, the information indicating the extracted phrase, and the information indicating the estimated scientific property in association with each other. Information processing system.

5. Obtaining text data indicating free-form text written by a consumer about the sensations the consumer felt with respect to a plurality of products used on the body or clothing, and measurement data indicating measurement results regarding predetermined scientific properties of the products; generating measurement conversion data by converting the measurement data into numerical values ​​or images for each of the products, and phrase conversion data by converting the words extracted from the text data into numerical values ​​for each of the products; generating a learning model that predicts the phrase conversion data from the measurement conversion data; Among the measurement conversion data, data that has a high contribution to the prediction of the phrase conversion data by the learning model is analyzed, thereby estimating the scientific properties of the product that contribute to the sensation. Information processing methods.

6. acquiring text data representing free-form text written by a consumer about the sensations the consumer felt with respect to a plurality of products used on the body or clothing of the consumer, and measurement data representing measurement results regarding predetermined scientific properties of the products; generating measurement conversion data by converting the measurement data into a numerical value or an image for each of the products, and phrase conversion data by converting the words extracted from the text data into a numerical value for each of the products; generating a learning model that predicts the phrase conversion data from the measurement conversion data; a step of estimating the scientific properties of the product that contribute to the sensation by analyzing data that has a high contribution to the prediction of the phrase conversion data by the learning model, among the measurement conversion data; A program that executes the following.

Images