Product concept design decision semantic representation method based on text sentiment analysis

Abstract

The invention discloses a text sentiment analysis-based product concept design decision semantic representation method, which comprises the following steps of: firstly, concluding products with similar functions by constructing a functional structure model of a target product; user comments of the similar products are crawled and preprocessed, and a text sentiment classification model based on a pre-training model BERT is trained and constructed; then, the trained data classification model is utilized to analyze target product comments, and positive, neutral and negative preference distribution of the comments is output; and finally, clustering preference distributions of all the comments, averaging the preference distributions, and constructing an intuitionistic fuzzy set to represent evaluation semantics of different criteria in a conceptual design decision process. According to the method, the defects of indefinite evaluation semantics, single granularity and relatively high subjectivity in a traditional concept design decision process are overcome, a reliable evaluation semantic set is given to a designer in a product concept scheme decision process, and the decision efficiency and the accuracy of a decision result are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of product concept design, and in particular to a semantic representation method for product concept design decisions based on text sentiment analysis. Background Art

[0002] In the context of Industry 4.0, improving product design capabilities has become an important part of accelerating the upgrading of China's manufacturing industry. Product concept design decisions, as an important manifestation of product design capabilities, have become a key factor affecting whether products can meet the rapidly changing market needs. Concept design decision semantic representation is a field that deserves great attention in the early stages of design. Different scholars have proposed various mathematical modeling methods and theories to support product design decision semantic representation. Such as triangular fuzzy sets, trapezoidal fuzzy sets, binary semantics, Z-numbers, intuitive fuzzy sets, etc. Among them, intuitive fuzzy sets are considered to be the most consistent with human expression and have been widely used in the evaluation process of various product concept designs. Faced with different product concept schemes, using appropriate intuitive fuzzy sets to judge the various evaluation criteria of the scheme is of great significance to improving decision efficiency and decision results.

[0003] At present, the semantic representation of intuitive fuzzy evaluation mainly converts the qualitative evaluation made by designers into intuitive fuzzy numbers through existing mathematical models, or the designer directly gives intuitive fuzzy number evaluation based on existing design experience. The former method does not produce any difference in intuitive fuzzy numbers when evaluating different groups of people or different criteria, and cannot well reflect the differences between individuals and the complexity of criteria. The latter method completely relies on the subjective experience of designers, and there is a large difference in the degree of evaluation between different individuals, which leads to reduced decision-making efficiency and inaccurate decision-making results, resulting in a waste of resources in the early stage of product design. Summary of the invention

[0004] In order to overcome the problems existing in semantic representation of existing concept design decisions, the purpose of the present invention is a semantic representation method for product concept design decisions based on text sentiment analysis. The invention expands the breadth and depth of evaluation semantic sources, provides objective data, can evaluate the subjectivity of semantic generation, improves the applicability of decision semantics in product group decision-making, and improves decision efficiency and the reliability of decision results.

[0005] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0006] A semantic representation method for product concept design decision based on text sentiment analysis includes the following steps:

[0007] S1. Construct a functional structure model of the target product to obtain products with similar functions.

[0008] S2. Use user reviews of similar products as data source to train the BERT model and obtain the optimal parameters.

[0009] S3. Perform text sentiment analysis on online user reviews of the target product and output the sentiment preference distribution of each review under the evaluation criteria.

[0010] S4. Construct a Gaussian mixture model to cluster the comment sentiment distribution obtained in S3, obtain the optimal number of clusters for each criterion sentiment preference distribution, and obtain the specific sentiment preference corresponding to each cluster.

[0011] S5. According to the clustering results obtained in S4, the mean of the sentiment preference is calculated to obtain the overall sentiment preference of each cluster under each criterion.

[0012] S6. Take the overall emotional preference as the intuitive fuzzy number, and its positive, neutral, and negative emotions correspond to the intuitive fuzzy number membership, hesitation, and non-membership respectively. Sort them to form an intuitive fuzzy evaluation semantic set.

[0013] Furthermore, in step S1, the functional structure model includes a decomposition of all functions realized by the product and a structural decomposition corresponding to the functions. Similar products and the target product have similar / identical functions and structures for realizing the functions, such as food waste disposers and wall breakers, which both realize the function of crushing objects through a crushing structure.

[0014] Furthermore, the specific steps of step S2 are:

[0015] S2.1. Crawl user review data of similar products and remove review data that is irrelevant to the evaluation criteria corresponding to the product features and characteristics;

[0016] S2.2. Label each comment according to its sentiment, where 0 represents negative sentiment, 1 represents neutral sentiment, and 2 represents negative sentiment. Use this as training data for the BERT model.

[0017] S2.3. Use the ten-fold crossover method to divide the training set and the validation set into a ratio of 8:2, obtain the training accuracy curve, and adjust the parameters including the attenuation weight, training batch, and regularization parameter.

[0018] Furthermore, the specific steps of step S3 are:

[0019] S3.1, obtain user reviews of the target product;

[0020] S3.2. Delete comments that are irrelevant to the evaluation criteria corresponding to the product features;

[0021] S3.3, classify the remaining comments according to the functional features and criteria of different aspects of the product corresponding to the topics they express;

[0022] S3.4. Use the trained BERT model to perform text sentiment analysis on the comments under each criterion, and use softmax to output the sentiment preference distribution. The specific calculation method is:

[0023]

[0024] S jlz S represents the positive, neutral, and negative scores obtained by the BERT model for the lth comment of evaluation criterion j, with z values ​​of 0, 1, and 2 respectively. jl0 represents the negative sentiment score of the sentiment of the lth comment of evaluation criterion j, S jl1 represents the neutral sentiment score of the sentiment of the lth comment of evaluation criterion j, S jl2 represents the positive sentiment score of the sentiment of the lth comment of evaluation criterion j; the sentiment preference distribution form is R jl [pos(P 2 jl ), neu(P 1 jl ), neg(P 0 jl )], where P 2 jl , P 1 jl , P 0 jl They represent the degree of preference of comment l under evaluation criterion j, namely positive, neutral, and negative. The positive, neutral, and negative preference values ​​are all between (0,1), and their sum is 1.

[0025] Furthermore, the specific steps of step S4 are:

[0026] S4.1. Vectorize the sentiment preference distribution corresponding to the comments under each criterion. The specific calculation method is:

[0027]

[0028] γ jlk Represents R jl The responsibility degree belonging to each Gaussian distribution, that is, R jl The degree to which each Gaussian distribution is included. N(R jl ∣μ k ,Σ k ) represents the kth (1≤k≤K) Gaussian distribution in R jl The probability density of occurrence, π k is the mixture weight of the kth Gaussian distribution, μ k is the mean of the Gaussian distribution, ∑ kis the covariance matrix. The denominator is the variance of all Gaussian components R jl The sum of the contributions of , which ensures that the total responsibility is 1;

[0029] S4.2. To avoid overfitting, the Bayesian Information Criterion (BIC) is used to evaluate the goodness of fit. Different cluster numbers K can be selected to calculate the corresponding BIC values, and the cluster number with the minimum BIC value is selected to obtain the evaluation semantic set granularity corresponding to the evaluation criterion. The specific calculation method is:

[0030] BIC = -2ln(E) + kln(n),

[0031] E is the log-likelihood estimate, which is calculated from the probability inside the model; k is the total number of parameters, which is the sum of the number of parameters of the mean, covariance matrix and cluster weight of each cluster; n compoments represents the number of clusters, d represents the data dimension. For intuitive fuzzy numbers, the dimension is 3; n is the total number of samples.

[0032] Furthermore, the specific steps of step S5 are:

[0033] Calculate the mean of each sentiment preference under each criterion and each cluster, which represents the overall preference of the criterion and the cluster. The calculation formula is:

[0034]

[0035] in and Respectively represent the sum of negative and positive preferences corresponding to each comment in the tth cluster in evaluation criterion j; μ jt represents the membership degree of the t-th cluster of intuitionistic fuzzy numbers in evaluation criteria j, ν jt It represents the non-membership degree of the t-th cluster of intuitionistic fuzzy numbers in evaluation criterion j.

[0036] Furthermore, the specific steps of step S6 are:

[0037] The intuitionistic fuzzy number score is used to defuzzify the intuitionistic fuzzy number, and the score of each intuitionistic fuzzy number in the intuitionistic fuzzy evaluation semantic set of different granularities is obtained. The intuitionistic fuzzy numbers are sorted from small to large, and the specific intuitionistic fuzzy number composition of the intuitionistic fuzzy evaluation semantic set of different criteria is obtained. The defuzzification sorting formula of the intuitionistic fuzzy number is:

[0038]

[0039] Where, π is the hesitation degree of the intuitionistic fuzzy number;

[0040] The design idea of ​​the present invention:

[0041] First, by constructing a functional structure model of the target product, products with similar functions are summarized; user comments of these similar products are crawled and preprocessed, and a text sentiment classification model based on the pre-trained model BERT is trained and constructed; then, the target product comments are analyzed using the trained data classification model, and the positive, neutral, and negative preference distributions of the comments are output; finally, the preference distributions of all comments are clustered, and the preference distributions are averaged, and intuitive fuzzy sets are constructed to characterize the evaluation semantics of different criteria in the concept design decision-making process. The present invention solves the shortcomings of unclear evaluation semantics, single granularity, and strong subjectivity in the traditional concept design decision-making process, and provides designers with a reliable evaluation semantic set in the product concept solution decision-making process, effectively improving the decision-making efficiency and the accuracy of the decision results.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1) The intuitive fuzzy evaluation semantic set corresponding to different evaluation criteria of the product described in the present invention effectively expresses the evaluation granularity that should be selected when making decisions based on different evaluation criteria, and effectively distinguishes the complexity between the evaluation criteria.

[0044] 2) The emotional preference characterization method described in the present invention pays attention to the three types of emotions, namely positive, neutral and negative, at the same time, effectively compensating for the problem of coarse granularity in the emotional analysis process, making the characterization of emotional preferences more delicate and accurate, and effectively characterizing the user's attention and satisfaction with different criteria of the product.

[0045] 3) The intuitive fuzzy evaluation semantic set described in the present invention solves the problem of relying on the designer's empirical knowledge to subjectively give evaluation semantics or relying on mathematical models to obtain regularized evaluation semantics in the traditional concept design decision-making process. The intuitive fuzzy evaluation semantic set given takes into account the opinions of many users, making it universal, and improves decision-making efficiency and reduces decision-making time. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a method for semantic representation of product concept design decisions for text sentiment analysis according to Embodiment 1 of the present invention;

[0047] Figure 2 BERT training accuracy curve of Example 1 of the present invention;

[0048] Figure 3 This is a cluster number diagram of sentiment preference distribution under different evaluation criteria in Example 1 of the present invention;

[0049] Figure 4 This is a distribution diagram of emotion preferences under different criteria according to Example 1 of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and beneficial effects of the present invention more clearly described, the technical scheme in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0051] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0052] Example 1

[0053] like Figure 1 As shown, a semantic representation method for product concept design decision based on text sentiment analysis in this embodiment includes the following steps:

[0054] S1. Construct a functional structure model of the target product to obtain products with similar functions.

[0055] Specifically, the similar product in step S1 should have similar or identical functions and corresponding structures for realizing the functions as the target product, and its functional structure model should have a certain (high) degree of similarity.

[0056] S2. Use user reviews of similar products as the data source, label each review according to the degree of emotional expression, where 0 represents negative emotion, 1 represents neutral emotion, and 2 represents positive emotion. Train the BERT model to obtain the optimal model parameters.

[0057] Specifically, Figure 2 As shown, the training accuracy should be higher than a certain value to verify the reliability of the model.

[0058] S3. Perform text sentiment analysis on online user reviews of the target product and output the sentiment preference distribution of each review under the evaluation criteria.

[0059] Specifically, the emotional preference distribution form is R jl [pos(P 2 jl ), neu(P 1 jl ), neg(P 0 jl )], where P 2 jl , P 1 jl , P 0 jl They represent the degree of preference of the comment l of evaluation criterion j as positive, neutral, and negative. j Indicates the evaluation criteria.

[0060] Table 1. Sentiment preference distribution of online reviews corresponding to each evaluation criterion

[0061]

[0062] Specifically, the specific steps of step S3 are:

[0063] S3.1. Obtain user reviews of the target product.

[0064] S3.2. Delete comments that are not related to product features.

[0065] S3.3. Classify the remaining comments according to the functional features and criteria of different aspects of the product corresponding to the topics they express.

[0066] S3.4. Use the trained BERT model to perform text sentiment analysis on the comments under each criterion, and use softmax to output the sentiment preference distribution. The specific calculation method is:

[0067]

[0068] S jlz It indicates that the BERT model analyzes the lth comment of evaluation criterion j and obtains positive, neutral, and negative scores, with z values ​​of 0, 1, and 2 respectively.

[0069] S4. Construct a Gaussian mixture model to cluster the comment sentiment distribution obtained in S3, obtain the optimal number of clusters for each criterion sentiment preference distribution, and obtain the specific sentiment preference corresponding to each cluster.

[0070] Specifically, the specific steps for calculating the domain distance of the analog source are as follows:

[0071] S4.1. Vectorize the sentiment preference distribution corresponding to the comments under each criterion. The specific calculation method is:

[0072]

[0073] In the formula, γ jlk Represents R jl The responsibility degree belonging to each Gaussian distribution, that is, R jl The degree to which each Gaussian distribution is included. N(R jl ∣μ k ,Σ k ) represents the kth (1≤k≤K) Gaussian distribution in R jl The probability density of occurrence, π k is the mixing weight of the kth Gaussian distribution (i.e., the proportion of the Gaussian distribution in the mixture model), μ k is the mean of the Gaussian distribution, ∑k is the covariance matrix. The denominator is the variance of all Gaussian components R jl The sum of the contributions of , which ensures that the total responsibility is 1. Estimate the mean μ of each Gaussian distribution k , covariance matrix∑ k and the mixing weight π k , and repeat this step until the change of clustering parameters is very small or reaches the preset number of iterations to determine the optimal number of clusters.

[0074]

[0075] S4.2. To avoid overfitting, the Bayesian Information Criterion (BIC) is used to evaluate the goodness of fit. Different cluster numbers K can be selected to calculate the corresponding BIC values, and the cluster number with the minimum BIC value is selected. The specific calculation method is:

[0076] BIC = -2ln(E) + kln(n),

[0077] E is the log-likelihood estimate, which is calculated from the probability inside the model; k is the total number of parameters, which is the sum of the number of parameters of the mean, covariance matrix and cluster weight of each cluster; n compoments represents the number of clusters, d represents the data dimension, for intuitive fuzzy numbers, the dimension is 3; n is the total number of samples

[0078] Specifically, the number of clusters, i.e., the granularity of semantic evaluation and the specific distribution of sentiment preference are as follows: Figure 3 and Figure 4 As shown; Figure 4 The A in the three-dimensional coordinates is the positive emotional preference, corresponding to the degree of membership; B is the neutral emotional preference, corresponding to the degree of hesitation; and C is the negative emotional preference, corresponding to the degree of non-membership.

[0079] S5. According to the clustering results obtained in S4, the mean of the sentiment preference is calculated, and the mean is used as the overall sentiment preference of each cluster under each criterion.

[0080] Specifically, the specific calculation method of the emotional preference mean in step S5 is:

[0081]

[0082] in and Respectively represent the sum of negative and positive preferences corresponding to each comment in the tth cluster in evaluation criterion j. jt represents the membership degree of the t-th cluster of intuitionistic fuzzy numbers in evaluation criteria j, ν jt It represents the non-membership degree of the t-th cluster of intuitionistic fuzzy numbers in evaluation criterion j.

[0083] S6. Take the overall sentiment preference as the intuitive fuzzy number, sort the intuitive fuzzy numbers, and form an intuitive fuzzy evaluation semantic set.

[0084] The intuitionistic fuzzy number score is used to defuzzify the intuitionistic fuzzy number, and the score of each intuitionistic fuzzy number in the intuitionistic fuzzy evaluation semantic set of different granularities is obtained. The intuitionistic fuzzy numbers are sorted from small to large to obtain the specific intuitionistic fuzzy number composition of the evaluation semantic set of different criteria. The defuzzification sorting formula of the intuitionistic fuzzy number is:

[0085]

[0086] Where π is the hesitation degree of the intuitionistic fuzzy number.

[0087] Table 2 Intuitive fuzzy evaluation semantic sets corresponding to different product criteria

[0088]

[0089] The above description has described in detail the preferred embodiments and principles of the present invention. For those skilled in the art, according to the ideas provided by the present invention, there may be changes in the specific implementation methods, and these changes should all be within the scope of protection of the present invention.

Claims

1. A semantic representation method for product concept design decision based on text sentiment analysis, characterized in that: The following steps are involved: S1. Build a functional structure model of the target product and obtain similar products; S2. Use user reviews of similar products as data source to train the BERT model and obtain the optimal parameters. S3, perform text sentiment analysis on online user reviews of the target product and output the sentiment preference distribution of each review under the evaluation criteria; S4, construct a Gaussian mixture model to cluster the comment sentiment distribution obtained in S3, obtain the optimal number of clusters for each criterion sentiment preference distribution, and obtain the specific sentiment preference corresponding to each cluster; S5. According to the clustering results obtained in S4, the mean of the sentiment preference is calculated to obtain the overall sentiment preference of each cluster under each criterion; S6. Take the overall sentiment preference as the intuitive fuzzy number, sort the intuitive fuzzy numbers, and form an intuitive fuzzy evaluation semantic set.

2. According to the method of semantic representation of product concept design decision based on text sentiment analysis in claim 1, it is characterized by: In the step S1, all functions realized by the target product are decomposed, and the structures corresponding to the functions are decomposed; based on the functions of the target product and the structures realizing the functions, similar products having the same structure and realizing the same functions as the target product are summarized.

3. The method for semantic representation of product concept design decision based on text sentiment analysis according to claim 1 is characterized in that: The specific steps in step S2 are: S2.

1. Crawl user review data of similar products and remove review data whose subject expression is irrelevant to product features; S2.2, label each comment according to the sentiment it contains, where 0 represents negative sentiment, 1 represents neutral sentiment, and 2 represents negative sentiment; S2.

3. Divide the ratio of training set and validation set by the ten-fold crossover method, obtain the training accuracy curve, and adjust the parameters including attenuation weight, training batch, and regularization parameter.

4. The method for semantic representation of product concept design decision based on text sentiment analysis according to claim 1 is characterized in that: The specific steps in step S3 are: S3.1, obtain user reviews of the target product; S3.

2. Delete comments that are not related to product features; S3.3, classify the remaining comments according to the functional features and criteria of different aspects of the product corresponding to the topics they express; S3.

4. Use the trained BERT model to perform text sentiment analysis on the comments under each criterion, and use softmax to output the sentiment preference distribution. The specific calculation method is: S jlz S represents the positive, neutral, and negative scores obtained by the BERT model for the lth comment of evaluation criterion j, with z values ​​of 0, 1, and 2 respectively. jl0 represents the negative sentiment score of the sentiment of the lth comment of evaluation criterion j, S jl1 represents the neutral sentiment score of the sentiment of the lth comment of evaluation criterion j, S jl2 represents the positive sentiment score of the sentiment of the lth comment of evaluation criterion j; the sentiment preference distribution form is R jl [pos(P 2 jl ), neu(P 1 jl ), neg(P 0 jl )],R jl represents the lth comment corresponding to the evaluation criterion j of the target product, pos(P 2 jl ) represents the positive preference for the lth comment of evaluation criterion j, neu(P 1 jl ) represents the neutral preference degree of the lth comment of evaluation criterion j, neg(P 0 jl ) represents the degree of negative preference for the lth comment of evaluation criterion j.

5. According to the method for semantic representation of product concept design decision based on text sentiment analysis as described in claim 1, it is characterized in that: The specific steps of step S4 are: S4.

1. Vectorize the sentiment preference distribution corresponding to the comments under each criterion. The specific calculation method is: γ jlk Represents R jl The responsibility degree belonging to each Gaussian distribution, that is, R jl The degree of belonging to each Gaussian distribution; N(R jl ∣μ k ,Σ k ) represents the kth Gaussian distribution in R jl Probability density of occurrence, 1≤k≤K, π k is the mixture weight of the kth Gaussian distribution, μ k is the mean of the Gaussian distribution, ∑ k is the covariance matrix; S4.

2. To avoid overfitting, the Bayesian Information Criterion (BIC) is used to evaluate the goodness of fit. Different cluster numbers K are selected to calculate the corresponding BIC values, and the cluster number with the minimum BIC value is selected. The specific calculation method is: E is the log-likelihood estimate; k is the total number of parameters, which is the sum of the number of parameters for the mean, covariance matrix, and cluster weights of each cluster; n compoments represents the number of clusters, and d represents the data dimension; n is the total number of samples.

6. According to the method for semantic representation of product concept design decision based on text sentiment analysis as described in claim 1, it is characterized in that: The specific calculation method of the emotional preference mean in step S5 is: in and Represents the sum of negative and positive preferences corresponding to each comment in the tth cluster of evaluation criteria j, μ jt represents the membership degree of the t-th cluster of intuitionistic fuzzy numbers in evaluation criteria j, ν jt It represents the non-membership degree of the t-th cluster of intuitionistic fuzzy numbers in evaluation criterion j.

7. According to the method for semantic representation of product concept design decision based on text sentiment analysis as described in claim 1, it is characterized in that: The defuzzification sorting formula of the intuitive fuzzy number in step S6 is: The intuitionistic fuzzy numbers are sorted from small to large according to the defuzzification scores to form an intuitionistic fuzzy evaluation set; where L(a jt ) represents the defuzzified score of the tth cluster of intuitionistic fuzzy numbers in evaluation criteria j, π jt It represents the hesitation degree of the t-th cluster intuitionistic fuzzy number in evaluation criterion j, and its value is 1-μ jt -ν jt .

Images