Method and system for identifying sentiment metaphor of low-quality data based on fuzzy granular ball modeling

Abstract

The application discloses a low-quality data sentiment metaphor recognition method and system based on fuzzy granular ball modeling, and relates to the field of natural language processing.The method comprises the following steps: S1, obtaining low-quality text data and performing labeling and semantic preprocessing to obtain text data containing linguistic information; S2, inputting the text data containing linguistic information into a word embedding model to obtain sentiment metaphor word vectors and form a word vector matrix; S3, generating a granularity list satisfying a containment threshold by using fuzzy granular ball calculation, performing feature reduction on the word vector matrix by using a variable precision dependency function, and obtaining a reduced matrix; and S4, dividing the reduced matrix into a training set and a test set, training and predicting a convolutional neural network, and obtaining a sentiment metaphor recognition result.The application selects features by using fuzzy granular ball calculation, deletes redundant features, improves the feature extraction efficiency of the convolutional neural network model, and solves the problem of mixed noise information in a large amount of text information acquisition.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing, and in particular to a low-quality data sentiment metaphor recognition method and system based on fuzzy granular ball modeling. BACKGROUND

[0002] Metaphor is a language phenomenon used for rhetoric, and also a cognitive phenomenon, which exists in various aspects of human thinking and language expression, and is one of the core problems of artificial intelligence, especially computational linguistics. In order to enable machines to understand human language and thus better perform intelligent information processing, it is necessary to deeply mine and analyze the characteristics of human language, and metaphor is a major means for humans to describe abstract concepts. In this context, the recognition and understanding of metaphor have important theoretical significance and practical application value. Understanding metaphor will help solve many natural language processing tasks, such as search engines, chat robots, machine translation, and public opinion analysis. In recent years, with the development of third-generation platform technologies such as intelligent terminals, mobile Internet, and Internet of Things, especially the rapid rise of Internet media such as WeChat and online shopping platforms, people are widely involved in comments in various fields of social life, generating a large amount of language information with implied emotional tendencies. These comments often contain a large amount of metaphors. Only by quickly and accurately analyzing this information can the needs of public opinion monitoring and intelligent data analysis be met. In life, emotions are abstract, fuzzy, and difficult to describe, and people often use metaphors to express emotions, so emotional metaphor computing is an important field of artificial intelligence and a key to promoting the development of emotional computing to a deeper level. As a result, there has been a wave of research on metaphor and emotional metaphor in the academic and industrial communities at home and abroad, and a large number of emotional metaphor computing models have emerged. As a result, emotional metaphor computing has begun a new exploration and has made a lot of valuable achievements, but it also faces great difficulties and challenges.

[0003] With the development of technology and the explosive growth of data, large-scale, high-dimensional data sets have emerged. High-dimensional data contains rich semantic information, but often also contains a large number of irrelevant and redundant features, posing great challenges to data analysis. On the one hand, redundant features increase the risk of overfitting of classifiers, easily leading to poor generalization ability of the trained model. On the other hand, a large number of irrelevant features lead to high computational complexity. Feature selection is a technique to reduce features, with the goal of finding an optimal feature subset to predict sample categories while minimizing information loss. Feature selection can solve the curse of dimensionality, reduce computational cost, and improve the interpretability of learning models. In recent years, feature selection has played a crucial role in data mining, machine learning, and data analysis. SUMMARY

[0004] In view of the problems that the obtained text data contains noise information, so that the emotional metaphor recognition result is not accurate enough, the redundant information exists in the high-dimensional data set, so that the recognition efficiency is low, and the feature extraction capability of the traditional machine learning model is weak, the present application provides a low-quality data emotional metaphor recognition method and system based on fuzzy granular ball modeling, which solves the technical problems that the existing emotional metaphor recognition technology obtains a large amount of text data with rich emotional tendency through crawling, and the text data contains noise information or redundant information, so that the emotional metaphor recognition result is not accurate enough, and the recognition efficiency is not high enough.

[0005] In one aspect, the low-quality data emotional metaphor recognition method based on fuzzy granular ball modeling has the following specific steps:

[0006] S1, obtaining low-quality text data and performing labeling and semantic preprocessing to obtain text data containing linguistic information;

[0007] S2, inputting the text data containing linguistic information into a word embedding model to obtain emotional metaphor word vectors, and composing a word vector matrix;

[0008] S3, generating a granularity list satisfying a containment threshold by using fuzzy granular ball calculation, and performing feature reduction on the word vector matrix by using a variable precision dependence function based on the granularity list to obtain a reduced matrix;

[0009] S4, dividing the reduced matrix into a training set and a test set, training a convolutional neural network using the training set to obtain a trained convolutional neural network, predicting the trained convolutional neural network using the test set, and recognizing metaphor information and emotional categories in the test set in parallel to obtain an emotional metaphor recognition result.

[0010] Preferably, the word embedding model is a Skip-Gram skip word model in a Word2Vec framework.

[0011] Preferably, the generation of the granularity list satisfying the containment threshold by using fuzzy granular ball calculation is as follows:

[0012] Fuzzy granular ball calculation is performed according to the containment threshold to obtain a granular ball list: based on the current feature set, k-means clustering is performed until the containment degree of each granular ball reaches the containment threshold, the granular balls that do not meet the condition are deleted, and the granular ball list is obtained after multiple iterations; the feature set is composed of column vectors of the word vector matrix.

[0013] Preferably, the feature reduction of the word vector matrix by using the variable precision dependence function based on the granularity list to obtain an optimal reduced matrix is as follows:

[0014] Initial granular ball calculation is performed on the word vector matrix as a data set, and an initial granular ball list GB0 and an initial dependence degree are obtained based on a feature set C the value of the dependency, i.e. the variable precision dependency function;

[0015] performing feature selection on the feature a in the feature set C s : if removing the feature a in the feature set C s , i.e. C = C-{a s}, otherwise, retaining the feature a s ;

[0016] performing feature selection on each feature in the feature set C to obtain a feature selection result, and restoring a reduction matrix based on the feature selection result;

[0017] wherein C represents a feature set composed of column vectors of a word vector matrix; D represents a decision set, D = (d1, d2), d1 represents metaphorical information, and d2 represents a sentiment category; represents a dependency based on the set C-{a s}.

[0018] Preferably, the variable precision dependency function of the decision set D with respect to the feature subset B is represented as Specifically as follows:

[0019]

[0020] wherein |·| represents the cardinality of a set; the feature subset C represents a feature set; U represents a sample set composed of row vectors of a word vector matrix; D represents a decision set, D = (d1, d2), d1 represents metaphorical information, and d2 represents a sentiment category; represents a variable precision positive region of the decision set D with respect to the feature subset B; represents a variable precision lower approximation of D t with respect to B; represents a sample set contained in a granular ball GB q that satisfies the inclusion degree ; x represents a sample, x ∈ U; ∪ represents a union set; GB represents a granular ball list; GB q represents the qth granular ball, and l represents the number of granular balls, q = 1, 2, …, l; represents the inclusion degree, represents a granular ball fuzzy neighborhood of the sample x, represents the tth fuzzy decision, and r represents the number of decisions, t = 1, 2, …, r; D t represents the tth sample decision; and α represents an inclusion degree threshold.

[0021] Preferably, for the sample x ∈ GB q , in inclusion degree is expressed as:

[0022]

[0023] wherein |·| represents the cardinality of a set; ∩ represents the intersection.

[0024] Preferably, the granular fuzzy neighborhood of sample x is a fuzzy set, and the granular fuzzy neighborhood of sample x is expressed as: contains all samples y and the membership degree corresponding to y is expressed as:

[0025]

[0026] wherein y∈U, R B (x, y) represents the fuzzy similarity between sample x and sample y, which is calculated based on normalized values.

[0027] Preferably, the convolutional neural network comprises a convolutional layer, a pooling layer, two fully connected layers and two softmax layers; the convolutional layer is connected to the pooling layer, and then is fully connected to the two parallel fully connected layers, each of which is connected to a softmax layer, one of which is used for metaphor category division and the other of which is used for sentiment category division.

[0028] Preferably, in S4, the reduced matrix is divided into a training set and a test set using a ten-fold cross-validation method.

[0029] On the other hand, the low-quality data sentiment metaphor recognition system based on fuzzy granular ball modeling comprises the following:

[0030] The data acquisition and preprocessing module is used for acquiring low-quality text data and performing labeling and semantic preprocessing to obtain text data containing linguistic information.

[0031] The word vector matrix acquisition module is used for inputting the text data containing linguistic information into a word embedding model to obtain sentiment metaphor word vectors and form a word vector matrix.

[0032] The feature reduction module is used for generating a granularity list satisfying an inclusion degree threshold by using fuzzy granular ball calculation, and performing feature reduction on the word vector matrix by using a variable precision dependence function based on the granularity list to obtain a reduced matrix.

[0033] The model training and testing module is used for dividing the reduced matrix into a training set and a test set, training the convolutional neural network using the training set to obtain a trained convolutional neural network, predicting the trained convolutional neural network using the test set, and recognizing metaphor information and sentiment categories in the test set in parallel to obtain a sentiment metaphor recognition result.

[0034] Compared with the prior art, the present application has the following beneficial effects:

[0035] The present application obtains a vector representation by performing word embedding processing on low-quality data, combines the vectors into a matrix, obtains a granular ball list satisfying a containment threshold by using fuzzy granular ball processing, determines whether a certain candidate feature can be reduced by calculating a variable precision dependence function, outputs a matrix after feature selection, and is used for training and prediction of a convolutional neural network; the process of obtaining the granular ball list removes noise data, and the feature selection process eliminates redundant features and leaves features with strong discrimination ability. The efficiency of feature extraction of the convolutional neural network model is improved, the problem of mixed noise information in a large amount of text information acquisition is solved, a higher-accuracy sentiment metaphor recognition result is obtained, and a more accurate solution is provided for the field related to sentiment metaphor. BRIEF DESCRIPTION OF DRAWINGS

[0036] The present application will be further described in detail below with reference to the accompanying drawings;

[0037] Figure 1 A flowchart of the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling of the embodiment of the present application;

[0038] Figure 2 A flowchart of the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling of the embodiment of the present application;

[0039] Figure 3 A granular ball generation process diagram of the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling of the embodiment of the present application; wherein (a) represents a granular ball list after the 2nd iteration; (b) represents a granular ball list after the 8th iteration; (c) represents a granular ball list after the 21st iteration; (d) represents a granular ball list after the 44th iteration; (e) represents a granular ball list after the 63rd iteration; (f) represents a granular ball list after the 71st iteration; (g) represents a granular ball list after the 97th iteration; and (h) represents a final granular ball list result after iteration ends;

[0040] Figure 4 A reduced matrix and convolution kernel diagram of the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling of the embodiment of the present application;

[0041] Figure 5 A single-layer convolutional neural network structure diagram of the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling of the embodiment of the present application;

[0042] Figure 6 A sentiment metaphor recognition process diagram of the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling of the embodiment of the present application;

[0043] Figure 7 Schematic diagram of the principle structure of the low-quality data emotion metaphor recognition system based on fuzzy granular sphere modeling according to an embodiment of the present invention;

[0044] Figure 8 Block diagram of the structure of the low-quality data emotion metaphor recognition system based on fuzzy granular sphere modeling according to an embodiment of the present invention. Specific implementation manners

[0045] The present invention will be further described below through specific implementation manners.

[0046] As Figure 1 and Figure 2 shown, the low-quality data emotion metaphor recognition method based on fuzzy granular sphere modeling is specifically as follows:

[0047] S1. Obtain low-quality text data, perform annotation and semantic preprocessing, and obtain text data containing linguistic information.

[0048] Quickly obtain a large amount of low-quality text information with rich metaphor information and emotional tendencies from some public network platforms (including Sina Weibo, Baidu Encyclopedia, Meituan Dianping, Zhihu Community, and / or Douban Books) through crawlers and the like. Among them, the low-quality text information refers to low-quality text data with a large amount of redundant and irrelevant information and low authenticity. Then, construct an emotion metaphor training data set through manual annotation. ​​​​​​​​​​​​

[0053] where x i is the word vector corresponding to the i-th word in the sentence S, k represents the dimension of the word vector, and each element m ij in the matrix represents the numerical value of the i-th word vector in the j-th dimension.

[0054] S3, generating a granularity list satisfying the containment threshold by using the fuzzy granular ball calculation, and performing feature reduction on the word vector matrix by using a variable precision dependency function based on the granularity list to obtain a reduced matrix.

[0055] Regarding each row of the matrix M d×k as a sample, all row vectors of the matrix M d×k can form a sample set U = (x1, x2,..., x d ); regarding each column of the matrix M d×k as a feature of the sample, all column vectors of the matrix M d×k can form a feature set C = (a1, a2,..., a k ); the decision of the sample is recorded as D = (d1, d2), where d1 and d2 respectively represent the metaphorical information and the emotional category of the sample, and according to D, the sample set U can be divided into multiple sets {D1, D2,..., D r}. For each feature a s ∈ C, the fuzzy similarity between samples x and y is calculated as where f(x, a s ) represents the normalized value of the sample x on the feature a s . Further, the fuzzy similarity based on the feature set C is calculated as Further, the fuzzy neighborhood membership degree of the sample [x] C (y) = R C (x, y), x, y ∈ U.

[0056] Generally, the decision of the sample is classified as "yes" or "no", and the fuzzy decision allows the existence of "maybe", that is, there is a range of [0, 1] to describe the membership degree of the sample to the category. The fuzzy decision membership degree of the sample x obtained from the decision set D is defined as

[0057]

[0058] where |·| represents the cardinality of the set, and [x] C is the fuzzy neighborhood of the sample x, which is a fuzzy set containing all samples y and their corresponding membership degrees [x] C (y), x, y ∈ U.

[0059] This allows for the calculation of fuzzy decisions. Here It is a fuzzy set that contains all samples x and their corresponding membership degrees.

[0060] Particle size distribution (PPD) is a data grouping method where samples with small differences in feature values ​​are grouped into the same particle size. Coverage is a quantitative metric describing particle size quality. PPD is calculated based on a coverage threshold α (0.5 ≤ α ≤ 1): It is based on the current feature subset. Perform k-means clustering until the inclusion degree of each cluster reaches α. Clusters that do not meet the condition are deleted. After multiple iterations, a list of clusters GB = {GB...} is obtained. q ,q=1,2,...,l}, such as Figure 3 As shown, Figure 3 The inclusion degree threshold is set to 0.7. For sample x∈GB q , exist The inclusion degree in is defined as:

[0061]

[0062] Where |·| represents the cardinality of the set, GB q For the q-th ball, It is the spherical fuzzy neighborhood of sample x, which is a fuzzy set containing all samples y and their corresponding membership degrees. The membership degree of the spherical fuzzy neighborhood of a sample is defined as:

[0063]

[0064] The dependency function represents the proportion of samples that can be accurately classified within feature subset B out of the total sample set. Therefore, the dependency function is an indicator of the ability of feature subset B to distinguish between samples, and its calculation can be used to determine whether a certain candidate feature can be reduced. The variable precision dependency function of decision set D relative to feature subset B is:

[0065]

[0066] Where |·| represents the cardinality of the set, and U is the total sample set. The variable-precision positive region of the decision set D with respect to the feature subset B is defined as:

[0067]

[0068] Where ∪ represents the union, D tThe approximation of variable precision with respect to B is defined as:

[0069]

[0070] where ∪ denotes the union, i.e. satisfying the inclusion degree of the granule GB q contains the set of samples, q = 1, 2,..., l.

[0071] Before feature selection, the initial granule calculation is performed on the data set, and the initial granule list GB0 and the initial dependence degree are obtained based on the feature set C For each feature a s in the feature set C s , the current granule list GB' and the dependence degree are obtained based on C-{a By comparing the dependence function, it is determined whether a candidate feature can be reduced, if This indicates that the discriminant sample ability of the feature subset C-{a s} is equal to the feature set C, so the feature a s can be removed, then let C = C-{a s}, otherwise, keep a s .

[0072] If all features in C have been checked, terminate the feature selection process, obtain the feature selection result, and output the reduction matrix

[0073]

[0074] S4, divide the reduction matrix into a training set and a test set, train the convolutional neural network using the training set, obtain the trained convolutional neural network, use the test set to predict the trained convolutional neural network, and identify the metaphor information and the sentiment category in the test set in parallel, and obtain the sentiment metaphor recognition result.

[0075] As shown in Figure 4 and Figure 5 , the convolutional neural network structure is designed, and multiple sizes of convolution kernels C1, C2, …, C m are used to extract the semantic features of the sentence. Specifically, the steps of using the convolution kernel to generate semantic features are as follows: define the convolution kernel The convolution kernel will extract the features of consecutive d' words in the sentence S, denoted as c i , the specific calculation is as follows:

[0076] c i = f(w·M'(i,j) + b)

[0077] where b∈R is a bias term and f is an activation function.

[0078] In the current model, the convolution kernel column dimension is consistent with the input matrix column dimension, so the output matrix degenerates into a vector, that is, The pooling layer will simplify the features extracted by the convolution layer, aggregate and filter the features. Here, the maximum pooling method is adopted, that is, the maximum value of each feature is extracted The most important features in the middle:

[0079]

[0080] After the convolution layer and the pooling layer operation, it is connected to two parallel full connection layers and a Softmax layer. The first full connection layer and the Softmax layer are used to calculate the scores of the input sentences as metaphor sentences and non-metaphor sentences, and divide the samples into one class with high scores. The second full connection layer and the Softmax layer are used to calculate the scores of the input sentences in the emotional categories, and divide the samples into the emotional categories (joy, good, anger, sadness, fear, evil, surprise) with high scores.

[0081] As shown in Figure 6 The low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling provided by the embodiment of the application carries out word embedding processing on low-quality text information to obtain semantic vector representation, combines these vectors into a matrix, processes the matrix by using fuzzy granular ball to obtain a granular ball list meeting a containment threshold, determines whether a candidate feature can be reduced by calculating a variable precision dependence function, and outputs a matrix after feature selection. The matrixes of the training set and the test set are input into a convolution neural network model in sequence for training and prediction, and the identification of whether it is a metaphor and the identification of the emotional category are performed in parallel to obtain a sentiment metaphor recognition result. Through such a technical solution, the process of obtaining the granular ball list removes noise data, the feature selection process eliminates redundant features, and leaves features with strong discrimination ability. The efficiency of the convolution neural network model in extracting features is improved, the problem of mixed noise information in a large amount of text information acquisition is solved, a higher-accuracy sentiment metaphor recognition result is obtained, and a more accurate solution is provided for the field related to sentiment metaphor.

[0082] As shown in Figure 7 and Figure 8 The application further discloses a low-quality data sentiment metaphor recognition system based on fuzzy granular ball modeling, which comprises:

[0083] The data acquisition and preprocessing module 801 comprises an acquisition module and a preprocessing module, and is used for acquiring low-quality text data and performing labeling and semantic preprocessing to obtain text data containing linguistic information.

[0084] The word vector matrix obtaining module 802 comprises a word embedding processing module, which is configured to input text data containing linguistic information into a word embedding model, obtain sentiment metaphor word vectors, and form a word vector matrix.

[0085] The feature reduction module 803 comprises a fuzzy granular ball calculation module, which is configured to generate a granularity list satisfying a containment threshold by using fuzzy granular ball calculation, and perform feature reduction on the word vector matrix by using a variable precision dependent function based on the granularity list to obtain a reduced matrix.

[0086] The model training and testing module 804 comprises a training module and a testing module, which are configured to divide the reduced matrix into a training set and a testing set, train a convolutional neural network by using the training set to obtain a trained convolutional neural network, use the testing set to make a prediction on the trained convolutional neural network, and in parallel, identify metaphor information and sentiment categories in the testing set to obtain a sentiment metaphor identification result.

[0087] The specific implementation of the low-quality data sentiment metaphor recognition system based on fuzzy granular ball modeling is the same as the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling, and thus will not be repeated here.

[0088] The above is only a specific embodiment of the present application, but the design concept of the present application is not limited thereto, and any non-essential modification of the present application by using this concept shall be deemed as an infringement of the protection scope of the present application.

Claims

1. A low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling, characterized in that, It comprises the following steps: S1, obtaining low-quality text data and labeling and semantic preprocessing to obtain text data containing linguistic information; S2, inputting the text data containing linguistic information into a word embedding model to obtain sentiment metaphor word vectors and forming a word vector matrix; S3, generating a granularity list satisfying a containment threshold by fuzzy granular ball calculation, and performing feature reduction on the word vector matrix by a variable precision dependency function based on the granularity list to obtain a reduced matrix; S4, dividing the reduced matrix into a training set and a test set, training a convolutional neural network using the training set to obtain a trained convolutional neural network, predicting the trained convolutional neural network using the test set, and identifying metaphor information and sentiment categories in the test set in parallel to obtain a sentiment metaphor recognition result; The variable precision dependency function based on the granularity list is used to perform feature reduction on the word vector matrix to obtain an optimal reduced matrix, which is as follows: performing initial granule computation on a word vector matrix as a data set, based on a feature set obtaining an initial granule list and an initial dependency ; the dependency is a value of a variable precision dependency function; performing feature selection, filtering features in the feature set : if , remove features in the feature set , i.e., let , otherwise, keep the features ;​​ For feature set Each feature in the matrix is ​​selected to obtain the feature selection result, and the reduced matrix is ​​restored based on the feature selection result. wherein, represents a feature set, consisting of column vectors of the word vector matrix; D represents a decision set, , represents metaphor information, represents an emotion category; represents a dependency degree based on the set of features. set of decisions variable accuracy dependency function with respect to a subset of features is represented as in particular as follows: in, Cardinality of a set; characteristic subset , Represents a set of features; The sample set represents the set of words, which consists of the row vectors of the word vector matrix; D represents the decision set. , To convey metaphorical information, Indicates the emotion category; Let D represent the variable-precision positive region of the decision set D relative to the feature subset B; express Approximation with variable precision relative to B; Indicates that the inclusion degree is satisfied. granules The set of samples included; Indicates a sample, ; Represents the union; Represents a list of particles; This represents the q-th ball. Indicates the number of pellets. ; Indicates the degree of inclusion. Indicates sample The fuzzy neighborhood of the sphere Indicates the first A vague decision, Indicates the number of decisions. ; Indicates the first Individual sample decision; This indicates the inclusion threshold. 2.The low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling according to claim 1, wherein, The word embedding model is a Skip-Gram model in the Word2Vec framework. 3.The low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling according to claim 1, wherein, The granularity list satisfying the containment threshold is generated by fuzzy granular ball calculation, which is as follows: Based on the current feature set, k-means clustering is performed until the containment of each granular ball reaches the containment threshold, and the granular balls that do not meet the condition are deleted, and the granular ball list is obtained after multiple iterations; the feature set is composed of column vectors of the word vector matrix. 4.The low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling according to claim 1, wherein, For the sample , In the inclusion degree is expressed as: wherein denotes the cardinality of a set; denotes the intersection.

5. The low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling according to claim 4, characterized in that, The granulocyte fuzzy neighborhood is a fuzzy set, sample The granulocyte fuzzy neighborhood Contains all samples y and the membership degree corresponding to y , denoted as: wherein, , denotes the fuzzy similarity of sample x and sample y, calculated based on normalized values. 6.The low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling according to claim 1, wherein, The convolutional neural network comprises a convolutional layer, a pooling layer, two fully connected layers and two softmax layers; the convolutional layer is connected to the pooling layer, and then connected to two parallel fully connected layers, each fully connected layer is connected to a softmax layer, one softmax layer is used for metaphor category division, and the other softmax layer is used for sentiment category division. 7.The low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling according to claim 1, wherein, In S4, the reduced matrix is divided into a training set and a test set using ten-fold cross-validation.

8. A low-quality data sentiment metaphor recognition device based on fuzzy granular ball modeling using the low-quality data sentiment metaphor recognition method based on fuzzy granular ball modeling according to any one of claims 1-7, comprising the following: A data acquisition and preprocessing module for obtaining low-quality text data and labeling and semantic preprocessing to obtain text data containing linguistic information; A word vector matrix acquisition module for inputting the text data containing linguistic information into a word embedding model to obtain sentiment metaphor word vectors and forming a word vector matrix; A feature reduction module for generating a granularity list satisfying a containment threshold by fuzzy granular ball calculation, and performing feature reduction on the word vector matrix by a variable precision dependency function based on the granularity list to obtain a reduced matrix; A model training and testing module for dividing the reduced matrix into a training set and a test set, training a convolutional neural network using the training set to obtain a trained convolutional neural network, predicting the trained convolutional neural network using the test set, and identifying metaphor information and sentiment categories in the test set in parallel to obtain a sentiment metaphor recognition result.

Images