Method for product feature extraction and sentiment analysis on product online reviews

Abstract

The application discloses a kind of product feature extraction and sentiment analysis for product online comment, including 1) data acquisition: using crawler tool or software in internet e-commerce platform obtains the comment text related to target product;2) data preprocessing;3) product feature extraction: the target product comment text after pre-processing is input into product feature extraction model, and the feature extraction result of target product is obtained;4) sentiment analysis: the target product feature obtained is fused with feature label, as the input data of sentiment analysis model, and the product feature expressed in comment text and corresponding sentiment analysis result are output.The application takes the online comment text of product in internet e-commerce platform as research object, solves the problem of insufficient product feature extraction in existing product feature sentiment analysis, and further considers the recognition of irony sentence, effectively improves the accuracy of sentiment analysis.

Description

Technical Field

[0001] This invention relates to the field of data mining and analysis technology, specifically to a method for extracting product features and performing sentiment analysis on online product reviews. Background Technology

[0002] As market competition intensifies, products need to quickly and accurately grasp consumers' most pressing needs to stand out in the industry. In recent years, due to the booming development of e-commerce and social media technologies, the vast majority of users choose to purchase goods on major e-commerce platforms and post related online product reviews to express their opinions and needs. Compared to traditional survey data, online product review data is large in volume, low in collection cost, and because it comes from consumers' active sharing rather than passive questioning, it can more realistically reflect consumers' satisfaction with related products and their emotional needs. Therefore, accurately analyzing the emotional polarity expressed by users regarding various aspects of a product in online reviews can not only improve product design efficiency but also strengthen user orientation and enhance the product's competitiveness in the market.

[0003] Current product feature sentiment analysis methods rely on part-of-speech tags as a selection criterion to extract product features from online product reviews. Furthermore, existing research on implicit product feature extraction suffers from computational complexity and poor generalization ability, leading to reduced feature extraction effectiveness and impacting subsequent sentiment analysis results. On the other hand, due to current online language habits, many users employ irony to emphasize negative emotions towards certain things or phenomena. Irony is also evident in online product reviews, but existing sentiment analysis studies rarely consider it. Correctly identifying the sentiment polarity of ironic comments could improve the overall accuracy of sentiment analysis and allow for a more precise understanding of user needs regarding the product. Summary of the Invention

[0004] The purpose of this invention is to address the problem of insufficient product feature extraction in existing product feature sentiment analysis, and to further improve the accuracy of sentiment analysis. Taking online review texts of products on internet e-commerce platforms as the research object, this invention provides a gated capsule network product feature sentiment analysis method that considers irony recognition.

[0005] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution.

[0006] A method for extracting product features and performing sentiment analysis on online product reviews includes the following steps: Step S1: Use web crawling tools or data crawling software to obtain review texts related to the target product from internet e-commerce platforms; Step S2: Preprocess the obtained target product review text, including removing useless review text, segmenting long sentences, deleting stop words, and adding tags; Step S3: Build a product feature extraction model. Input the preprocessed target product review text into the product feature extraction model to obtain the feature extraction results of the target product. Step S4: Build a sentiment analysis model, integrate the acquired target product features and feature tags as input data for the sentiment analysis model, and output the product features expressed in the comment text and the corresponding sentiment analysis results.

[0007] Specifically, the steps in S2, including removing useless comment text, segmenting long sentences, deleting stop words, and adding tags, are as follows: Remove useless comment text: Remove meaningless or too short and irrelevant text from the comment text; Long sentence segmentation: Based on the set word count requirement and the location of the product features involved, the text exceeding the set word count requirement is segmented without destroying the semantics of the original comment text; Remove stop words: Remove stop words from the comment text based on the existing stop word list; Add tags: Add tags to product review text, including product feature tags, sentiment polarity tags, and ironic semantic tags; product feature tags are categorized according to the product's basic functions and characteristics; sentiment tags are divided into positive sentiment and negative sentiment; ironic tags are divided into non-ironic and ironic categories.

[0008] Specifically, step S3 involves building a product feature extraction model, which consists of three parts: The first part is the word embedding layer, which is used to transform the preprocessed target product review text into 128-dimensional word embedding data, which serves as the input data for the product feature extraction model. ; The second part is the computation layer, which includes capsule channels and gated channels. The input data obtained after word embedding layer processing will enter the two channels simultaneously for computation. The capsule channel consists of a convolutional layer, an initial capsule layer, and a digital capsule layer. After the data enters the channel, it will undergo the following series of calculations. First, the input data will undergo convolution operations in the convolutional layer: (1); In the above formula, * represents the convolution operation; The first convolution operation represents the local features obtained after the first convolution operation, and is a scalar matrix; I represents the input data of the model, and c represents the convolution kernel used in the current convolutional layer. Secondly, the obtained local features Further analysis of the activation function: (2); In the above formula, Represents the local features after activation. This indicates that the activation function used is the ReLU function; After initial activation, The data will then be further transferred to the initial capsule layer for calculation. (3); In the above formula, This represents the feature information further extracted from the initial capsule layer; d represents the convolution kernel used in the initial capsule layer; different from The attributes of each element in the array are scalars. All elements in the vector are vectors; Finally, calculate the output result of the capsule channel. : After the initial capsule layer calculations, the output results of the capsule channels are obtained through dynamic routing operations of the digital capsule layer. : 1) Normalization (4); In the above formula, The coupling coefficient is... is the prior probability of capsule i in layer L to capsule j in layer L+1, and k represents the total number of capsules in layer L; 2) Calculate the linear combination input of the upper capsule. (5); In the above formula, For the input vector, For each L-layer capsule i, the prediction vector for L+1-layer capsule j; The output of capsule i in layer L, i.e. The elements in; 3) Calculate the output of the upper capsule and compress it. (6); In the above formula, Output for L+1 layer capsule; 4) Weight update (7); 5) Repeat steps 1) to 4) to finally obtain the output of the capsule channel. : (8); In the above formula, The value is in the interval [0,1]. The closer the value is to 1, the higher the probability that the current input information contains n features. The gated channel consists of a gated unit layer and a max pooling layer. After the input data enters the channel, the following calculations will be performed: First, determine what the unit needs to update, including deciding which previously stored information to forget and how much current input information to absorb, as shown in the following formula: (9); In the above formula, This indicates the proportion of information read from the previous cell unit; This indicates the proportion of input information read from the current cell unit; and These are the weight matrices for the update gate and the reset gate, respectively; This refers to the information data from the previous gate control unit; Input data for the current gate control unit; Calculate the Sigmoid function; Secondly, the information to be passed to the next sequence unit and the output information of this unit are determined by the following formula: (10); In the above formula, This is the current gate control unit information data; Input information data for candidates; This is the hidden layer weight coefficient matrix; Finally, calculate the output of the gated channel: (11); In the above formula, The output gate weights; Output results for the gated channel; This indicates the probability that the current input information contains n features; The third part is the integration layer. After the calculations are performed through the two channels and the corresponding results are obtained, the results are integrated and categorized to obtain the final extraction result of the product feature extraction model: (12); In the above formula, This represents the calculation result after initial fusion; β and γ are weight coefficients, both with an initial value of 0.5. During the model training process, the weight coefficients will be automatically adjusted according to the formula based on the ratio of the calculated loss values ​​of the capsule channel and the gated channel. and These are the calculated loss values ​​for the capsule channel and the gated channel, respectively. The features contained in the current input information.

[0009] Specifically, step S4 describes building a sentiment analysis model, which consists of three parts: The first part is the labeling layer, used to label... Feature label fusion involves inserting a single column vector containing the corresponding feature label for each comment sentence before the comment sentence vector data. This vector represents the product feature expressed in the comment sentence. The fused data is denoted as... , as input data for the constructed sentiment analysis model; The second part is the computation layer, which includes the irony recognition channel and the sentiment analysis channel; The irony recognition channel consists of a convolutional layer, an initial capsule layer, and a digital capsule layer. This part performs binary classification to complete the irony semantic recognition of the input data, and adds an irony semantic judgment value to each comment. If the irony recognition channel determines that the comment in the current operation contains ironic semantics, then... =1, otherwise =0; The sentiment analysis channel consists of a gating unit layer and a Softmax layer. It is responsible for performing sentiment polarity binary classification on the currently input comment sentence based on the mined sentiment semantic features, and at the same time, it adds a sentiment polarity judgment value to each currently input comment sentence. If the sentiment analysis channel determines that the sentiment polarity of the comment in the current operation is positive, then take... =1, otherwise =0; The third part is the fusion layer, which fuses the ironic semantic judgment value and sentiment polarity judgment value from the second part using the following discriminant formula, and finally outputs the features expressed by the comment sentence and the corresponding final sentiment polarity analysis results: (13); In the above formula, =0 indicates that the final sentiment polarity of the current comment is negative. =1 indicates that the final sentiment polarity of the current comment is positive.

[0010] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses a large amount of real data crawled from the Internet as research data. The model established by this invention has high credibility. It can not only realistically explore consumers' evaluation of various aspects of product features, but also provide reasonable data support for product design and improvement.

[0011] 2. Based on a comprehensive consideration of the shortcomings of current research on product feature sentiment analysis and a comparison of the characteristics and advantages of different representative deep learning technologies, a novel deep learning model combining gating mechanisms and capsule networks is proposed to improve the efficiency and accuracy of feature extraction and sentiment analysis in product feature sentiment analysis.

[0012] 3. Considering the impact of irony semantics on the sentiment analysis results of product features, a sentiment analysis model that combines irony recognition with gated capsule networks is proposed. Based on traditional binary sentiment analysis, an irony semantic recognition channel is added to further improve the accuracy of product feature sentiment analysis. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this disclosure and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the overall method in this invention; Figure 2 This is a structural diagram of the product feature extraction model in this invention; Figure 3 This is a structural diagram of the sentiment analysis model in this invention. Detailed Implementation

[0015] To facilitate understanding and implementation of the present invention by those skilled in the art, the various steps of the method proposed in this invention are described in detail below, and experimental procedures are given. It should be understood that the implementation examples described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] Example like Figure 1 As shown, this invention provides a method for extracting product features and performing sentiment analysis on online product reviews, including the following steps: Step S1: Use web crawling tools or data crawling software to obtain review texts related to the target product from internet e-commerce platforms; In this embodiment, Octopus web crawler software is used to randomly collect online user reviews of a certain model of mobile phone from the JD.com e-commerce platform in China. Step S2: Preprocess the obtained target product review text, including removing useless review text, segmenting long sentences, deleting stop words, and adding tags; Step S3: Build a product feature extraction model. Input the preprocessed target product review text into the product feature extraction model to obtain the feature extraction results of the target product. Step S4: Build a sentiment analysis model, integrate the acquired target product features and feature tags as input data for the sentiment analysis model, and output the product features expressed in the comment text and the corresponding sentiment analysis results.

[0017] It should be noted that after building the above product feature extraction model and sentiment analysis model, in order to better train and verify the effectiveness of the model, the dataset is divided into three parts according to the structure of training set, test set and validation set, to further verify the generalization ability of the model proposed in this embodiment. The dataset is divided into training set, test set and validation set in a ratio of 2:2:6.

[0018] Specifically, the steps in S2, including removing useless comment text, segmenting long sentences, deleting stop words, and adding tags, are as follows: Remove useless comment text: Remove meaningless or too short and irrelevant text from the comment text; Long sentence segmentation: Based on the set word count requirement and the location of the product features involved, the text exceeding the set word count requirement is segmented without destroying the semantics of the original comment text; Remove stop words: Remove stop words from the comment text based on the existing stop word list; Add tags: Add tags to product review text, including product feature tags, sentiment polarity tags, and ironic semantic tags; product feature tags are categorized according to the product's basic functions and characteristics; sentiment tags are divided into positive sentiment and negative sentiment; ironic tags are divided into non-ironic and ironic categories.

[0019] Specifically, step S3 involves building a product feature extraction model. The main parameters of the product feature extraction model are shown in Table 1 below: Table 1. Parameters of GRU-CAP Product Feature Extraction Model ; In this embodiment, the product feature extraction model consists of three parts: The first part is the word embedding layer, which is used to transform the preprocessed target product review text into 128-dimensional word embedding data, which serves as the input data for the product feature extraction model. ; The second part is the computation layer, which includes capsule channels and gated channels. The input data obtained after word embedding layer processing will enter the two channels simultaneously for computation. The capsule channel consists of a convolutional layer, an initial capsule layer, and a digital capsule layer. After the data enters the channel, it will undergo the following series of calculations. First, the input data will undergo convolution operations in the convolutional layer: (1); In the above formula, * represents the convolution operation; The first convolution operation represents the local features obtained after the first convolution operation, and is a scalar matrix; I represents the input data of the model, and c represents the convolution kernel used in the current convolutional layer. Secondly, the obtained local features Further analysis of the activation function: (2); In the above formula, Represents the local features after activation. This indicates that the activation function used is the ReLU function; After initial activation, The data will then be further transferred to the initial capsule layer for calculation. (3); In the above formula, This represents the feature information further extracted from the initial capsule layer; d represents the convolution kernel used in the initial capsule layer; different from The attributes of each element in the array are scalars. All elements in the vector are vectors; Finally, calculate the output result of the capsule channel. : After the initial capsule layer calculations, the output results of the capsule channels are obtained through dynamic routing operations of the digital capsule layer. : 1) Normalization (4); In the above formula, The coupling coefficient is... is the prior probability of capsule i in layer L to capsule j in layer L+1, and k represents the total number of capsules in layer L; 2) Calculate the linear combination input of the upper capsule. (5); In the above formula, For the input vector, For each L-layer capsule i, the prediction vector for L+1-layer capsule j; The output of capsule i in layer L, i.e. The elements in; 3) Calculate the output of the upper capsule and compress it. (6); In the above formula, Output for L+1 layer capsule; 4) Weight update (7); 5) Repeat steps 1) to 4) to finally obtain the output of the capsule channel. : (8); In the above formula, The value is in the interval [0,1]. The closer the value is to 1, the higher the probability that the current input information contains n features. The gated channel consists of a gated unit layer and a max pooling layer. After the input data enters the channel, the following calculations will be performed: First, determine what the unit needs to update, including deciding which previously stored information to forget and how much current input information to absorb, as shown in the following formula: (9); In the above formula, This indicates the proportion of information read from the previous cell unit; This indicates the proportion of input information read from the current cell unit; and These are the weight matrices for the update gate and the reset gate, respectively; This refers to the information data from the previous gate control unit; Input data for the current gate control unit; Calculate the Sigmoid function; Secondly, the information to be passed to the next sequence unit and the output information of this unit are determined by the following formula: (10); In the above formula, This is the current gate control unit information data; Input information data for candidates; This is the hidden layer weight coefficient matrix; Finally, calculate the output of the gated channel: (11); In the above formula, The output gate weights; Output results for the gated channel; This indicates the probability that the current input information contains n features; The third part is the integration layer. After the calculations are performed through the two channels and the corresponding results are obtained, the results are integrated and categorized to obtain the final extraction result of the product feature extraction model: (12); In the above formula, This represents the calculation result after initial fusion; β and γ are weight coefficients, both with an initial value of 0.5. During the model training process, the weight coefficients will be automatically adjusted according to the formula based on the ratio of the calculated loss values ​​of the capsule channel and the gated channel. and These are the calculated loss values ​​for the capsule channel and the gated channel, respectively. The features contained in the current input information.

[0020] Specifically, the main parameters of the sentiment analysis model described in step S4 are shown in Table 2 below: Table 2. Parameters of GRU-CAP Sentiment Analysis Model ; In this embodiment, the sentiment analysis model consists of three parts: The first part is the labeling layer, used to label... Feature label fusion involves inserting a single column vector containing the corresponding feature label for each comment sentence before the comment sentence vector data. This vector represents the product feature expressed in the comment sentence. The fused data is denoted as... , as input data for the constructed sentiment analysis model; The second part is the computation layer, which includes the irony recognition channel and the sentiment analysis channel; The irony recognition channel consists of a convolutional layer, an initial capsule layer, and a digital capsule layer. This part performs binary classification to complete the irony semantic recognition of the input data, and adds an irony semantic judgment value to each comment. If the irony recognition channel determines that the comment in the current operation contains ironic semantics, then... =1, otherwise =0; The sentiment analysis channel consists of a gating unit layer and a Softmax layer. It is responsible for performing sentiment polarity binary classification on the currently input comment sentence based on the mined sentiment semantic features, and at the same time, it adds a sentiment polarity judgment value to each currently input comment sentence. If the sentiment analysis channel determines that the sentiment polarity of the comment in the current operation is positive, then take... =1, otherwise =0; The third part is the fusion layer, which fuses the ironic semantic judgment value and sentiment polarity judgment value from the second part using the following discriminant formula, and finally outputs the features expressed by the comment sentence and the corresponding final sentiment polarity analysis results: (13); In the above formula, =0 indicates that the final sentiment polarity of the current comment is negative. =1 indicates that the final sentiment polarity of the current comment is positive.

[0021] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for extracting product features and performing sentiment analysis on online product reviews, characterized in that, Includes the following steps: Step S1: Use web crawling tools or data crawling software to obtain review texts related to the target product from internet e-commerce platforms; Step S2: Preprocess the obtained target product review text, including removing useless review text, segmenting long sentences, deleting stop words, and adding tags; Step S3: Build a product feature extraction model. Input the preprocessed target product review text into the product feature extraction model to obtain the feature extraction results of the target product. The product feature extraction model consists of three parts: The first part is the word embedding layer, which is used to transform the preprocessed target product review text into 128-dimensional word embedding data, which serves as the input data for the product feature extraction model. ; The second part is the computation layer, which includes capsule channels and gated channels. The input data obtained after word embedding layer processing will enter the two channels simultaneously for computation. The third part is the integration layer. After the calculations are performed through the two channels and the corresponding calculation results are obtained, the results are integrated and classified to obtain the final extraction result of the product feature extraction model. Step S4: Build a sentiment analysis model, integrate the acquired target product features and feature tags as input data for the sentiment analysis model, and output the product features expressed in the comment text and the corresponding sentiment analysis results. The sentiment analysis model consists of three parts: The first part is the labeling layer, used to label... Feature label fusion involves inserting a single column vector containing the corresponding feature label for each comment sentence before the comment sentence vector data. This vector represents the product feature expressed in the comment sentence. The fused data is denoted as... , as input data for the constructed sentiment analysis model; The second part is the computation layer, which includes the irony recognition channel and the sentiment analysis channel; The third part is the fusion layer, which integrates the ironic semantic judgment value and the sentiment polarity judgment value from the second part using a discriminant formula, and finally outputs the features expressed by the comment sentence and the corresponding final sentiment polarity analysis results.

2. The method for extracting product features and performing sentiment analysis on online product reviews according to claim 1, characterized in that, The specific details of removing useless comment text, segmenting long sentences, deleting stop words, and adding tags in step S2 are as follows: Remove useless comment text: Remove meaningless or too short and irrelevant text from the comment text; Long sentence segmentation: Based on the set word count requirement and the location of the product features involved, the text exceeding the set word count requirement is segmented without destroying the semantics of the original comment text; Remove stop words: Remove stop words from the comment text based on the existing stop word list; Add tags: Add tags to product review text, including product feature tags, sentiment polarity tags, and irony tags; Product feature tags are categorized based on the product's basic functions and characteristics; emotional tags are divided into positive and negative emotions; and ironic tags are divided into non-ironic and ironic tags.

3. The method for extracting product features and performing sentiment analysis on online product reviews according to claim 1, characterized in that, The capsule channel described in step S3 consists of a convolutional layer, an initial capsule layer, and a digital capsule layer. After the data enters this channel, it will undergo the following series of calculations. First, the input data will undergo convolution operations in the convolutional layer: （1）； In the above formula, * represents the convolution operation; The first convolution operation represents the local features obtained after the first convolution operation, and is a scalar matrix; I represents the input data of the model, and c represents the convolution kernel used in the current convolutional layer. Secondly, the obtained local features Further analysis of the activation function: （2）； In the above formula, Represents the local features after activation. This indicates that the activation function used is the ReLU function; After initial activation, The data will then be further transferred to the initial capsule layer for calculation. （3）； In the above formula, This represents the feature information further extracted from the initial capsule layer; d represents the convolution kernel used in the initial capsule layer; different from The attributes of each element in the array are scalars. All elements in the vector are vectors; Finally, calculate the output result of the capsule channel. : After the initial capsule layer calculations, the output results of the capsule channels are obtained through dynamic routing operations of the digital capsule layer. : 1) Normalization （4）； In the above formula, The coupling coefficient is... is the prior probability of capsule i in layer L to capsule j in layer L+1, and k represents the total number of capsules in layer L; 2) Calculate the linear combination input of the upper capsule. （5）； In the above formula, For the input vector, For each L-layer capsule i, the prediction vector for L+1-layer capsule j; The output of capsule i in layer L, i.e. The elements in; 3) Calculate and compress the output of the upper capsule. （6）； In the above formula, Output for L+1 layer capsule; 4) Weight update （7）； 5) Repeat steps 1) to 4) to finally obtain the output of the capsule channel. : （8）； In the above formula, The value is in the interval [0,1]. The closer the value is to 1, the higher the probability that the current input information contains n features. The gated channel consists of a gated unit layer and a max pooling layer. After the input data enters the channel, the following calculations will be performed: First, determine what the unit needs to update, including deciding which previously stored information to forget and how much current input information to absorb, as shown in the following formula: （9）； In the above formula, This indicates the proportion of information read from the previous cell unit; This indicates the proportion of input information read from the current cell unit; and These are the weight matrices for the update gate and the reset gate, respectively; This refers to the information data from the previous gate control unit; Input data for the current gate control unit; Calculate the Sigmoid function; Secondly, the information to be passed to the next sequence unit and the output information of this unit are determined by the following formula: （10）； In the above formula, This is the current gate control unit information data; Input information data for candidates; This is the hidden layer weight coefficient matrix; Finally, calculate the output of the gated channel: （11） In the above formula, The output gate weights; Output results for the gated channel; This indicates the probability that the current input information contains n features; After processing through two channels and obtaining the corresponding calculation results, the results are integrated and categorized to obtain the final extraction result of the product feature extraction model: （12）； In the above formula, This represents the calculation result after initial fusion; β and γ are weight coefficients, both with an initial value of 0.

5. During the model training process, the weight coefficients will be automatically adjusted according to the formula based on the ratio of the calculated loss values ​​of the capsule channel and the gated channel. and These are the calculated loss values ​​for the capsule channel and the gated channel, respectively. The features contained in the current input information.

4. The method for extracting product features and performing sentiment analysis on online product reviews according to claim 1, characterized in that, The irony recognition channel in step S4 consists of a convolutional layer, an initial capsule layer, and a digital capsule layer. This part performs binary classification to complete the irony semantic recognition of the input data, and adds an irony semantic judgment value to each comment. If the irony recognition channel determines that the comment in the current operation contains ironic semantics, then... =1, otherwise =0; The sentiment analysis channel consists of a gating unit layer and a Softmax layer. It is responsible for performing sentiment polarity binary classification on the currently input comment sentence based on the mined sentiment semantic features, and at the same time, it adds a sentiment polarity judgment value to each currently input comment sentence. If the sentiment analysis channel determines that the sentiment polarity of the comment in the current operation is positive, then take... =1, otherwise =0; The irony semantic judgment value and sentiment polarity judgment value in the second part are fused using the following discriminant formula to finally output the features expressed by the comment sentence and the corresponding final sentiment polarity analysis results: （13）； In the above formula, =0 indicates that the final sentiment polarity of the current comment is negative. =1 indicates that the final sentiment polarity of the current comment is positive.

Images