KR20220118703A - Machine Learning based Online Shopping Review Sentiment Prediction System and Method - Google Patents

Abstract

The present invention relates to a machine learning-based online shopping review sentiment prediction system and method. The machine learning-based online shopping review sentiment prediction system and method include a prediction model generation unit generating a sentiment prediction model based on review data after collecting pre-written review data and a server outputting a sentiment prediction result on real-time reviews to be transmitted to a user terminal when the real-time reviews are input to the sentiment prediction model.

Description

Machine Learning based Online Shopping Review Sentiment Prediction System and Method}

The present invention relates to a machine learning-based online shopping review emotion prediction system and method. More specifically, after collecting previously written review data, an emotion prediction model is generated based on the review data, and a real-time review is applied to the emotion prediction model. It relates to a machine learning-based online shopping review emotion prediction system and method for outputting emotion prediction results for real-time reviews when inputted.

An online shopping mall or an internet shopping mall refers to a virtual store created to allow products to be bought and sold using the Internet. Unlike offline shopping malls, online shopping malls do not have time and spatial restrictions, so consumers can purchase products at relatively low prices anytime, anywhere. From the standpoint of an online shopping mall operator, it is possible to provide product information that meets the tastes of consumers without restrictions on the commercial district. The operator of the online shopping mall may directly sell the product to the consumer or relay the sale by connecting the seller and the consumer. The online shopping market is rapidly growing in recent years due to the development of the logistics industry, the development of information and communication technology, the convenience of purchasing products, and a global pandemic such as Corona 19.

Online shopping mall managers are incorporating various technologies to provide better services to consumers in line with the growth of online shopping malls. Related Document 1 relates to a method for recommending online shopping products based on customer's offline activity data. By accurately analyzing customers' interests, tendencies, and consumption patterns, it is possible to recommend products based on similar purchase histories of other customers. Related Document 2 relates to an online shopping system using artificial intelligence, and by learning offline shopping information prior to online shopping, product arrangement, appearance order, price adjustment, etc. can be performed.

However, it is very important for consumers to make purchase decisions by referring to actual reviews and reviews of previous buyers as the most important indicator when purchasing products from an online shopping mall. . In other words, it is necessary for online shopping mall managers to develop technologies that can provide better services by indirectly receiving feedback from consumers by using actual reviews and reviews of previous buyers.

KR 10-1740148 KR 10-2167345

The present invention is to solve the above problems, and after collecting pre-written review data so as to automatically analyze the emotions of consumers included in a large amount of reviews written on a web page, emotion prediction based on the review data A machine learning-based online shopping review appraisal comprising a predictive model generator for generating a model, and outputting an emotion prediction result for the real-time review when a real-time review is input to the emotion prediction model, and then transmitting the emotion prediction result to the user's terminal It aims to obtain a prediction system and method.

In order to achieve the above object, the machine learning-based online shopping review emotion prediction system of the present invention includes a prediction model generator that collects pre-written review data and then generates an emotion prediction model based on the review data, and predicts the emotion. When a real-time review is input to the model, the server outputs the emotion prediction result for the real-time review and then transmits the emotion prediction result to the user's terminal.

In order to achieve the above object, the machine learning-based online shopping review emotion prediction method of the present invention is a predictive model in which, by a predictive model generator, pre-written review data is collected, and then the review data is used as the basis to generate an emotion prediction model. creation step; an emotion prediction result output step of outputting, by the server, an emotion prediction result for the real-time review when a real-time review is input to the emotion prediction model; and an emotion prediction result transmission step in which the emotion prediction result is transmitted to the user's terminal by the server.

As described above, according to the present invention, a prediction model generator for generating an emotion prediction model based on the review data after collecting previously written review data is included, and when a real-time review is input to the emotion prediction model, emotion prediction for a real-time review By transmitting the emotion prediction result to the user's terminal after outputting the result, it is possible to automatically analyze the consumer's emotions included in a large number of reviews written on the web page. In addition, the online shopping mall manager can easily identify the strengths and weaknesses, responsiveness, preference, and improvement points of products and services to be sold, and based on this, it is possible to provide better products and services to future consumers.

1 is a configuration diagram of a machine learning-based online shopping review emotion prediction system of the present invention.
2 is a diagram illustrating a GRU-based artificial neural network model according to an embodiment of the present invention.
3 is a view showing a terminal screen to which an emotion prediction result is transmitted according to an embodiment of the present invention.
4 is a flowchart of a method for predicting emotion in a machine learning-based online shopping review of the present invention.
5 is a diagram showing a detailed flowchart of a predictive model generation step according to an embodiment of the present invention.
6 is a view showing review data for each length of review data according to an embodiment of the present invention.

The terms used in this specification have been selected as currently widely used general terms as possible while considering the functions in the present invention, which may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

Machine learning-based online shopping review sentiment prediction system

Hereinafter, an embodiment according to the present invention will be described in detail with reference to the accompanying drawings. 1 is a configuration diagram of a machine learning-based online shopping review emotion prediction system of the present invention. 2 is a diagram illustrating a GRU-based artificial neural network model according to an embodiment of the present invention. 3 is a view showing the screen of the terminal 200 to which the emotion prediction result is transmitted according to an embodiment of the present invention.

First, referring to FIG. 1, the machine learning-based online shopping review emotion prediction system of the present invention includes a prediction model generator 110 that collects pre-written review data and then generates an emotion prediction model based on the review data, When a real-time review is input to the emotion prediction model, the server 100 outputs the emotion prediction result for the real-time review and then transmits the emotion prediction result to the user's terminal 200.

2, most preferably, the emotion prediction model is a GRU (Gated Recurrent Unit) recurrent neural network model having a simpler structure while maintaining similar performance to the LSTM (Long Short Term Memory) recurrent neural network model, which is a model of a recurrent neural network. Can be used. This is because the LSTM is a model or structure designed to overcome the problem of a recurrent neural network model in which the analysis ability is greatly reduced as the back propagation gradient gradually decreases when the distance between the data to be analyzed and the point at which the data is used increases. There are disadvantages. That is, since the GRU model is used in the emotion prediction model of the present invention, the performance of the analysis capability is maintained and the structure is simple, so that the emotion prediction result is output more quickly.

Next, when the emotion prediction model is generated, the prediction model generator 110 may collect the review data. When the URL address including the review data is input, the predictive model generator 110 may preferentially check whether the URL address is a valid address activated on the Internet network.

And if the address is a valid address, the review data in the URL address can be crawled through parsing using the library. Here, crawling refers to a technology that collects documents distributed and stored in countless computers and includes them as an index for a search target.

If a large amount of review data exists in multiple pages, the review data can be saved as a CSV file after importing the review data on the last page.

In addition, after reading the CSV file, the predictive model generator 110 may perform normalization through a regular expression on the review data if the review data in the CSV file includes text. For example, spaces, tab keys, special characters, consonants, alphabets, consonants or vowel elements in the review data may be removed. In addition, it is possible to generate a new data frame (Data Frame) for the pre-processed review data.

Next, the predictive model generator 110 may separate the review data into positive data and negative data based on the rating included in the review data.

For example, the predictive model generating unit 110 assigns 1 as a label to review data having a score of 4 or more with a perfect score of 5, and 0 as a label for review data having a score of 3 or less. can be given

On the other hand, in order to generate the emotion prediction model, the review data must be divided into training data and test data, and the predictive model generator 110 most preferably divides the review data into 80% training data and 20% test data. can be separated, and the positive data and the negative data can be divided in a ratio of 5:5 to each of the training data and the test data.

Also, the predictive model generator 110 may tokenize the review data in units of morphemes based on a pre-stored stopword dictionary. Here, stopwords are words that are not used as search terms when searching the Internet.

That is, the predictive model generation unit 110 may perform a pre-processing of primarily removing all characters and blank data except for Hangul using the regular expression as described above for smooth analysis, Through morpheme tokenization, meaningless words such as articles, prepositions, surveys, and conjunctions in Hangul can be removed secondary. Accordingly, the review data may be refined so that only morphemes having meaning as words remain.

On the other hand, if the frequency of the morphemes generated after tokenizing the review data is less than a preset frequency, the predictive model generator 110 may be excluded from integer encoding and the corresponding morpheme may be deleted. The preset frequency may be within the range of 2 to 5 times, and most preferably 2 times.

Also, the predictive model generator 110 may integer-encode the tokenized review data. This is to change the morpheme into a number so that the predictive model generator 110 can recognize the morpheme in an analog format as a digital format.

For example, if there is review data 'I like the product too much', the predictive model generating unit 110 may return {'product', 'this', 'too', 'heart', 'e', 'listen'} to can be tokenized. And if you can encode integers as {102, 332, 131, 221, 442, 113}, and remove low frequency 'this' and 'e', the integer-encoded review data is {102, 131, 221, 113} days can And the length of the integer-encoded review data may be 4. Here, the length may be an integer-encoded size of the review data.

In addition, the predictive model generator 110 may select an optimal length based on the maximum length and the average length of the integer-encoded review data, and pad the integer-encoded review data to the optimal length. The length of the integer-encoded review data may be irregular, which may affect the speed at which an artificial neural network is trained in the future. Therefore, by unifying the length of the integer-encoded review data through the padding process, the integer-encoded review data can be viewed as one matrix and processed at once, thereby improving the learning speed of the artificial neural network.

6 is a view showing review data for each length of review data according to an embodiment of the present invention. Referring to FIG. 6 , in selecting the optimal length, the predictive model generator 110 calculates the maximum length and the average length for the whole of the integer-encoded review data, and then based on this, each review data length You can list the number of review data. And by using the ratio function, it is possible to check what % the ratio below the maximum length is. In FIG. 6 , since the ratio of the length of 55 or less is 99.9%, most of the review data may be included, and accordingly, the optimal length may be selected as 55. As shown in FIG.

Most preferably, the predictive model generator 110 divides the review data into positive data and negative data based on the rating, and each maximum length for positive data and negative data in the review data that is integer-encoded And after calculating the average length, it is possible to select the optimal length for each.

Also, when padding is completed, the predictive model generator 110 may train a GRU-based artificial neural network model using the padded review data. In this case, after training the GRU-based artificial neural network model, when the accuracy (D) exceeds the preset accuracy (D _ref ), it may be stored as the emotion prediction model.

Next, the server 100 may be connected to communicate with the terminal 200 . In addition, the server 100 may implement a review input box in a web page of a shopping mall to receive the real-time review input from the consumer's terminal on the Internet network. And since the server 100 includes the emotion prediction model generated by the predictive model generator 110, when the real-time review is input in the review input box, the real-time review can be input to the emotion prediction model, The emotion prediction result output from the emotion prediction model may be transmitted to the terminal 200 . Here, the terminal 200 may be a terminal of an administrator who operates a shopping mall, and may be a smartphone, a tablet PC, or the like capable of installing applications.

In addition, referring to FIG. 3 , the emotion prediction result most preferably includes either positive emotion or negative emotion, and may include prediction accuracy for the corresponding emotion. For example, when the server 100 inputs a real-time review written 'delivery is slow~' into the emotion prediction model, the emotion prediction result is transmitted to the terminal 200 as either a positive emotion or a negative emotion as shown in FIG. 'Negative review' and '88.57%', which is the prediction accuracy for 'negative review', can be provided at the same time.

As described above, according to the present invention, by providing the ability to automatically analyze the emotions of consumers included in a large number of reviews written on the web page, the online shopping mall manager can determine the strengths and weaknesses, responsiveness, preference, and improvement points of products and services to be sold. It has the effect of being able to provide better products and services to future consumers based on this.

Machine Learning-based Online Shopping Review Sentiment Prediction Method

Hereinafter, an embodiment according to the present invention will be described in detail with reference to the accompanying drawings. 4 is a flowchart of a method for predicting emotion in a machine learning-based online shopping review of the present invention. 5 is a diagram showing a detailed flowchart of the predictive model generation step (S100) according to an embodiment of the present invention.

First, referring to FIG. 4 , the machine learning-based online shopping review emotion prediction method of the present invention may include a prediction model generation step (S100), an emotion prediction result output step (S200), and an emotion prediction result transmission step (S300).

More specifically, in the predictive model generating step (S100), after the previously written review data is collected by the predictive model generating unit 110, an emotion prediction model is generated based on the review data.

In the predictive model generating step (S100), in collecting the previously written review data, when the URL address including the review data is input to the predictive model generating unit 110, whether the URL address is a valid address activated on the Internet network can be identified first.

And if the address is a valid address, the library is utilized and the review data in the URL address can be crawled through parsing. Here, crawling refers to a technology that collects documents distributed and stored in countless computers and includes them as an index for a search target.

If a large amount of review data exists in a plurality of pages, the review data may be saved as a CSV file after importing even the review data in the last page.

On the other hand, in the predictive model generation step (S100), after the CSV file is read, if text is included in the review data in the CSV file, normalization through a regular expression may be performed on the review data. For example, spaces, tab keys, special characters, consonants, alphabets, consonants or vowel elements in the review data may be removed. In addition, a new data frame (Data Frame) for the pre-processed review data may be generated.

In addition, the predictive model generation step (S100) includes a labeling step (S110), a tokenization step (S120), an integer encoding step (S130), and padding so that the review data can be pre-processed and learned after the previously written review data is collected. It may include a learning step (S140) and a learning step (S150).

More specifically, referring to FIG. 5 , the labeling step ( S110 ) may be labeled after the review data is divided into positive data and negative data based on the rating included in the review data.

For example, in the labeling step (S110), 1 is given as a label for review data with a score of 4 or more, which is out of 5, and 0 is given as a label for review data with a score of 3 or less. can

On the other hand, in order for the GRU-based artificial neural network model to be learned, the review data must be separated into training data and test data. The data can be separated at a ratio of 80% and test data 20%, and the positive data and the negative data can be divided in a ratio of 5:5 to each of the training data and test data. can

Next, in the tokenization step ( S120 ), the review data is tokenized based on the pre-stored stopword dictionary, and only morphemes can be extracted. Here, stopwords are words that are not used as search terms when searching the Internet.

That is, in the present invention, the regular expression is used from the predictive model generation step (S100) to perform preprocessing in which all characters except for Hangul and blank data are primarily removed, and the tokenization step (S120) ), through morpheme tokenization, meaningless words such as articles, prepositions, propositions, and conjunctions in Hangul can be removed secondarily. Accordingly, the review data may be refined so that only morphemes having meaning as words remain.

Meanwhile, in the tokenization step (S120), if the frequency among morphemes generated after the review data is tokenized is less than a preset frequency, the integer encoding step (S130) may exclude the morpheme and delete the corresponding morpheme. The preset frequency may be within the range of 2 to 5 times, and most preferably 2 times.

Next, in the integer encoding step (S130), the tokenized review data may be integer-encoded. This is to change the morpheme into a number so that the predictive model generator 110 can recognize the morpheme in an analog format as a digital format.

For example, when review data 'I like the product too much' is input by the predictive model generating unit 110, the tokenization step (S120) determines that the review data is {'product', 'this', 'too much' ', 'mind', 'e', 'listen'}, and low frequency 'this' and 'e' can be removed. And in the integer encoding step (S130), the tokenized review data may be integer-encoded as {102, 332, 131, 221, 442, 113}. In this case, since 'this' and 'e', which are low in frequency, are removed, the integer-encoded review data may be most preferably {102, 131, 221, 113}. And the length of the integer-encoded review data may be 4. Here, the length may be an integer-encoded size of the review data.

Next, in the padding step (S140), an optimal length is selected based on the maximum length and average length of the integer-encoded review data, and the integer-encoded review data may be padded to the optimal length.

The length of the integer-encoded review data may be irregular, which may affect the speed at which an artificial neural network is trained in the future. Accordingly, in the padding step (S140), the integer-encoded length of the review data is uniform.

Referring to FIG. 6 , in selecting the optimal length, in the padding step ( S140 ), the whole of the integer-encoded review data is a target, and the maximum length and the average length are calculated, and then based on this, each review data length is determined. A number of review data may be listed. And the ratio function can be used to check what % the ratio below the maximum length is. In FIG. 6 , since the ratio of the length of 55 or less is 99.9%, most of the review data may be included, and accordingly, the optimal length may be selected as 55. As shown in FIG.

Most preferably, in the padding step ( S140 ), the maximum length for positive data and negative data in the review data that is integer-encoded based on the rating as a basis and the review data is divided into positive data and negative data After calculating the and average length, each optimal length may be selected.

Next, in the learning step ( S150 ), when padding is completed, the padded review data may be used to learn a GRU-based artificial neural network model. In this case, when the accuracy (D) exceeds the preset accuracy (D _ref ) after the GRU-based artificial neural network model is trained, it may be stored as the emotion prediction model.

Next, in the emotion prediction result output step (S200), when a real-time review is input to the emotion prediction model by the server 100, an emotion prediction result for the real-time review is output.

After the emotion prediction model is generated as described above, the emotion prediction result output step (S200) is the real-time review inputted in real time in the review input field in the web page of the online shopping mall implemented from the server 100 is applied to the test prediction model. can be entered. In this case, the emotion prediction result output step ( S200 ) may include an emotion prediction result including either positive emotion or negative emotion, and prediction accuracy for the corresponding emotion.

Next, in the emotion prediction result transmission step S300 , the emotion prediction result is transmitted to the user's terminal 200 by the server 100 . Here, the terminal 200 is a terminal of an administrator who operates a shopping mall, and may be a smart phone, a tablet PC, or the like capable of installing an application.

And the emotion prediction result transmission step (S300) is the prediction accuracy for 'negative review' and 'negative review', which are either positive or negative emotions, as shown in FIG. 3, in which the emotion prediction result is transmitted to the terminal 200, '88.57 %' may be provided simultaneously.

As described above, although the embodiments have been described with reference to the limited embodiments and drawings, various modifications and variations are possible by those skilled in the art from the above description. For example, the described techniques are performed in an order different from the described method, and/or the components of a system, structure, apparatus, circuit, etc. described as are combined or combined in a different manner than the described method, or other configuration. Substituted or substituted for elements or equivalents may achieve appropriate results.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

100.. Server
110. Prediction model generation unit
200.. Terminal

Claims (5)

and a predictive model generator for generating an emotion prediction model based on the review data after collecting pre-written review data, and when a real-time review is input to the emotion prediction model, after outputting an emotion prediction result for the real-time review A machine learning-based online shopping review emotion prediction system comprising a; a server that transmits the emotion prediction result to a terminal. a predictive model generation step of generating, by the predictive model generating unit, an emotion prediction model based on the review data after the previously written review data is collected;
an emotion prediction result output step of outputting, by the server, an emotion prediction result for the real-time review when a real-time review is input to the emotion prediction model; and
A machine learning-based online shopping review emotion prediction method comprising a; by the server, an emotion prediction result transmitting step of transmitting the emotion prediction result to the user's terminal. 3. The method of claim 2,
The predictive model generation step is,
A labeling step of labeling after the review data is divided into positive data and negative data based on the rating included in the review data;
a tokenization step of tokenizing the review data in units of morphemes based on a pre-stored stopword dictionary, and removing the tokenized review data if the frequency is less than a preset frequency;
an integer encoding step in which the tokenized review data is integer-encoded;
a padding step in which an optimal length is selected based on a maximum length and an average length of the integer-encoded review data, and the integer-encoded review data is padded to the optimal length; and
A learning step of learning a GRU-based artificial neural network model by using the padded review data. A machine learning-based online shopping review emotion prediction method comprising a. 4. The method of claim 3,
The learning step is
After the GRU-based artificial neural network model is trained, when the accuracy (D) exceeds the preset accuracy (D _ref ), the machine learning-based online shopping review emotion prediction method, characterized in that it is stored as the emotion prediction model. 3. The method of claim 2,
The emotion prediction result is
Machine learning-based online shopping review emotion prediction method, characterized in that one of positive emotions or negative emotions is included, and the prediction accuracy for the corresponding emotion is included.

Images