KR102053858B1 - Method for calculating rating of content - Google Patents

Abstract

Content rating method according to an embodiment of the present invention comprises the steps of collecting documents containing a rating for the content; Classifying the collected documents into a plurality of groups according to a predetermined classification condition; Calculating vector values for correlations between documents included in some of the classified groups and predetermined keywords using a text mining algorithm; Generating a rating calculation model based on some of the calculated vector values using a pattern recognition algorithm; And calculating a rating for the content output as a result of matching the documents included in the remaining groups among the classified groups to the rating calculation model.

Description

How to calculate your content rating {METHOD FOR CALCULATING RATING OF CONTENT}

The present invention relates to a method for calculating a content rating, wherein the reviews are morphologically analyzed using TF-IDF (Term Frequency-Inverse Document Frequency), and a method of calculating a content rating using a support vector machine (Support Vector Machine). It is about.

Grading is a rating that is valued. It has become popular to rate media contents such as movies, music and dramas due to the development of social network services.

In addition, the ratings given by netizens, critics, etc. are affecting the popularity of media content. For example, it is undeniable that the online ratings of audiences who have previewed a movie ahead of time affect the performance of the movie after the release.

In the related art, a dynamic noise removing method and a content recommendation system for content recommendation for removing content noise of user rating data on media content and improving reliability of data have been disclosed.

However, the prior art only discloses a configuration for filtering noise by analyzing the distribution of ratings by reflecting the number and time points of ratings by users as weights, and does not disclose any configuration for estimating ratings by analyzing reviews. Not.

A review is a description of the user's feelings about the media content, and contains the user's intention to evaluate the rating alone.

Therefore, it is necessary to develop a technique for estimating a more accurate rating of media content by analyzing reviews, rather than simply analyzing the rating of media content by synthesizing the ratings.

KR 10-2017-0079423 A

The present invention relates to a method of morphologically analyzing documents for content using TF-IDF based on documents collected for content of interest, and calculating an average rating of the content using a support vector machine.

Content rating method according to an embodiment of the present invention comprises the steps of collecting documents containing a rating for the content; Classifying the collected documents into a plurality of groups according to a predetermined classification condition; Calculating vector values for correlations between documents included in some of the classified groups and predetermined keywords using a text mining algorithm; Generating a rating calculation model based on some of the calculated vector values using a pattern recognition algorithm; And calculating a rating for the content that is output as a result of matching the documents included in the remaining groups among the classified groups to the rating calculation model.

After the calculating of the rating, the method may further include repeating the step of calculating the rating to the group to the groups of the collected documents.

The repeating may be repeated until the number of repetitions set in proportion to the number of collected documents is satisfied.

After the repeating step, the method may further include calculating an average of the ratings calculated in the calculating of the rating.

After collecting the documents, the method may further include classifying the collected documents by the category based on a classification keyword preset for each category.

The classifying or calculating the rating may be performed for each classified category, and a rating for each category may be calculated.

After generating the calculation model, the accuracy of the classification performance of the generated rating calculation model may be calculated by inputting the rest of the calculated vector values into the generated rating calculation model.

The calculating of the vector value may include correlation between documents included in first and second groups of the classified groups and preset keywords using a term frequency-inverse document frequency (TF-IDF) algorithm. Calculating vector values for.

The generating of the rating calculation model may include generating a rating calculation model from the vector values calculated for the first group by using a support vector machine.

The classifying may include adjusting, in documents included in a first group of the classified groups, the number of documents having a rating greater than or equal to a predetermined score based on the number of documents having a rating less than a predetermined score. Can be.

The predetermined classification condition may include a condition of classifying the collected documents based on a rating of the content, and classifying the number of classified documents into a plurality of groups according to a predetermined ratio. .

According to the content rating calculation method according to an embodiment of the present invention, since a large rating is calculated for the content from documents collected using the support vector machine together with the F-IDF algorithm, the user simply ranks the rating with the review. There is an advantage that can provide a more specific and reliable rating than the synthesis method.

In addition, in addition to the overall rating of the media content, there is an advantage that can be provided by calculating the rating for each category, such as image beauty, story or OST.

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, which are included as part of the detailed description in order to provide a thorough understanding of the present invention, provide examples of the present invention and together with the description, describe the technical features of the present invention.
1 is a view briefly illustrating a flowchart of a method for calculating a content rating according to an embodiment of the present invention.
2 is a flowchart illustrating a content rating method according to another embodiment of the present invention briefly.

In this specification, terms such as first and / or second are used only for the purpose of distinguishing one component from another component. In other words, it is not intended to limit the components by the above terms.

Components, features, and steps that are referred to herein as "comprising" mean that such components, features, and steps exist and are intended to exclude one or more other components, features, steps, and equivalents. This is not it.

Unless otherwise specified and stated in the singular, the plural forms are included. That is, the components and the like mentioned herein may mean the presence or addition of one or more other components.

Unless defined otherwise, 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. to be.

In other words, terms such as those defined in the commonly used dictionaries should be construed as meanings consistent with the meanings in the context of the related art, and, unless expressly defined herein, are construed in ideal or overly formal meanings. It doesn't work.

Hereinafter, with reference to the accompanying drawings will be described in detail a content rating calculation method according to an embodiment of the present invention.

1 is a view briefly illustrating a flowchart of a method for calculating a content rating according to an embodiment of the present invention.

Referring to FIG. 1, in the method of calculating a content rating according to an embodiment of the present invention, the method may include collecting documents (S101), classifying them into groups (S103), calculating vector values (S105), and calculating a rating. The method may include generating a model (S107), calculating an accuracy and a rating (S109), repeating the operation (S111), and calculating an average of the calculated accuracy and ratings (S113).

Collecting the document (S101) is a step of collecting documents containing a rating for the content through a portal site or a social network service, it may be performed using a crawling (Crawling) technique.

For example, documents including a rating for an A movie in a portal site may be collected using a crawling technique. Here, the collected documents include a rating and a user review for the A movie, and all the documents posted on the plurality of portal sites may be collected.

In the step of collecting documents S101, when documents including ratings for a specific movie are collected from a plurality of portal sites, the range of ratings may be different for each portal site, and thus, the range of ratings of the plurality of portal sites is determined. Scaling the range of ratings to match.

For example, if the rating of a site is 1 to 10 points, while the rating of b site is 1 to 100 points, the rating of b site is scaled to the rating of a site, and the rating of a site is a site. And the range of the rating of the b site can be the same from 1 to 10 points.

However, unlike the above example, if the rating of the site a ranges from 1 to 10 points, and the rating of the site b ranges from 0 to 10 points, users can rate the same criteria, Scaling the range can optionally be performed.

In step S103, the collected documents are classified into a plurality of groups according to a predetermined classification condition.

The preset classification condition is a condition for classifying the collected documents based on the rating of the content, and then classifying the number of classified documents according to a predetermined ratio and classifying the documents into a plurality of groups.

Documents can be classified into positive, negative, or neutral documents based on their grade.

For example, documents having a score less than 3 points may be classified as negative documents, documents having 4 points or more and 6 points or less, neutral documents, documents having 7 points or more and 10 points or less may be classified as positive documents.

Thereafter, the positive documents and the negative documents can be randomly divided according to a predetermined ratio, classified into the first group and the second group, and the neutral documents can be divided into the third group according to the predetermined ratio.

For example, when a predetermined ratio is set at a ratio of 7 to 3, 70% of each of the positive documents and the negative documents may be classified into a first group, and the remaining 30% may be classified into a second group. In addition, 30% of the neutral documents may be classified into a third group.

Documents included in 70% or 30% of positive, negative and neutral documents may be randomly selected and classified into the first to third groups.

Positive and negative documents are documents for generating a rating model, while neutral documents are documents for calculating a rating for content through the generated rating model. Therefore, when all neutral documents among the collected documents are used, only 30% of the neutral documents can be classified into the third group because the calculated scores are likely to be distributed between 4 and 6 points.

The step S103 of classifying into groups may include classifying the first to third groups and then adjusting the number of positive documents included in the first group.

For example, if the ratio of positive documents in the collected documents is higher than the ratio of negative documents, there is a high probability that the average of the calculated ratings can be biased towards the higher and calculated.

Therefore, the number of positive documents included in the first group may be adjusted based on the number of negative documents included in the first group.

For example, if 100 positive documents and 30 negative documents are included in the first group, and the number of positive documents is set to include 3 times or less of the number of negative documents in the first group, 10 positive documents are random. Can be selected and deleted from the first group.

The calculating of the vector values (S105) is a step of calculating vector values for correlations between documents included in some groups among the groups classified using text mining and predetermined keywords.

As a text mining algorithm, a TF-IDF (Term Frequency-Inverse Document Frequency) algorithm may be used, but the type of algorithm is not limited thereto, and an algorithm capable of calculating vector values of correlations between documents and keywords. Can be used without limitation.

The preset keyword may be set to a word that is frequently used in a review of the content. For example, keywords such as “real”, “fun” and “random” may be set.

Since TF-IDF is a product of TF (Term Frequency) and IDF (Inverse Document Frequency), the TF-IDF between the documents included in each of the first group and the second group and preset keywords Can be calculated.

Since the TF-IDF for each of the documents and the keywords is calculated, a vector value of the TF-IDF values of the keywords preset for the documents can be calculated.

For example, in the first document, if the TF-IDF value of the keyword "real" is 0.3, the TF-IDF value of the keyword "funny" is 0, and the TF-IDF value of the keyword "random" is 0.7, A vector value for correlation between the first document and the predetermined keywords may be calculated as (0.3, 0, 0.7).

Using the TF-IDF algorithm, vector values for correlations between predetermined keywords for all documents included in the first group and the second group may be calculated.

Generating a rating calculation model (S107) is a step of generating a rating calculation model based on some of the calculated vector values using a pattern recognition algorithm.

As a pattern recognition algorithm, a support vector machine (SVM) algorithm may be used, but the type of pattern recognition algorithm is not limited thereto.

As a result of calculating the vector values (S105), vector values for documents included in the first group and the second group have been calculated, but vector values for generating a rating calculation model are included in the document included in the first group. Only vector values are used, and the vector values for the documents included in the second group are used to measure the accuracy of the generated rating calculation model.

The support vector machine is a supervised learning model for pattern recognition and data analysis. Given a set of data belonging to one of two categories, the probability of determining which category the new data belongs to is based on the given data set. A binary binary classification model can be created, and a kernel trick can be used to create a nonlinear classification model.

The generating of the rating calculation model (S107) may generate a hyperplane, which is a rating calculation model that distinguishes vector values for documents included in the first group by using the support vector machine.

That is, a rating calculation model may be generated for distinguishing a vector value between positive documents and negative documents included in the first group and predetermined keywords.

The calculating of the accuracy and the rating (S109) is a step of calculating the rating for the content using the accuracy of the classification performance of the generated rating calculation model and the generated rating calculation model.

First, the accuracy of the classification performance of the rating calculation model may be calculated by inputting the remaining vector values, ie, vector values of documents included in the second group, to the rating calculation model.

Since the second group includes the positive documents and the negative documents like the first group, when the vector values for the documents included in the second group are input to the scoring model, the second group includes the documents included in the second group. According to the classification result, the accuracy of the classification performance of the rating calculation model may be calculated.

For example, if the rating calculation model receives vector values for documents included in the second group and classifies the documents included in the second group into positive documents and negative documents without errors, the classification performance of the rating calculation model is determined. Accuracy can be calculated as 100%.

Next, a rating of the content output as a result of matching the documents included in the remaining groups among the classified groups to the rating calculation model may be calculated.

That is, as a result of correlating the documents included in the third group that does not calculate the vector value of the correlation between the predetermined keywords to the rating calculation model, the rating for the content may be calculated.

The result corresponding to the rating calculation model is a probability that each of the documents included in the third group is a positive document, and the average of probabilities output for the documents included in the third group is averaged to calculate a rating for the content. have.

For example, if the average of the probabilities output as a result of mapping the documents included in the third group to the rating calculation model is 80%, the rating for the corresponding content may be calculated as 80 points.

In step S111, after calculating the accuracy and the rating, the step of classifying the collected documents into groups S103 to calculating the accuracy and the rating S109 is performed repeatedly. .

According to a predetermined classification condition, the collected documents are classified into positive, negative and neutral documents according to the ratings, and randomly classified into the first, second and third groups according to a predetermined ratio, and thus, the first to third documents. Documents included in the group can be changed.

That is, in step S111, the documents included in the first to third groups may be changed to generate a new rating calculation model, and the accuracy of the classification performance of the generated rating calculation model may be calculated. Ratings for the content may be calculated by matching the documents included in the third group with a new rating calculation model.

The repeating step S111 may be performed until the number of repetitions set in proportion to the number of collected documents is satisfied.

Since the number of documents collected in the first to third groups is reduced each time the number of collected documents is repeated, the step of repeating until the number of repetitions set in proportion to the number of collected texts is satisfied (S111). ) May be performed.

In addition, the repeating step (S111) may be performed until a predetermined number of repetitions is satisfied regardless of the number of collected documents.

For example, when the content rating calculation method according to an embodiment of the present invention is set to perform 100 times, the step of performing the repetition (S111) is repeated 99 times regardless of the amount of the collected document, thereby calculating 100 calculated ratings. The accuracy of the classification performance of the model and the rating for the content may be respectively calculated.

Computing the average of the calculated accuracy and the ratings (S113) is a step of calculating the average of the accuracy of the classification performance of the rating calculation model and the ratings for the content if the predetermined number of repetitions is satisfied so far .

For example, if the accuracy for the classification performance of the 100 calculated rating models and the rating for the content have been calculated, respectively, an average of the 100 accuracy and the rating is calculated to calculate the accuracy for the classification performance of the rating model and the content. Can be output as a rating.

2 is a flowchart illustrating a content rating method according to another embodiment of the present invention briefly.

Referring to FIG. 2, the method for calculating a content rating according to another embodiment of the present invention may further include classifying the collected documents by category after performing the step of collecting documents (S101). .

In step S102, the collected documents are classified by category, the collected documents are classified based on a classification keyword preset for each category.

For example, a category for acting power, story, director, original sound track (OST) and visual beauty may be set for the movie genre. Classification keywords such as story, story, content, probability, and material may be set for the story category, and classification keywords such as OST, song, music, and sound may be set for the OST category.

Since some documents may include a plurality of classification keywords preset for each category, documents including classification keywords set in a plurality of categories may be included in all of the corresponding categories.

After classifying the collected documents by category, the step of classifying the groups independently by category (S103) or calculating the average of the calculated accuracy and ratings (S113) may be performed.

For example, since a rating calculation model for the story category is generated in the story category, a rating for the story category is calculated. In the OST category, a rating for the OST category may be calculated because the rating calculation model for the OST category is generated.

That is, the content rating calculation method may calculate a rating for the content of all the collected documents, and may calculate a rating for each category based on a preset classification keyword.

Therefore, when a user needs to purchase, purchase, or view one of the contents of interest, the user can compare the contents of interest in detail by calculating and confirming a rating and a category-specific rating of the contents of interest. have.

Although the description herein has been shown in several illustrative aspects, various modifications and changes can be made from the scope defined by the following claims, and the technical protection scope of the invention is set forth in the following claims. It must be decided by.

Claims (11)

Collecting documents including a rating of the content;
Classifying the collected documents into a plurality of groups according to a predetermined classification condition;
Calculating vector values of correlations between documents included in some of the classified groups and predetermined keywords using a text mining algorithm;
Generating a rating calculation model based on some of the calculated vector values using a pattern recognition algorithm; And
Calculating a rating for the content output as a result of matching the documents included in the remaining groups among the classified groups to the rating calculation model,
The classifying step,
In the documents included in the first group of the classified groups,
Adjusting the number of documents whose rating is above a certain score based on the number of documents having a rating below a certain score,
How content is calculated. The method of claim 1,
After calculating the rating,
Further comprising: classifying the collected documents into groups and calculating the ratings.
How content is calculated. The method of claim 2,
Repeating the step,
Repeatedly performed until the number of repetitions set in proportion to the number of collected documents is satisfied;
How content is calculated. The method of claim 2,
After the repeating step,
Calculating an average of ratings calculated in the calculating of the rating;
How content is calculated. The method of claim 1,
After collecting the documents,
The method may further include classifying the collected documents by the category based on a classification keyword preset for each category.
How content is calculated. The method of claim 5,
The classifying or calculating the rating may include:
Each of the classified categories is performed to calculate a rating for each of the categories,
How content is calculated. The method of claim 1,
After generating the calculation model,
Computing the accuracy of the classification performance of the generated rating calculation model by inputting the rest of the calculated vector value to the generated rating calculation model,
How content is calculated. The method of claim 1,
Computing the vector value,
Calculating vector values of correlations between documents included in the first and second groups of the classified groups and predetermined keywords by using a term frequency-inverse document frequency (TF-IDF) algorithm. Included,
How content is calculated. The method of claim 8,
Generating the rating calculation model,
Generating a rating calculation model from the vector values calculated for the first group using a support vector machine;
How content is calculated. delete The method of claim 1,
The preset classification condition is,
Classify the collected documents based on a rating for the content,
Is a condition of dividing the number of the classified documents according to a predetermined ratio and classifying the plurality of groups.
How content is calculated.

Images