JP2020047229A - Article analyzer and article analysis method - Google Patents

Abstract

To provide a beneficial proposal for an advertisement plan newly inputted on the basis of knowledge obtained from past advertisement articles.SOLUTION: An article analyzer 1 has: a mechanical learning part 21 which stores a result of totalizing frequency of appearance of respective words acquired as character data 101 of past article data 11 to a feature DB 13 as character feature data; and an improvement calculation part 23 which calculates scores of respective words by calculating evaluation functions showing that the more similar the respective words acquired as character data 201 of inputted prediction article data 12 are to words in the character feature data and the higher frequency of appearance of the character feature data is, the higher the score becomes, and presents that rewrite of words in the prediction article data 12 is promoted with words in the character feature data of high score as rewrite candidates with respect to the respective words in the prediction article data 12.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an article analysis device and an article analysis method.

  2. Description of the Related Art In recent years, attention has been paid to the development of artificial intelligence (AI) by utilizing deep learning. With the advancement of network technology, a large amount of data can be efficiently collected. Therefore, it is expected to construct a computation model that performs work equivalent to humans by machine learning some knowledge from the data.

  A main application field of artificial intelligence is image recognition processing in which image data is used as input data and the image data is classified into any category. For example, Non-Patent Document 1 describes an attempt to model a visual model showing how a human visually perceives an artistic image as a deep neural network. A deep neural network is a convolutional neural network composed of layers of small calculation units that hierarchically process visual information.

Leon A. Gatys et al., "A Neural Algorithm of Artistic Style", [online], August 26, 2015, [Search August 3, 2018], Internet <URL: https://arxiv.org/abs/ 1508.06576>

  There is also a need to mechanically create an advertisement recognition model indicating how a human perceives an advertisement article such as a CM, a signboard, and a poster. Then, by using an advertisement recognition model, a user on the advertisement providing side, such as an advertising agency, expects the effects of the advertisement before the advertisement created by the advertisement is actually published to the public in CM or the like and the reputation is obtained. There is expected.

  In the research of the conventional artificial intelligence such as Non-Patent Literature 1, general machine recognition such as grouping similar advertisements is limited. Therefore, there has not been a system for an advertisement provider to provide a useful proposal that goes into the content of how to evaluate a prototype advertisement proposal and how to improve it.

  Therefore, the main object of the present invention is to make a useful proposal for the currently input advertisement plan based on knowledge obtained from past advertisement articles.

In order to solve the above-described problems, an article analysis device according to the present invention has the following features.
The present invention provides a machine learning unit that stores, in a storage unit, a result of counting the frequency of appearance of each word acquired as character data of past article data as character feature data,
For each word acquired as the character data of the input predicted article data, the higher the degree of similarity with the word of the character feature data, and the higher the score of the appearance function of the character feature data, the higher the evaluation function that is scored. By calculating, the score of each word is calculated,
An improvement calculation unit that presents a message that prompts rewriting of the word of the predicted article data, with the word of the character feature data having the higher score as a rewriting candidate for each word of the predicted article data. .
Other means will be described later.

  ADVANTAGE OF THE INVENTION According to this invention, a useful proposal with respect to the advertisement plan input this time can be made based on the knowledge obtained from the past advertisement articles.

It is a lineblock diagram of an article analysis device concerning one embodiment of the present invention. It is a flowchart which shows the process of the article analysis apparatus concerning one Embodiment of this invention. FIG. 3 is an explanatory diagram showing a neural network learning process for character data and image data of past article data according to an embodiment of the present invention. FIG. 5 is an explanatory diagram showing an inference process of a neural network for character data and image data of predicted article data according to an embodiment of the present invention. FIG. 8 is an explanatory diagram showing a feature DB learning process for character data of past article data according to an embodiment of the present invention. It is explanatory drawing which shows the application process of the feature DB for character data of the prediction article data concerning one Embodiment of this invention. 7 is a specific example relating to the character data of FIGS. 5 and 6 according to an embodiment of the present invention.

  Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a configuration diagram of the article analysis device 1.
The article analysis device 1 stores past article data 11, predicted article data 12, and a feature DB 13 in a storage unit. The article analysis device 1 includes a machine learning unit 21, an effect prediction unit 22, an improvement calculation unit 23, a morphological analysis unit 31, a score calculation unit 32, and a score total unit 33 as processing units.
The article analysis device 1 is configured as a computer having a CPU (Central Processing Unit), a memory, storage means (storage unit) such as a hard disk, and a network interface.
In this computer, a CPU executes a program (also called an application or an abbreviation for an application) read into a memory, thereby operating a control unit (control unit) including each processing unit.

The past article data 11 is data in which articles and their evaluation points are associated with each other, and is input in advance as labeled teacher data in machine learning. The article of the past article data 11 is, for example, an article indicating an advertisement such as a CM, a signboard, a poster, and the like, and is published on a net such as a social networking service (SNS).
The evaluation score of the past article data 11 is a label that quantitatively indicates the evaluation (effect) of the published article by the reader from the reader. The evaluation points include, for example, the number of accesses to the article, the number of likes, which is the number of responses from readers to the article, the number of retweets, the number of replies, and the like, as access data from the Internet. “Like” is an operation that evaluates an article with one click (one touch) for an article judged by the reader to have a good impression. Depending on the application, “like” or “favorite” may be used. be called.
Alternatively, the evaluation point may be sales data as a response from the market to the product introduced in the article.
Hereinafter, with reference to FIG. 2, components other than the past article data 11 in the article analysis apparatus 1 will be described.

FIG. 2 is a flowchart showing the processing of the article analysis device 1.
The machine learning unit 21 performs machine learning using the past article data 11 as teacher data, and creates a regression prediction model as a result. The prediction model is stored in the feature DB 13 as, for example, a word (keyword) indicating a feature of the past article data 11 (S11).
The prediction model is configured as, for example, a neural network that connects an input layer, an intermediate layer, and an output layer in order and transmits information to each layer. When the intermediate layer has multiple layers (multiple layers), the machine learning unit 21 creates a prediction model by deep learning.

  The article analysis device 1 receives an input of the predicted article data 12 (S12). The input predicted article data 12 is an article whose evaluation point is undecided at an undisclosed prototype stage or the like, and is an article to be predicted using a prediction model. It is desirable that the prediction article data 12 be prepared in advance with a plurality of data such as an advertisement A plan, an advertisement B plan, and an advertisement C plan that are to be used in the advertisement.

  The effect prediction unit 22 predicts an evaluation point of each predicted article data 12 input in S12 using the prediction model read from the feature DB 13 (S13). It is desirable that not only articles with high evaluation scores but also articles with low evaluation scores be used as the past article data 11. Thus, the effect prediction unit 22 can appropriately predict, for example, the predicted article data 12 including a word that is offensive to the reader as “low evaluation score”.

  The improvement calculation unit 23 refers to the words described in the past article data 11 from the feature DB 13 and creates an improvement plan for improving the article content of the predicted article data 12 (S14). Therefore, in order to create an improvement plan that brings the predicted article data 12 input this time closer to the highly evaluated past article data 11, as the past article data 11, an evaluation score such as an article that acquired a large number of likes in the past is used. It is desirable to collect expensive articles in advance.

The article analysis device 1 displays the advertisement effect (evaluation score) of the predicted article data 12 predicted in S13 and the improvement plan for the predicted article data 12 created in S14 (S15). With this display, the user can make a decision such as adopting the advertisement B having the highest evaluation score, or can rewrite words written in the advertisement B so as to resemble popular articles in the past. Or adopt a plan. That is, the user can know the predicted article data 12 that contributes to sales based on the evaluation points and the improvement plans.
Note that the article analysis device 1 displays the evaluation points of the effect prediction unit 22 and the improvement proposal of the improvement calculation unit 23 at the same time, but only one of them is useful for the user. Therefore, the article analysis device 1 may include only one of the effect prediction unit 22 and the improvement calculation unit 23.

FIG. 3 is an explanatory diagram showing a neural network learning process for character data and image data of the past article data 11. The learning process in FIG. 3 corresponds to the process in S11 in FIG.
The machine learning unit 21 extracts character data 101 and image data 102 by analyzing the syntax of the past article data 11. As the character data 101, text data described in the past article data 11 may be extracted, or text data may be mechanically extracted from voice data or moving image data attached to the past article data 11 by voice recognition. The image data 102 may be configured from one image file at a time, or may be configured as a set of image files extracted from the moving image data of the past article data 11.

  The machine learning unit 21 creates CNN (Convolutional Neural Network) data 112 that is a convolutional neural network (convolution layer) using the image data 102 as an input. Then, as described in Non-Patent Document 1, the machine learning unit 21 extracts style information from the CNN data 112 and registers the style information in the feature DB 13. The style information is a learning result of the image data 102 and is feature information of an image for generating an advice image (improved image data 232 in FIG. 4).

The machine learning unit 21 registers the word frequency data 111 extracted from the character data 101 in the feature DB 13 (see FIG. 5 for details). The machine learning unit 21 creates a neural network by combining the word frequency data 111 and the CNN data 112 with the fully connected layer data 121. The machine learning unit 21 associates the output destination of the fully connected layer data 121 with the article effect data 131.
That is, the first layer (input layer) of the neural network is character data 101 and image data 102, and the second layer (intermediate layer) is character characteristic data (word frequency data 111) and image characteristic data (CNN data 112). The third layer (output layer) is the all-connected layer data 121.
The article effect data 131 is a label (evaluation score of an article such as the number of likes) of teacher data associated with the past article data 11. The machine learning unit 21 learns the neural network by inputting (propagating) the past article data 11 one after another to the neural network generated in this way.

FIG. 4 is an explanatory diagram showing the inference process of the neural network for the character data and the image data of the predicted article data 12. The inference process in FIG. 4 corresponds to the processes in S13 and S14 in FIG.
The effect prediction unit 22 extracts the character data 201 and the image data 202 by analyzing the syntax of the input predicted article data 12.
The effect prediction unit 22 predicts a label (evaluation score of an article such as the number of likes) for the predicted article data 12 by inputting the extracted character data 201 and image data 202 to the neural network created in FIG. Specifically, the effect predicting unit 22 inputs the character data 201 and the image data 202 to an intermediate layer suitable for each data format.

  As a result, information is propagated from the layer of the word frequency data 111 read from the feature DB 13 and the layer of the CNN data 112 read as style information from the feature DB 13 to the all-connected layer data 121, respectively. The result is propagated to the effect prediction data 231 having the same format as the article effect data 131. Therefore, the user can know the prediction of how many like articles the prediction article data 12 can obtain.

Further, as described later with reference to FIG. 6, the improvement calculation unit 23 generates improved character data 233 in which another word stored in the feature DB 13 is applied to the character data 201 as a learning result, and presents it to the user.
Further, as described in Non-Patent Document 1, the improvement calculation unit 23 generates improved image data 232 in which style information stored in the feature DB 13 is applied to the image data 202 as a learning result, and an advice image is generated. To the user.
For example, it is assumed that bright color style information is extracted from an article having a large number of likes from the past article data 11. Then, when the style information of the dark color scheme is extracted from the image data 202 of the predicted article data 12, the improvement calculation unit 23 refers to the style information of the bright color scheme to improve the image data 232 by brightening the image data 202. Generate

  The analysis processing of the predicted article data 12 using the neural network for the character data and the image data has been described above with reference to FIGS. In Non-Patent Document 1, only a neural network for image data is constructed. In the present embodiment, by incorporating a new character data layer into this neural network of image data, it is possible to accurately perform prediction processing of the predicted article data 12 including both character data and image data.

FIG. 5 is an explanatory diagram illustrating a learning process of the feature DB 13 for character data of the past article data 11. The learning process in FIG. 5 corresponds to the process in S11 in FIG.
The machine learning unit 21 instructs the morphological analysis unit 31 to extract the word frequency data 111 based on the character data 101 extracted from the past article data 11. The morphological analysis unit 31 generates word frequency data 111 (detailed in FIG. 7) by counting the number of appearances (use frequency) of each word divided by the morphological analysis from the character data 101, and generates the word frequency. The data 111 is registered in the feature DB 13 as feature data learned from a set of past article data 11.

FIG. 6 is an explanatory diagram showing a process of applying the feature DB 13 to the character data of the predicted article data 12. The application process in FIG. 6 corresponds to the processes in S13 and S14 in FIG.
The morphological analysis unit 31 divides the character data 201 of the predicted article data 12 into a set of words by morphological analysis, as in the learning process of FIG.
The score calculation unit 32 scores each word based on the word frequency data 111 obtained in the learning step of FIG. 5 and the set of words of the predicted article data 12 by the morphological analysis unit 31. This score indicates the degree to which the higher the score of each word appearing in the predicted article data 12, the better it is to rewrite it to another word. The improved character data 233 is a list of another word to be rewritten.

  Further, the score totaling unit 33 totals the results of the scoring for each of the predicted article data 12 (for each page of the article), and uses the total value as the input to the all connected layer data 121. As described with reference to FIG. 4, the effect prediction unit 22 predicts an evaluation score of an article such as the number of likes for each predicted article data 12 by associating the data with the effect prediction data 231 from the total connected layer data 121.

FIG. 7 is a specific example of the character data of FIGS.
In the learning step, the morphological analysis unit 31 extracts word frequency data 111 from the past article data 11. The word frequency data 111 is the frequency of appearance of each word such as “200 occurrences after unprecedented abortion”, “150 romantic occurrences”, and so on.
In the application step, first, the morphological analysis unit 31 extracts a set of words from the predicted article data 12. Here, it is assumed that words such as unheard of, romantic, refreshing,... Are extracted.

Next, the score totaling unit 33 obtains a total score obtained by adding the scores for each word calculated by the score calculating unit 32 using the following formula for each predicted article data 12.
(Score of word) = Similarity between (word of predicted article data 12) and (word of word frequency data 111) × (frequency of use of word frequency data 111)
For example, if the similarity between “unheard of before” and “after unprecedented” is 2.8, the score of “unheard of before” = 2.8 × 200 = 560. The total score for each word is input to the all connected layer data 121.

Then, the improvement calculation unit 23 generates a list of similar words (first, second, third,...) Of the word frequency data 111 for each word of the predicted article data 12 in the descending order of the score, in the improved character data 233. Extract as
For example, the word "unprecedented" was described in the prototype of the advertisement B produced this time. However, in a popular article (past article data 11), not the word "unheard of" but a lot of words with similar meanings "unprecedented". Therefore, the improvement calculation unit 23 presents the user with the improved character data 233 indicating that “unprecedented unheard of” is rewritten to “high-scoring” “after unprecedented” or “innovative”.
Thereby, by rewriting “unheard of before” to “after all” before releasing the advertisement B plan, the impression of the advertisement B plan can be improved without largely changing the purpose of the advertisement B plan.

In the present embodiment described above, the effect of the predicted article data 12 can be predicted by the effect prediction unit 22 machine learning the neural network from the past article data 11. Here, by using information on the word-of-mouth effect of the SNS such as the number of likes as the learning data of the past article data 11, data indicating the reputation of the market can be easily obtained.
Further, the improvement calculation unit 23 proposes to rewrite the words of the predicted article data 12 based on the word frequency data 111 extracted from the past article data 11, so that the user can make an improvement in the content of the predicted article data 12. Can be encouraged.

Note that the present invention is not limited to the above-described embodiment, and includes various modifications. For example, the above-described embodiment has been described in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to one having all the described configurations.
Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of one embodiment can be added to the configuration of another embodiment.
Also, for a part of the configuration of each embodiment, it is possible to add, delete, or replace another configuration. In addition, each of the above-described configurations, functions, processing units, processing means, and the like may be partially or entirely realized by hardware by, for example, designing an integrated circuit.
In addition, the above-described configurations, functions, and the like may be implemented by software by a processor interpreting and executing a program that implements each function.

Information such as programs, tables, and files that realize each function is stored in a memory, hard disk, recording device such as SSD (Solid State Drive), IC (Integrated Circuit) card, SD card, DVD (Digital Versatile Disc), etc. Recording media.
In addition, control lines and information lines are shown as necessary for the description, and do not necessarily indicate all control lines and information lines on a product. In fact, almost all components may be considered to be interconnected.

DESCRIPTION OF SYMBOLS 1 Article analysis apparatus 11 Past article data 12 Predicted article data 13 Feature DB
Reference Signs List 21 machine learning unit 22 effect prediction unit 23 improvement calculation unit 31 morphological analysis unit 32 score calculation unit 33 score total unit 101 character data 102 image data 111 word frequency data 112 CNN data 121 fully connected layer data 131 article effect data 201 character data 202 Image data 231 Effect prediction data 232 Improved image data 233 Improved character data

Claims (4)

A machine learning unit that stores, in the storage unit, the result of counting the appearance frequencies of the words acquired as the character data of the past article data as character feature data,
For each word acquired as the character data of the input predicted article data, the higher the degree of similarity with the word of the character feature data, and the higher the score of the appearance function of the character feature data, the higher the evaluation function that is scored. By calculating, the score of each word is calculated,
An improvement calculation unit that presents a message that prompts rewriting of the word of the predicted article data, with the word of the character feature data having the higher score as a rewriting candidate for each word of the predicted article data. Article analysis device. The machine learning unit stores, in addition to the character feature data of the past article data, image feature data acquired from the image data of the past article data in a storage unit, and stores the character data and the image of the past article data. A first layer for receiving data input, a second layer corresponding to the character feature data and the image feature data, and combining the character feature data and the image feature data of the second layer with the past article data; Form a neural network in which the third layer that outputs the evaluation points of is connected in order from the input side,
The article analysis device further inputs the character data and the image data of the predicted article data to the first layer of the neural network, so that the score of the predicted article data output from the third layer is calculated. The article analysis device according to claim 1, further comprising an effect prediction unit that presents the effect data of the prediction article data as prediction data. The article analysis device according to claim 2, wherein the machine learning unit uses the number of responses from readers to the past article data posted to a social network service as an evaluation score for each piece of past article data. . The article analysis device has a machine learning unit and an improvement calculation unit,
The machine learning unit stores, in the storage unit, a result of counting the appearance frequency of each word acquired as character data of past article data as character feature data,
The improvement calculation unit includes:
For each word acquired as the character data of the input predicted article data, the higher the degree of similarity with the word of the character feature data, and the higher the score of the appearance function of the character feature data, the higher the evaluation function that is scored. By calculating, the score of each word is calculated,
An article analysis method, wherein for each word of the predicted article data, a word indicating that the word of the character feature data having a high score is a rewriting candidate is presented to encourage rewriting of the word of the predicted article data.

Images