JP6623186B2 - Content evaluation prediction system and content evaluation prediction method - Google Patents

Description

  The present invention relates to a content evaluation prediction system and a content evaluation prediction method for predicting evaluation of content of a viewer who has viewed content.

2. Description of the Related Art In recent years, product advertisement contents in mail order sales consisting of images and sounds have been distributed over communication media such as television, radio, and the Internet. (For example, see Patent Document 1)
When performing online shopping for products using this TV shopping, a program producer (director) performs a program that creates a willingness to purchase viewers based on past experience from a creative viewpoint.

  The reaction to the viewer's content is the number of inquiries by telephone or fax after the advertisement is notified to the viewer, or the number of clicks in the Web (Web) browser. Based on this inquiry or the number of clicks (the number of accesses), the program producer determines whether the produced program has an effect on the sale of the product. Thereby, the program producer accumulates experience in program production and accumulates as know-how what kind of program configuration will increase the number of accesses leading to product sales.

JP 2009-139777 A

However, even for program producers, when creating advertising content, even if it is created based on know-how, the number of accesses is an overall evaluation of the content, and which part of the content contributes to sales It cannot be recognized. For this reason, the evaluation of the content by the viewer is different for each content, and high access is not always obtained.
In addition, for a program producer who creates advertisement content for the first time, it is completely unknown how many viewers can get access to the product from the advertisement content that he creates.

  The present invention has been made in view of such circumstances, and by supplying a predicted value of the number of accesses to the created content to the program producer who produces the content, the content can be obtained regardless of the accumulated amount of know-how. It is an object of the present invention to provide a content evaluation prediction system and a content evaluation prediction method that enable the production of content.

The present invention has been made to solve the above-described problems, and the content evaluation prediction system of the present invention extracts information features, which are features of the information of content whose information changes in time series, at predetermined intervals. a feature extraction unit, and the information feature, the machine learning using the sum of the number of access as the evaluation value of the content in each of the periods just after the predetermined period and the predetermined period, corresponding to the information And a prediction model generation unit that generates a content evaluation prediction model for predicting the number of accesses to the information feature by inputting the information feature.

  The content evaluation prediction system according to the present invention is characterized in that the information includes at least one of a moving image and a dialogue.

  In the content evaluation prediction system according to the present invention, when the feature extraction unit performs feature extraction of the moving image, the feature extraction function up to the previous stage of all connected layers in another convolutional neural network (CNN) after learning of image feature extraction It is characterized by using.

  In the content evaluation prediction system of the present invention, when the feature extraction unit performs feature extraction of the dialogue, it registers a decomposition restriction word, which is a word that restricts decomposition in morphological analysis, in the dictionary, and refers to the dictionary for the morpheme. It is characterized by performing analysis.

The content evaluation prediction system of the present invention is characterized in that the prediction model generation unit uses a sparse modeling method when generating the content evaluation prediction model.

In the content evaluation prediction method of the present invention, the feature extraction unit extracts a feature of the information of the content whose information changes in time series for each predetermined period, and a prediction model generation unit includes: said information feature, by machine learning using the sum of the number of access as the evaluation value of the content in each of the periods just after the predetermined period and the predetermined period, and inputs the information features corresponding to the information And a prediction model generation process for generating a content evaluation prediction model for predicting the number of accesses to the information feature.

  According to the present invention, a content evaluation prediction system that enables content production regardless of the amount of know-how accumulated by supplying a predicted value of the number of accesses to the created content to a program producer who produces content. And a content evaluation prediction method can be provided.

It is a figure which shows the structural example of the content evaluation prediction system by one Embodiment of this invention. It is a conceptual diagram explaining acquiring an image feature extraction function from CNN in AlexNet. 5 is a diagram illustrating a configuration example of an image feature vector table stored in an extracted feature storage unit 18. FIG. 4 is a diagram illustrating a configuration example of a dialogue feature vector table stored in an extracted feature storage unit 18. FIG. FIG. 6 is a diagram illustrating a configuration example of an access number table stored in an extracted feature storage unit 18. It is a conceptual diagram which integrates an image feature vector and a dialogue feature vector as a content comprehensive vector. It is a figure which shows the structural example of the learning content data set table memorize | stored in the extraction feature memory | storage part. It is a flowchart which shows the operation example of the production | generation process of the content evaluation prediction model in the content evaluation prediction system 1 of this embodiment. It is a graph which shows the comparison with the prediction number of incoming calls of each evaluation period of the object content which is the advertisement of goods predicted with the contents evaluation prediction model in this embodiment, and the number of incoming calls actually obtained. 10 is a graph showing a comparison between the predicted number of incoming calls in each evaluation cycle of the target content selected from the linked content in FIG. 9 and the actually obtained number of incoming calls.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a diagram illustrating a configuration example of a content evaluation prediction system according to an embodiment of the present invention. In FIG. 1, the content evaluation prediction system 1 according to the present embodiment includes a content data input unit 11, a feature extraction unit 12, an access number accumulation unit 13, a prediction model generation unit 14, an access number prediction unit 15, a database 16, and a decomposition restriction word. Each of the dictionary storage unit 17, the extracted feature storage unit 18, and the evaluation result storage unit 19 is provided.

  The content data input unit 11 reads learning content data used when generating a content evaluation prediction model, which will be described later, and target content data for predicting the number of accesses as an evaluation value for evaluating the content, and writes it to the database 16. Remember me. Here, the number of accesses refers to viewers viewing content at each evaluation cycle (for example, every minute) when the content is a moving image of an advertisement for a product that is being sold through mail-order shopping or the like. The number of incoming calls as the number of calls made to the call center of the company that sells the product. The greater the number of incoming calls, the higher the degree of contribution to the sales of the content product because there are more viewers interested in the product while viewing the content. Hereinafter, in the present embodiment, the number of accesses will be described as the number of incoming calls from the viewer.

The feature extraction unit 12 extracts a feature (information feature) for each evaluation period from information that changes in time series of the content (the moving image, dialogue, etc.). In the present embodiment, since the content is described as a moving image, an image feature and a dialogue feature are extracted from each of the image and the dialogue.
Therefore, the feature extraction unit 12 includes an image feature extraction unit 121 that extracts image features and a dialogue feature extraction unit 122 that extracts dialogue features.
The image feature extraction unit 121 is formed by CNN (Convolutional Neural Network) learned by machine learning using deep learning, and has an image feature extraction function for extracting image features. A feature vector (image feature vector) of an image having a predetermined number of dimensions corresponding to the obtained image is extracted.

  In this embodiment, a scene is sampled every second from the moving image of the content, the pixels of the scene are divided into 227 × 227 blocks in a grid pattern, and the average of the RGB gradation in the pixels is calculated for each block. The numerical values obtained are used as a data string for extracting features. Since 227 × 227 = 51529 and there are three pieces of data for each of RGB for 51529 blocks, the data string of the input image is composed of 154587 data.

In this embodiment, the image feature extraction unit 121 extracts an image feature vector composed of 4096-dimensional image features from the above-described data sequence of 154587.
Further, in the present embodiment, the CNN having the image feature extraction function in the image feature extraction unit 121 is used as other learning CNN after learning of image feature amount extraction. Creating a CNN from a clean state requires a very large number of images and is not practical when the number of learning content data that can be used for learning is limited. For this reason, in the present embodiment, a part from the layer in the CNN network in AlexNet to the part obtained by removing the last all combined layer is used as the image feature extraction function.

FIG. 2 is a conceptual diagram for explaining acquisition of an image feature extraction function from a CNN in AlexNet. The CNN has a configuration in which a plurality of convolution layers, a maximum pooling layer, and a normalization layer are repeated, and finally a plurality of all coupling layers are provided. Here, in the present embodiment, up to the first fully connected layer (preceding to the fully connected layer) is used as an image feature extractor that extracts image feature values. In general, only the last fully connected layer is removed and used. However, in order to use without reducing the amount of information in the later stage, the first fully connected layer is used.
Then, the output of the first fully connected layer is connected to a predictor (sparse modeling or the like) which is a content evaluation prediction model described later in the present embodiment.

  Returning to FIG. 1, the image feature extraction unit 121 uses the CNN having the image feature extraction function transferred from the other shown in FIG. 2 to calculate the feature amount of each image from the learning content image or the evaluation content image. The image feature vector shown is extracted and written and stored in the extracted feature storage unit 18. At this time, the image feature extraction unit 121 averages the image feature amount vector for each predetermined time (for example, 1 second), the feature amount of each dimension of 4096 dimensions for each evaluation period (for example, 1 minute), and finally Specifically, the image feature vector is composed of 4096 dimensions. Then, the image feature extraction unit 121 writes and stores the obtained image feature vector in the extraction feature storage unit 18 for each evaluation period.

  FIG. 3 is a diagram illustrating a configuration example of an image feature vector table stored in the extracted feature storage unit 18. The image feature vector table is provided for each content. In each record, for each time (evaluation period), the content name (eg, video A in the present embodiment), description of time, video frame, and one to m dimensions (in this embodiment, m = 4096) Each feature amount is shown, and an image feature vector is shown for each record. The image feature vectors are represented as P1 (X1_1, X2_1,..., Xm_1), P2 (X1_2, X2_2,..., Xm_2),. Each of X1_1, X2_1,..., Xm_1 is an image feature amount in each dimension. For example, the feature quantity X1_1 is an average value of the first dimensional feature quantity in 4096 dimensions of each of the 60 feature vectors obtained in the evaluation period.

  Returning to FIG. 1, the dialogue feature extraction unit 122 performs morphological analysis on the speech dialogue text data that flows along with the image, and performs feature extraction of each obtained word. At this time, in the present embodiment, a word that is not desired to be divided into words smaller than necessary, such as a coined word created by concatenating a plurality of general words such as a product name of content, is set as a decomposition restriction word, and is limited in decomposition. It is written in advance and stored in the word dictionary storage unit 17 (registered in the dictionary).

  Accordingly, the line feature extraction unit 122 performs morpheme analysis with reference to this dictionary when performing input morpheme analysis, and even connected words that are normally decomposed are registered as decomposition limited words in the dictionary. Is treated as the simplest word, and no further division is performed. Further, the dialogue feature extraction unit 122 outputs a feature feature vector indicating the feature amount of the word subjected to morphological analysis using Word2vec or Doc2vec. At this time, the dialogue feature extraction unit 122 extracts a 100-dimensional feature vector of each of all words included in the dialogue for each evaluation period (for example, one minute). Further, instead of the above Word2vec and Doc2vec, the word feature extraction may be performed using another method as long as it is an algorithm for extracting the feature of a word for which unsupervised learning is performed.

Then, the dialogue feature extraction unit 122 extracts each of the maximum feature amount and the minimum feature amount of all the words in each of the 100 dimensions of the feature vector for each evaluation period, and all the words The average value of the feature quantities is calculated. For example, the maximum value and the minimum value are extracted from the first dimension feature quantities in 100 dimensions in the feature quantity vectors of all words, and the first dimension feature quantities of all words are aggregated and one dimension Extend eye features to two dimensions. Further, the average value of the first-dimensional feature values in 100 dimensions in the feature value vectors of all words is calculated, and further one-dimensional expansion is performed. As a result, the one dimension of the feature amount is expanded to three dimensions.
Thereby, the line feature extraction unit 122 aggregates the feature values of all words, the feature values of all the dimensions of all words, and sets the feature values of each of the 100 dimensions to the maximum value for each dimension of 100 dimensions. A dialogue feature vector having 300-dimensional feature values, which is composed of data expanded to three types (three-dimensional) of the minimum value and the average value, is written and stored in the extracted feature storage unit 18.

  FIG. 4 is a diagram illustrating a configuration example of a dialogue feature vector table stored in the extracted feature storage unit 18. The dialogue feature vector table is provided for each content. For each record, for each time (evaluation cycle), the content name (for example, video A in the present embodiment), the description of the time and scene dialogue, and the first dimension to the nth dimension (n = 300 in the present embodiment). Each feature amount is shown, and a dialogue feature vector is shown for each record. The dialogue feature vectors are expressed as Q1 (Y1_1, Y2_1, Y3_1,..., Ym_1), Q2 (Y1_2, Y2_2,..., Ym_2),. Here, the dialogue feature vector is a dialogue feature vector used in a frame corresponding to the image feature vector in FIG. Each of Y1_1, Y2_1, Y3_1,..., Ym_1 is a line feature amount in each dimension. Here, for example, the feature quantity Y1_1 is the maximum value of the first dimension feature quantity in the feature vectors of all words, the feature quantity T1_2 is the minimum value of the first dimension feature quantity in the feature vectors of all words, and the feature quantity Y1_3 is the average value of the first-dimensional feature values in the feature vectors of all words.

  Returning to FIG. 1, the access number accumulating unit 13 inputs the number of accesses for each evaluation period of the learning content from an external device (such as a server that counts each incoming call of the operator), and extracts it according to the evaluation period. The data is written and stored in the storage unit 18.

  FIG. 5 is a diagram illustrating a configuration example of the access number table stored in the extracted feature storage unit 18. Time (evaluation period) and the number of incoming calls that are the number of accesses in the evaluation period are associated with each record.

  Returning to FIG. 1, the prediction model generation unit 14 obtains a content evaluation prediction model including a predetermined multiple regression model by machine learning. Here, the prediction model generation unit 14 combines the above-described image feature vector and dialogue feature vector to generate a content comprehensive feature vector. The content total feature vector is a content total feature vector having a dimension of 4396 dimensions, by combining an image feature vector composed of 4096 dimensions and a dialogue feature vector composed of 300 dimensions. Further, instead of the multiple regression model, multiple regression (linear regression) may be configured by a neural network, and a simple linear regression problem may be learned by the neural network to configure a content evaluation prediction model.

FIG. 6 is a conceptual diagram for integrating an image feature vector and a dialogue feature vector as a content total vector. Content vector vector 100 (for example, an image-based shopping program vector space) and dialogue feature vector vector space 102 (for example, a scenario vector space) are integrated, and the total content input to the content evaluation prediction model A feature vector space 104 (eg, a total feature vector space) is generated. In FIG. 6, each of the vector spaces 100, 102, and 104 is shown as three dimensions, but has 4096 dimensions, 300 dimensions, and 4396 dimensions, respectively.
For example, in the image feature vector vector space 100, the content name in the image feature vector table of FIG. 3 is the movie A, the feature vector at time 00:01 is the movie A_00: 01, and the feature is time 00:02. The vector is displayed as a moving image A_00: 02. Also, in the dialogue feature vector vector space 102, the content name in the dialogue feature vector table of FIG. 4 is the video A, the feature vector of time 00:01 is dialogue A_00: 01, and the feature is time 00:02. The vector is displayed as line A_00: 02.

  That is, the prediction model generation unit 14 integrates the image feature vector table of FIG. 3, the dialogue feature vector table of FIG. 4, and the access number table shown in FIG. 5 to learn the content evaluation prediction model and learn content data Create a pair. And the prediction model production | generation part 14 writes and memorize | stores the data set of learning content in the extraction feature memory | storage part 18 for every evaluation period.

  FIG. 7 is a diagram illustrating a configuration example of the learning content data set table stored in the extracted feature storage unit 18. The learning content data set table corresponds to time (evaluation period) for each record, and each image feature amount from one dimension to m dimension (in this embodiment, m = 4096) and m + 1 dimension to m + n dimension ( In this embodiment, each line feature amount of n = 300) and the number of received powers in the evaluation period are shown as a learning content data set. Here, the number of incoming calls corresponding to each dimension is the number of calls until the viewer views the content introducing the product, the viewer is interested in the product, moves to the place where the phone is located, and calls the call center In consideration of the delay time, the number of incoming calls in the predicted evaluation cycle and the immediately following evaluation cycle is set as the evaluation (number of incoming calls) for the image and dialogue of the predicted evaluation cycle. That is, the number of incoming calls in the predicted evaluation cycle and the immediately following evaluation cycle are added to form a learning content data set. Therefore, the content evaluation prediction model to be described later predicts the number of incoming calls in the evaluation cycle to be predicted and the immediately following evaluation cycle as the predicted number of incoming calls.

  Returning to FIG. 1, the prediction model generation unit 14 sequentially inputs the content comprehensive feature vectors in the learning content data set in time order in the learning content data set table with respect to the content evaluation prediction model. The predicted incoming call predicted from the content comprehensive feature vector is output. At this time, the prediction model generation unit 14 performs the content evaluation prediction so that the predicted incoming call output in the learning content data sets in all the evaluation cycles of the content approaches the number of incoming calls in the learning content data set in the same evaluation cycle. Adjust the weights of the neural network connection layer in the model. Moreover, the prediction model production | generation part 14 performs the process mentioned above also in the learning content data set table of several learning content.

  At this time, as a problem of regression, predicting one parameter (number of incoming calls) from a plurality of parameters (content comprehensive feature vector) is a multiple regression model. In this multiple regression model (content evaluation prediction model), the number of data twice the number of dimensions used for prediction is required. However, in this embodiment, since the content comprehensive feature vector has a dimension of 4396, 8792 pieces of data are simply required. However, when the number of data (learning content data sets) is insufficient or the calculation is simplified. For this purpose, a sparse modeling technique is used to calculate the coefficient in the combination. Specifically, an L1 minimization algorithm called LASSO (least absolute shrinkage and selection operator) is used. In the present embodiment, parameters are determined by 10 cross-validation to prevent overlearning. Further, as described above, the same processing can be performed when the content evaluation prediction model is generated not by the multiple regression model but by a neural network model.

The prediction model generation unit 14 obtains a weighting regression coefficient w (p) in the multiple regression model constituting the content evaluation prediction model by the following equation (1). In the following formula (1), n is an evaluation cycle number indicating the order of the evaluation cycles. N is the total number of evaluation periods in the content. Zn is the number of accesses (number of incoming calls) in the nth evaluation cycle number. w (p) is a weighting coefficient (regression coefficient) for the data of the p-th feature amount in the content comprehensive feature vector. x _n (p) is data of the p-th feature amount in the content comprehensive feature vector of the n-th evaluation cycle number. b is intercept data. λ is a regularization coefficient.

In the above equation (1), it is necessary to determine the coefficient w (p) in order to estimate the predicted incoming power number z from the x _n (p) by multiple regression. At this time, since the normalized term is included in the first order in the equation (1), the coefficient w (p) for x _n (p) that is not necessary for estimating the predicted incoming power z is set to 0 ( Ie sparse). This makes it possible to generate a content evaluation prediction model that predicts at least the predicted number of incoming calls z with respect to the rules described above for the number of learning content data sets.

  The access number prediction unit 15 uses the content evaluation prediction model generated by the prediction model generation unit 14 to obtain the predicted number of incoming calls for each evaluation period of the target content that is the content to be evaluated. Here, the access number prediction unit 15 causes the image feature extraction unit 121 to extract an image feature vector for each evaluation period of the target content. In addition, the access number prediction unit 15 causes the dialogue feature extraction unit 122 to extract dialogue feature vectors for each evaluation period of the target content.

  Then, the access number prediction unit 15 integrates each of the image feature vector and the dialogue feature vector supplied from the feature extraction unit 12 to generate a content comprehensive feature vector. The number-of-accesses prediction unit 15 inputs the content comprehensive feature vector to the content evaluation prediction model in time series for each evaluation period, and obtains the predicted number of incoming calls z.

As described above, according to the present embodiment, how much evaluation the created content receives from the viewer can be obtained in advance by the predicted number of incoming calls, so that the program creation without know-how in advertising program production Even a viewer can be created in consideration of viewer evaluation.
As a result, according to the present embodiment, since the program creator (even a program creator who does not have know-how in advertising program production) can create a content by predicting the viewer's evaluation, The user's evaluation is completely unknown, and there is no useless broadcast on TV shopping programs.

  Further, according to the present embodiment, in the created content, when the predicted number of incoming calls expected in a predetermined evaluation cycle cannot be obtained, the content of the portion corresponding to the evaluation cycle in the image and dialogue created with reference to past cases Since the content can be evaluated again, even a program producer who does not have the know-how of creating an advertising program can easily create content that can contribute to an increase in the predicted number of incoming calls.

  Next, a flow of processing for generating a content evaluation prediction model in the content evaluation prediction system 1 according to the present embodiment will be described with reference to FIG. FIG. 8 is a flowchart illustrating an operation example of the content evaluation prediction model generation process in the content evaluation prediction system 1 of the present embodiment.

Step S1:
The content data input unit 11 inputs learning content data for generating a content evaluation prediction model from an external device, and writes and stores it in the database 16.

Step S2:
The dialogue feature extraction unit 122 sequentially reads out text sentences from the database 16 in units of evaluation periods in the learning content. Then, the line feature extraction unit 122 extracts words from the text sentences in the evaluation period by performing morphological analysis while referring to the already described dictionary.

Step S3:
The dialogue feature extraction unit 122 extracts a 100-dimensional feature vector for each extracted word using Word2vec or Doc2vec. Then, the line feature extraction unit 122 extracts the maximum value and the minimum value of the feature amount in each dimension in the feature vectors of all words. Further, the line feature extraction unit 122 obtains an average value of feature amounts in each dimension in the feature vectors of all words. Accordingly, the dialogue feature extraction unit 122 has the maximum value and the minimum value of the feature values of all the words in the evaluation period and the average value of the feature values of all the words in each of the 100 dimensions, and thus the 300-dimensional feature. A dialogue feature vector having a quantity is generated.
Then, the dialogue feature extraction unit 122 writes and stores the generated dialogue feature vector in the dialogue feature vector table of the extraction feature storage unit 18 in association with the evaluation period.

Step S4:
The dialogue feature extraction unit 122 determines whether or not the extraction of dialogue feature vectors for each evaluation cycle of the learning content is completed in all the evaluation cycles of the learning content. At this time, if the dialogue feature extraction unit 122 completes the extraction of dialogue feature vectors for each evaluation cycle of the learning content in all the evaluation cycles of the learning content, the process proceeds to step S5. On the other hand, if the dialogue feature vector extraction for each learning content evaluation cycle has not been completed in all the evaluation cycles of the learning content, the dialogue feature extraction unit 122 learns the portion corresponding to the next evaluation cycle. In order to extract a dialogue feature vector from the content, the process proceeds to step S2.

Step S5:
The image feature extraction unit 121 sequentially reads out moving images from the database 16 in units of evaluation periods in the learning content. In the case of television shopping, 30 frames of images are used per second (predetermined period). For this reason, the image feature extraction unit 121 reads, for example, the first frame in one second as sampling.
Then, the image feature extraction unit 121 extracts the image feature vector of this image from the sampled image by the feature extraction function configured with the CNN transferred from the other CNN.

Step S6:
The image feature extraction unit 121 determines whether an image feature vector (feature vector for 60 seconds) having a predetermined period (1 second) within the evaluation period (1 minute) has been extracted (extraction has been completed). . At this time, if the image feature extraction unit 121 extracts an image feature vector having a predetermined period within the evaluation period, the process advances to step S7. On the other hand, if the image feature extraction unit 121 has not extracted an image feature vector of a predetermined period within the evaluation period, the process proceeds to step S5.

Step S7:
The image feature extraction unit 121 calculates an average value of feature values for each of 4096 dimensions of the 60 image feature vectors, that is, 60 image feature vectors, and calculates an image feature vector composed of the average value of each dimension as an evaluation cycle. As an image feature vector. Then, the dialogue feature extraction unit 122 writes and stores the obtained image feature vector in the image feature vector table of the extraction feature storage unit 18 in association with the evaluation period.

Step S8:
The image feature extraction unit 121 determines whether or not the extraction of the image feature vector for each evaluation period of the learning content is completed in all the evaluation periods of the learning content. At this time, the image feature extraction unit 121 advances the process to step S9 when the extraction of the image feature vector for each evaluation period of the learning content is completed in all the evaluation periods of the learning content. On the other hand, if the extraction of image feature vectors for each evaluation period of the learning content has not been completed in all the evaluation periods in the learning content, the image feature extraction unit 121 learns the part corresponding to the next evaluation period. In order to extract the image feature vector from the content, the process proceeds to step S5.

Step S9:
The access number accumulating unit 13 inputs the number of accesses for each evaluation period of the learning content from an external device, and writes and stores it in the access number table of the extracted feature storage unit 18 in correspondence with the evaluation period.
Next, the prediction model generation unit 14 generates a learning content data set used for generating a content evaluation prediction model that is a multiple regression model. At this time, the prediction model generation unit 14 reads the image feature vector table, the dialogue feature vector table, and the access number table from the extracted feature storage unit 18 and integrates them to form a learning content data set table. The extracted feature storage unit 18 is written and stored.

Step S10:
The prediction model generation unit 14 refers to the learning content data set table in the extracted feature storage unit 18, inputs the content comprehensive feature vector of the learning content data set to the multiple regression model, and the output number of predicted inputs is the learning content data set. A process of adjusting the regression coefficient of the multiple regression model is performed so as to approach the number of incoming calls in the table. At this time, the prediction model generation unit 14 sparses the regression coefficient for each feature amount of the multiple regression model by machine learning using an L1 minimization algorithm called LASSO, and adjusts content evaluation prediction based on the multiple regression model Generate a model. Further, the prediction model generation unit 14 writes and stores the generated content evaluation prediction model in the extracted feature storage unit 18.

Through the above-described processing, a content evaluation prediction model is generated, and the access number prediction unit 15 uses this content evaluation prediction model to obtain a predicted power reception number as an evaluation of the viewer of the target content to be evaluated.
That is, the access number prediction unit 15 reads the content evaluation prediction model from the extracted feature storage unit 18. Then, the access number predicting unit 15 inputs the content comprehensive feature vector generated from the image feature vector and the speech feature vector for each evaluation period of the target content to the content evaluation prediction model, and obtains the predicted number of incoming calls for each evaluation period. .
Then, the number-of-accesses prediction unit 15 adds information for identifying the target content that has been evaluated to the calculated predicted number of incoming calls, and sets each of the evaluation periods and the predicted number of incoming calls corresponding to the evaluation period as the target. The content evaluation result is written and stored in the evaluation result storage unit 19.

FIG. 9 is a graph showing a comparison between the predicted number of incoming calls in each evaluation period of the target content that is the advertisement of the product predicted by the content evaluation prediction model in the present embodiment and the actually obtained number of incoming calls. In the graph of FIG. 9, the horizontal axis indicates time (evaluation cycle (1 minute)), and the vertical axis indicates the predicted number of incoming calls and the number of incoming calls (number of calls). This graph is generated by connecting the predicted number of incoming calls and the actual number of incoming calls of a plurality of contents.
Further, in this graph, the predicted number of incoming calls is indicated by a broken line, and the actual number of incoming calls is indicated by a solid line. The error (RSME: root mean squared error) between the predicted number of incoming calls and the actual number of incoming calls in the graph of FIG. 9 is in the range of ± 1.0 as an average value of all evaluation periods.

FIG. 10 is a graph showing a comparison between the predicted number of incoming calls for each evaluation period of the target content selected from the linked content in FIG. 9 and the number of incoming calls actually obtained. FIG. 10A shows the target content for which the product A is advertised in the television shopping. FIG. 10B shows target content for which the product B is advertised in the television shopping.
In each of FIG. 10A and FIG. 10B, the horizontal axis indicates time (evaluation cycle (1 minute)), and the vertical axis indicates the predicted number of incoming calls and the number of incoming calls (number of calls). Also, in both FIG. 10A and FIG. 10B, the content length is 30 minutes. Further, in the graphs of FIGS. 10A and 10B, the predicted number of incoming calls is indicated by a broken line, and the actual number of incoming calls is indicated by a solid line.

The error (RSME) between the measured power input number and the actual power input number of the product A in FIG. 10A is 0.96477 as an average value of all evaluation periods in 30 minutes. In addition, the error (RSME) between the measured power input number and the actual power input number of the product B in FIG. 10B is 0.77674 as an average value of all evaluation periods in 30 minutes.
From this result, the predicted number of incoming calls predicted by the content evaluation prediction model according to the present embodiment is almost the same as the number of incoming calls indicating the evaluation of the viewer who actually viewed the content (the change tendency of the number of incoming calls is also the same). It turns out that it is obtained.

Therefore, according to the present embodiment, the newly created content is the target content, and the content of the advertisement is divided into a plurality of parts and evaluated by obtaining the predicted number of incoming calls in each evaluation cycle by the content evaluation prediction model. It is possible to separate and evaluate the part that introduces the product to be advertised in the content (evaluation cycle) and the part that improves the viewer's willingness to purchase the product (evaluation cycle). It is also possible to estimate the relationship between the part that introduces products and the part that improves purchase motivation.
In addition, according to the present embodiment, in the newly created content, the predicted number of incoming calls in the evaluation cycle created by giving an impact to the viewer does not increase as expected in order to increase the number of incoming calls. In this case, it is possible to change the image and dialogue in this evaluation period in the content and try again before the broadcast, and support the work of creating content that can contribute to an increase in the predicted number of incoming calls.

In the above-described embodiment, the content is TV shopping, and the feature of the content input to the content evaluation prediction model is the image feature and the dialogue feature. However, if the content is an advertisement for a product such as a radio, in addition to the dialogue features, features such as speech frequency and intensity (speech signal amplitude value) of the speaker who speaks dialogue are extracted and used as a content evaluation prediction model. An input configuration may be used.
Further, image features may be extracted and used in the same manner for content in which still images change in time series instead of moving images in television shopping.

  Further, in the above embodiment, feature quantities extracted from words by Word2vec or Doc2vec are used as feature quantities of words in the dialogue. However, in order to learn the long-term dependencies of the words given in time series within the evaluation period in the time-series arrangement of each word, the future takes into account recurrent neural network, past information LSTM (long short-term memory), which is an advanced type of a neural network that predicts the above, may be used.

  In the case of this configuration, feature vectors extracted from words by Word2vec or Doc2vec described above are input to the LSTM, and for each feature vector, an array feature vector (word feature vector) that considers the time-series arrangement of words. Forward and backward vectors in a different time series). Thereby, the arrangement in the context of each word, which is not reflected as a parameter in the word feature vector, can be reflected, and the contribution to the evaluation of the dialogue in the content with higher accuracy can be predicted.

Then, for each of the array vectors output for each word from the LSTM, an importance level (Attention) corresponding to the time-series word arrangement (word arrangement: line context) is calculated. That is, by calculating the importance of the array vector, it is determined which array vector is important for predicting the predicted number of incoming calls (number of accesses). Then, in the content evaluation prediction model that predicts the predicted number of incoming calls based on this importance, the predicted number of incoming calls is predicted by calculating a weighted average of the array vectors.
In this embodiment, the number of incoming calls by telephone is used as the reaction (number of accesses) indicating the evaluation from the viewer, but the number of received faxes or the product in the web browser (Internet browser) The number of clicks selected may be used as a user evaluation reaction.

  Further, a program for realizing the function of predicting the number of accesses from the feature vector of the input content information in the content evaluation prediction system shown in FIG. 1 is recorded on a computer-readable recording medium and recorded on this recording medium. The number of accesses may be predicted from the feature vector of the content information by causing the computer system to read and execute the program. Here, the “computer system” includes an OS and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design and the like within a scope not departing from the gist of the present invention.

DESCRIPTION OF SYMBOLS 1 ... Content evaluation prediction system 11 ... Content data input part 12 ... Feature extraction part 13 ... Access number accumulation part 14 ... Prediction model production | generation part 15 ... Access number prediction part 16 ... Database 17 ... Decomposition limit word dictionary memory | storage part 18 ... Extraction feature Storage unit 19 ... Evaluation result storage unit

Claims (6)

A feature extraction unit for extracting information features, which are features of the information of content whose information changes in time series, at predetermined intervals;
Said information feature, by machine learning using the sum of the number of access as the evaluation value of the content in each of the periods just after the predetermined period and the predetermined period, and inputs the information features corresponding to the information Thus, a content evaluation prediction system comprising: a prediction model generation unit that generates a content evaluation prediction model for predicting the number of accesses to the information feature. The content evaluation prediction system according to claim 1, wherein the information includes at least one of a moving image and a dialogue. The feature extraction unit
The feature extraction function up to the first stage of all connected layers in another convolutional neural network (CNN) after learning of image feature extraction is used when performing feature extraction of the moving image. Content evaluation prediction system. The feature extraction unit
4. The feature extraction of the dialogue is performed by registering a decomposition restriction word, which is a word that restricts decomposition in morpheme analysis, in a dictionary, and performing morpheme analysis with reference to the dictionary. The content evaluation prediction system described in 1. The prediction model generation unit
The content evaluation prediction system according to any one of claims 1 to 4, wherein a sparse modeling method is used when generating the content evaluation prediction model. A feature extraction process in which a feature extraction unit extracts an information feature, which is a feature of the information of content whose information changes in time series, at predetermined intervals;
Prediction model generation unit, and the information characteristic, the machine learning using the sum of the number of access as the evaluation value of the content in each of the periods just after the predetermined period and the predetermined period, corresponding to the information And a prediction model generation process for generating a content evaluation prediction model for predicting the number of accesses to the information feature by inputting the information feature.

Images