JP6839001B2 - Model learning device, information judgment device and their programs - Google Patents

Description

The present invention relates to an information determination technique for determining whether or not social media information is information in a certain field.

In recent years, the development of social networking services (SNS) has made it possible for individuals to easily and in real time transmit information. For example, a lot of sighting information from a person who happened to be present at the scene of a fire, an accident, etc. is posted on SNS. Some of these posts have photos and videos of the scene added to them, and are often used in news programs as images and videos that capture the moment when a fire or accident occurs. Therefore, the broadcasting station extracts this information by manual work such as constantly monitoring the SNS.

A keyword search is often used as a method for manually extracting necessary information from such an SNS. However, there are various ways to express the post, such as when the line name such as "The XX line is delayed" or when the station name such as "The train is delayed due to an accident at the △△ station" is written. There is. Therefore, it is difficult to create a keyword that covers all of these expressions.

Therefore, in order to solve these problems, many post extraction methods using machine learning have been studied. For example, there is a method of learning words and phrases that can be dangerous expressions depending on a specific theme by a neural network, and extracting words and phrases that can be dangerous expressions depending on a specific theme from SNS social big data. It is disclosed (see Patent Document 1). Further, for example, a method of calculating the n-gram of posts to SNS and the degree of relevance of meteorological conditions and extracting useful posts related to meteorological events by machine learning is disclosed (see Non-Patent Document 1). ).

JP-A-2015-72614

Masatsugu Hagiyuki, “Extracting Useful Posts from Social Media Using Selective Weather Information”, Natural Language Processing Society, Proceedings of the 22nd Annual Conference, pp.397-400, March 2016

In the conventional method using machine learning, the learning accuracy using machine learning has a great influence on the amount of learning data. Therefore, in order to improve the accuracy, it is necessary to prepare a lot of learning data. However, in order to create learning data, it is necessary to distinguish whether the social media information (posted text) is correct answer data or incorrect answer data. For this reason, there is a problem that costs such as manpower and work time are required.

Therefore, the present invention utilizes not only learning data that distinguishes social media information but also characteristics of text data whose fields such as news manuscripts are known, so that social media information is information in the same field as news manuscripts and the like. An object of the present invention is to provide a model learning device, an information determination device, and a program thereof for determining whether or not the information is present.

In order to solve the above problems, the model learning device according to the present invention learns a feature extraction model for extracting a feature amount from a plurality of manuscript sentences whose fields are known, and is social media information whose fields are known. It is a model learning device that learns an information judgment model for judging whether or not the posted text to be judged is information indicating the above field by using the posted text as teacher data, and is a vectorization means and a feature extraction model learning means. , A feature extraction means and an information determination model learning means.

In such a configuration, the model learning device uses the vectorizing means to average the distributed expression vector of the word unit stored in the storage means in advance with respect to the manuscript sentence or the posted sentence which is the teacher data, and then the manuscript unit or the posted sentence. Generates a sentence-based distributed representation vector, which is a unit distributed representation vector. The distributed expression vector for each word is given a numerical vector that is closer to the word that shows the similar meaning and content from the distribution of the words. This distributed representation vector can be learned and generated by a method such as word2vec.
As a result, the vectorization means generates a vector in which the meaning and content of the manuscript sentence and the submitted sentence are added.

Then, the model learning device learns the feature extraction model that extracts the feature vector by compressing the dimension of the sentence unit distributed expression vector generated from the manuscript sentence by the feature extraction model learning means. This feature extraction model can be configured as a part for performing feature extraction of the autoencoder of the neural network.
Then, the model learning device extracts the feature vector from the sentence-based distributed representation vector generated from the teacher data by the feature extraction means using the feature extraction model.

Then, the model learning device inputs the feature vector for the teacher data extracted by the feature extraction means by the information determination model learning means and the sentence unit distributed expression vector of the teacher data from which the feature vector is extracted. The information judgment model is learned by machine learning.
In this way, the information determination model learning means uses the feature vector extracted by the feature extraction model learned from the manuscript sentence with respect to the sentence-based distributed representation vector of the teacher data, so that the past manuscript such as a news manuscript can be used. Therefore, the learning effect can be enhanced.
The model learning device can be operated by a model learning program for operating the computer as each of the above-mentioned means.

Further, in order to solve the above-mentioned problems, the information determination device according to the present invention uses the feature extraction model and the information determination model learned by the model learning device to generate unknown data which is a posted sentence of social media information whose field is unknown. This is an information determination device for determining whether or not the information indicates a learned field, and is configured to include a vectorization means, a feature extraction means, and a determination means.

In such a configuration, the information determination device averages the distributed expression vector of the word unit stored in the storage means in advance with respect to the unknown data by the vectorizing means, and the sentence unit distributed which is the distributed expression vector of the posting unit. Generate a representation vector.
Then, the information determination device extracts the feature vector from the sentence-based distributed representation vector generated by the vectorization means by the feature extraction means using the feature extraction model. Since the feature extraction model is pre-learned from a manuscript whose field is known, the feature vector generated from unknown data includes features that should be extracted if it is information about the field.
Then, the information determination device has learned unknown data from the sentence unit distributed expression vector generated by the vectorization means and the feature vector extracted by the feature extraction means by using the information determination model. Judge whether the information indicates the field.

Further, in order to solve the above-mentioned problems, the information determination device according to the present invention learns a feature extraction model for extracting a feature amount from a plurality of manuscript sentences whose fields are known, and social media information whose fields are known. Using the posted text as teacher data, the information judgment model for determining whether or not the posted text to be judged is information indicating the above field is learned, and the unknown data which is the posted text of social media information whose field is unknown is , An information determination device for determining whether or not the information indicates a learned field, the vectorization means, the feature extraction model learning means, the feature extraction means, the information determination model learning means, and the determination means. It was configured to be prepared.
The information determination device can be operated by an information determination program for operating the computer as each of the above-mentioned means.

The present invention has the following excellent effects.
According to the present invention, it is possible to determine whether or not the social media information is information in the same field by utilizing the characteristics of the manuscript text (text data) whose field is known.
Thereby, the present invention can enhance the learning effect by using a large amount of existing manuscripts without learning a large amount of social media information in SNS, and can improve the accuracy of information determination.
Thereby, the present invention can effectively utilize social big data transmitted by an individual on SNS as an information source such as news.

It is a block block diagram which shows the structure of the information determination apparatus which concerns on embodiment of this invention. It is a figure for demonstrating the processing content of the vectorization means, (a) is an example of dividing media information into words, (b) calculates the distributed expression vector of a manuscript sentence or a post sentence from the distributed expression vector of a word. It is explanatory drawing for demonstrating the example. It is explanatory drawing for demonstrating the structure of the feature extraction model which a feature extraction model learning means learns. It is explanatory drawing for demonstrating the structure of the information judgment model which the information judgment model learning means learns. It is a flowchart which shows the operation of the 1st learning mode of the information determination apparatus which concerns on embodiment of this invention. It is a flowchart which shows the operation of the 2nd learning mode of the information determination apparatus which concerns on embodiment of this invention. It is a flowchart which shows the operation of the evaluation mode of the information determination apparatus which concerns on embodiment of this invention. It is a block block diagram which shows the structure of the model learning apparatus which concerns on other embodiment of this invention. It is a block block diagram which shows the structure of the information determination apparatus which concerns on other embodiment of this invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[Configuration of information judgment device]
First, the configuration of the information determination device 1 according to the embodiment of the present invention will be described with reference to FIG.

The information determination device 1 includes a control unit 10 and a storage unit 20.
The information determination device 1 determines whether or not the information transmitted by the SNS (tweet [registered trademark], etc., which is text data for each posting) is information in a specific field.

As shown in FIG. 1, the control unit 10 includes a word dividing means 11, a vectorizing means 12, a feature extraction model learning means 13, a feature extraction means 14, an information determination model learning means 15, and a determination means 16. , Equipped with.
The control unit 10 controls the operation of the information determination device 1. The control unit 10 operates in three operation modes. The first operation mode is a first learning mode in which a feature extraction model for extracting a feature amount of information from information (text data) whose field is known is learned. The second operation mode is the second learning to learn an information judgment model for determining whether or not unknown media information is information in the field from social media information in which the field is known (hereinafter, simply media information). The mode. The third operation mode is an evaluation mode for determining whether or not the unknown media information is information in the learned field.

In the present embodiment, as a specific field, "news" that can be used in news programs and the like will be described as an example. Of course, this field may be a field other than news, and may be, for example, sports, music, or the like.

In the first embodiment, in the first learning mode, the control unit 10 inputs a large amount (for example, hundreds of thousands) of news manuscripts as input information. For example, the control unit 10 inputs a news title in units of manuscripts.
Further, in the second learning mode, the control unit 10 inputs a large amount (for example, tens of thousands) of media information known to be news in the field as input information. For example, the control unit 10 inputs a tweet [registered trademark] in units of posts.
Further, in the evaluation mode, the control unit 10 inputs media information whose field is unknown as input information, and outputs a determination result of whether or not the field is news information.

The word dividing means 11 divides a news manuscript (manuscript sentence) or media information (posted sentence), which is text data, into words. Specifically, the word dividing means 11 divides the text data into words by morphological analysis and outputs the text data to the vectorizing means 12.

The vectorization means 12 vectorizes the manuscript sentence or the posted sentence, which is the text data, from the words divided by the word dividing means 11. The vectorizing means 12 averages the distributed expression vectors of the words constituting the manuscript sentence or the posted sentence from the distributed expression vector for each word stored in advance in the distributed expression vector storage means 21, and the manuscript unit or the post unit. Generate a distributed representation vector.
The variance representation vector is a vector in which words having similar meanings (similar characteristics of variance) are associated with close vectors, and the words are represented by a finite high-dimensional (for example, 200-dimensional) numerical vector. This distributed representation vector is generated by a general method such as word2vec or GloVe (Global Vectors for Word Representation).

The vectorization means 12 reads out the distributed representation vector corresponding to the divided words and adds them. Then, the vectorizing means 12 normalizes the vector by dividing the added distributed expression vector by the number of words included in the manuscript sentence or the posted sentence, and the distributed expression vector (sentence unit distributed expression) of the manuscript sentence or the posted sentence. Vector) is generated.

In the first learning mode, the vectorization means 12 outputs the generated sentence-based distributed expression vector to the feature extraction model learning means 13. Further, in the second learning mode, the vectorization means 12 outputs the generated sentence unit distributed expression vector to the feature extraction means 14 and the information determination model learning means 15. Further, in the evaluation mode, the vectorization means 12 outputs the generated sentence-based distributed expression vector to the feature extraction means 14 and the determination means 16.

Here, the sentence-based distributed representation vector will be described with reference to FIG. 2 (see FIG. 1 as appropriate).
As shown in FIG. 2A, when an example of text data (manuscript sentence or posted sentence) is "the XX line is delayed due to an accident", the word dividing means 11 sets the text data as "○". ○ / line / ga / accident / de / delayed / / is /. ”Is divided into word units.

Then, the vectorization means 12 reads out the corresponding distributed expression vector from the distributed expression vector storage means 21 for each divided word. For example, as shown in FIG. 2B, the distributed representation vector “0.1, 0.3, 0.4, 0.” Has n (for example, 200 dimensions) of dimensions corresponding to the word “○○”. 1,0.8, 0.9, 0.2, ..., 0.9 ”is read out.
Then, the vectorization means 12 adds the distributed expression vectors by the number of words constituting the text data, and totals all the words (in the example of FIG. 2B, “7.2, 1.8, 2.7, 3.6, 3.6, 7.2, 4.5, ..., 6.3 ") is calculated.

After that, the vectorization means 12 divides the total of all the words of the distributed expression vector by the number of words constituting the text data (9 in the example of FIG. 2), thereby dividing the sentence-based distributed expression vector (FIG. 2 (b). ), “0.8, 0.2, 0.3, 0.4, 0.4, 0.8, 0.5, ..., 0.7”) is calculated.
As a result, the vectorizing means 12 generates a sentence-based distributed representation vector for each manuscript sentence or posted sentence.
Returning to FIG. 1, the configuration of the information determination device 1 will be described.

The feature extraction model learning means 13 learns a feature extraction model for extracting a feature amount from a sentence-based distributed expression vector generated by the vectorization means 12 in the first learning mode. The feature quantity of the sentence-based distributed representation vector is obtained by compressing the number of dimensions of the sentence-based distributed representation vector.

Specifically, the feature extraction model learning means 13 uses an autoencoder AE, which is a neural network composed _{of an input layer A L1} , a hidden layer A _L2 , and an output layer A _{L3 shown in FIG.} learning the feature extraction model M ₁ for extracting.
Autoencoder AE shown in FIG. 3, the input layer A _L1, and inputs the sentence distributed representation vector Vs generated by vectorization unit 12. The input layer A _L1 and the output layer A _L3 have the same number of dimensions (for example, 200 dimensions) as the sentence-based distributed representation vector. The hidden layer A _L2 has a smaller number of dimensions than the input layer A _L1 and the output layer A _L3 (for example, 100 dimensions).

Feature extraction model learning unit 13, in Autoencoder AE, the input layer _{A L1} and an output layer _{A L3} is such that the same sentence distributed representation vector Vs, the encoding type from the input layer _{A L1} to the hidden layer _{A L2} , The parameters such as the coefficient of the decoding formula from the hidden layer A _L2 to the output layer A _{L3 are learned.} As described above, the autoencoder AE provides the _{hidden layer A L2} having a smaller number of dimensions than _{the input layer A L1} and the output layer A _{L3 in} the intermediate layer, so that the dimension of the sentence unit distributed expression vector is provided in the hidden layer A _L2. It is possible to extract the feature amount compressed by.

The feature extraction model learning means 13 learns the autoencoder AE so that the sentence-based distributed expression vectors that are sequentially input become the same in _{the input layer A L1} and the output layer A _L3, and hides from _{the input layer A L1.} the parameters to a _L2, learning as a feature extraction model _{M 1.} For learning the autoencoder AE, for example, an error back propagation method is used.
The feature extraction model learning means 13 ends learning when the learning using the news manuscript is performed a predetermined number of times or when the parameter error converges within a predetermined error.
The feature extraction model learning means 13 _{writes and stores the learned feature extraction model M 1} in the feature extraction model storage means 22.

In the second learning mode or the evaluation mode, the feature extraction means 14 uses the feature extraction model stored in the feature extraction model storage means 22 to obtain a feature vector from the sentence-based distributed representation vector generated by the vectorization means 12. (Feature amount) is extracted.
The feature extraction means 14 uses the feature extraction model M ₁ (see FIG. 3) stored in the feature extraction model storage means 22 to convert a sentence-based distributed representation vector (for example, 200 dimensions) into a vector having a small number of dimensions. (For example, 100 dimensions) is calculated and a feature vector is extracted.
The feature extraction means 14 outputs the extracted feature vector to the information determination model learning means 15 in the second learning mode. Further, the feature extraction means 14 outputs the extracted feature vector to the information determination model learning means 15 in the evaluation mode.

In the second learning mode, the information determination model learning means 15 learns an information determination model that determines whether or not the media information is information in a certain field from the sentence unit distributed expression vector generated by the vectorization means 12. It is a thing.

Specifically, the information determining the model learning unit 15, an input layer _{F L1} shown in FIG. 4, the hidden layer _{F L2,} constituted order propagation neural network output layer _{F L3:} the (Feed Forward Neural Network FFNN), information The determination model M ₂ is learned.
FFNN shown in Figure 4, the input layer F _L1, and inputs a sentence unit variance representation vectors Vs generated by vectorization unit 12, a feature vector Vf that is extracted by the feature extraction model M _1. Then, FFNN, in the hidden layer _{F L2,} and is propagated by adding a weight to the value of each element of the input to the input layer _{F L1} vector (sentence distributed representation vector Vs + feature vector Vf), the output layer _{F L3} Outputs the judgment result from. Here, the output layer _FL3 outputs, for example, by setting the number of dimensions to 2 and normalizing the probability indicating that the sentence unit distributed expression vector Vs is the information of the field learned in the first learning mode from one node. To do. Further, from the other node, the probability indicating that the sentence unit distributed expression vector Vs is not the information of the field learned in the first learning mode is normalized and output.

Then, when the teacher data is a positive example, the information determination model learning means 15 has a probability value that the output of one node is a vector of posted sentences in the field in which the sentence unit distributed expression vector Vs is learned in the first learning mode. 1 ", the output of the other node probability value" so that the 0 "to learn the weights of each layer as a parameter information determining the model M _2. Also, if the teacher data is negative example, the output of one node is "0", so that the output of the other nodes becomes "1", to learn the weights of each layer as a parameter information determining the model M _2. For learning FFNN, for example, an error back propagation method is used.
The information determination model learning means 15 ends learning when the learning using the teacher data is performed a predetermined number of times or when the parameter error converges within a predetermined error.
The information determination model learning means 15 writes and stores the learned information determination model in the information determination model storage means 23.

In the evaluation mode, the determination means 16 determines whether or not the media information is information in the field learned in the learning modes (first learning mode and second learning mode).
In the evaluation mode, the determination means 16 inputs the sentence unit distributed expression vector from the vectorization means 12, and inputs the feature vector extracted from the sentence unit distributed expression vector from the feature extraction means 14.

The determination means 16 uses the information determination model stored in the information determination model storage means 23, and the sentence-based distributed expression vector and the feature vector input are vectors corresponding to the information in the field learned in the learning mode. Judge whether or not. Specifically, the determination means 16 determines based on a result of inputs the Buntan'i distributed representation vector Vs and the feature vector Vf to the input layer F _L1 of FFNN shown in FIG. 4, is outputted from the output layer F _L3 I do. In the example of FIG. 4, the determination means 16 subtracts the probability value output from the other node from the probability value which is the information of the event in the learned field which is the output of one node of _{the output layer FL3, and is positive.} If so, it is determined that the media information is the information of the learned field. On the other hand, if it is negative, the determination means 16 determines that the media information is not the information in the learned field.
Thereby, the determination means 16 can determine whether or not the media information is the information of the learned field. The determination means 16 outputs this determination result to the outside.

The storage unit 20 includes a distributed expression vector storage means 21, a feature extraction model storage means 22, and an information determination model storage means 23. The storage unit 20 stores various data used or generated by the operation of the information determination device 1.
Each of these storage means can be configured by a general storage device such as a hard disk or a semiconductor memory. Here, although each storage means is individually provided in the storage unit 20, the storage area of one storage device may be divided into a plurality of storage means as each storage means. Further, the storage unit 20 may be used as an external storage device and may be omitted from the configuration of the information determination device 1.

The distributed expression vector storage means 21 stores the distributed expression vector in association with a word. Here, it is assumed that the distributed expression vector storage means 21 stores the distributed expression vector in advance, but the information determination device 1 may be configured to include the distributed expression vector generation means (not shown) in the control unit 10. I do not care. In that case, the distributed expression vector generation means generates a distributed expression vector for each word from a large amount of learning data such as existing media information by word2vec or the like, and stores it in the distributed expression vector storage means 21.

The feature extraction model storage means 22 stores the feature extraction model learned by the feature extraction model learning means 13. The feature extraction means 14 refers to the feature extraction model stored in the feature extraction model storage means 22.

The information determination model storage means 23 stores the information determination model learned by the information determination model learning means 15. The information determination model stored in the information determination model storage means 23 is referred to by the determination means 16.

By configuring the information determination device 1 as described above, the information determination device 1 can learn a feature extraction model that extracts features of a manuscript sentence in the news field from a news manuscript. Further, the information determination device 1 can learn an information determination model from media information whose information is known to be information in a predetermined field (here, news field) which is teacher data.
Then, the information determination device 1 can determine whether or not the unknown media information is information in the learned field by using the information determination model.
The information determination device 1 can be operated by a program (information determination program) that causes a general computer to function as each means of the control unit 10 described above.

[Operation of information judgment device]
Next, the operation of the information determination device 1 according to the embodiment of the present invention will be described with reference to FIGS. 5 to 7. It is assumed that the distributed expression vector storage means 21 stores the distributed expression vector in advance in association with the word. Here, the operation of the information determination device 1 will be described separately for the first learning mode, the second learning mode, and the evaluation mode.

(1st learning mode)
First, the operation of the first learning mode for learning the feature extraction model of the information determination device 1 will be described with reference to FIG. 5 (see FIG. 1 for the configuration as appropriate).
In step S1, the word dividing means 11 of the information determination device 1 inputs a news manuscript (for example, a news title) which is text data for each manuscript.
Then, in step S2, the word dividing means 11 of the information determination device 1 divides the manuscript sentence input in step S1 into words by morphological analysis.

Then, in step S3, the vectorizing means 12 of the information determination device 1 reads out the distributed expression vector corresponding to the words divided in step S2 from the distributed expression vector storage means 21 and adds the number of words.
Further, in step S4, the vectorizing means 12 of the information determination device 1 divides the distributed expression vector added in step S3 by the number of words included in the manuscript sentence, thereby normalizing the vector for each manuscript sentence ( Sentence unit distributed representation vector) is generated.

In step S5, the feature extraction model learning unit 13 of the information determination device 1, a sentence unit variance representation vectors generated in step S4, the input to the input layer A _L1 of Autoencoder AE shown in FIG. 3, and an output layer A _The feature extraction model is learned as the output from L3.

Then, in step S6, the feature extraction model learning means 13 of the information determination device 1 determines whether or not the learning is completed depending on whether the learning is performed a predetermined number of times or the parameter error of the feature extraction model has converged.
If it is determined in step S6 that the learning has not been completed (No), the information determination device 1 returns to step S1 and continues the learning operation.
On the other hand, when it is determined in step S6 that the learning is completed (Yes), the information determination device 1 writes the learned feature extraction model in the feature extraction model storage means 22 in step S7.

By the above operation, the information determination device 1 can generate a feature extraction model for extracting a feature amount whose text data is information in the field of news, using a large amount of news manuscripts as teacher data.

(Second learning mode)
Next, the operation of the second learning mode for learning the information determination model of the information determination device 1 will be described with reference to FIG. 6 (see FIG. 1 for the configuration as appropriate). This second learning mode is performed after the operation of the first learning mode described with reference to FIG.
In step S10, the word dividing means 11 of the information determination device 1 inputs media information (teacher data) known to be news in the field for each post.
Then, in step S11, the word dividing means 11 of the information determination device 1 divides the posted sentence input in step S10 into words by morphological analysis.

Then, in step S12, the vectorizing means 12 of the information determination device 1 reads out the distributed expression vector corresponding to the words divided in step S11 from the distributed expression vector storage means 21 and adds the number of words.
Further, in step S13, the vectorizing means 12 of the information determination device 1 divides the distributed expression vector added in step S12 by the number of words included in the posted sentence, thereby normalizing the vector for each manuscript sentence ( Sentence unit distributed representation vector) is generated.

In step S14, the feature extraction means 14 of the information determination device 1 uses the feature extraction model stored in the feature extraction model storage means 22 and uses the feature extraction model generated in step S13 to obtain features that are feature quantities from the sentence-based distributed representation vector. Extract the vector and use it as the output of the feature extraction model.
Then, in step S15, the information determination model learning means 15 of the information determination device 1 converts the sentence-based distributed representation vector of the teacher data generated in step S13 and the feature vector generated in step S14 into the FFNN shown in FIG. as inputs to the input layer F _L1, it is supervised information determination model.

Then, in step S16, whether or not the information determination model learning means 15 of the information determination device 1 has completed the learning depending on whether the learning using the teacher data is performed a predetermined number of times or the parameter error of the information determination model has converged. To judge.
If it is determined in step S16 that the learning has not been completed (No), the information determination device 1 returns to step S10 and continues the learning operation.
On the other hand, when it is determined in step S16 that the learning is completed (Yes), the information determination device 1 writes the learned information determination model in the information determination model storage means 23 in step S17.

By the above operation, the information determination device 1 can generate an information determination model for determining whether or not the media information whose field is unknown is the information in the news field from the teacher data.

(Evaluation mode)
Next, the operation of the evaluation mode of the information determination device 1 will be described with reference to FIG. 7 (see FIG. 1 for the configuration as appropriate). The operation of this evaluation mode is performed after the operation of the second learning mode described with reference to FIG.
The operation of steps S20 to S24 is the same as the operation of steps S10 to S14 described with reference to FIG. 6, except that the input information is media information of teacher data or media information of unknown field. Therefore, the description thereof will be omitted.

In step S25, the determination means 16 of the information determination device 1 uses the information determination model stored in the information determination model storage means 23 to obtain media information whose field is unknown from the sentence unit distributed expression vector and the feature vector. , Judge whether it is the information of the field learned in the learning mode. This sentence-based distributed representation vector is generated in step S23, and the feature vector is generated in step S24.
Further, in step S26, the determination means 16 of the information determination device 1 outputs the result of determination in step S25 to the outside.

In step S27, the information determination device 1 determines the end of the operation of the evaluation mode depending on whether or not media information is further input.
If the media information is further input in step S27 and the evaluation mode operation is not completed (No), the information determination device 1 returns to step S20 and continues the determination operation.
On the other hand, in step S27, when new media information is not input and the operation of the evaluation mode ends (Yes), the operation ends.
By the above operation, the information determination device 1 can determine whether or not the unknown media information is the information of the field learned in the learning mode.

Although the configuration and operation of the information determination device 1 according to the embodiment of the present invention have been described above, the present invention is not limited to this embodiment.
Here, the information determination device 1 uses a learning operation (first learning mode and second learning mode) for learning the feature extraction model and the information determination model, and the feature extraction model and the information determination model to obtain unknown media information. , It is assumed that two operations, a determination operation (evaluation mode) for determining whether or not the information is in the learned field, are performed by one device.
However, these operations may be operated by separate devices.

Specifically, the device that realizes the learning operation for learning the feature extraction model and the information determination model can be configured as the model learning device 3 shown in FIG.
As shown in FIG. 8, the model learning device 3 may be configured by omitting the determination means 16 from the information determination device 1 described with reference to FIG. This configuration performs only the learning operation of learning the feature extraction model and the information determination model, which is the same as the information determination device 1 described with reference to FIG. The operation of the model learning device 3 is the same as the operation described with reference to FIGS. 5 and 6.
The model learning device 3 can be operated by a program (model learning program) for operating the computer as each of the above-mentioned means.

Further, a device that realizes a determination operation for determining whether or not the unknown media information is information in the learned field by using the feature extraction model and the information determination model is configured as the information determination device 1B shown in FIG. can do.
As shown in FIG. 9, the information determination device 1B may be configured by omitting the feature extraction model learning means 13 and the information determination model learning means 15 from the information determination device 1 described with reference to FIG. This configuration performs only the determination operation of determining whether or not the unknown media information is the information of the field learned in advance, which is the same as the information determination device 1 described with reference to FIG. The operation of the information determination device 1B is the same as the operation described with reference to FIG. 7.
The information determination device 1B can be operated by a program (information determination program) for operating the computer as each of the above-mentioned means.
By operating the learning operation and the determination operation on different devices in this way, the feature extraction model and the information determination model learned by one model learning device 3 can be used by a plurality of information determination devices 1B. become.

Further, here, the information determination model learned by the information determination model learning means 15 is a neural network learned by supervised learning. However, this supervised learning can use other common machine learning. For example, a support vector machine (SVM), a conditional random field (CRF), or the like can be used.

Further, here, the word dividing means 11 divides the input news manuscript (manuscript sentence) and media information (posted sentence) into words by morphological analysis. However, when the input to the information determination device 1 is divided into words in advance, the word dividing means 11 can be omitted from the configuration of the information determination device 1.

1,1B Information judgment device 11 Word division means 12 Vectorization means 13 Feature extraction model learning means 14 Feature extraction means 15 Information judgment model learning means 16 Judgment means 21 Distributed expression vector storage means 22 Feature extraction model storage means 23 Information judgment model memory Means 3 Model learning device

Claims (6)

A feature extraction model for extracting the feature amount of the information contained in the manuscript sentence is learned from a plurality of manuscript sentences whose fields are known, and the posted sentence which is the social media information whose field is known is judged as teacher data. It is a model learning device that learns an information judgment model for determining whether or not the posted text that is the target of the above field is information indicating the above field.
A sentence-based distributed representation vector that is a manuscript-based or post-based distributed representation vector by averaging the word-based distributed representation vectors stored in advance in the storage means with respect to the manuscript text or the posted text that is the teacher data. Vectorization means to generate
A feature extraction model learning means for learning a feature extraction model that extracts a feature vector by compressing the dimension of the sentence unit distributed expression vector generated from the manuscript sentence, and
A feature extraction means for extracting the feature vector from the sentence-based distributed representation vector generated from the teacher data using the feature extraction model.
The information determination model is obtained by inputting a feature vector for the teacher data extracted by the feature extraction means and a sentence-based distributed expression vector of the teacher data that is the source of extracting the feature vector and performing machine learning. Information judgment model to learn Learning means and
A model learning device characterized by being equipped with. The model learning device according to claim 1, wherein the feature extraction model learning means learns parameters from an input layer to a hidden layer as the feature extraction model by an autoencoder which is a neural network. Using the feature extraction model and the information determination model learned by the model learning device according to claim 1 or 2, the unknown data, which is a post of social media information whose field is unknown, indicates the learned field. It is an information judgment device that determines whether or not it is.
A vectorization means for averaging the word-based distributed representation vector stored in the storage means in advance with respect to the unknown data to generate a sentence-based distributed representation vector, which is a post-unit distributed representation vector.
A feature extraction means that extracts a feature vector from the sentence-based distributed representation vector generated by this vectorization means using the feature extraction model, and a feature extraction means.
Whether or not the unknown data indicates the learned field by using the information determination model from the sentence-based distributed representation vector generated by the vectorization means and the feature vector extracted by the feature extraction means. Judgment means to determine whether
An information determination device comprising. A feature extraction model for extracting the feature amount of the information contained in the manuscript sentence is learned from a plurality of manuscript sentences whose fields are known, and the posted sentence which is the social media information whose field is known is judged as teacher data. The information judgment model for determining whether or not the posted text to be the target of the field is information indicating the above-mentioned field is learned, and the unknown data which is the posted text of the social media information whose field is unknown indicates the learned field. It is an information judgment device that determines whether or not it is.
A sentence-based distributed representation vector, which is a manuscript-based or submission-based distributed representation vector, is obtained by averaging the word-based distributed representation vectors stored in advance in the storage means with respect to the manuscript sentence, the teacher data, or the unknown data. Vectorization means to generate
A feature extraction model learning means for learning a feature extraction model that extracts a feature vector by compressing the dimensions of the sentence-based distributed expression vector generated from the manuscript sentence.
A feature extraction means for extracting the feature vector from the teacher data or the sentence-based distributed representation vector generated from the unknown data using the feature extraction model.
The information determination model is obtained by inputting a feature vector for the teacher data extracted by the feature extraction means and a sentence-based distributed expression vector of the teacher data that is the source of extracting the feature vector and performing machine learning. Information judgment model to learn Learning means and
Using the information determination model, the sentence-based distributed representation vector generated from the unknown data generated by the vectorizing means and the feature vector extracted from the sentence-based distributed representation vector by the feature extracting means are used. , A determination means for determining whether or not the unknown data is information indicating the learned field,
An information determination device comprising. A model learning program for operating a computer as the model learning device according to claim 1 or 2. An information determination program for causing a computer to function as the information determination device according to claim 3 or 4.

Images