JP2019133565A - News material classification apparatus, program and learning model - Google Patents

Abstract

To accurately classify news posts extracted from social media information.SOLUTION: A series generating unit 10 of the news material classification apparatus 1-1 generates a character vector and a word vector from a news posts information extracted by a news material extraction apparatus 100. A sentence classification unit 11 performs calculation using the outputted data of a character vector and a word vector of the news posts information and the input data as the out and non-out information using a learning model and generates and outputs the out and non-out information. Thus, the news posts information is classified into non-out or out information. The learning model is constituted of a character NN21 of a character vector of the news post information as the input data, a word NN22 of a word vector of news post information as the input data, and an output NN23 of the vector which combines the calculation result of the character NN21 and the calculation result of the word NN22 as the input data, and the output data of the out and the non-out information as the output data.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a news material classification device and program for extracting information that can be news material from a large number of social media information, and classifying the extracted information, and a learning model used for the news material classification device.

  2. Description of the Related Art Conventionally, social media such as Twitter (registered trademark), which forms various information exchange services by an individual sending information using the Internet, has been actively used. In social media, post information with high immediacy (hereinafter referred to as “social media information”) is transmitted from an individual who is present at a scene such as a fire or an accident.

  In broadcasting stations, by extracting information that can be used as news material from individual social media information (hereinafter referred to as “news posting information”), the number of cases where news posting information is used in the production of news or programs has increased. Yes.

  If the work of extracting news posting information from social media information is performed manually, a large amount of labor is required, which is not realistic. Therefore, in order to reduce labor by manpower, a method of automatically performing a work of extracting news posting information from social media information has been proposed (see, for example, Patent Document 1 and Non-Patent Document 1).

  FIG. 22 is a diagram for explaining the outline of a conventional news material extracting apparatus. This news material extraction apparatus is an apparatus incorporating the technique of Patent Document 1, for example.

  In the learning phase, the news material extracting apparatus 100 inputs social media information as posting information, generates a feature as a feature amount of a feature vector, and generates a learning model by machine learning. Moreover, the news material extraction apparatus 100 inputs social media information in a determination phase, and determines whether the input social media information is news posting information using a learning model (learned model). Then, the news material extracting apparatus 100 outputs the news posting information and discards information other than the news posting information.

  The news material extracting apparatus 100 is based on extracting language information from input social media information and generating a learning model by machine learning using the language information as input data. Such a method using machine learning includes a case where a word series is used as input data, a case where a character series is used as input data, and a case where a combination of a word series and a character series is used as input data (for example, non-patent). Reference 2-6).

  Furthermore, a method for detecting a predetermined event such as a train delay from social media information has also been proposed (see, for example, Patent Documents 2 and 3).

JP 2017-201437 A JP 2013-168021 A Japanese Patent Application Laid-Open No. 2014-2551

Miyazaki et al., “Multi-classification of learning data for extracting useful information from Twitter”, IFAT, 2017, pages 1-6 Yoon Kim, et al. “Character-Aware Neural Language Models”, AAAI 2016 Rafal Jozefowicz, et al. “Exploring the limits of language modeling.”, CoRR, 2016, abs / 1602.02410. Lyan Verwimp, et al., “Character-Word LSTM Language Models”, EACL2017, pages 417-427 Rupesh K. Srivastava, et al., “Training Very Deep Networks.” NIPS, 2015, pages 2377-2385. X. Ma and E. Hovy.End-to-end sequence labeling via bi-directional lstm-cnns-crf.In Proceedings of the 54th ACL, pages 1064-1074, Berlin, Germany, August 2016.

  In general, there are many social media information transmitted to social media, and there are also many news posting information extracted from social media information. For this reason, even if the news material extraction device 100 shown in FIG. 22 automatically extracts news posting information from social media information, it is difficult for the user to monitor all news posting information.

  Further, the utilization method of the news posting information varies greatly depending on the user, purpose, application, situation, and the like. It is considered that the user wants to obtain only desired posting information among a large number of news posting information extracted from a large number of social media information. For this reason, it is desirable to classify the news posting information so as to adapt to the user's request.

  For example, if the user wants to obtain the first news posting information, the posting information related to the news agency or the summary site and the posting information that cites it are already posted (posts that have already appeared). Since it is not the posting information of the first report, it is not required. Here, the posted post means a post that has already appeared on social media, that is, a post that has already been hit on social media, or a post that cites it.

  In general, once a posting is made on news that many people want to send their opinions, the number of posted posts increases significantly. For this reason, it is difficult for the user to monitor the news posting information in real time and find a non-existing posting that is the first news posting information. Here, the non-existing post means a post that has not appeared in the past in social media but has appeared for the first time.

  Similarly, when the user wants to obtain the news-related posting information of the opinion and the follow-up news, the non-existing posting that is the news-related posting information of the first report is not required.

  In this way, in order for a user to obtain a non-existing post that is the first news posting information, after the news posting information is automatically extracted from the social media information, the news posting information is already displayed. It is desirable to automatically classify whether the posting is a post or a non-existing post. The same applies to the case where the user acquires an existing post which is an opinion on news and subsequent news posting information.

  Already posted posts include posts whose citation source is specified and can be easily classified by filtering using the name as a keyword. In addition, the posted posts are classified as having been published by the style of the human eye, but include posts that are difficult to classify by simple processing such as keyword filtering. In social media, even if it is a citation, it is possible to easily change the sentence, and the source may be deleted by the citation.

  For this reason, when the user obtains the first news posting information, classification is not performed by simple keyword filtering, but by machine learning that considers various sentence conditions. It is desirable that

  Here, from the viewpoint of machine learning, the process of extracting news posting information from social media information and the news posting information are classified into non-existing posts that are the first information and other postings that are other information Assume that processing is performed in a series of processing. The former process is called an extraction process, and the latter process is called a classification process.

  In this series of processing, not only the news posting information extracted by the extraction processing, but also information that should be discarded by the extraction processing (information that is not news posting information, that is, information that does not need to be classified) is classified. Will be done. For this reason, when machine learning is applied to a series of processes, the learning data increases, the processing load increases, and the classification accuracy decreases.

  In addition, when changing the classification configuration to perform classification processing other than posted postings and non-posted postings, the entire system that performs a series of processing is rebuilt and re-created including large-scale learning of extraction. Learning is necessary.

  As described above, when the extraction process and the classification process are performed in a series of processes, it is necessary to perform the entire learning again every time the classification process according to the user's request is increased, which is not realistic.

  On the other hand, when a learning model is generated by machine learning using linguistic information as input data, there are different advantages and disadvantages depending on whether a character sequence is used as input data or a word sequence as follows.

  When character sequences are used as input data, because characters have fewer variations than words, it is easier to assign each character to a node in the learning model input layer with fewer nodes, and fewer unknown words that are not input as nodes constitute a sentence. Can be handled as input data accurately, and the determination accuracy using a learning model is improved. Comparing characters and words, the word is composed of a plurality of characters and has various combinations, so the word has more unknown words than characters. For this reason, in the case of a word sequence, it is difficult to assign each word to a node in the learning model input layer. On the other hand, since there are no combinations of characters, characters have fewer unknown words than words.

  On the other hand, characters have more meaning than a word, and there is less information for the whole sentence, and it is difficult to specify the meaning, so the same judgment is made even if the meaning of the letters is different for the whole sentence There is a disadvantage that a learning model for performing the above is generated and the determination accuracy is lowered. This is because the meaning of a single character is ambiguous, and it is difficult to distinguish the meaning of the character used in one sentence from the meaning of the character used in the other sentence, even if the same character is used. Is assigned to one node in the input layer of the learning model.

  On the other hand, when a word sequence is used as input data, the meaning of a word is limited to one element rather than a character, and it is easy to specify the meaning. Therefore, a learning model that makes a different determination when the word is different is generated. There is an advantage that the determination accuracy becomes high.

  On the other hand, since words have more variations than letters and the number of unknown words increases when the number of words used is limited (or the number of nodes used is limited), each word is assigned to a node in the input layer of the learning model. This makes it difficult to handle words that make up a sentence accurately as input data, resulting in lower accuracy of determination using a learning model.

  In addition, when using a language such as Japanese that does not leave a space between words, the performance of the sequence generator that converts sentences into word sequences is greatly affected. Does not exist, the learning and determination accuracy is low.

  For this reason, it is desirable to perform machine learning that takes advantage of both the advantages of character sequences and the advantages of word sequences. In the above Non-Patent Documents 3-6, machine learning is performed using a character sequence and a word sequence as input data, but either one complements the other, and the character sequence and the word sequence are treated as one sequence. Therefore, it is not a machine learning that fully utilizes each advantage.

  Therefore, the present invention has been made to solve the above-mentioned problems, and its purpose is to provide a news posting extracted from social media information when a user acquires desired information from a large number of social media information. An object is to provide a news material classification device, a program, and a learning model capable of classifying information with high accuracy.

  In order to solve the above-described problem, the news material classification device according to claim 1 inputs post information that can be a news material among a large number of social media information as news post information, and uses the news post information as a user's post information. In the news material classification device for classifying on demand, the news posting information is input, the characters and words included in the news posting information are respectively extracted, and a character vector composed of the character series is generated. A sequence generation unit that generates a word vector composed of the sequence of words, a learning model storage unit that stores a machine-learned learning model, and reads out the learning model from the learning model storage unit, and uses the learning model, Based on the character vector and the word vector generated by the series generation unit, A classification unit that generates and outputs classification information according to the user's request for information, and the learning model uses the character vector of the news posting information as input data, and calculates a vector of an operation result. Character NN (neural network) to be output data, word NN having the word vector of the news posting information as input data and a vector of the operation result as output data, vector of the operation result of the character NN and the word It is comprised by the output NN which uses the vector which combined the vector of the calculation result of NN as input data, and uses the said classification information which is a calculation result as output data, The said classification | category part uses the said character NN, and is the said news contribution NN is calculated based on the character vector of the information, and the word NN is used to calculate the news posting information. An NN operation is performed based on a word vector, and an NN operation is performed based on a vector obtained by combining the operation result vector of the character NN and the operation result vector of the word NN using the output NN. Is output as the classification information.

  Further, the news material classification apparatus according to claim 2 is the news material classification apparatus according to claim 1, wherein the classification information is an already-displayed / not-displayed information indicating whether the news posting information has been published or has not been published. It is characterized as non-existing information.

  Further, the news material classification device according to claim 3 is the news material classification device according to claim 2, further comprising a second classification unit and a determination unit, wherein the classification unit stores the already-present / non-existing information as first information. The classification result is output, and the second classification unit inputs the news posting information, extracts agent information added to the news posting information for identifying the posting source device, and the agent information The generated / non-exposed information is generated and output as the second classification result, and the determination unit is output by the first classification result output by the classification unit and the second classification unit. Based on the second classification result, new appearance / non-occurrence information is generated and output.

  The news material classification apparatus according to claim 4 is the news material classification apparatus according to claim 2 or 3, wherein the series generation unit inputs the news posting information, and characters included in the news posting information. And one-hot column vector corresponding to the character are arranged to generate a character one-hot vector sequence, and one-hot column vector corresponding to the word is arranged to generate a word one-hot vector sequence. The learning model uses the character one-hot vector sequence generated by the sequence generation unit as input data, and the FFNN (feed forward neural network) calculation result vector as output data, FFNN for character input layer, The vector of the calculation result of the character input layer FFNN is used as input data, and the RNN (Recurrent Neural Network) FFNN for word input layer using as input data the character RNN having the vector of the calculation result as output data, the word one-hot vector sequence generated by the sequence generation unit as the output data, and the vector of the calculation result of FFNN as the output data, The word RNN having the vector of the FFNN calculation result for the word input layer as input data and the vector of the calculation result of the RNN, the vector of the calculation result of the character RNN, and the vector of the calculation result of the word RNN are combined. An intermediate layer FFNN having a vector as input data and an FFNN calculation result vector as output data, and an intermediate layer FFNN calculation result vector as input data, and the above-mentioned and non-existing FFNN calculation results. It is composed of FFNN for output layer whose information is output data, and the classification unit is for the character input layer Based on the character one-hot vector sequence generated by the sequence generation unit using FNN, FFNN is calculated, and using the character RNN, based on the calculation result vector of the character input layer FFNN, RNN calculation is performed, FFNN calculation is performed based on the word one-hot vector sequence generated by the sequence generation unit using the word input layer FFNN, and the word input is performed using the word RNN. RNN calculation is performed based on the vector of the calculation result of the layer FFNN, and based on a vector obtained by combining the vector of the calculation result of the character RNN and the vector of the calculation result of the word RNN using the intermediate layer FFNN. FFNN is calculated, and the output layer FFNN is used to calculate the FFNN based on the calculation result vector of the intermediate layer FFNN. Performed, and outputs the operation result as the foregoing, non foregoing information, characterized in that.

  The news material classification apparatus according to claim 5 is the news material classification apparatus according to claim 2 or 3, wherein the series generation unit inputs the news posting information, and characters included in the news posting information. And one-hot column vector corresponding to the character are arranged to generate a character one-hot vector sequence, and one-hot column vector corresponding to the word is arranged to generate a word one-hot vector sequence. The learning model uses the character one-hot vector sequence generated by the sequence generation unit as input data, and the character input layer FFNN using the vector of the FFNN calculation result as output data, the character input layer FFNN The character CNN having the vector of the result of the operation as input data and the vector of the result of the operation of the CNN (convolutional neural network) as output data The character pooling layer having the vector of the calculation result of the character CNN as input data, the vector of the pooling calculation result as output data, and the word one-hot vector sequence generated by the sequence generation unit as input data, and FFNN FFNN for the word input layer using the vector of the result of the calculation of the above as output data, the vector of the calculation result of the FFNN for the word input layer as the input data, the word CNN using the vector of the calculation result of the CNN as the output data, and the word CNN A word pooling layer having an operation result vector as input data and a pooling operation result vector as output data, and a vector obtained by combining the character pooling layer operation result vector and the word pooling layer operation result vector. Use as input data, and use FFNN operation result vector as output data The classification unit includes the FFNN for the intermediate layer and the FFNN for the output layer using the vector of the calculation result of the intermediate layer FFNN as input data and the output / non-existing information that is the calculation result of the FFNN as output data. Performs an FFNN operation based on the character one-hot vector sequence generated by the sequence generation unit using the character input layer FFNN, and uses the character CNN to calculate the character input layer FFNN. The CNN is calculated based on the calculation result vector, the pooling calculation is performed based on the character CNN calculation result vector using the character pooling layer, and the word input layer FFNN is used to calculate the sequence. Based on the word one-hot vector sequence generated by the generation unit, the FFNN is calculated, and the word CNN is used for the word input layer Based on the vector of the calculation result of FNN, CNN is calculated, using the word pooling layer, calculating the pooling based on the vector of the calculation result of the word CNN, and using the intermediate layer FFNN, Based on a vector obtained by combining the vector of the result of the character pooling layer and the vector of the result of the word pooling layer, the FFNN is calculated and the output FFNN is used to calculate the result of the calculation of the intermediate FFNN. An FFNN operation is performed based on the vector, and the operation result is output as the previous / non-existing information.

  The news material classification apparatus according to claim 6 is the news material classification apparatus according to claim 2 or 3, wherein the series generation unit inputs the news posting information, and characters included in the news posting information. And a word are extracted, a character BOW vector corresponding to all the extracted characters is generated, a word BOW vector corresponding to all the extracted words is generated, and the learning model is generated by the sequence generation unit The input character BOW vector is used as input data, the character input layer FFNN using the FFNN operation result vector as output data, the word BOW vector generated by the sequence generation unit as input data, and the FFNN operation result FFNN for word input layer using vector as output data, vector of calculation result of FFNN for character input layer and word input The input data is a vector obtained by combining the vector of the operation result of FFNN for the input data, the FFNN for the intermediate layer using the vector of the operation result of the FFNN as the output data, and the vector of the operation result of the FFNN for the intermediate layer, The character generated by the sequence generation unit using the character input layer FFNN, the FFNN for the output layer having the output / non-existing information that is the calculation result of the FFNN as output data. FFNN is calculated based on the BOW vector, FFNN is calculated based on the word BOW vector generated by the sequence generator using the word input layer FFNN, and the intermediate layer FFNN is used. The vector of the calculation result of the character input layer FFNN and the vector of the calculation result of the word input layer FFNN are combined. FFNN is calculated based on the vector, and FFNN is calculated based on the calculation result vector of the intermediate layer FFNN using the output layer FFNN, and the calculation result is output as the previous / non-existing information. It is characterized by that.

  The news material classification device according to claim 7 inputs the posting information that can be news material among a large number of social media information as news posting information, and classifies the news posting information as published or not published. In the classification device, the news posting information is input, characters or words included in the news posting information are extracted, and a character vector consisting of the character series or a word vector consisting of the word series is generated. A series generator and the character vector or the word vector of the news posting information as input data, and whether the news posting information as a calculation result has already been shown or not shown This is a learning model composed of NN that uses the existing information as output data. A learning model storage unit that stores a learning model, and reads out the learning model from the learning model storage unit, and uses the learning model to generate the previously-explained data based on the character vector or the word vector generated by the sequence generation unit. A classification unit that generates non-existing information and outputs it as a first classification result, and agent information for inputting the news posting information and identifying the posting source device added to the news posting information A second classifying unit that extracts and outputs the second / non-existing information based on the agent information, and outputs the second classification result; the first classification result output by the classification unit; and the second Based on the second classification result output by the classification unit, new previous / non-existing information is generated and output.

  Furthermore, a program according to claim 8 causes a computer to function as the news material classification device according to any one of claims 1 to 7.

  Further, the learning model of claim 9 is a learning model for causing a computer to function so as to classify news posting information extracted from a large number of social media information according to a user's request and output classification information. A character vector consisting of a character sequence extracted from the news posting information and a character NN using a calculation result vector as output data, and a word sequence extracted from the news posting information. A word NN having a word vector as input data and a vector of the operation result as output data, and a vector obtained by combining the vector of the operation result of the character NN and the vector of the operation result of the word NN are input data, and the operation result The character vector of the news posting information by the output NN using the classification information as output data And the word vector as the input data of the learning model, and the classification information of the news posting information as the output data of the learning model, the weight coefficient of the character NN, the weight coefficient of the word NN, and the output NN The NN operation is performed on the character vector input to the input layer of the character NN based on the weight coefficient of the character NN, and an operation result vector is output from the output layer. NN operation is performed on the word vector input to the input layer of the word NN based on the weight coefficient of the word NN, and a vector of the operation result is output from the output layer and input to the input layer of the output NN The operation result vector output from the character NN output layer and the operation result vector output from the word NN output layer are combined. To vector performs a calculation of NN based on the weighting factor of the output NN, to output the classification information is an arithmetic result from the output layer, and characterized by causing the computer to function.

  As described above, according to the present invention, when a user acquires desired information from a large number of social media information, it is possible to classify news posting information extracted from social media information with high accuracy. .

It is a figure explaining the outline | summary of the whole structural example containing the news material classification | category apparatus of Example 1. FIG. It is a block diagram which shows the structural example of the news material classification | category apparatus of Example 1. FIG. It is a figure explaining the outline | summary of the whole structural example containing the news material classification | category apparatus of Example 2. FIG. It is a block diagram which shows the structural example of the news material classification | category apparatus of Example 2. FIG. It is a flowchart which shows the process example of an agent classification | category part. It is a flowchart which shows the 1st process example (process example 1) of a determination part. It is a flowchart which shows the 2nd process example (process example 2) of a determination part. It is a block diagram which shows the structural example of a learning apparatus. It is a figure explaining the schematic structure of a learning model. It is a figure explaining the schematic structure of the learning model (specific example 1) using FFNN and LSTM. It is a figure explaining the number of nodes and input-output data in the specific example 1. It is a figure explaining the schematic structure of the learning model (specific example 2) using FFNN and CNN. It is a figure explaining the number of nodes and input-output data in the specific example 2. It is a figure explaining the schematic structure of the learning model (specific example 3) using FFNN. It is a figure explaining the number of nodes and input-output data in the specific example 3. FIG. It is a figure which shows the example of the character series produced | generated from news property posting information, and the example of a word series. It is a figure which shows the example of the character one-hot vector series { _xchar }. It is a figure which shows the example of the word one-hot vector series { _xword }. Is a diagram illustrating an example of a character BOW vector x _char. It is a figure which shows the example of the word BOW vector _xword . It is a figure explaining an experimental result. It is a figure explaining the outline | summary of the conventional news material extraction apparatus.

  Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings. In the present invention, apart from the process of extracting news posting information from social media information, the news posting information is divided into information according to the user's request (for example, non-existing information in the first report) and other information (for example, Information). The already-present information refers to information that has already appeared, and the non-existing information refers to information that has not appeared in the past but has appeared for the first time.

  The process when a user acquires desired information from a large number of social media information is divided into two stages. The first process is a process of extracting news posting information from social media information, and is realized using existing technology.

  The second process is according to the present invention, and is a process of classifying the news posting information extracted by the first process into information according to the user's request and other information by machine learning. .

  In the second processing, in the second process, a character vector and a word vector are generated from the news posting information, and the character vector and the word vector are input to the input layer as input data by using a learning model that has been previously machine-learned. The present invention combines the intermediate data of the character vector and the intermediate data of the word vector in the intermediate layer, determines information according to the user's request and other information, and uses the determination result as output data. Output from the output layer.

  Hereinafter, the present invention will be specifically described with reference to Examples 1 and 2. In the first and second embodiments, the user can acquire the first non-existing information from a large number of social media information. Example 1 uses a learning model to classify news posting information into non-existing information of the first report and other already-existing information based on the text (text) of news posting information and generate a determination result To do.

  In the second embodiment, in addition to the processing of the first embodiment, the news posting information is filtered by using the sender information other than the text of the news posting information and the agent information such as the OS, and the first report is not published. The information is classified into information and other previously-described information, and a determination result is generated. And Example 2 produces | generates a final determination result based on the determination result using the learning model by the process of Example 1, and the determination result by filtering.

  Hereinafter, the types of learning models used in the first and second embodiments will be described with specific examples 1, 2, and 3. Specific example 1 is an example using FFNN (FeedForward Neural Network) and LSTM (Long Short Term Memory) as learning models. Specific example 2 is an example using FFNN and CNN (Convolutional Neural Network), and specific example 3 is an example using FFNN.

[Example 1]
First, Example 1 will be described. As described above, the first embodiment classifies the news posting information into the first report non-existing information and the other existing information using the learning model based on the text of the news posting information, and determines the determination result. This is an example of generation.

  FIG. 1 is a diagram for explaining an overview of an overall configuration example including the news material classification apparatus according to the first embodiment. This system is for a user to acquire the first non-existing information from a large number of social media information, and includes a news material extracting device 100 and a news material classifying device 1-1.

  As described with reference to FIG. 22, the news material extraction apparatus 100 inputs social media information that is posting information, and extracts news posting information from the social media information. Then, the news material extraction device 100 outputs the news posting information extracted from the social media information to the news material classification device 1-1 and discards other information.

  The news material extraction apparatus 100 is realized by existing techniques such as the technique of Non-Patent Document 1 and the keyword matching technique described above using a neural network learning model. For example, the news material extraction apparatus 100 performs machine learning using learning data in which social media information with news properties actually extracted manually in a news report site is a positive example and social media information randomly sampled is a negative example. Generate a learning model. Machine learning is performed using a character one-hot vector sequence of social media information as input data and using, for example, Bi-directional LSTM. The news material extracting apparatus 100 automatically extracts news posting information from social media information using a machine-learned learning model.

  The news material classification device 1-1 receives the news posting information from the news material extraction device 100, and classifies the news posting information into the first non-existing information and the other existing information using the learning model. To do.

  The news material classifying apparatus 1-1 sets the classified first non-existing information as non-existing labeled news posting information, and sets the other classified information as the leveled news posting information already described. Then, the news material classification device 1-1 outputs label information indicating the appearance or non-existing for the news posting information as the appearance / non-existing information. Details of the news material classification device 1-1 will be described later.

  FIG. 2 is a block diagram illustrating a configuration example of the news material classification device 1-1 according to the first embodiment illustrated in FIG. The news material classification device 1-1 includes a sequence generation unit 10, a sentence classification unit 11, and a learning model storage unit 12.

  The sequence generation unit 10 inputs the news posting information, extracts characters included in the news posting information as a character sequence based on the news posting information, and sets the words included in the news posting information as a word sequence. Extract. Then, the sequence generation unit 10 generates a character vector composed of the character sequence and also generates a word vector composed of the word sequence. Then, the sequence generation unit 10 outputs the character vector and the word vector of the news posting information to the text classification unit 11.

  For example, the sequence generation unit 10 extracts characters included in the news posting information, arranges one-hot column vectors corresponding to the characters, and generates a character one-hot vector sequence. Further, the sequence generation unit 10 extracts words included in the news posting information, arranges one-hot sequence vectors corresponding to the words, and generates a word one-hot vector sequence. Further, for example, the sequence generation unit 10 generates a character BOW (Bags Of Words) vector composed of the extracted character sequence, and generates a word BOW vector composed of the extracted word sequence. Details of the character one-hot vector series, the word one-hot vector series, the character BOW vector, and the word BOW vector will be described later.

  The text classification unit 11 inputs a character vector and a word vector of news posting information from the sequence generation unit 10. Then, the sentence classification unit 11 uses the learning model stored in the learning model storage unit 12 to generate label information indicating the appearance or non-occurrence based on the character vector and the word vector, and outputs the label information indicating the appearance / non-appearance information. Output as. This published / not-published information is information indicating whether the news posting information input by the sequence generation unit 10 has been published or has not been published.

  The learning model storage unit 12 stores a learning model machine-learned by the learning device 2 shown in FIG.

  FIG. 9 is a diagram illustrating a schematic structure of a learning model stored in the learning model storage unit 12 illustrated in FIG. This learning model includes a character NN (Neural Network) 21, a word NN22, and an output NN23, and the character NN21 and the word NN22 so that the output data of the character NN21 and the output data of the word NN22 are input to the output NN23. And the output NN23 are combined.

  This learning model is generated by machine learning of the learning device 2 shown in FIG. 8 to be described later, and the character vector and the word vector of the news posting information are used as input data. This is for causing the computer to function so as to output the appearing / not appearing information indicating whether it is present or not appearing. In addition, this learning model holds the weight coefficient of machine-learned character NN21, the weight coefficient of word NN22, and the weight coefficient of output NN23.

  The sentence classification unit 11 reads the learning model shown in FIG. 9 stored in the learning model storage unit 12 and stores it in a memory not shown in FIG. The sentence classification unit 11 inputs the character vector generated by the sequence generation unit 10 to the input layer of the character NN21, and adds a machine-learned weighting factor to the value of each element of the character vector. I do. Then, the sentence classification unit 11 outputs the vector of the calculation result from the output layer of the character NN21 and propagates it to the output NN23.

  In addition, the sentence classification unit 11 inputs the word vector generated by the sequence generation unit 10 to the input layer of the word NN22, and adds a machine-learned weighting factor to the value of each element of the word vector. Perform the operation. Then, the sentence classification unit 11 outputs the vector of the calculation result from the output layer of the word NN22 and propagates it to the output NN23.

  In this case, the calculation result vector output from the character NN21 and the calculation result vector output from the word NN22 are combined, and the combined vector is propagated to the output NN23.

  The text classification unit 11 inputs a calculation result vector from the character NN21 to the input layer of the output NN23, and inputs a calculation result vector from the word NN22 to the input layer of the output NN23. Then, the text classification unit 11 performs an NN calculation such as adding a machine-learned weighting factor to the value of each element of the input vector. The sentence classification unit 11 calculates a probability value using the softmax function for the calculation result, calculates existing or non-existing binary information using the argmax function for the probability value, and outputs the binary information. Output from the output layer of the output NN 23 as non-existing information.

  Note that the sentence classification unit 11 calculates an activation function RELU (REctified Linear Unit) when outputting calculation results from each layer of the character NN21, the word NN22, and the output NN23. The same applies to the learning models shown in FIGS.

  As described above, according to the news material classification device 1-1 of the first embodiment, the sequence generation unit 10 generates a character vector and a word vector from the news posting information extracted by the news material extraction device 100. Using the learning model, the sentence classification unit 11 performs an operation using the character vector and the word vector of the news posting information as input data and the existing / non-existing information as output data to generate the existing / non-existing information. Output. In this way, the news posting information is classified as non-existing information or existing information.

  The learning model combines the character NN21 having the character vector of the news posting information as input data, the word NN22 having the word vector of the news posting information as input data, and the calculation result of the character NN21 and the calculation result of the word NN22. The output NN 23 having the generated vector as input data and the output / non-existing information as output data.

  As a result, the text classification unit 11 performs independent calculation of the character vector and the word vector, so that it is possible to make a determination utilizing the superiority of the character information and the word information of the news posting information, and the high accuracy Non-existing information can be obtained. The superiority in this case is that, as described above, for character information, for example, when the number of nodes in the input layer of character NN21 is the same as the number of nodes in the input layer of word NN22, the number of unknown words is less than that of word information. In terms of word information, the meaning that an element can have is limited compared to character information.

  That is, the learning model used by the news material classification device 1-1 is a model that is machine-learned so as to reflect the advantages of characters with few unknown words and the advantages of words that have a limited meaning that one element can have. Therefore, a determination that the superiority of characters and words is reflected from the viewpoint of the sentence is performed on the news posting information, and as a result, the high-precision existing / non-existing information can be obtained.

  Thus, when the user acquires the first non-existing information or the pre-existing information from a large number of social media information, by using the news material classification device 1-1, the preexisting information and the non-existing information are accurately obtained. High classification is possible.

  Also, the news material classification device 1-1 of the first embodiment operates uncorrelated independently of the news material extraction device 100. For this reason, the learning model of the news material classification device 1-1 does not need to use a large number of social media information input by the news material extraction device 100, and is smaller than the number of social media information input by the news material extraction device 100. News posting information may be used.

  Therefore, since less learning data is required, the burden of hardware for machine learning of the learning model is reduced, and it is possible to perform complicated processing even with limited computer resources.

  Furthermore, since the news material classification device 1-1 according to the first embodiment operates independently of the news material extraction device 100, the extraction processing by the news material extraction device 100 and the classification processing by the news material classification device 1-1 are described. Is almost uncorrelated. Therefore, the contents of the classification process of the news material classification apparatus 1-1 can be changed without much concern about the relationship with the extraction process by the news material extraction apparatus 100, and the work for changing the classification process is easy. It becomes.

  In addition, according to the experiment by the inventors of the present application, when the correct ratio of the existing information and the non-existing information is set to 20% and 80%, in the simple conventional filtering process, the existing information and The result of classification into non-existing information was obtained. That is, when 20% of the previously published information is extracted from the news posting information, 34% of the original previously posted information can be acquired with an accuracy of 94.1%. On the other hand, in the news material classification apparatus 1-1 of Example 1, the result of classifying into the existing information and the non-existing information with an accuracy of about 80% or more was obtained. That is, when 20% of the previously published information is to be extracted from the news posting information, 80% of the original previously posted information can be acquired with 90% accuracy. Details will be described later with reference to FIG.

[Example 2]
Next, Example 2 will be described. As described above, in the second embodiment, a determination result using a learning model is generated by the processing of the first embodiment, and news posting information is converted into non-existing information in the first report by filtering using agent information. This is an example in which a classification result is generated and a determination result is generated, and a final determination result is generated based on a determination result using a learning model and a determination result by filtering.

  In general, not only the text but also agent information such as the application of the posting or the OS, and the posting time information are added to the social media posting, and the user can use the information. There is a case.

  Posts related to news transmitted from portable communication devices such as portable terminals are likely to be non-existing information in the first report. On the other hand, posts sent from a stationary server or personal computer that is not a portable communication device such as a portable terminal are automatically posted from the news bureau or homepage, or something posted on an indoor personal computer. Many are posted with information. The post information in this case is unlikely to be the first non-existing information, and is likely to be the pre-existing information. Although the type of agent information is enormous, it is possible to determine whether it has been posted or not by using only information regarding whether or not the posting is from a mobile phone.

  Therefore, in the second embodiment, non-existing information and existing information are determined by filtering using agent information, and a final determination result is generated based on the determination result using the learning model and the determination result by filtering. I tried to do it.

  FIG. 3 is a diagram for explaining an overview of an overall configuration example including the news material classification device according to the second embodiment. This system is for a user to acquire first non-existing information from a large number of social media information, and includes a news material extraction device 100 and a news material classification device 1-2.

  The news material extracting apparatus 100 is the same as the news material extracting apparatus 100 shown in FIG.

  The news material classification device 1-2 receives the news posting information from the news material extraction device 100, and classifies the news posting information into non-existing information and existing information based on the text using a learning model. Moreover, the news material classification device 1-2 classifies the news posting information into non-existing information and existing information based on the agent information by filtering using the agent information. Then, the news material classification device 1-2 generates a final determination result based on the non-existing information and the preexisting information classified based on the text, and the non-existing information and the preexisting information classified based on the agent information. .

  As a final determination result, the news material classification device 1-2 outputs label information indicating the appearance or non-existence of the news posting information as the appearance / non-appearance information. Details of the news material classification device 1-2 will be described later.

  FIG. 4 is a block diagram illustrating a configuration example of the news material classification apparatus 1-2 according to the second embodiment illustrated in FIG. The news material classification apparatus 1-2 includes a sequence generation unit 10, a sentence classification unit 11, a learning model storage unit 12, an agent classification unit 13, and a determination unit 14.

  The sequence generation unit 10, the sentence classification unit 11, and the learning model storage unit 12 are the same as the sequence generation unit 10, the sentence classification unit 11, and the learning model storage unit 12 illustrated in FIG. The sentence classification unit 11 outputs the already-present / non-existing information a (first determination result) to the determination unit 14.

  The agent classification unit 13 inputs the news posting information from the news material extracting apparatus 100, and based on the agent information added to the news posting information, whether the news posting information has been published or has not been published. And the label information is generated. Then, the agent classification unit 13 outputs the generated label information to the determination unit 14 as the previous / non-existing information b (second determination result).

  The agent information is information for identifying the posting source device, and includes, for example, the type of the posting source device, information such as an application or OS used for posting in the posting source device.

  FIG. 5 is a flowchart illustrating a processing example of the agent classification unit 13. The agent classification unit 13 inputs news posting information from the news material extraction device 100 (step S501), and extracts agent information from the news posting information (step S502). Agent information may be added to the news posting information, and the agent classification unit 13 can extract the agent information added to the news posting information from the news posting information.

  Based on the agent information, the agent classification unit 13 determines whether the news posting information has been published or has not been published (step S503), and generates the label information. The agent classification unit 13 outputs the label information to the determination unit 14 as the previous / non-existing information b (step S504).

  Specifically, the agent classification unit 13 determines that the posting source is the mobile terminal when the agent information is the device name of the mobile terminal, the name of the application or OS used in the mobile terminal, and the news property It is determined that the posted information is not published. On the other hand, when the agent information is not the device name of the mobile terminal, the application used on the mobile terminal, or the name of the OS, the agent classification unit 13 (the device name of the installed computer, the application used on the installed computer, and If it is the name of the OS), it is determined that the posting source is not a mobile terminal, and it is determined that the news posting information has already been issued.

  For example, when the agent information is “Twitter for iPhone (registered trademark)” or “Twitter for Android (registered trademark)”, the agent classifying unit 13 determines that the information is not yet published. On the other hand, when the agent information is other than “Twitter for iPhone (registered trademark)” and “Twitter for Android (registered trademark)”, the agent classification unit 13 determines that the agent information has already been issued.

  Returning to FIG. 4, the determination unit 14 inputs the already-noted / not-to-present information a from the sentence classification unit 11, and the already-not-to-not-present information b from the agent classification unit 13, and the previously-not-to-not-present information a, b Based on the above, it generates and outputs the previous / non-existing information.

  FIG. 6 is a flowchart illustrating a first processing example (processing example 1) of the determination unit 14. The determination unit 14 inputs the already-present / non-existing information a and b from the sentence classifying unit 11 and the agent classifying unit 13 (step S601).

  The determination unit 14 determines whether or not the previous / non-existing information a indicates the previous appearance (step S602). The determination unit 14 proceeds to step S603 when the previous / non-existing information “a” indicates the previous appearance (step S602: Y), and when the existing / non-existing information “a” does not indicate the previous appearance (shows the non-existing state) (Step S602: N), the process proceeds to Step S605.

  The determination unit 14 shifts from step S602 (Y) to determine whether or not the previous / non-existing information b indicates the previous (step S603). When the previous / non-previous information b indicates the previous appearance (step S603: Y), the determination unit 14 determines the previous appearance and generates label information indicating the previous appearance (step S604). On the other hand, the determination part 14 transfers to step S605, when the previous / non-existing information b does not indicate the previous (when it indicates the non-existing) (step S603: N).

  The determination unit 14 proceeds from step S602 (N) or step S603 (N), determines non-existing, and generates label information indicating non-existing (step S605). The determination unit 14 proceeds from step S604 or step S605, and outputs the label information as the already-present / non-existing information (step S606).

  As described above, in the processing example 1 shown in FIG. 6, the determination unit 14 outputs the previous / non-existing information indicating the existing when both the existing / non-existing information a and b indicate the existing, In cases other than, information on the appearance / non-appearance that indicates non-appearance is output.

  FIG. 7 is a flowchart illustrating a second processing example (processing example 2) of the determination unit 14. Steps S701 and S706 shown in FIG. 7 are the same as steps S601 and S606 shown in FIG.

  The determination unit 14 determines whether or not the previous / non-existing information a indicates the previous appearance (step S702). The determination unit 14 proceeds to step S704 when the previous / non-existing information “a” indicates the previous appearance (step S702: Y), and when the existing / non-existing information “a” does not indicate the previous appearance (shows the non-existing state) (Step S702: N), the process proceeds to Step S703.

  The determination unit 14 shifts from step S702 (N) to determine whether or not the previous / non-existing information b indicates the previous (step S703). The determination unit 14 proceeds to step S704 when the previous / non-existing information b indicates the previous (step S703: Y). On the other hand, the determination part 14 transfers to step S705, when the previous / non-existing information b does not indicate the previous (when it indicates the non-existing) (step S703: N).

  The determination unit 14 proceeds from step S702 (Y) or step S703 (Y), determines the appearance, and generates label information indicating the appearance (step S704). The determination unit 14 proceeds from step S703 (N), determines non-existing, and generates label information indicating non-existing (step S705).

  In this way, in the processing example 2 shown in FIG. 7, the determination unit 14 outputs the appearing / non-appearing information indicating the appearance when one of the appearing / non-appearing information a and b indicates the appearance. In other cases, information on the appearance / non-appearance indicating non-appearance is output.

  As described above, according to the news material classification device 1-2 of the second embodiment, the sequence generation unit 10 generates a character vector and a word vector from the news posting information extracted by the news material extraction device 100. Using the learning model, the sentence classification unit 11 performs an operation using the character vector and the word vector of the news posting information as input data and the existing / non-existing information as output data, and generates the existing / non-existing information a. . The agent classification unit 13 generates the already-not-present / not-to-not-present information b based on the agent information included in the news posting information. The determination unit 14 generates the appearing / non-appearing information for the news posting information based on the appearing / non-appearing information a and b.

  Thereby, there exists an effect similar to Example 1. In particular, based on the text of the news posting information, in addition to the determination utilizing the superiority of the character information and the word information, the determination based on the agent information included in the news posting information is performed, so that more accurate Non-existing information can be obtained.

  Thus, when the user acquires the first non-existing information or the pre-existing information from a large number of social media information, the news material classification device 1-2 is used to further combine the pre-existing information and the non-existing information. It becomes possible to classify with high accuracy.

[Modification of Example 2]
Next, a modification of the second embodiment will be described. The modification of the second embodiment uses a learning model machine-learned using only character vectors or only word vectors as input data, instead of using a machine model learning using character vectors and word vectors as input data. . That is, the modified example of the second embodiment generates a determination result using a learning model machine-learned only with a character vector or only a word vector, generates a determination result by filtering using agent information, This is an example of generating a final determination result based on the determination result used and the determination result by filtering.

  The news material classification device 1-2 in the modification of the second embodiment includes a sequence generation unit 10, a sentence classification unit 11, and a learning model storage unit 12 different from the second embodiment in the configuration of FIG. The same agent classification unit 13 and determination unit 14 are provided.

  The series generation unit 10 according to the modification of the second embodiment inputs news posting information and extracts characters included in the news posting information as a character series based on the news posting information or news posting information. The words included in are extracted as a word series. Then, the sequence generation unit 10 generates a character vector composed of a character sequence or a word vector composed of a word sequence, and outputs the character vector or the word vector to the text classification unit 11.

  The sentence classification unit 11 inputs a character vector or a word vector from the sequence generation unit 10, and uses the learning model stored in the learning model storage unit 12 to obtain the previous / non-existing information a based on the character vector or the word vector. This is generated and output to the determination unit 14.

  The learning model storage unit 12 stores a conventional learning model machine-learned by a conventional learning device. This learning model is composed of, for example, a character NN and an output NN, or is composed of a word NN and an output NN.

  As described above, according to the news material classification device 1-2 of the modified example of the second embodiment, the sentence classification unit 11 uses the learning model as a character vector or a word vector of news posting information, as input data, A calculation is performed using the previously-noted / not-present information as output data to generate the previously-not-present information a. The agent classification unit 13 generates the already-not-present / not-to-not-present information b based on the agent information included in the news posting information. The determination unit 14 generates the appearing / non-appearing information for the news posting information based on the appearing / non-appearing information a and b.

  As a result, in addition to the determination based on the text of the news posting information, the determination based on the agent information included in the news posting information is performed. Non-existing information can be obtained.

  As described above, when the user acquires the first non-existing information or the preexisting information from a large number of social media information, by using the news material classification device 1-2 according to the modification of the second embodiment, the preexisting information is obtained. And non-existing information can be classified with high accuracy.

[Learning device]
Next, a learning apparatus that performs machine learning of the learning model stored in the learning model storage unit 12 illustrated in FIGS. 2 and 4 will be described.

  FIG. 8 is a block diagram illustrating a configuration example of the learning device. The learning device 2 includes a sequence generation unit 10, a learning unit 20, and a learning model storage unit 12.

  Since the sequence generation unit 10 and the learning model storage unit 12 are the same as the sequence generation unit 10 and the learning model storage unit 12 illustrated in FIGS. 2 and 4, description thereof is omitted here. The learning model stored in the learning model storage unit 12 is generated by the learning unit 20.

  The learning unit 20 inputs the character vector and the word vector of the news posting information from the sequence generation unit 10 and inputs the appearing / non-appearing information indicating whether the news posting information has been published or has not been published. To do.

  The learning unit 20 uses the character vector and the word vector of the news posting information as input data, uses the previous / non-existing information as output data, and performs machine learning using these learning data. Then, the learning unit 20 generates the learning model illustrated in FIG. 9 and stores the learning model in the learning model storage unit 12. For example, the learning unit 20 generates a learning model shown in FIGS. Learning by the learning unit 20 is performed by, for example, a back-propagation method, and a weighting coefficient between nodes is updated every time machine learning is performed.

  The learning data of the already-present / not-present information is generated by a human annotator. The annotator determines whether or not a news article is included in the news posting information extracted by the news material extracting apparatus 100 shown in FIGS. When the annotator determines that the news posting information includes the citation of the news article, the annotator performs the labeling of the news posting information and generates the appearing / non-appearing information indicating the appearing. On the other hand, if the annotator determines that the news posting information does not include the citation of the news article, the annotator performs non-existing labeling on the news posting information and generates the existing / non-exposed information indicating the non-existing .

  In this way, the learning device 2 generates, for example, a learning model shown in FIGS. 10 to 15 described later, and this learning model is stored in the learning model storage unit 12 shown in FIG. 2 or FIG.

[Learning model]
Next, the news material classification apparatus 1-1 of the first embodiment shown in FIGS. 1 and 2 or the news material of the second embodiment shown in FIGS. 3 and 4 generated by the learning device 2 shown in FIG. A learning model used for the classification device 1-2 will be described.

  The learning model generated by the learning device 2 uses the character vector and the word vector of the news posting information as input data, and determines whether the news posting information has already been given or not given to the input data. This is for causing the computer to function so as to output the displayed / not-shown information, and holds the machine-learned weighting coefficient. The sentence classification unit 11 shown in FIGS. 2 and 4 functions as an arithmetic unit that reads out the learning model stored in the learning model storage unit 12 and stores it in the memory, and performs an operation using the learning model.

  Specific example 1 is a learning model using FFNN and LSTM, specific example 2 is a learning model using FFNN and CNN, and specific example 3 is a learning model using FFNN. These specific examples 1 to 3 specifically show the learning model shown in FIG.

(Specific example 1)
First, the learning model of specific example 1 will be described. FIG. 10 is a diagram illustrating a schematic structure of a learning model (specific example 1) using FFNN and LSTM, and FIG. 11 is a diagram illustrating the number of nodes and input / output data in specific example 1.

  The learning model includes a character input layer FFNN 24, a word input layer FFNN 25, a character LSTM 26, a word LSTM 27, an intermediate layer FFNN 28, and an output layer FFNN 29. This learning model also combines the character input layer FFNN 24 and the character LSTM 26 so that the output data of the character input layer FFNN 24 is input to the character LSTM 26, and the output data of the word input layer FFNN 25 to the word LSTM 27. The character LSTM26, the word LSTM27, and the intermediate layer are input so that the word input layer FFNN25 and the word LSTM27 are combined and the output data of the character LSTM26 and the output data of the word LSTM27 are input to the intermediate layer FFNN28. The intermediate layer FFNN 28 and the output layer FFNN 29 are combined so that the output layer FFNN 28 is coupled to the output layer FFNN 28 and the output data of the intermediate layer FFNN 28 is input to the output layer FFNN 29.

This learning model uses each of a character one-hot vector sequence {x _char } and a word one-hot vector sequence {x _word } as input data, and combines both in the intermediate layer to obtain binary appearing / not appearing information. Is a neural network whose output data is. Further, this learning model holds the weighting factors of the character input layer FFNN24 to the output layer FFNN29 machine-learned by the learning device 2 shown in FIG.

The sentence classification unit 11 reads the learning model shown in FIG. 10 stored in the learning model storage unit 12. The text classification unit 11 inputs the character one-hot vector sequence {x _char } generated by the sequence generation unit 10 into the character input layer FFNN 24 and sets the value of each element of the character one-hot vector sequence {x _char }. On the other hand, an FFNN calculation such as adding a weighting coefficient learned by machine learning is performed. Then, the text classification unit 11 outputs the vector sequence {h _char ⁱⁿ } of the operation result from the character input layer FFNN 24 and propagates it to the character LSTM 26. In addition, since the calculation of FFNN is known, description is abbreviate | omitted here.

  FIG. 16 is a diagram illustrating an example of a character sequence and a word sequence generated from news posting information. When the news posting information is “November 15 near Shinjuku Station West Exit”, the series generation unit 10 causes the news generation posting information to generate a character sequence “New, Inn, Station, West, Exit, Tsuke, Near, Thus, “fire, thing, 1, 1, month, 1, 5, day” is extracted. This character sequence is composed of the characters “new”, “hotel”, and “day”.

  The series generation unit 10 extracts the word series “Shinjuku station west exit, near, fire, November 15” from the news posting information. This word sequence is made up of the words “Shinjuku Station West Exit”, “Nearby”... “November 15”.

FIG. 17 is a diagram illustrating an example of a character one-hot vector sequence {x _char }. When the character sequence is “new, inn, station, west, mouth, date, near, fire, thing, 1, 1, month, 1, 5, day” shown in FIG. A character one-hot vector sequence {x _char } of 6300 rows × 150 columns shown in FIG. 17 is generated. The number of columns 150 of this character one-hot vector sequence {x _char } corresponds to the maximum value 150 of the number of characters in a post. The number of lines 6300 corresponds to the maximum character type used in the first and second embodiments, and is set in advance by the designer of the learning model in consideration of the frequency of characters appearing as learning data.

The column data of the character one-hot vector sequence {x _char } corresponds to each character “new”, “hotel”... “Day” constituting the character sequence, and corresponds to the character in 6300 row positions. “1” is set in one row position, and “0” is set in the other. If the character series is less than 150 characters, “0”, which is null data, is set in the remaining column data.

Referring to FIG. 11, in the examples of FIGS. 16 and 17, the sentence classification unit 11, when the news posting information is “November 15 near the west exit of Shinjuku Station”, 6300 lines × 150 columns of characters. The one-hot vector series {x _char } is input to the character input layer FFNN 24 and FFNN is calculated. In this case, the number of nodes in the input layer in the character input layer FFNN 24 is 6300.

Then, the sentence classification unit 11 generates a 200-row × 150-column vector sequence {h _char ⁱⁿ } as the FFNN calculation result, outputs this from the character input layer FFNN 24, and propagates it to the character LSTM 26. In this case, the number of nodes in the output layer in the character input layer FFNN 24 is 200.

Returning to FIG. 10, the text classification unit 11 inputs the vector sequence {h _char ⁱⁿ } of the calculation result from the character input layer FFNN 24 to the character LSTM 26, and for each element value of the vector sequence {h _char ⁱⁿ }, LSTM operations such as adding machine-learned weighting coefficients are performed. Then, the text classification unit 11 outputs the vector h _char ^LSTM of the operation result from the character LSTM ²⁶ , combines the vector h _char ^LSTM and a vector h _word ^LSTM described later, and propagates them to the intermediate layer FFNN 28. Since the LSTM calculation is known, the description thereof is omitted here.

Referring to FIG. 11, in the examples of FIGS. 16 and 17, the text classification unit 11 inputs a vector sequence {h _char ⁱⁿ } of 200 rows × 150 columns from the character input layer FFNN 24 to the character LSTM 26, and Perform the operation. In this case, the number of nodes in the input layer in the character LSTM26 is 200.

Then, the sentence classification unit 11 generates a vector h _char ^LSTM of 200 rows × 1 column as an LSTM calculation result, outputs this from the character LSTM 26, and propagates it to the intermediate layer FFNN 28. In this case, the number of nodes in the output layer in the character LSTM26 is 200.

Returning to FIG. 10, wording classification unit 11, the word one-hot vector sequence generated by sequence generation unit 10 {x _word} entered word input layer FFNN25, word one-hot vector sequence {x _word} FFNN operations such as adding machine-learned weighting coefficients to the values of the elements are performed. Then, the sentence classification unit 11 outputs the vector sequence {h _word ⁱⁿ } of the operation result from the word input layer FFNN 25 and propagates it to the word LSTM 27.

FIG. 18 is a diagram illustrating an example of a word one-hot vector sequence {x _word }. When the word sequence is “Shinjuku station west exit, near, fire, November 15” shown in FIG. 16, the sequence generation unit 10 causes the word one-hot vector sequence of 72000 rows × 150 columns shown in FIG. {X _word } is generated. The number of columns 150 of the word one-hot vector sequence {x _word } corresponds to the maximum value 150 of the number of characters in the post. The number of lines 72000 corresponds to the maximum value of the word type used in the first and second embodiments, and is set in advance by the learning model designer in consideration of the frequency of words appearing as learning data.

The column data of the word one-hot vector sequence {x _word } corresponds to the words “Shinjuku Station West Exit”, “Near”, and “November 15” constituting the word sequence, and the row position of 72000 Of these, “1” is set in one line position corresponding to the word, and “0” is set in the other lines. If the word series is less than 150 characters, “0”, which is null data, is set in the remaining column data.

Referring to FIG. 11, in the examples of FIGS. 16 and 18, the sentence classification unit 11 has a word of 72000 rows × 150 columns when the news posting information is “November 15 near Shinjuku Station West Exit”. The one-hot vector series {x _word } is input to the word input layer FFNN 25 and FFNN is calculated. In this case, the number of nodes in the input layer in the word input layer FFNN 25 is 72,000.

Then, the text classification unit 11 generates a 200-row × 150-column vector sequence {h _word ⁱⁿ } as the FFNN calculation result, outputs this from the word input layer FFNN 25, and propagates it to the word LSTM 27. In this case, the number of nodes in the output layer in the word input layer FFNN 25 is 200.

Returning to FIG. 10, the sentence classification unit 11 inputs the vector sequence {h _word ⁱⁿ } of the operation result from the word input layer FFNN 25 to the word LSTM 27, and for each element value of the vector sequence {h _word ⁱⁿ }, LSTM operations such as adding machine-learned weighting coefficients are performed. Then, the text classification unit 11 outputs the vector h _word ^LSTM of the operation result from the _word ^{LSTM 27} , combines the vector h _char ^LSTM and the vector h _word ^LSTM , and propagates them to the intermediate layer FFNN 28.

Referring to FIG. 11, in the examples of FIGS. 16 and 18, the text classification unit 11 inputs a vector sequence {h _word ⁱⁿ } of 200 rows × 150 columns from the word input layer FFNN 25 to the word LSTM 27, and Perform the operation. In this case, the number of nodes in the input layer in the word LSTM27 is 200.

Then, the sentence classification unit 11 generates a vector h _word ^LSTM of 200 rows × 1 column as an LSTM calculation result, outputs this from the word LSTM 27, and propagates it to the intermediate layer FFNN 28. In this case, the number of nodes in the output layer in the word LSTM27 is 200.

Returning to FIG. 10, wording classifying unit 11 receives the vector h _char ^LSTM and vector h ^LSTM where the vector h _word ^LSTM bound from words LSTM27 from characters LSTM26 the intermediate layer FFNN28, vector h _char ^LSTM and FFNN operations such as adding machine-learned weighting coefficients are performed on the values of each element of the vector h _word ^LSTM . Then, the text classification unit 11 outputs the calculation result vector h ^int from the intermediate layer FFNN 28 and propagates it to the output layer FFNN 29.

Referring to FIG. 11, in the examples of FIGS. 16 to 18, the sentence classification unit 11 includes a vector h _char ^LSTM of 200 rows × 1 column from the character LSTM ²⁶ and a vector h _word of 200 rows × 1 column from the word LSTM 27. ^The vector h ^{LSTM combined} with the LSTM is input to the intermediate layer FFNN 28 and the FFNN is calculated. In this case, the number of nodes in the input layer in the intermediate layer FFNN 28 is 400.

Then, the sentence classification unit 11 generates a vector h ^int of 200 rows × 1 column as the calculation result of FFNN, outputs this from the intermediate layer FFNN 28, and propagates it to the output layer FFNN 29. In this case, the number of nodes in the output layer in the intermediate layer FFNN 28 is 200.

Returning to FIG. 10, the sentence classification unit 11 inputs the vector h ^int from the intermediate layer FFNN 28 to the output layer FFNN 29, and adds a machine-learned weight coefficient to the value of each element of the vector h ^int. FFNN is calculated. Then, the text classification unit 11 calculates a probability value using the softmax function with respect to the vector h ^{out of the} calculation result, calculates binary information using the argmax function with respect to the probability value, Output as previous information.

Referring to FIG. 11, in the examples of FIGS. 16 to 18, the text classification unit 11 inputs a vector h ^int of 200 rows × 1 column from the intermediate layer FFNN 28 to the output layer FFNN 29 and performs an FFNN operation. . In this case, the number of nodes in the input layer in the output layer FFNN 29 is 200.

Then, the text classification unit 11 generates a vector h ^out of 2 rows × 1 column as a calculation result of FFNN, performs a function process on the vector h ^out , and outputs the previous / non-existing information. In this case, the number of nodes in the output layer in the output layer FFNN 29 is two.

  In the learning model of Example 1 using FFNN and LSTM shown in FIGS. 10 and 11, an attention mechanism may be added to each of the character LSTM26 and the word LSTM27. By adding an attention mechanism to the character LSTM 26, a high weighting factor is given to characters that affect the entire character sequence of the news posting information, and a low weighting factor is given to characters that do not have much influence. Is given. Further, by adding an attention mechanism to the word LSTM27, a high weighting factor is given to a word that affects the entire word sequence of the news posting information, and a low weight is given to a word that does not have a great influence. A weighting factor is given.

Attention mechanisms are known, see the following references for details.
Miyazaki et al., “Automatic extraction method of tweet useful for news production”, Linguistic Processing Society of Japan, 23rd Annual Conference, Presentation Papers (March 2017), P.418-421

  Further, in the learning model of Example 1 using FFNN and LSTM shown in FIGS. 10 and 11, other RNNs such as GRU (Gated Recurrent Unit) instead of LSTM ( A Recurrent Neural Network may be used. Specifically, a character RNN and a word RNN are used instead of the character LSTM26 and the word LSTM27, respectively. That is, the present invention is applicable to a learning model using FFNN and RNN.

(Specific example 2)
Next, the learning model of specific example 2 will be described. FIG. 12 is a diagram illustrating a schematic structure of a learning model (specific example 2) using FFNN and CNN, and FIG. 13 is a diagram illustrating the number of nodes and input / output data in specific example 2.

  This learning model includes FFNN30 for character input layer, FFNN31 for word input layer, character CNN32, word CNN33, pooling 40, 41 (character pooling layer 40, word pooling layer 41), FFNN34 for intermediate layer, and FFNN35 for output layer. Configured. This learning model also combines the character input layer FFNN 30 and the character CNN 32 so that the output data of the character input layer FFNN 30 is input to the character CNN 32, and inputs the output data of the character CNN 32 to the pooling 40 layer. Thus, the character input layer FFNN31 and the word CNN33 are combined and the output of the word CNN33 so that the character CNN32 and the pooling 40 are combined and the output data of the word input layer FFNN31 is input to the word CNN33. The words CNN 33 and the pooling 41 are combined so that data is input to the pooling 41 layer, and the pooling 40 and 41 are output so that the output data of the pooling 40 and the output data of the pooling 41 are input to the FFNN 34 for the intermediate layer. And the intermediate layer FFNN 34 are combined, and the output of the intermediate layer FFNN 34 As over data is input to the output layer for FFNN35, configured as an intermediate layer for FFNN34 output layer FFNN35 combine to. Note that the calculation is repeated for the characters CNN32 and pooling 40, and the calculation is also repeated for the words CNN33 and pooling 41.

This learning model uses each of a character one-hot vector sequence {x _char } and a word one-hot vector sequence {x _word } as input data, and combines both in the intermediate layer to obtain binary appearing / not appearing information. Is a neural network whose output data is. Further, this learning model holds the weight coefficients of the character input layer FFNN30 to the output layer FFNN35 and the pooling 40, 41 that are machine-learned by the learning device 2 shown in FIG.

The sentence classification unit 11 reads the learning model shown in FIG. 12 stored in the learning model storage unit 12. The text classification unit 11 inputs the character one-hot vector sequence {x _char } generated by the sequence generation unit 10 to the character input layer FFNN 30 and sets the value of each element of the character one-hot vector sequence {x _char }. On the other hand, an FFNN calculation such as adding a weighting coefficient learned by machine learning is performed. Then, the sentence classification unit 11 outputs the vector sequence {h _char ⁱⁿ } of the operation result from the character input layer FFNN 30 and propagates it to the character CNN 32.

Referring to FIG. 13, in the examples of FIGS. 16 and 17, the text classification unit 11 converts a character one-hot vector sequence {x _char } of 6300 rows × 150 columns into a character as in the first specific example of FIG. 11. The data is input to the input layer FFNN 30 and the FFNN is calculated. In this case, the number of nodes in the input layer in the character input layer FFNN 30 is 6300.

Then, the sentence classification unit 11 generates a 200-row × 150-column vector sequence {h _char ⁱⁿ } as the FFNN calculation result, and outputs this from the character input layer FFNN 30, as in the first specific example of FIG. 11. , And propagate to the character CNN32. In this case, the number of nodes in the output layer in the character input layer FFNN 30 is 200.

Returning to FIG. 12, the sentence classification unit 11 inputs the vector sequence {h _char ⁱⁿ } of the operation result from the character input layer FFNN 30 to the character CNN 32, and for each element value of the vector sequence {h _char ⁱⁿ }, CNN operations such as adding machine-learned weighting coefficients are performed. Then, the text classification unit 11 outputs a vector of the calculation result from the character CNN 32 and propagates it to the pooling 40. The sentence classification unit 11 inputs a vector of calculation results from the character CNN 32 to the pooling 40 layer, and performs pooling calculations such as adding machine-learned weighting factors to the values of each element of the vector. In this case, the sentence classification unit 11 repeats the calculation of the character CNN 32 and the pooling 40 a predetermined number of times, outputs the vector h _char ^CNN of the calculation result from the pooling 40, and combines the vector h _char ^CNN and a vector h _word ^CNN described later. To the intermediate layer FFNN 34.

Here, the CNN calculation refers to a convolution calculation, and a local feature of the input vector sequence {h _char ⁱⁿ } is extracted by the calculation. The pooling process is a process for collecting CNN calculation results, and reduction is performed while maintaining local features of the vector sequence {h _char ⁱⁿ }. The details of the calculation of CNN and pooling are already known, and the description thereof is omitted here.

Referring to FIG. 13, in the example of FIGS. 16 and 17, the text classification unit 11 inputs a vector sequence {h _char ⁱⁿ } of 200 rows × 150 columns from the character input layer FFNN 30 to the character CNN 32, and An operation is performed, a vector of the operation result is input to the pooling 40 layer, the pooling operation is performed, and the CNN and pooling operations are repeated. In this case, the number of nodes in the input layer in the character CNN 32 is 200.

Then, the sentence classification unit 11 generates a vector h _char ^CNN of 200 rows × 1 column as a calculation result of CNN and pooling, outputs this from the pooling 40 layer, and propagates it to the intermediate layer FFNN 34. In this case, the number of nodes in the output layer in the pooling 40 is 200.

Returning to FIG. 12, wording classification unit 11, the word one-hot vector sequence generated by sequence generation unit 10 {x _word} entered word input layer FFNN31, word one-hot vector sequence {x _word} FFNN operations such as adding machine-learned weighting coefficients to the values of the elements are performed. Then, the text classification unit 11 outputs the vector sequence {h _word ⁱⁿ } of the calculation result from the word input layer FFNN 31 and propagates it to the word CNN 33.

Referring to FIG. 13, in the examples of FIGS. 16 and 18, the text classification unit 11 converts a word one-hot vector sequence {x _word } of 72000 rows × 150 columns into a word, as in the specific example 1 of FIG. 11. The data is input to the input layer FFNN 31 and FFNN is calculated. In this case, the number of nodes in the input layer in the word input layer FFNN 31 is 72,000.

Then, the sentence classification unit 11 generates a 200-row × 150-column vector sequence {h _word ⁱⁿ } as the FFNN calculation result, outputs this from the word input layer FFNN 31, and propagates it to the word CNN 33. In this case, the number of nodes in the output layer in the word input layer FFNN 31 is 200.

Returning to FIG. 12, the text classification unit 11 inputs the vector sequence {h _word ⁱⁿ } of the operation result from the word input layer FFNN 31 to the word CNN 33, and for each element value of the vector sequence {h _word ⁱⁿ }, CNN operations such as adding machine-learned weighting coefficients are performed. Then, the text classification unit 11 outputs a vector of the calculation result from the word CNN 33 and propagates it to the pooling 41. The sentence classification unit 11 inputs a vector of calculation results from the word CNN 33 to the pooling 41 layer, and performs a pooling calculation such as adding a machine-learned weighting factor to the value of each element of the vector. In this case, the sentence classification unit 11 repeats the calculation of the _word ^CNN 33 and the pooling 41 a predetermined number of times, outputs the vector h _word ^CNN of the calculation result from the pooling 41, and combines the vector h _char ^CNN and the vector h _word ^CNN Propagate to the FFNN 34 for layers.

Referring to FIG. 13, in the examples of FIGS. 16 and 18, the text classification unit 11 inputs a vector sequence {h _word ⁱⁿ } of 200 rows × 150 columns from the word input layer FFNN 31 to the _word CNN 33, and An operation is performed, a vector of the operation result is input to the pooling 41 layer, the pooling operation is performed, and the CNN and pooling operations are repeated. In this case, the number of nodes in the input layer in the word CNN33 is 200.

Then, the sentence classification unit 11 generates a vector h _word ^CNN of 200 rows × 1 column as a calculation result of CNN and pooling, outputs this from the pooling 41 layer, and propagates it to the intermediate layer FFNN 34. In this case, the number of nodes in the output layer in the pooling 41 is 200.

Returning to FIG. 12, the sentence classification unit 11 inputs the vector h ^CNN obtained by combining the vector h _char ^CNN from the pooling 40 and the vector h _word ^CNN from the pooling 41 to the intermediate layer FFNN 34. The sentence classification unit 11 performs the same processing as the intermediate layer FFNN 28 and the output layer FFNN 29 shown in FIG. 10 and outputs the previous / non-existing information.

  The intermediate layer FFNN 34 and the output layer FFNN 35 are the same as the intermediate layer FFNN 28 and the output layer FFNN 29 shown in FIG. 10, respectively, and the number of nodes, vector size, and calculation are also the same. Is omitted.

(Specific example 3)
Next, the learning model of specific example 3 will be described. FIG. 14 is a diagram illustrating a schematic structure of a learning model (specific example 3) using FFNN, and FIG. 15 is a diagram illustrating the number of nodes and input / output data in specific example 3.

  This learning model includes a character input layer FFNN 36, a word input layer FFNN 37, an intermediate layer FFNN 38, and an output layer FFNN 39. The learning model also includes the character input layer FFNN 36, the word input layer FFNN 37, and the intermediate layer so that the output data of the character input layer FFNN 36 and the output data of the word input layer FFNN 37 are input to the intermediate layer FFNN 38. The intermediate layer FFNN 38 and the output layer FFNN 39 are combined so that the output FFNN 38 is coupled to the output layer FFNN 38 and the output data of the intermediate layer FFNN 38 is input to the output layer FFNN 39.

This learning model is a neural network in which each of a character BOW vector x _char and a word BOW vector x _word is input data, both are combined in an intermediate layer, and binary appearing / not appearing information is output data. Further, this learning model holds the weight coefficients of the character input layer FFNN 36 to the output layer FFNN 39 machine-learned by the learning device 2 shown in FIG.

The text classification unit 11 reads the learning model shown in FIG. 14 stored in the learning model storage unit 12. The text classification unit 11 inputs the character BOW vector x _char generated by the sequence generation unit 10 to the character input layer FFNN 36, and adds machine-learned weighting factors to the values of each element of the character BOW vector x _char. FFNN is calculated. Then, the sentence classification unit 11 outputs the calculation result vector h _char ⁱⁿ from the character input layer FFNN 36 and propagates it to the intermediate layer FFNN 38.

FIG. 19 is a diagram illustrating an example of the character BOW vector x _char . When the character sequence is “new, inn, station, west, mouth, date, near, fire, thing, 1, 1, month, 1, 5, day” shown in FIG. A character BOW vector x _char of 6300 rows × 1 column shown in FIG. 19 is generated. The number of lines 6300 of the character BOW vector x _char corresponds to the maximum number of characters appearing in any character series.

  Among the 6300 line positions, “1” is set to the line position corresponding to each character “new”, “hotel”, “day”, and “0” is set to the other line positions. The That is, “1” is set to the line position corresponding to each character among the line positions of 6300.

Referring to FIG. 15, in the example of FIGS. 16 and 19, the text classification unit 11 inputs a character BOW vector x _char of 6300 rows × 1 column to the character input layer FFNN 36 and performs an FFNN operation. In this case, the number of nodes in the input layer in the character input layer FFNN 36 is 6300.

Then, the text classification unit 11 generates a vector h _char ⁱⁿ of 200 rows × 1 column as a calculation result of FFNN, outputs this from the character input layer FFNN 36, and outputs a vector h _char ⁱⁿ and a vector h _word ⁱⁿ described later. Are propagated to the intermediate layer FFNN 38. In this case, the number of nodes in the output layer in the character input layer FFNN 36 is 200.

Returning to FIG. 14, the sentence classification unit 11 inputs the word BOW vector x _word generated by the sequence generation unit 10 to the word input layer FFNN 37, and performs machine learning on the value of each element of the word BOW vector x _word. The FFNN is calculated by adding the weighting factor. Then, the sentence classification unit 11 outputs the vector sequence h _word ⁱⁿ of the operation result from the word input layer FFNN 37 and propagates it to the intermediate layer FFNN 38.

FIG. 20 is a diagram illustrating an example of the word BOW vector _xword . Word sequence is shown in FIG. 16 "Shinjuku Station West Exit, around in, fire, November 15" If it is, the sequence generating unit 10, the word BOW vector x _word of 72000 rows × 1 column shown in FIG. 20 Generated. The number of rows 72000 of the word BOW vector x _word corresponds to the maximum number of words appearing in any word series.

  Among the 72000 line positions, “1” is set to the line position corresponding to each word “Shinjuku Station West Exit”, “Near”, “November 15” constituting the word series, and “ “0” is set. That is, “1” is set to the line position corresponding to each word among the 72000 line positions.

Referring to FIG. 15, in the examples of FIGS. 16 and 20, the sentence classification unit 11, when the news posting information is “November 15 near the west exit of Shinjuku Station”, 72,000 rows × 1 column of words The BOW vector x _word is input to the word input layer FFNN 37 and FFNN is calculated. In this case, the number of nodes in the input layer in the word input layer FFNN 37 is 72,000.

Then, the sentence classification unit 11 generates a vector h _word ⁱⁿ of 200 rows × 1 column as the calculation result of FFNN, outputs this from the word input layer FFNN 37, and propagates it to the intermediate layer FFNN 38. In this case, the number of nodes in the output layer in the word input layer FFNN 37 is 200.

Returning to FIG. 14, wording classifying unit 11 inputs the vector h ⁱⁿ which the vector h _word ⁱⁿ from the vector h _char ⁱⁿ the word input layer FFNN37 from the character input layer FFNN36 is bonded to the intermediate layer for FFNN38 . The sentence classification unit 11 performs the same processing as the intermediate layer FFNN 28 and the output layer FFNN 29 shown in FIG. 10 and outputs the previous / non-existing information.

  The intermediate layer FFNN 38 and the output layer FFNN 39 are the same as the intermediate layer FFNN 28 and the output layer FFNN 29 shown in FIG. 10, respectively, and the number of nodes, the vector size, and the calculation are also the same. Is omitted.

〔Experimental result〕
Next, experimental results by simulation using a computer will be described. FIG. 21 is a diagram for explaining experimental results. (1) shows the experimental results of the prior art using the keyword filtering technique, (2) shows the experimental results of the prior art using NN using only the characters as input data, and (3) shows the word The experimental result of the prior art using NN which used only as input data is shown. (4) shows the experimental results of Example 1 using NN using characters and words as input data.

  In (2), (3), and (4), experimental results when using learning models of LSTM, CNN, and FFNN are shown. In (4), LSTM is the learning model of the specific example 1 shown in FIGS. 10 and 11, CNN is the learning model of the specific example 2 shown in FIGS. 12 and 13, and FFNN is It is a learning model of the specific example 3 shown in FIG.14 and FIG.15.

  The precision (Precision) is an index relating to accuracy indicating how much the determination result, which is the information that has already appeared or not, matches the actual correct data. In addition, the recall (Recall) is an index relating to completeness that represents how much the determination result covers actual correct answer data. Further, the F value is a harmonic average of the precision and the recall.

  As learning data for machine learning of each learning model, 44670 pieces of news posting information for a predetermined period were used. Of these, the number of previously-exposed information is 9300, and the number of non-existing information is 35370. As test data for obtaining the experimental results shown in FIG. 21, 10,000 news posting information randomly sampled from news posting information for a predetermined period was used. Of these, the number of previously-exposed information is 2028 and the number of non-existing information is 7972.

  According to the experimental results of FIG. 21, the F value of (1) is 50.0%, and the F values of LSTM, CNN and FFNN of (2) are 87.3%, 86.1% and 85.0, respectively. The F values of LSTM, CNN, and FFNN in (3) are 86.2%, 85.4%, and 84.0%, respectively. The F values of LSTM, CNN, and FFNN in (4) are 88.1%, 88.4%, and 85.9%, respectively. Thus, it can be seen that the results are better when the learning models (2) to (4) are used than the conventional technique (1).

  In addition, when the learning model of only the character of (2) is used or when the learning model of only the word of (3) is used, the learning model of Example 1 of (4) is used in LSTM. It can be seen that the results are improved in each of the learning models of CNN and FFNN.

  The present invention has been described with reference to Examples 1 and 2 and Specific Examples 1, 2, and 3. However, the present invention is not limited to Examples 1 and 2 and Specific Examples 1, 2, and 3, Various modifications can be made without departing from the technical idea.

  For example, in the first and second embodiments and the first, second, and third specific examples, the news material classifying devices 1-1 and 1-2 use the learning model to convert the news posting information to the first report non-existing information. It was classified into other existing information. The present invention is not limited to this, and can be applied to other classifications according to user requirements. For example, the news material classification devices 1-1 and 1-2 may classify the news posting information into information indicating whether the poster is male or female, and includes a dialect. It may be classified into information indicating whether or not. In this case, the learning device 2 performs machine learning on the learning model using classification information such as information indicating male or female, information indicating whether or not a dialect is included, instead of the previously displayed / not-present information.

  As the hardware configuration of the news material classification device 1-1 according to the first embodiment of the present invention, the news material classification device 1-2 according to the second embodiment, and the learning device 2, a normal computer can be used. The news material classification devices 1-1 and 1-2 and the learning device 2 are configured by a computer having a CPU (or GPU), a volatile storage medium such as a RAM, a non-volatile storage medium such as a ROM, an interface, and the like. Is done.

  Each function of the sequence generation unit 10, the sentence classification unit 11, and the learning model storage unit 12 included in the news material classification apparatus 1-1 is realized by causing the CPU to execute a program describing these functions. In addition, each function of the sequence generation unit 10, the sentence classification unit 11, the learning model storage unit 12, the agent classification unit 13 and the determination unit 14 included in the news material classification device 1-2 also has a program describing these functions as a CPU. This is realized by executing the above. Furthermore, the functions of the sequence generation unit 10, the learning unit 20, and the learning model storage unit 12 included in the learning device 2 are also realized by causing the CPU to execute a program describing these functions.

  These programs are stored in the storage medium and read out and executed by the CPU. These programs can also be stored and distributed in a storage medium such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), optical disk (CD-ROM, DVD, etc.), semiconductor memory, etc. You can also send and receive.

  The learning models shown in FIGS. 9 to 15 are learned models that have been machine-learned, and the computer displays news posting information that indicates whether or not news posting information extracted from a large number of social media information has already appeared. It is for functioning to classify the information into the existing information and output the existing / non-existing information. The configuration is as shown in each figure.

  These learning models are used as program modules that are a part of artificial intelligence software, and are used in a computer having a CPU and a memory. Specifically, for example, in the case of the learning model shown in FIG. 9, the CPU of the computer inputs a character vector to the character NN21 in accordance with a command from the learning model stored in the memory, and is input to the input layer of the character NN21. An NN operation is performed on the character vector based on the machine-learned weighting coefficient, and an operation result vector is output from the output layer to the output NN23. Further, the CPU of the computer inputs a word vector to the word NN22, performs an NN operation on the word vector input to the input layer of the word NN22, based on a machine-learned weight coefficient, and outputs an operation result from the output layer. It operates to output the vector to the output NN23. Then, the CPU of the computer combines the vector of the operation result output from the output layer of the character NN21 and the vector of the operation result output from the output layer of the word NN22, which are input to the input layer of the output NN23. On the other hand, an NN calculation based on the machine-learned weighting coefficient is performed, and the operation result is output from the output layer. The learning model shown in FIGS. 10 to 15 operates similarly.

DESCRIPTION OF SYMBOLS 1 News material classification device 2 Learning apparatus 10 Sequence generation part 11 Text classification part 12 Learning model memory | storage part 13 Agent classification part 14 Determination part 20 Learning part 21 Character NN (neural network)
22 words NN
23 output NN
24, 30, 36 FFNN for character input layer (forward propagation neural network)
25, 31, 37 FFNN for word input layer
26 character LSTM (Long Short Term Memory Unit)
27 words LSTM
FFNN for 28, 34, 38 intermediate layer
29, 35, 39 FFNN for output layer
32 character CNN (convolutional neural network)
33 words CNN
40, 41 Pooling 100 News material extraction device a, b Previous / non-existing information {x _char } character one-hot vector sequence {x _word } word one-hot vector sequence x _char character BOW vector x _word word BOW vector

Claims (9)

In a news material classification device that inputs post information that can be news material among a large number of social media information as news post information, and classifies the news post information according to a user's request,
A sequence that inputs the news posting information, extracts characters and words included in the news posting information, generates a character vector composed of the character sequence, and generates a word vector composed of the word sequence A generator,
A learning model storage unit for storing a machine-learned learning model;
Reading the learning model from the learning model storage unit, and using the learning model, based on the character vector and the word vector generated by the sequence generation unit, to the user's request for the news posting information A classification unit that generates and outputs corresponding classification information,
The learning model is
Character NN (neural network) having the character vector of the news posting information as input data and the vector of the operation result as output data,
The word NN having the word vector of the news posting information as input data and the vector of the operation result as output data, and
A vector obtained by combining the vector of the calculation result of the character NN and the vector of the calculation result of the word NN as input data, and an output NN using the classification information as the calculation result as output data;
The classification unit includes:
Using the character NN, NN is calculated based on the character vector of the news posting information,
Using the word NN, NN is calculated based on the word vector of the news posting information,
Using the output NN, performing an NN operation based on a vector obtained by combining the operation result vector of the character NN and the operation result vector of the word NN, and outputting the operation result as the classification information; Feature news material classification device. The news material classification device according to claim 1,
A news material classification device, characterized in that the classification information is used / not-shown information indicating whether the news posting information has been published or has not been published. The news material classification device according to claim 2,
Furthermore, a second classification unit and a determination unit are provided,
The classification unit includes:
Outputting the above-mentioned information on non-existing information as the first classification result,
The second classification unit includes:
Input the news posting information, extract agent information added to the news posting information for identifying the posting source device, and generate the previous / non-existing information based on the agent information Output as the second classification result,
The determination unit
Based on the first classification result output by the classification unit and the second classification result output by the second classification unit, new output / non-existing information is generated and output. News material classification device. In the news material classification device according to claim 2 or 3,
The sequence generation unit
Input the news posting information, extract characters and words included in the news posting information, generate one-hot vector sequences by arranging one-hot column vectors corresponding to the characters, and the word Generates a word one-hot vector sequence by arranging one-hot column vectors corresponding to
The learning model is
FFNN for a character input layer having the character one-hot vector sequence generated by the sequence generation unit as input data and a vector of a calculation result of FFNN (feed forward neural network) as output data;
A character RNN having a vector of calculation results of the character input layer FFNN as input data and a vector of calculation results of RNN (recurrent neural network) as output data,
FFNN for a word input layer having the word one-hot vector sequence generated by the sequence generation unit as input data and a vector of the operation result of FFNN as output data,
A word RNN having a vector of the calculation result of the word input layer FFNN as input data and a vector of the calculation result of the RNN as output data;
FFNN for intermediate layer having a vector obtained by combining the vector of the calculation result of the character RNN and the vector of the calculation result of the word RNN as input data and the vector of the calculation result of FFNN as output data, and
The intermediate layer FFNN operation result vector is used as input data, and the FFNN operation result is the output layer FFNN using the above-mentioned previous / non-existing information as output data.
The classification unit includes:
Based on the character one-hot vector sequence generated by the sequence generation unit using the character input layer FFNN, FFNN is calculated,
Using the character RNN, the RNN is calculated based on the vector of the calculation result of the character input layer FFNN,
Based on the word one-hot vector sequence generated by the sequence generation unit using the word input layer FFNN, FFNN is calculated,
Using the word RNN, RNN is calculated based on a vector of calculation results of the word input layer FFNN,
Using the intermediate layer FFNN, based on a vector obtained by combining the vector of the character RNN and the vector of the word RNN, the FFNN is calculated.
A news material classification device, wherein the output layer FFNN is used to perform an FFNN operation based on a vector of operation results of the intermediate layer FFNN, and the operation result is output as the previous / non-existing information. . In the news material classification device according to claim 2 or 3,
The sequence generation unit
Input the news posting information, extract characters and words included in the news posting information, generate one-hot vector sequences by arranging one-hot column vectors corresponding to the characters, and the word Generates a word one-hot vector sequence by arranging one-hot column vectors corresponding to
The learning model is
FFNN for a character input layer having the character one-hot vector sequence generated by the sequence generation unit as input data and a vector of the operation result of FFNN as output data,
A character CNN having a vector of calculation results of the character input layer FFNN as input data and a vector of calculation results of CNN (convolutional neural network) as output data;
A character pooling layer having the vector of the calculation result of the character CNN as input data and the vector of the calculation result of pooling as output data;
FFNN for a word input layer having the word one-hot vector sequence generated by the sequence generation unit as input data and a vector of the operation result of FFNN as output data,
A word CNN having a vector of calculation results of the word input layer FFNN as input data and a vector of calculation results of CNN as output data,
A word pooling layer having a vector of calculation results of the word CNN as input data and a vector of calculation results of pooling as output data;
FFNN for an intermediate layer having a vector obtained by combining the vector of the result of operation of the character pooling layer and the vector of the result of operation of the word pooling layer as input data, and FFNN for the output of the vector of the operation result of FFNN, and
The intermediate layer FFNN operation result vector is used as input data, and the FFNN operation result is the output layer FFNN using the above-mentioned previous / non-existing information as output data.
The classification unit includes:
Based on the character one-hot vector sequence generated by the sequence generation unit using the character input layer FFNN, FFNN is calculated,
Using the character CNN, the CNN is calculated based on the vector of the calculation result of the character input layer FFNN,
Using the character pooling layer, a pooling operation is performed based on a vector of the operation results of the character CNN,
Based on the word one-hot vector sequence generated by the sequence generation unit using the word input layer FFNN, FFNN is calculated,
Using the word CNN, CNN is calculated based on the calculation result vector of the word input layer FFNN,
Using the word pooling layer, a pooling calculation is performed based on a vector of calculation results of the word CNN,
Based on a vector obtained by combining the vector of the result of the character pooling layer and the vector of the result of the operation of the word pooling layer using the intermediate layer FFNN, the FFNN is calculated.
A news material classification device, wherein the output layer FFNN is used to perform an FFNN operation based on a vector of operation results of the intermediate layer FFNN, and the operation result is output as the previous / non-existing information. . In the news material classification device according to claim 2 or 3,
The sequence generation unit
Input the news posting information, extract characters and words included in the news posting information, generate character BOW vectors corresponding to all the extracted characters, and correspond to all extracted words Generate a word bow vector,
The learning model is
FFNN for character input layer having the character BOW vector generated by the sequence generation unit as input data and the vector of the operation result of FFNN as output data,
FFNN for word input layer having the word BOW vector generated by the sequence generation unit as input data and the vector of the operation result of FFNN as output data,
An intermediate layer FFNN having a vector obtained by combining the vector of the calculation result of the character input layer FFNN and the vector of the calculation result of the word input layer FFNN as input data and an output data of the vector of the calculation result of FFNN; ,
The intermediate layer FFNN operation result vector is used as input data, and the FFNN operation result is the output layer FFNN using the above-mentioned previous / non-existing information as output data.
The classification unit includes:
Based on the character BOW vector generated by the sequence generation unit using the character input layer FFNN, the FFNN is calculated,
Based on the word BOW vector generated by the sequence generation unit using the word input layer FFNN, FFNN is calculated,
Using the intermediate layer FFNN, based on a vector obtained by combining the vector of the character input layer FFNN and the word input layer FFNN, the FFNN is calculated.
A news material classification device, wherein the output layer FFNN is used to perform an FFNN operation based on a vector of operation results of the intermediate layer FFNN, and the operation result is output as the previous / non-existing information. . In a news material classification device that inputs post information that can be news material among a lot of social media information as news post information, and classifies the news post information as published or not published,
A sequence generator that inputs the news posting information, extracts characters or words included in the news posting information, and generates a character vector composed of the character sequence or a word vector composed of the word sequence; ,
The character vector or the word vector of the news posting information is used as input data, and the previous / non-existing information indicating whether the news posting information has been or has not been output, which is a calculation result, is output data. And a learning model storage unit that stores the learning model that has been machine-learned,
The learning model is read from the learning model storage unit, and using the learning model, the previously-noted / non-existing information is generated based on the character vector or the word vector generated by the sequence generation unit, and the first classification is performed. A classification part to be output as a result;
Input the news posting information, extract agent information added to the news posting information for identifying the posting source device, and generate the previous / non-existing information based on the agent information A second classification unit for outputting as a second classification result,
Based on the first classification result output by the classification unit and the second classification result output by the second classification unit, new output / non-existing information is generated and output. News material classification device.   A program for causing a computer to function as the news material classification device according to any one of claims 1 to 7. A learning model for causing a computer to function so as to classify news posting information extracted from a large number of social media information according to a user's request and output classification information,
A character NN having a character vector composed of a character sequence extracted from the news posting information as input data, and a vector of the operation result as output data;
A word NN having a word vector consisting of a series of words extracted from the news posting information as input data, and a vector of an operation result as output data; and
By an output NN using a vector obtained by combining the vector of the calculation result of the character NN and the vector of the calculation result of the word NN as input data, and using the classification information that is the calculation result as output data,
The weighting coefficient of the character NN machine-learned using the character vector and the word vector of the news posting information as input data of the learning model, and the classification information of the news posting information as output data of the learning model, A weighting factor for the word NN and the weighting factor for the output NN;
An NN operation based on the weight coefficient of the character NN is performed on the character vector input to the input layer of the character NN, and an operation result vector is output from the output layer.
An NN operation is performed on the word vector input to the input layer of the word NN based on the weight coefficient of the word NN, and an operation result vector is output from the output layer,
For a vector obtained by combining the vector of the operation result output from the output layer of the character NN and the vector of the operation result output from the output layer of the word NN, which are input to the input layer of the output NN A learning model that causes the computer to function so as to perform an NN operation based on a weight coefficient of the output NN and to output the classification information as an operation result from an output layer.

Images