JP6899367B2 - Learning device, validation device, learning method, validation method, learning program, and validation program - Google Patents

Description

The present invention relates to the technical field of machine learning, and in particular to the technical field of methods for classifying texts.

In recent years, research and development of machine learning for classifying a large number of texts has been actively carried out. For example, Non-Patent Document 1 discloses a classification method using a neural network. Specifically, the feature vectors of the words that make up the sentence are concatenated to express the sentence. A filter is applied in the convolution layer to the concatenated feature vectors to generate new features. Max pooling is applied to the new features and the probability distribution is output from the fully coupled softmax layer.

Yoon Kim, Convolutional Neural Networks for Sentence Classification, Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing), Association for Computational Linguistics, 2014, p.1746-1751

For example, assume that the text created or entered by people is valid. In a situation where there is no fixed correct answer as a text, there is a limit to determining validity on a rule basis, for example. Therefore, a human confirms the text and makes a judgment. In this case, if the number of texts to be judged rises to a large number, the work load becomes enormous.

For example, when a free-answer questionnaire is conducted, various answers can be entered. Answers that are not valid for a question, such as misguided answers or random answers, are of low value as information. Therefore, it is desirable to appropriately judge the validity of the answer.

It is conceivable to use machine learning to flexibly judge validity on behalf of humans. However, in the conventional technique, learning and classification have not been performed from the viewpoint of the validity of the text.

The present invention has been made in view of the above points, and is a learning device, a validity judgment device, a learning method, a validity judgment method, a learning program, which makes it possible to improve the validity judgment accuracy of a text. And to provide a validation program.

In order to solve the above problems, the invention according to claim 1 has one or more word information indicating one or more words constituting the first text, and one or more part lyrics information indicating the part of the one or more words. A training data acquisition means for acquiring the plurality of training data from a storage means for storing a plurality of training data including a label indicating the validity of the first text as an answer to a question in a questionnaire, and the above. For each of the acquired plurality of training data, a feature information generation means for generating two-dimensional array word information including one or more word feature vectors indicating the features of each of the one or more words from the one or more word information. By machine learning from the plurality of acquired training data, a model that outputs second validity information indicating the validity of a given second text as an answer to a question in a questionnaire and includes a convolution layer is obtained. The model generation means for generating, for each of the one or more words included in each of the acquired plurality of training data, the word feature vector corresponding to the word and the part code information are arranged side by side. The dimension array word information and the one or more part word information are input to the convolution layer to acquire the feature information indicating the feature of the first text, and the validity of the first text is obtained based on the acquired feature information. The first validity information indicating the above is output from the model, and the model generation means for training the model by error back propagation based on comparing the first validity information with the label is provided and is the same as each other. The size of the filter of the convolution layer in the direction in which the word feature vector corresponding to the word and the part word information are arranged is characterized in that the size of the filter of the convolution layer matches the sum of the size of the word feature vector and the size of the part word information. ..

According to the present invention, a model is generated by machine learning using part-speech information in addition to word information. Part-of-speech information can give the model hints about the words contained in the text and the grammar used in the text. Therefore, by using the generated model, the accuracy of determining the validity of the text can be improved.
Further, according to the present invention, the feature vector of the word and the information of the part of speech of the word are convoluted in a state of being associated with each other, and the feature of the text is output. The model is optimized based on the validation results using the characteristics of this text. Therefore, the accuracy of determining the validity of the text can be improved.
Further, according to the present invention, the validity of the answer to the question in the questionnaire can be determined.

In the invention according to claim 2, the model generating means has the two-dimensional array word information in which the one or more word feature vectors are arranged according to the order of the one or more words in the first text, and the arrangement. The one or more part-of-speech information arranged in order is input to the convolution layer, and the convolution layer is input.
The size of the filter in the direction in which one or more word information is arranged is characterized in that it is a size corresponding to a plurality of words .

In the invention according to claim 3, the feature information generating means further obtains two-dimensional array part of speech information including one or more part of speech feature vectors indicating the characteristics of each of the one or more part of speech from the one or more part of speech information. The model generation means is characterized in that the three-dimensional arrangement information obtained by superimposing the two-dimensional arrangement word information and the two-dimensional arrangement part of speech information is input to the convolution layer.

According to the present invention, a three-dimensional filter is applied in the convolution layer to a three-dimensional array composed of a two-dimensional array composed of word feature vectors and a two-dimensional array composed of part-word feature vectors. The text features are output. Therefore, the accuracy of determining the validity of the text can be improved.

In the invention according to claim 4, the feature information generating means further obtains two-dimensional array part code information including one or more part code feature vectors indicating the characteristics of each of the one or more part words from the one or more part code information. The model generation means is characterized in that the two-dimensional array information formed by connecting the two-dimensional array word information and the two-dimensional array part part information is input to the convolution layer.

According to the present invention, a two-dimensional filter is applied in the convolution layer to a two-dimensional array composed of a word feature vector and a two-dimensional array composed of a part-word feature vector. The text features are output. Therefore, the accuracy of determining the validity of the text can be improved.

In the invention according to claim 5, the model generating means adds a numerical value indicating each of the one or more part of speech to the word feature vector of the word corresponding to the part of speech among the one or more word feature vectors. It is characterized in that the two-dimensional array word information is input to the convolution layer.

According to the present invention, each word feature vector is convoluted with a numerical value indicating a part of speech added. Therefore, the accuracy of determining the validity of the text can be improved.

The invention according to claim 6 is characterized in that each of the plurality of training data includes a plurality of the part of speech information indicating a plurality of part of speech layered by a predetermined part of speech system for each of the one or more words. To do.

According to the present invention, it is possible to give a hint about a word contained in a text and a grammar used in the text hierarchically by using a plurality of part-speech information. Therefore, the accuracy of determining the validity of the text can be improved.

The invention according to claim 7 is characterized in that each of the plurality of training data further includes type information indicating what kind of content is required to be shown in the first text.

According to the present invention, it is possible to improve the accuracy of determining the validity of a text when the part of speech of the word contained in the text and the grammar used in the text are different depending on the type of content required. ..

In the invention according to claim 8, each of the plurality of training data includes one or more second word information indicating one or more second words constituting the third text indicating the question, and the first or more second word information. It is characterized by further including one or more second part of speech information indicating the part of speech of a word.

According to the present invention, a model is generated by machine learning using word information and part-speech information of the text indicating the required content as the content indicated by the text. Therefore, since the text indicating the request content and the text created for this request are learned in pairs, the accuracy of determining the validity of the text can be improved.

The invention according to claim 9 is characterized in that each of the plurality of training data further includes one or more character number information indicating the number of characters of each of the one or more words.

According to the present invention, the number of characters information can provide a model with further hints about the words contained in the text. Therefore, the accuracy of determining the validity of the text can be improved.

The invention according to claim 10 is characterized in that each of the plurality of training data further includes one or more character type information indicating the type of characters included in each of the one or more words.

According to the present invention, the character type information can provide a model with further hints about the words contained in the text. Therefore, the accuracy of determining the validity of the text can be improved.

In the invention according to claim 11, the label indicates the validity of the first text as an answer to the first question, and the model generating means is of the same type as the type of answer required by the first question. It is characterized by generating the model that outputs the second validity information indicating the validity of the second text as indicating the answer to the second question requesting the answer of.

The invention according to claim 12 uses a reading means for reading the model from a model storage means for storing the model generated by the learning device according to any one of claims 1 to 11 and a given text. A text information acquisition means for acquiring one or more determination word information indicating one or more words constituting one or more, and one or more determination part of speech information indicating the part of speech of the one or more words, and the acquired one or more determinations. It is characterized by including an output means for outputting validity information indicating the validity of the given text by inputting word information and one or more determination part of speech information into the read model. ..

According to the present invention, the validity of a text is determined by using a model generated by machine learning using not only word information but also part of speech information. Therefore, the accuracy of determining the validity of the text can be improved.

The invention according to claim 13 is, in a learning method executed by a computer, one or more word information indicating one or more words constituting the first text and one or more part words indicating the part of the one or more words. A training data acquisition step of acquiring the plurality of training data from a storage means for storing a plurality of training data including information and a label indicating the validity of the first text as an answer to a question in a questionnaire. With respect to each of the acquired plurality of training data, a feature information generation step of generating two-dimensional array word information including one or more word feature vectors indicating the features of each of the one or more words from the one or more word information. , A model that outputs second validity information indicating the validity of a given second text as an answer to a question in a questionnaire by machine learning from the acquired plurality of training data and includes a convolution layer. In the model generation step for generating the above, for each of the one or more words included in each of the acquired plurality of training data, the word feature vector corresponding to the word and the part code information are arranged side by side. Two-dimensional array word information and one or more part word information are input to the convolution layer to acquire feature information indicating the characteristics of the first text, and based on the acquired feature information, the validity of the first text is obtained. A model generation step of training the model by error backpropagation based on outputting the first validity information indicating the property from the model and comparing the first validity information with the label is included. The size of the filter of the convolution layer in the direction in which the word feature vector and the part word information are arranged is characterized in that the size of the filter of the convolution layer matches the sum of the size of the word feature vector and the size of the part word information .

The invention according to claim 14 is a model storage means for storing a model generated by the learning device according to any one of claims 1 to 11 in a validity determination method executed by a computer. Text information acquisition to acquire a reading step of reading, one or more judgment word information indicating one or more words constituting a given text, and one or more judgment part information indicating the part of the one or more words. By inputting the step, the acquired one or more judgment word information and one or more judgment part word information into the read model, the validity information indicating the validity of the given text is output. It is characterized by including an output step.

The invention according to claim 15 uses a computer to obtain one or more word information indicating one or more words constituting the first text, one or more part lyrics information indicating the part of the one or more words, and a question in a questionnaire. A training data acquisition means for acquiring the plurality of training data from a storage means for storing a plurality of training data including a label indicating the validity of the first text as an answer to the above , and a plurality of acquired training data. For each of the training data, a feature information generating means for generating two-dimensional array word information including one or more word feature vectors indicating the features of each of the one or more words from the one or more word information, and the acquired plurality of A model generation means that outputs a second validity information indicating the validity of a given second text as an answer to a question in a questionnaire and generates a model including a convolution layer by machine learning from the training data of. For each of the one or more words included in each of the acquired plurality of training data, the two-dimensional array word information and the two-dimensional array word information and the word information in a state where the word feature vector corresponding to the word and the part-word information are arranged side by side. The one or more part word information is input to the convolution layer to acquire the feature information indicating the feature of the first text, and the first validity indicating the validity of the first text is based on the acquired feature information. The sexual information is output from the model and functions as a model generation means for training the model by error back propagation based on comparing the first validity information with the label, and the word feature vector and the label are described. The size of the filter of the convolution layer in the direction in which the part-of-speech information is arranged is characterized by matching the sum of the size of the word feature vector and the size of the part-of-speech information .

The invention according to claim 16 is given a reading means for reading a computer from a model storage means for storing a model generated by the learning device according to any one of claims 1 to 11. A text information acquisition means for acquiring one or more determination word information indicating one or more words constituting the text, and one or more determination part of speech information indicating the part of speech of the one or more words, and the acquired 1 By inputting the above determination word information and one or more determination part of speech information into the read model, it functions as an output means for outputting validity information indicating the validity of the given text. It is characterized by.

According to the present invention, the accuracy of determining the validity of a text can be improved.

It is a block diagram which shows an example of the outline structure of the validity determination apparatus 1 which concerns on one Embodiment. It is a figure which shows an example of the functional block of the system control unit 11 and GPU 18 of the validity determination apparatus 1 which concerns on one Embodiment. It is a figure which shows the structural example of the model 2. It is a figure which shows an example of the information generated by the conversion unit 21. It is a figure which shows the configuration example of a filter. It is a flowchart which shows an example of the learning process by the system control unit 11 and GPU 18 of the validity determination apparatus 1. It is a flowchart which shows an example of the validity determination processing by the system control unit 11 and GPU 18 of the validity determination apparatus 1. It is a figure which shows the structural example of the model 2. It is a figure which shows the information tail Nichirei generated by the conversion part 21-2. It is a figure which shows the structural example of the model 2. It is a figure which shows an example of the information generated by the conversion unit 21-3. It is a figure which shows the structural example of the model 2. It is a figure which shows an example of the information generated by the conversion part 21-4. It is a figure which shows an example of the information generated by the conversion part 21-4. It is a figure which shows the structural example of the model 2. It is a figure which shows an example of the information generated by the determination unit 23-5. It is a figure which shows the structural example of the model 2. It is a figure which shows an example of the information generated in the conversion part 21-6. It is a figure which shows the structural example of the model 2. It is a figure which shows an example of the information generated in the conversion part 21-7. It is a figure which shows the structural example of the model 2. It is a figure which shows an example of the information generated by the conversion unit 21-8.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The embodiment described below is an embodiment in which the present invention is applied to learning using text and a validity determination device for determining the validity of text using a model generated by this learning. .. The text to be validated is, for example, a sentence, a sentence, a phrase, a noun, or the like. These texts are usually created, entered, etc. under some circumstances. Examples of situations include answering questionnaire questions, creating texts for product or service advertisements, writing business emails, writing texts for given themes, issues, etc. Be done. For example, even in the case of a questionnaire, the situation varies depending on the content of the question. The validation device determines the validity of the text for a given situation, for example. The validity is judged, for example, whether the answer matches the question, whether the text of the advertisement contains inappropriate wording, whether the email contains inappropriate wording, and whether the text contains inappropriate wording. Whether or not it matches the theme, etc. The device that executes learning and the device that executes validity determination may be separate devices.

[1. First Embodiment]
[1-1. Configuration of validation device]
First, the configuration of the validity determination device 1 will be described with reference to FIG. FIG. 1 is a block diagram showing an example of an outline configuration of the validity determination device 1 according to the present embodiment. As shown in FIG. 1, the validity determination device 1 includes a system control unit 11, a system bus 12, an input / output interface 13, a storage unit 14, a communication unit 15, an input unit 16, and a display unit 17. , GPU (Graphics Processing Unit) 18 is provided. The system control unit 11 and the input / output interface 13 are connected via the system bus 12. The validity determination device 1 may be, for example, a server device or a personal computer.

The system control unit 11 is composed of a CPU (Central Processing Unit) 11a, a ROM (Read Only Memory) 11b, a RAM (Random Access Memory) 11c, and the like.

The input / output interface 13 performs interface processing between the storage units 14 to GPU 18 and the system control unit 11.

The storage unit 14 is composed of, for example, a hard disk drive, a solid state drive, or the like. The generated model 2 and a plurality of training data used for generating the model 2 are stored in the storage unit 14. The storage unit 14 further stores an operating system, a program for generating a model, a program for determining validity, and the like. The training data and various programs may be acquired from a predetermined computer via a network, or may be recorded on a recording medium such as an optical disk, a memory card, or a magnetic tape and read via a drive device. It may be. When the device for generating the model 2 and the device for determining the validity are separate devices, the generated model 2 may be delivered via the network or via a recording medium. May be good.

The communication unit 15 is composed of, for example, a network interface controller or the like. The communication unit 15 connects to another computer via a network such as the Internet or a LAN (Local Area Network), and controls the communication state with the computer.

The input unit 16 receives an operation by the operator and outputs a signal corresponding to the operation content to the system control unit 11. Examples of the input unit 16 include a keyboard, a mouse, a touch panel, and the like.

The display unit 17 is composed of, for example, a graphic controller, a display, and the like. The display unit 17 displays information such as images and characters under the control of the system control unit 11. Examples of display panels include liquid crystal panels, organic EL (Light Emitting) panels, and the like.

The GPU 18 executes matrix operations and the like in machine learning under the control of the system control unit 11. The GPU 18 pipelines a plurality of operations in parallel. A VRAM (Video RAM) is connected to the GPU 18. The CPU 11a may perform this matrix operation and the like.

[1-2. Functional overview of the system control unit of the validation device]
Next, the functional outline of the system control unit 11 and the GPU 18 will be described with reference to FIGS. 2 to 5. FIG. 2 is a diagram showing an example of functional blocks of the system control unit 11 and the GPU 18 of the validity determination device 1 according to the present embodiment. As shown in FIG. 2, the system control unit 11 and the GPU 18 read and execute various codes and the like included in the program stored in the storage unit 14 by the CPU 11a, so that the training data acquisition unit 111 and the model generation unit 112 , The model reading unit 113, the determination target data acquisition unit 114, the validity information output unit 115, and the like.

[1-2-1. Model generation]
The training data acquisition unit 111 acquires a plurality of training data from the storage unit 14. Each training data contains one or more word information and one or more part of speech information. The word information of one or more indicates one or more words constituting the text. For example, when training data is prepared for texts created for questionnaires, themes, tasks, etc., a plurality of texts created for only one question, one theme, one task, etc. of the questionnaire are used. Training data may be created. Alternatively, a plurality of training data may be created from a plurality of texts created for a plurality of questions, a plurality of themes, a plurality of tasks, and the like. The part-speech information of one or more indicates the part-speech of each one or more words constituting the text. The morphological analysis extracts words from the text and identifies the part of speech of each word. The word information and part of speech information are, for example, one-hot vectors. The order of the word information and the order of the part of speech information match the order of the words in the text. Padding is performed as necessary so that the numbers of word information and part of speech information match among all the training data. Therefore, even when the text is composed of only one word, a plurality of word information and a plurality of part-speech information are actually generated. The morphological analysis, one-hot vectorization, and padding may be performed by the training data acquisition unit 111, or may be performed by another computer and stored in the storage unit 14 in advance.

Each training data may further include a validity label. The validity label is a label that indicates the validity of the text. For example, if the text is valid, the validity label indicates 1, and if the text is not valid, the validity label indicates 0. For example, when the text to be created, such as a questionnaire question, theme, or assignment, is required to indicate what the text should indicate, the content indicated by the text matches the required content. A validity label may be created in terms of whether or not it is done. For example, if the answer to the questionnaire matches the question, the validity label indicates 1, and if the answer does not match the question, the validity label indicates 0. Also, depending on the situation in which the text is used, such as an advertisement or a business email, a validity label may be created from the viewpoint of whether or not the text contains an inappropriate phrase. Validity labels are created, for example, manually.

The model generation unit 112 generates a model 2 that outputs validity information indicating the validity of a given text by machine learning from a plurality of training data acquired by the training data acquisition unit 111.

Model 2 may be a neural network such as a convolutional neural network. As each layer in the model 2, the model generation unit 112 executes arithmetic processing corresponding to the layer, calculates the validity of the text, and outputs the validity information. The model generation unit 112 compares the validity information with the validity label and executes learning. The model generation unit 112 generates a word feature two-dimensional array including one or more word feature vectors showing the features of each of the one or more words constituting the text from one or more word information for each of the plurality of training data. The word feature vector is, for example, a one-dimensional array and includes a plurality of numerical values. The model generation unit 112 has a two-dimensional array of word features and a state in which each of the one or more part-speech information is associated with the word feature vector of the word corresponding to the part of speech indicated by the part-speech information among the one or more word feature vectors. Input one or more part-speech information into the convolution layer, and let the convolution layer perform learning. As a result, the model generation unit 112 generates the model 2.

The model generation unit 112 further generates a two-dimensional array of part of speech features including one or more part of speech feature vectors indicating the features of each of the one or more part of speech corresponding to each of the words constituting the text from one or more part of speech information. May be good. The part-speech feature vector is also, for example, a one-dimensional array and includes a plurality of numerical values. Then, as an example of associating the part-speech information with the word feature vector, the model generation unit 112 may input the feature three-dimensional array obtained by superimposing the word feature two-dimensional array and the part-speech feature two-dimensional array into the convolution layer.

FIG. 3 is a diagram showing a configuration example of the model 2. As shown in FIG. 3, the model 2 includes a conversion unit 21, a feature extraction unit 22, and a determination unit 23. The conversion unit 21 is a layer group for converting word information and part of speech information obtained from the text into information in a form suitable for feature extraction of the text. The feature extraction unit 22 is a layer group for outputting feature information of text. The determination unit 23 is a layer group for outputting validity information indicating the validity of the text based on the feature information of the text.

The conversion unit 21 includes input layers 201-1 and 201-2, embedded layers 202-1 and 202-2, a connecting layer 203, and a reshape layer 204. A plurality of word information is input to the input layer 2011-1. The embedded layer 202-1 generates a word feature two-dimensional array composed of a plurality of word feature vectors corresponding to each of the plurality of word information output from the input layer 211-1. For example, the Word2vec model may be used to generate the feature vector, or another model may be used. FIG. 4 is a diagram showing an example of information generated by the conversion unit 21. In the word feature two-dimensional array 301 shown in FIG. 4, the direction in which the numerical values constituting the feature vector of one word are arranged is defined as the x-axis, and the size of the word feature two-dimensional array 301 in the x-axis direction is referred to as width. The width of the word feature two-dimensional array 301 matches the number of dimensions of the feature vector. In the word feature two-dimensional array 301, the direction in which a plurality of word feature vectors are arranged according to the word order is defined as the y-axis, and the size of the word feature two-dimensional array 301 in the y-axis direction is referred to as height. The order of the words constituting the text and the order of the word feature vectors of those words in the word feature two-dimensional array 301 match. A plurality of part of speech information is input to the input layer 201-2. The embedded layer 202-2 outputs a two-dimensional array of part-speech features composed of a plurality of part-speech feature vectors corresponding to each of the plurality of part-speech information output from the input layer 201-2. For example, the feature vector of the word indicating each part of speech may be output as the part of speech feature vector. In this case, the model applied to the embedded layer 202-1 can be applied to the embedded layer 202-2 as it is. In the part-speech feature two-dimensional array 302 shown in FIG. 4, the direction in which the numerical values constituting the feature vector of one part of speech are arranged is defined as the x-axis, and the direction in which a plurality of part-speech feature vectors are arranged in word order is defined as the y-axis.

The connecting layer 203 connects the word feature two-dimensional array 301 and the part-speech feature two-dimensional array 302 output from the embedded layers 202-1 and 202-2 in the height direction. The reshape layer 204 converts the shape of the concatenated array into a three-dimensional shape by halving the number of elements in the x-axis direction of the concatenated array and setting the number of elements in the z-axis direction to 2. Therefore, the feature three-dimensional array 310 shown in FIG. 4 is output. The word feature two-dimensional array 301 and the part-speech feature two-dimensional array 302 may be first converted into a three-dimensional array by the reshape layer. The number of elements in the z-axis direction of each of the three-dimensional array of word features and the three-dimensional array of part-speech features output by this conversion is 1. Then, the three-dimensional array of word features and the three-dimensional array of part-speech features may be connected in the z-axis direction by a connecting layer, and the feature three-dimensional array 310 may be output. In the feature three-dimensional array 310, the word feature two-dimensional array 301 and the part-speech feature two-dimensional array 302 are stacked in the z-axis direction. That is, the reshape layer 204 separates the word feature two-dimensional array 301 and the part-speech feature two-dimensional array 302 into separate channels. Features in the z-axis direction The size of the three-dimensional array 310 is called the depth. Features The depth of the three-dimensional array 310 matches the number of channels. For each word constituting the text, the y-coordinate of the feature vector of the word in the feature three-dimensional array 310 and the y-coordinate of the feature vector of the part of speech of the word match. For example, the feature vector of the first word "luxury" in the text and the feature vector of the part of speech "noun" of the "luxury" overlap in the z-axis direction.

The feature extraction unit 22 includes a plurality of convolution layers 205, a plurality of MAX pooling layers 206, and a connecting layer 207. Each convolution layer 205 is a convolution layer capable of accepting the feature three-dimensional array 310 for each text. For example, the convolution layer 205 may be a multi-channel 2D convolution layer or a single-channel 3D convolution layer. Each convolution layer 205 convolves the feature three-dimensional array 310 output from the conversion unit 21 and outputs a text feature map. In each convolution layer 205, a three-dimensional filter is applied to the feature three-dimensional array 310. The size of the filter varies between the convolution layers 205. FIG. 5 is a diagram showing a configuration example of the filter. The width and depth of each of the filters 410 to 430 shown in FIG. 5 coincide with the width and depth of the feature three-dimensional array 310. The heights of the filters 410-430 are different from each other. Each of the filters 410 to 430 slides in the height direction with respect to the feature three-dimensional array 310. In FIG. 4, three convolution layers 205 are shown. However, the number of convolution layers 205 may be two or four or more. The number of MAX pooling layers 206 matches the number of convolution layers 205. Each MAX pooling layer 206 outputs the maximum value among the numerical values in the feature map output from the convolution layer 205 corresponding to the MAX pooling layer 206. The connecting layer 207 concatenates the maximum values output from the plurality of MAX pooling layers 206 and outputs text feature information. This feature information is a one-dimensional array and reflects the words that make up the text and their order, as well as the part of speech and their order.

The determination unit 23 includes one or more fully connected layers 208 and an output layer 209. The fully connected layer 208 performs affine transformation of the feature information of the text output from the feature extraction unit 22 and outputs the validity information. When there are a plurality of fully connected layers 208, the fully connected layers 208 are connected in series and the affine transformation is sequentially executed. Regularization such as dropout may be performed in each fully connected layer 208. In addition, batch normalization may be performed. The output layer 209 outputs the validity information generated by the fully connected layer 208. In the output layer 209, for example, the validity information from the fully connected layer 208 may be converted into a numerical value in the range of 0 to 1 by a sigmoid function.

[1-2-2. Judgment of text validity]
The model reading unit 113 reads the model 2 stored in the storage unit 14 and loads it into the RAM 14c or the VRAM connected to the GPU 18.

The determination target data acquisition unit 114 acquires one or more word information indicating one or more words constituting a given text to be determined, and one or more part of speech information indicating the part of speech of the one or more words. .. For example, when judging the validity of texts created for questionnaires, themes, assignments, etc., the questions, themes or assignments of the questionnaire corresponding to the texts to be judged and the texts collected for creating training data The questions, themes or assignments in the corresponding survey may or may not match. Since there is not always a fixed correct answer as the content indicated by the text, from the viewpoint of judging the validity of the text, what kind of content is required by the text rather than the question, theme or task itself. is important. Therefore, it is desirable that the type of content requested for the text to be determined matches the type of content requested for at least some of the texts collected for the creation of training data. .. In the case of questionnaires, examples of the types of answers required include reasons, pure recollections, impressions, positives, negatives, and noun answers. The text to be determined may be obtained from another computer via a network, for example. Alternatively, the text may be recorded on the recording medium and read into the storage unit 14 via the drive device. The method of generating word information and part of speech information is the same as the method of generating word information and part of speech information in the training data.

The validity information output unit 115 is given by inputting one or more word information and one or more part of speech information acquired by the determination target data acquisition unit 114 into the model 2 read by the model reading unit 113. Outputs validity information indicating the validity of the text. The validity information output unit 115, as each layer in the model 2, executes arithmetic processing corresponding to the layer, calculates the validity of the text, and outputs the validity information. The validity information output unit 115 may store the validity information in, for example, the storage unit 14. Alternatively, the validity information output unit 115 may transmit the validity information to another computer via the network. Alternatively, the validity information output unit 115 may display the validity information on the display unit 17. The validity information is originally a numerical value in the range of 0 to 1. For example, if the validity information is 0.5 or more, the text may be interpreted as valid, and if the validity information is less than 0.5, the text may be interpreted as invalid. The validity information output unit 115 may actually convert the validity information into a numerical value of 0 or 1 and output it. Alternatively, validity may be interpreted in three stages, for example, valid, cautionary, and invalid. Attention indicates text that should not be unequivocally determined as valid or not.

By using the information indicating the part of speech of the words constituting the text for the learning by the model 2 and the judgment of the validity, the accuracy of the judgment of the validity of the text can be improved. One of the reasons for improving the determination accuracy is that hints for words contained in the text are given. For texts containing words that do not appear at all or appear infrequently in the training data, it may be difficult to properly judge the validity by machine learning using only word information. On the other hand, the part-speech information takes into account the relationship between the part-speech of each word and the part-speech of the words before and after it, that is, the context based on the part-speech. It is presumed that it will be possible to judge the validity. Another reason may be that the grammar used in the text is taken into account. It is expected that the grammar used for valid text will be limited to some extent depending on the situation or the type of content required. It is presumed that the validity of the text is judged not from the viewpoint of whether the text is grammatically correct, but from the viewpoint of whether the grammar used for the text matches the situation or the type of content required. Will be done.

[1-3. Operation of validation device]
Next, the operation of the validity determination device 1 will be described with reference to FIGS. 6 and 7. FIG. 6 is a flowchart showing an example of learning processing by the system control unit 11 and the GPU 18 of the validity determination device 1. The system control unit 11 and the GPU 18 execute the learning process according to the model generation program. As shown in FIG. 6, the training data acquisition unit 111 acquires a plurality of texts as training data and the validity label corresponding to each text from the storage unit 14 (step S11). Next, the training data acquisition unit 111 extracts words constituting each text by morphological analysis, and identifies the part of speech of each extracted word. Then, the training data acquisition unit 111 generates one or more word information and one or more part of speech information for each text (step S12). Next, the model generation unit 112 generates the model 2 by learning using a plurality of training data including word information, part of speech information, and validity label (step S13). For example, the model generation unit 112 inputs word information and part-of-speech information to the convolutional neural network that serves as the model 2, and executes arithmetic processing in each layer in the network. The model generator 112 calculates the validity of each text, compares the validity information with the validity label, and updates the weights and biases of each layer by error backpropagation. The model generation unit 112 ends learning when, for example, the number of epochs reaches a predetermined number, and generates the model 2. The model generation unit 112 stores the generated model 2 in the storage unit 14 (step S14), and ends the learning process.

FIG. 7 is a flowchart showing an example of validation processing by the system control unit 11 and GPU 18 of the validation device 1. The system control unit 11 and the GPU 18 execute the validation process according to the validation program. As shown in FIG. 7, the model reading unit 113 reads the model 2 from the storage unit 14 (step S21). Next, the determination target data acquisition unit 114 acquires a text for determining validity (step S22). Next, the determination target data acquisition unit 114 extracts the words constituting this text by morphological analysis, and identifies the part of speech of each extracted word. Then, the determination target data acquisition unit 114 generates one or more word information and one or more part of speech information (step S23). Next, the validity information output unit 115 generates a word feature two-dimensional array 301 including one or more word feature vectors from one or more word information in the conversion unit 21 of the model 2, and from one or more part of speech information, A part-of-speech feature two-dimensional array 302 including one or more part-speech feature vectors is generated. The validity information output unit 115 generates a feature three-dimensional array 310 in which the word feature two-dimensional array 301 and the part-speech feature two-dimensional array 302 are superimposed (step S24). Next, the validity information output unit 115 convolves the feature three-dimensional array 310 with a plurality of filters in the feature extraction unit 22 of the model 2 to generate a plurality of feature maps (step S25). The validity information output unit 115 concatenates the maximum values of each of the plurality of feature maps to generate text feature information (step S26). Next, the validity information output unit 115 performs affine transformation of the feature information of the text in the determination unit 23 of the model 2 (step S27). The validity information output unit 115 outputs the validity information generated by this conversion (step S28), and ends the validity determination process.

As described above, according to the present embodiment, the validity determination device 1 has one or more word information indicating one or more words constituting the text, and one or more part of speech indicating the part of speech of the one or more words. A plurality of training data are acquired from the storage unit 14 that stores the information and a plurality of training data including the information. Further, the validity determination device 1 generates a model 2 that outputs validity information indicating the validity of a given text by machine learning from a plurality of acquired training data. Therefore, by using the generated model 2, the accuracy of determining the validity of the text can be improved.

Further, even if the validity determination device 1 generates a word feature two-dimensional array including one or more word feature vectors indicating the features of each of the one or more words from one or more word information for each of the plurality of training data. Good. Further, the validity determination device 1 associates one or more part of speech information with the word feature vector of the word corresponding to the part of speech indicated by the part of speech information among the one or more word feature vectors, and the word feature is two-dimensional. Model 2 may be generated by inputting an array and one or more part-speech information into the convolution layer 205 for machine learning. In this case, the feature vector of the word and the information of the part of speech of the word are convoluted in an associative state, and the feature of the text is output. Model 2 is optimized based on the validity judgment result using the characteristics of this text. Therefore, the accuracy of determining the validity of the text can be improved.

Further, the validity determination device 1 may further generate a two-dimensional array of part of speech features including a plurality of part of speech feature vectors indicating the features of each of one or more part of speech from one or more part of speech information. Further, the validity determination device 1 may input the feature three-dimensional array formed by superimposing the word feature two-dimensional array and the part-speech feature two-dimensional array into the convolution layer 205. In this case, a three-dimensional filter is applied in the convolution layer 205 to the three-dimensional array composed of the two-dimensional array including the word feature vector and the two-dimensional array including the part-speech feature vector, and the text feature is obtained. It is output. Therefore, the accuracy of determining the validity of the text can be improved.

The text may also indicate the answer to the question in the questionnaire. The validation device 1 may output validity information indicating the validity of the text as an answer to a question in a given questionnaire. In this case, the validity of the answer to the question in the questionnaire can be judged.

Further, the validity determination device 1 may read the model 2 from the storage unit 14 that stores the generated model 2. Further, even if the validity determination device 1 acquires one or more word information indicating one or more words constituting the text to be determined, and one or more part of speech information indicating the part of speech of the one or more words. Good. Further, the validity determination device 1 inputs the acquired one or more word information and one or more part of speech information into the read model 2, thereby providing validity information indicating the validity of a given text. It may be output. In this case, the validity of the text is determined by using the model 2 generated by machine learning using the part of speech information in addition to the word information. Therefore, the accuracy of determining the validity of the text can be improved.

[2. Second Embodiment]
Next, the second embodiment will be described with reference to FIGS. 8 and 9. At the time of learning, the model generation unit 112 inputs a feature two-dimensional array formed by connecting the word feature two-dimensional array and the part-speech feature two-dimensional array into the convolution layer as another example of associating the part-speech information with the word feature vector. You may.

FIG. 8 is a diagram showing a configuration example of the model 2. In FIG. 8, the same elements as those in FIG. 3 are designated by the same reference numerals. As shown in FIG. 8, the model 2 generated in the present embodiment includes a conversion unit 21-2, a feature extraction unit 22-2, and a determination unit 23.

The conversion unit 21-2 includes input layers 201-1 and 201-2, embedded layers 202-1 and 202-2, and a connecting layer 211. FIG. 9 is a diagram showing an example of information generated by the conversion unit 21-2. In FIG. 9, the same elements as those in FIG. 4 are designated by the same reference numerals. The connecting layer 211 connects the word feature two-dimensional array 301 and the part-speech feature two-dimensional array 302 output from the embedded layers 202-1 and 202-2 in the width direction, and outputs the feature two-dimensional array 320. That is, a one-channel two-dimensional array is output from the connection layer 211. For each word constituting the text, the y-coordinate of the feature vector of the word in the feature two-dimensional array 320 and the y-coordinate of the feature vector of the part of speech of the word match. For example, the feature vector of the first word "luxury" in the text and the feature vector of the part of speech "noun" of the "luxury" are connected in the x-axis direction.

The feature extraction unit 22-2 includes a plurality of convolution layers 212, a plurality of MAX pooling layers 206, and a connecting layer 207. Each convolution layer 212 is a convolution layer capable of accepting the feature two-dimensional array 320 for each text. That is, each convolution layer 212 is a 1-channel 2D convolution layer. In each convolution layer 212, a two-dimensional filter is applied to the feature two-dimensional array 320. The filter may slide only in the height direction with respect to the feature two-dimensional array 320, or may slide in both the height direction and the width direction. The height of the filter differs between the convolution layers 212. The width of each filter may be the same as or smaller than the width of the feature two-dimensional array 320. The width of the filter may be the same or different between the convolution layers 212. The feature map output from each convolution layer 212 is input to the MAX pooling layer 206 corresponding to the convolution layer 212.

Similarly, in the determination of the validity of the text, the validity information output unit 115 also inputs the feature two-dimensional array formed by connecting the word feature two-dimensional array and the part-speech feature two-dimensional array to the convolution layer.

As described above, according to the present embodiment, the validity determination device 1 further includes one or more part of speech feature vectors indicating the characteristics of each of one or more part of speech from one or more part of speech information. Generate an array. Further, the validity determination device 1 inputs a two-dimensional array formed by connecting the two-dimensional array of word features and the two-dimensional array of part-speech features to the convolution layer 212. Therefore, a two-dimensional filter is applied in the convolution layer to the two-dimensional array composed of the two-dimensional array including the word feature vector and the two-dimensional array including the part-speech feature vector, and the text feature is output. To. Therefore, the accuracy of determining the validity of the text can be improved.

[3. Third Embodiment]
Next, the third embodiment will be described with reference to FIGS. 10 and 11. At the time of learning, the model generation unit 112 generates one numerical value indicating each part of speech of one or more words constituting the text as yet another example of associating the part of speech information with the word feature vector, and generates a numerical value of each part of speech. One or more word feature vectors may be added to the word feature vector of the word corresponding to the part of speech, and a two-dimensional array may be input to the convolution layer.

FIG. 10 is a diagram showing a configuration example of the model 2. In FIG. 10, the same elements as those in FIG. 8 are designated by the same reference numerals. As shown in FIG. 10, the model 2 generated in the present embodiment includes a conversion unit 21-3, a feature extraction unit 22-2, and a determination unit 23.

The conversion unit 21-3 includes input layers 201-1 and 201-2, an embedded layer 202-1, an embedded layer 213, and a connecting layer 211. FIG. 11 is a diagram showing an example of information generated by the conversion unit 21-3. In FIG. 11, the same elements as those in FIG. 4 are designated by the same reference numerals. The embedded layer 213 outputs a one-dimensional array 303 composed of a plurality of numerical values corresponding to each of the plurality of part of speech information output from the input layer 201-2. For example, the numerical value corresponding to the part of speech may be a numerical value corresponding to the position where 1 is stored in the one-hot vector of the part of speech, or may be a numerical value calculated by another method. The connecting layer 211 connects the word feature two-dimensional array 301 output from the embedded layer 202-1 and the one-dimensional array 303 output from the embedded layer 213 in the width direction, and outputs the feature two-dimensional array 330. .. That is, a one-channel two-dimensional array is output from the connection layer 211. For each word constituting the text, the y-coordinate of the feature vector of the word in the feature two-dimensional array 330 and the y-coordinate of the numerical value of the part of speech of the word match. For example, the feature vector of the first word "luxury" in the text and the numerical value of the part of speech "noun" of the "luxury" are connected in the x-axis direction. The feature two-dimensional array 320 is input to each convolution layer 212 of the feature extraction unit 22-2.

In determining the validity of the text, the validity information output unit 115 also adds a numerical value indicating each of one or more part of speech to the word feature vector of the word corresponding to the part of speech among the one or more word feature vectors. , Input the two-dimensional array into the convolution layer.

As described above, according to the present embodiment, the validity determination device 1 adds a numerical value indicating each of one or more part of speech to the word feature vector of the word corresponding to the part of speech among the one or more word feature vectors. The two-dimensional array generated in the above process is input to the convolution layer 212. Therefore, each word feature vector is convoluted with a numerical value indicating a part of speech added. Therefore, the accuracy of determining the validity of the text can be improved.

[4. Fourth Embodiment]
Next, the fourth embodiment will be described with reference to FIGS. 12 to 14. The training data acquisition unit 111 may acquire training data including a plurality of part-speech information indicating a plurality of part-speech layers layered by a predetermined part-speech system for each of one or more words constituting the text. In the layered part of speech, the lower part of speech is the part of speech classification corresponding to the target word among the plurality of part of speech classifications that divide the upper part of speech. The number of layers of part of speech may be 2 or more. In the case of Japanese, the part-of-speech system may be, for example, an IPA (Information-technology Promotion Agency) part-of-speech system.

The model generation unit 112 has a word feature two-dimensional array and a plurality of part-speech information in a state in which a plurality of part-speech information corresponding to the word is associated with the word feature vector of the word for each of one or more words constituting the text. Is entered in the convolution layer. For example, the model generation unit 112 may input a feature three-dimensional array formed by superimposing a word feature two-dimensional array and a plurality of part-speech feature two-dimensional arrays into the convolution layer.

FIG. 12 is a diagram showing a configuration example of the model 2. In FIG. 12, the same elements as those in FIG. 3 are designated by the same reference numerals. As shown in FIG. 12, the model 2 generated in the present embodiment includes a conversion unit 21-4, a feature extraction unit 22, and a determination unit 23. The conversion unit 21-4 includes an input layer 201-1, an input layer 2012-1 to 201-2-N, an embedded layer 202-1 and an embedded layer 2022-1 to 202-2-N. , The connecting layer 203 and the reshape layer 204 are included. N is the number of layers of part of speech. The number of input layers 2012-1 to 201-2-N and the number of embedded layers 2022-1 to 202-2-N correspond to the number of layers of part of speech.

Of the plurality of layered part of speech, the part of speech information of the layer corresponding to the input layer is input to each of the input layers 201-2-1 to 201-2-N. The embedded layer 2022-1 to 202-2-N corresponds to each of a plurality of part of speech information output from the input layer corresponding to the embedded layer among the input layers 2012-1-2 to 201-2-N. Outputs a two-dimensional array of part-speech features that includes multiple part-speech feature vectors. FIG. 13 is a diagram showing an example of information generated by the conversion unit 21-4. In FIG. 13, the same elements as those in FIG. 4 are designated by the same reference numerals. Part of speech features The two-dimensional array 302-1 is a two-dimensional array output from the embedded layer 2022-1, and corresponds to the part of speech of the first layer. Part of speech features The two-dimensional array 302-2 is a two-dimensional array output from the embedded layer 202-2-2, and corresponds to the part of speech of the second layer. Part of speech features The two-dimensional array 302-N is a two-dimensional array output from the embedded layer 202-2-N, and corresponds to the part of speech of the Nth layer.

FIG. 14 is a diagram showing an example of information generated by the conversion unit 21-4. The connecting layer 203 has the heights of the word feature two-dimensional array 301 output from the embedded layers 202-1 and 2022-1 to 202-2-N and the part-speech feature two-dimensional array 302-1 to 302-N. Connect in the direction. The reshape layer 204 outputs the feature three-dimensional array 340 shown in FIG. 14 by converting the shape of the connected array into a three-dimensional shape. That is, the reshape layer 204 separates the word feature two-dimensional sequences 301, 2022-1 to 202-2-N into separate channels. The feature three-dimensional array 340 is input to each convolution layer 205 of the feature extraction unit 22. The number of channels in each convolution layer 205 is N + 1.

Similarly, in the determination of the validity of the text, the determination target data acquisition unit 114 also has a plurality of part of speech information indicating a plurality of part of speech layered by a predetermined part of speech system for each of one or more words constituting the determination target text. To get. The validity information output unit 115 is a word feature two-dimensional array in a state in which a plurality of part-speech information corresponding to each of one or more words constituting the text to be determined is associated with the word feature vector of the word. And input multiple part of speech information into the convolution layer.

12 to 14 show an example in which the layering of part of speech is applied to the first embodiment. However, the layering of part of speech may be applied to the second or third embodiment. For example, a feature two-dimensional array formed by connecting a word feature two-dimensional array and a plurality of part-speech feature two-dimensional arrays in the width direction may be input to the convolution layer. In addition, a plurality of numerical values indicating a part of speech of a plurality of layers corresponding to each of one or more words constituting the text are generated, and the numerical values of the plurality of part of speech are words corresponding to those part of speech in the word feature vector of one or more. A two-dimensional array added to the word feature vector of may be input to the convolution layer.

As described above, according to the present embodiment, each of the plurality of training data includes a plurality of part of speech information indicating a plurality of part of speech layered by a predetermined part of speech system for each of one or more words. Therefore, the model 2 can be given hierarchically the hints about the words included in the text and the grammar used in the text by the plurality of part-speech information. Therefore, the accuracy of determining the validity of the text can be improved.

[5. Fifth Embodiment]
Next, the fifth embodiment will be described with reference to FIGS. 15 and 16. The training data acquisition unit 111 may acquire training data further including a request content type indicating the type of content required to be shown in the text. The required content varies depending on the content of the questions, themes, issues, etc. in the questionnaire. For example, in the case of a questionnaire, the request content type may be a questionnaire type indicating the type of answer required for a question in the questionnaire. Examples of questionnaire types include those that ask you to answer reasons, those that ask you to answer pure recollections of something, those that give you impressions, those that ask you to answer positive things, those that ask you to answer negative things, and those that ask you to answer with nouns. And so on. Each text may correspond to only one request content type, or may correspond to a plurality of request content types.

FIG. 15 is a diagram showing a configuration example of the model 2. In FIG. 15, the same elements as those in FIG. 3 are designated by the same reference numerals. As shown in FIG. 15, the model 2 generated in the present embodiment includes a conversion unit 21, a feature extraction unit 22, and a determination unit 23-5.

The determination unit 23-5 includes an input layer 221, a connection layer 222, one or more fully connected layers 208, and an output layer 209. Request content types such as questionnaire types are input to the input layer 221. FIG. 16 is a diagram showing an example of information generated by the determination unit 23-5. The questionnaire type 520 shown in FIG. 16 is, for example, a one-dimensional array having 6 elements. The six elements of the questionnaire type 520 correspond to the above-mentioned types, respectively. For example, each element may contain 1 if the text falls under that type, and 0 if the text does not fall under that type. The connection layer 222 connects the feature information 510 of the text output from the feature extraction unit 22 and the questionnaire type 520 output from the input layer 221 to output a one-dimensional array 530. The one-dimensional array 530 is input to the fully connected layer 208. Therefore, the determination unit 23-5 outputs validity information indicating the validity of the text based on the feature information 510 of the text and the questionnaire type 520.

In the determination of the validity of the text, the determination target data acquisition unit 114 also similarly acquires the request content type indicating the type of content required to be shown in the text. The validity information output unit 115 outputs the validity information by further using the request content type, as in the case of the model generation unit 112.

15 and 16 show an example when the use of the request content type is applied to the first embodiment. However, the use of the request content type may be applied to the second to fourth embodiments.

As described above, according to the present embodiment, each of the plurality of training data further includes a required content type indicating the type of content required to be shown in the text. Therefore, it is possible to improve the accuracy of determining the validity of the text when the part of speech of the word included in the text and the grammar used in the text are different depending on the type of content required.

[6. 6th Embodiment]
Next, the sixth embodiment will be described with reference to FIGS. 17 and 18. The training data acquisition unit 111 constitutes one or more request content word information indicating one or more words constituting the request content text indicating what kind of content is required to be shown in the text, and one request content text. Training data including one or more required content part-speech information indicating the part-speech of the above-mentioned words may be acquired. The request content text is, for example, a text indicating a questionnaire question, theme, task, or the like.

FIG. 17 is a diagram showing a configuration example of the model 2. In FIG. 17, the same elements as those in FIG. 3 are designated by the same reference numerals. As shown in FIG. 17, the model 2 generated in the present embodiment includes a conversion unit 21-6, a feature extraction unit 22, and a determination unit 23. FIG. 18 is a diagram showing an example of information generated by the conversion unit 21-6. The conversion unit 21-6 includes an input layer 201-1 to 201-4, an embedded layer 202-1 to 202-4, a connecting layer 231 to 233, and a reshape layer 234. In the input layer 201-3, as an example of the request content word information, question word information indicating one or more words constituting the text indicating the question of the questionnaire is input. In the input layer 201-4, as an example of the request content part of speech information, question part of speech information indicating the part of speech of one or more words constituting the text indicating the question of the questionnaire is input. The embedded layer 202-3 generates a question word feature two-dimensional array 304 composed of a plurality of question word feature vectors corresponding to each of the plurality of question word information output from the input layer 201-3. The embedded layer 202-4 generates a question part speech feature two-dimensional array 305 composed of a plurality of question part speech feature vectors corresponding to each of the plurality of question part speech information output from the input layer 201-4. The connecting layer 231 connects the word feature two-dimensional array 301 output from the embedded layers 202-1 and 202-3 and the question word feature two-dimensional array 304 in the width direction, and outputs the first two-dimensional array. .. The connecting layer 232 connects the part-speech feature two-dimensional array 302 and the question part-speech feature two-dimensional array 305 output from the embedded layers 202-2 and 202-4 in the width direction, and outputs a second two-dimensional array. .. The connecting layer 233 outputs a third two-dimensional array by connecting the first and second two-dimensional arrays output from the connecting layers 231 and 232 in the height direction. The reshape layer 234 outputs the feature three-dimensional array 360 shown in FIG. 18 by converting the shape of the third two-dimensional array output from the connecting layer 233 into a three-dimensional shape. In the feature three-dimensional array 360, the word feature two-dimensional array 301 and the question word feature two-dimensional array 304 are connected by the same channel. Further, the part-of-speech feature two-dimensional array 302 and the question part-of-speech feature two-dimensional array 305 are connected by the same channel. A two-dimensional array is formed in separate channels for words and part of speech. The feature three-dimensional array 360 is input to each convolution layer 205 of the feature extraction unit 22.

In the determination of the validity of the text, the determination target data acquisition unit 114 also acquires the request content word information and the request content part of speech information in the same manner. The validity information output unit 115 outputs the validity information by further using the request content word information and the request content part of speech information, as in the case of the model generation unit 112.

17 and 18 show an example in which the addition of the required content word information and the required content part of speech information is applied to the first embodiment. However, the addition of the required content word information and the required content part of speech information may be applied to the second to fourth embodiments. For example, the word feature two-dimensional array 301 and the part lyrics feature two-dimensional array 302 are connected in the width direction to generate a two-dimensional array for text, and the question word feature two-dimensional array 304 and the question part lyrics feature two-dimensional array are generated. A two-dimensional array for the text for the question is generated by connecting the 305 and the 305 in the width direction, and a two-dimensional array formed by connecting these generated two-dimensional arrays in the width direction is input to the convolution layer 212. You may. Further, a numerical value indicating the part of speech of the word constituting the text is added to the word feature two-dimensional array 301, and a numerical value indicating the part of speech of the word constituting the text indicating the question is added to the question part of speech feature two-dimensional array 305. Then, a two-dimensional array formed by connecting these two-dimensional arrays in the width direction may be input to the convolution layer 212. In addition, a plurality of part-speech information indicating a plurality of layered part of speech is acquired for each of one or more words constituting the text, and a plurality of layered parts of speech are obtained for each of the one or more words constituting the request content text. A plurality of request contents part-speech information indicating the part-speech of the above may be acquired.

As described above, according to the present embodiment, each of the plurality of training data indicates one or more words constituting the request content text indicating what kind of content is required to be shown in the text. It further includes request content word information and one or more request content part of speech information indicating the part of speech of the one or more words. Therefore, the model 2 is generated by machine learning using the word information and the part-speech information of the text indicating what kind of content is required to be shown. Therefore, since the text indicating what kind of content is required to be shown and the text created for this request are learned in pairs, the accuracy of determining the validity of the text can be improved. ..

[7. Seventh Embodiment]
Next, the seventh embodiment will be described with reference to FIGS. 19 and 20. The training data acquisition unit 111 may acquire training data further including one or more character number information indicating the number of characters of each one or more words constituting the text. The training data acquisition unit 111 may count the number of characters of each word to generate character number information, or the storage unit 14 may store the character number information in advance.

Similar to the case of part of speech information, the model generation unit 112 associates each of one or more character number information with the word feature vector of the word corresponding to the number of characters indicated by the character number information among the one or more word feature vectors. Then, the word feature two-dimensional array and one or more character number information may be input to the convolution layer.

FIG. 19 is a diagram showing a configuration example of the model 2. In FIG. 19, the same elements as those in FIG. 3 are designated by the same reference numerals. As shown in FIG. 19, the model 2 generated in the present embodiment includes a conversion unit 21-7, a feature extraction unit 22, and a determination unit 23. FIG. 20 is a diagram showing an example of information generated by the conversion unit 21-7. The conversion unit 21-7 includes input layers 201-1, 201-2, and 201-5, embedded layers 202-1, 202-2, and 202-5, a connecting layer 203, and a reshape layer 204. .. Character number information is input to the input layer 201-5. The embedded layer 202-5 generates a character number feature two-dimensional array 306 composed of a plurality of character number feature vectors corresponding to each of the plurality of character number information output from the input layer 201-5. For example, a word feature vector indicating the number of characters may be output as a character count feature vector. The connecting layer 203 has the word feature two-dimensional array 301, the part-speech feature two-dimensional array 302, and the character number feature two-dimensional array 306 output from the embedded layers 202-1, 202-2, and 202-5 in the height direction. Connect to. The reshape layer 204 outputs the feature three-dimensional array 370 shown in FIG. 20 by converting the shape of the connected array into a three-dimensional shape. In the feature three-dimensional array 370, the word feature two-dimensional array 301, the part-speech feature two-dimensional array 302, and the character number feature two-dimensional array 306 are separated into separate channels. The feature three-dimensional array 370 is input to each convolution layer 205 of the feature extraction unit 22.

In the determination of the validity of the text, the determination target data acquisition unit 114 also acquires the character number information further. The validity information output unit 115 outputs the validity information by further using the character number information, as in the case of the model generation unit 112.

19 and 20 show an example in which the addition of character number information is applied to the first embodiment. However, the addition of character number information may be applied to the second to sixth embodiments. For example, a two-dimensional array formed by connecting the word feature two-dimensional array 301, the part-speech feature two-dimensional array 302, and the character number feature two-dimensional array 306 in the width direction may be input to the convolution layer 212. Further, a two-dimensional array in which a numerical value indicating a part of speech and a numerical value indicating the number of characters are added to the word feature two-dimensional array 301 may be input to the convolution layer 212. Further, along with the character number information, a plurality of part-speech information indicating a plurality of layered part-speech may be acquired. Further, the request content type may be acquired together with the character number information. In addition, word information, part lyrics information, and number of characters information are acquired for the words that make up the text, and for the words that make up the request content text, the request content word information, the request content part lyrics information, and the words that make up the request content text are obtained. Character number information indicating the number of characters of may be acquired.

As described above, according to the present embodiment, each of the plurality of training data further includes one or more character number information indicating the number of characters of each one or more words. Therefore, the number of characters information can provide model 2 with further hints about the words contained in the text. Therefore, the accuracy of determining the validity of the text can be improved.

[8. 8th Embodiment]
Next, the eighth embodiment will be described with reference to FIGS. 21 and 22. The training data acquisition unit 111 may acquire training data further including one or more character type number information indicating the type of characters included in each of one or more words constituting the text. The training data acquisition unit 111 may specify the type of the character included in each word to generate the character type information, or the storage unit 14 may store the character type information in advance. For example, in the case of Japanese, the character type information may include kanji, hiragana, katakana, romaji, numbers, etc. Good. The training data acquisition unit 111 may acquire a plurality of character type information for each of one or more words constituting the text.

Similar to the case of part of speech information, the model generation unit 112 associates each of one or more character type information with the word feature vector of the word corresponding to the character type indicated by the character type information among the one or more word feature vectors. Then, the word feature two-dimensional array and one or more character type information may be input to the convolution layer.

FIG. 21 is a diagram showing a configuration example of the model 2. In FIG. 21, the same elements as those in FIG. 3 are designated by the same reference numerals. As shown in FIG. 21, the model 2 generated in the present embodiment includes a conversion unit 21-8, a feature extraction unit 22, and a determination unit 23. FIG. 22 is a diagram showing an example of information generated by the conversion unit 21-8. The conversion unit 21-8 includes input layers 201-1, 201-2, and 201-6, embedded layers 202-1, 202-2, and 202-6, a connecting layer 203, and a reshape layer 204. .. Character type information is input to the input layer 201-6. This character type information indicates, for example, whether or not the word contains Chinese characters. The embedded layer 202-6 generates a character type feature two-dimensional array 307 composed of a plurality of character type feature vectors corresponding to each of the plurality of character type information output from the input layer 201-6. For example, the feature vector of the word indicating the character type may be output as the character type feature vector. When a plurality of character type information exists for each word, the conversion unit 21-8 may include a plurality of input layers 201-6 and a plurality of embedded layers 202-6 corresponding to the plurality of character type information. The connecting layer 203 has the word feature two-dimensional array 301, the part-speech feature two-dimensional array 302, and the character type feature two-dimensional array 307 output from the embedded layers 202-1, 202-2, and 202-6 in the height direction. Connect to. The reshape layer 204 outputs the feature three-dimensional array 380 shown in FIG. 22 by converting the shape of the connected array into a three-dimensional shape. In the feature three-dimensional array 380, the word feature two-dimensional array 301, the part-speech feature two-dimensional array 302, and the character type feature two-dimensional array 307 are separated into separate channels. The feature three-dimensional array 380 is input to each convolution layer 205 of the feature extraction unit 22.

In the determination of the validity of the text, the determination target data acquisition unit 114 also acquires the character type information further. The validity information output unit 115 outputs the validity information by further using the character type information, as in the case of the model generation unit 112.

21 and 22 show an example in which the addition of character type information is applied to the first embodiment. However, the addition of character type information may be applied to the second to seventh embodiments. The application example of adding the character type information to the second embodiment to the sixth embodiment is the same as the application example of adding the number of characters. Further, the character number information may be acquired together with the character type information.

As described above, according to the present embodiment, each of the plurality of training data further includes one or more character type information indicating the type of character included in each of one or more words. Therefore, according to the present invention, the character type information can provide the model 2 with further hints about the words contained in the text. Therefore, the accuracy of determining the validity of the text can be improved.

In each of the above embodiments, the present invention has been used to determine the validity of the text. However, the present invention may be applied to the classification of texts. The present invention may also be applied to various text classifications other than texts for which validity matters. Either binary classification or multi-class classification may be applied. For text classification, the training data includes a label indicating the category to which the text belongs, instead of a validity label. Model 2 outputs information indicating the classification result. The determination of the validity of the text is an example of text classification.

1 Validity judgment device 11 System control unit 12 System bus 13 Input / output interface 14 Storage unit 15 Communication unit 16 Input unit 17 Display unit 18 GPU
111 Training data acquisition unit 112 Model generation unit 113 Model reading unit 114 Judgment target data acquisition unit 115 Validity information output unit 2 Model 21, 21-1 to 4, 21-6 to 21-8 Conversion unit 22, 22-2 Features Extraction unit 23, 23-5 Judgment unit 205, 212 Convolution layer

Claims (16)

One or more word information indicating one or more words constituting the first text, one or more part of speech information indicating the part of speech of the one or more words, and the validity of the first text as an answer to a question in a questionnaire. A training data acquisition means for acquiring the plurality of training data from a storage means for storing a plurality of training data including each of the indicated labels.
With respect to each of the acquired plurality of training data, a feature information generating means for generating two-dimensional array word information including one or more word feature vectors indicating the features of each of the one or more words from the one or more word information. ,
By machine learning from the plurality of acquired training data, a model that outputs second validity information indicating the validity of a given second text as an answer to a question in a questionnaire and includes a convolution layer is obtained. The model generation means for generating, for each of the one or more words included in each of the acquired plurality of training data, the word feature vector corresponding to the word and the part code information are arranged side by side. The dimension array word information and the one or more part word information are input to the convolution layer to acquire the feature information indicating the feature of the first text, and the validity of the first text is obtained based on the acquired feature information. A model generation means for training the model by error backpropagation based on outputting the first validity information indicating the above from the model and comparing the first validity information with the label.
With
The size of the filter of the convolution layer in the direction in which the word feature vector corresponding to the same word and the part of speech information are arranged is characterized in that the size of the filter of the convolution layer matches the sum of the size of the word feature vector and the size of the part of speech information. Learning device. The model generating means includes the two-dimensional array word information in which the one or more word feature vectors are arranged according to the order of the one or more words in the first text, and the one or more arranged according to the order. And the part of speech information of, are input to the convolution layer,
The learning device according to claim 1, wherein the size of the filter in the direction in which one or more word information is arranged is a size corresponding to a plurality of words. The feature information generation means further generates two-dimensional array part of speech information including one or more part of speech feature vectors indicating the features of each of the one or more part of speech from the one or more part of speech information.
The learning device according to claim 2, wherein the model generation means inputs three-dimensional array information obtained by superimposing the two-dimensional array word information and the two-dimensional array part of speech information into the convolution layer. The feature information generation means further generates two-dimensional array part of speech information including one or more part of speech feature vectors indicating the features of each of the one or more part of speech from the one or more part of speech information.
The learning device according to claim 2, wherein the model generation means inputs two-dimensional array information formed by connecting the two-dimensional array word information and the two-dimensional array part-speech information to the convolution layer. The model generation means adds a numerical value indicating each of the one or more part of speech to the word feature vector of the word corresponding to the part of speech in the one or more word feature vector, and convolves the two-dimensional array word information. The learning device according to claim 2, wherein the learning device is input to a layer. Any of claims 1 to 5, wherein each of the plurality of training data includes a plurality of said part of speech information indicating a plurality of part of speech layered by a predetermined part of speech system for each of the one or more words. The learning device according to paragraph 1. The invention according to any one of claims 1 to 6, wherein each of the plurality of training data further includes type information indicating what kind of content is required to be shown in the first text. Learning device. Each of the plurality of training data includes one or more second word information indicating one or more second words constituting the third text indicating the question, and one or more second words indicating the part of speech of the one or more second words. 2. The learning device according to any one of claims 1 to 7, further comprising part-speech information. The learning device according to any one of claims 1 to 8, wherein each of the plurality of training data further includes one or more character number information indicating the number of characters of each of the one or more words. The learning apparatus according to any one of claims 1 to 9, wherein each of the plurality of training data further includes one or more character type information indicating a character type included in each of the one or more words. .. The label indicates the validity of the first text as an answer to the first question.
The model generating means indicates the validity of the second text as indicating an answer to a second question requesting an answer of the same type as the answer requested by the first question. The learning device according to any one of claims 1 to 10, wherein the model is generated. A reading means for reading the model from a model storage means for storing the model generated by the learning device according to any one of claims 1 to 11.
A text information acquisition means for acquiring one or more determination word information indicating one or more words constituting a given text, and one or more determination part of speech information indicating the part of speech of the one or more words.
An output means for outputting the validity information indicating the validity of the given text by inputting the acquired one or more judgment word information and one or more judgment part of speech information into the read model. ,
A validation device comprising. In the learning method performed by the computer
One or more word information indicating one or more words constituting the first text, one or more part of speech information indicating the part of speech of the one or more words, and the validity of the first text as an answer to a question in a questionnaire. A training data acquisition step of acquiring the plurality of training data from a storage means for storing a plurality of training data including the labeled label and
With respect to each of the acquired plurality of training data, a feature information generation step of generating two-dimensional array word information including one or more word feature vectors indicating the features of each of the one or more words from the one or more word information. ,
By machine learning from the plurality of acquired training data, a model that outputs second validity information indicating the validity of a given second text as an answer to a question in a questionnaire and includes a convolution layer is obtained. In the model generation step to be generated, for each of the one or more words included in each of the acquired plurality of training data, the word feature vector corresponding to the word and the part code information are arranged side by side. The dimension array word information and the one or more part word information are input to the convolution layer to acquire the feature information indicating the feature of the first text, and the validity of the first text is obtained based on the acquired feature information. A model generation step of training the model by error backpropagation based on outputting the first validity information indicating the above from the model and comparing the first validity information with the label.
Including
A learning method characterized in that the size of the filter of the convolution layer in the direction in which the word feature vector and the part of speech information are aligned matches the sum of the size of the word feature vector and the size of the part of speech information. In the validation method performed by the computer
A read-out step of reading the model from the model storage means for storing the model generated by the learning device according to any one of claims 1 to 11.
A text information acquisition step for acquiring one or more determination word information indicating one or more words constituting a given text, and one or more determination part of speech information indicating the part of speech of the one or more words.
With the output step of outputting the validity information indicating the validity of the given text by inputting the acquired one or more judgment word information and one or more judgment part of speech information into the read model. ,
A validation method comprising. Computer,
One or more word information indicating one or more words constituting the first text, one or more part of speech information indicating the part of speech of the one or more words, and the validity of the first text as an answer to a question in a questionnaire. A training data acquisition means for acquiring the plurality of training data from a storage means for storing a plurality of training data including each of the indicated labels, and a training data acquisition means for acquiring the plurality of training data.
With respect to each of the acquired plurality of training data, a feature information generating means for generating two-dimensional array word information including one or more word feature vectors indicating the features of each of the one or more words from the one or more word information. ,
By machine learning from the plurality of acquired training data, a model that outputs second validity information indicating the validity of a given second text as an answer to a question in a questionnaire and includes a convolution layer is obtained. The model generation means for generating, for each of the one or more words included in each of the acquired plurality of training data, the word feature vector corresponding to the word and the part code information are arranged side by side. The dimension array word information and the one or more part word information are input to the convolution layer to acquire the feature information indicating the feature of the first text, and the validity of the first text is obtained based on the acquired feature information. A model generation means for training the model by error backpropagation based on outputting the first validity information indicating the above from the model and comparing the first validity information with the label.
To function as
A learning program characterized in that the size of the filter of the convolution layer in the direction in which the word feature vector and the part of speech information are aligned matches the sum of the size of the word feature vector and the size of the part of speech information. Computer,
A reading means for reading the model from a model storage means for storing the model generated by the learning device according to any one of claims 1 to 11.
A text information acquisition means for acquiring one or more determination word information indicating one or more words constituting a given text, and one or more determination part of speech information indicating the part of speech of the one or more words.
An output means for outputting the validity information indicating the validity of the given text by inputting the acquired one or more judgment word information and one or more judgment part of speech information into the read model. ,
A validation program characterized by functioning as.

Images