JP6802332B1 - Information processing method and information processing equipment - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technique for easily and accurately classifying a monitor's response to a free-answer questionnaire in online research. An information processing method according to the present invention is an information processing method for classifying answers to free-answer questions included in a questionnaire into any of a plurality of categories, and is for questions created by respondents. A step to obtain an answer consisting of text information, a step to divide the character string included in the answer into characters and encode each, and a trained answer generated by a character-level convolutional neural network in the encoded answer. It has the steps of applying a model to determine which of the multiple categories the respondent's response falls into. [Selection diagram] Fig. 1

Description

The present invention relates to an information processing method and an information processing device.

Traditionally, surveys have been conducted in which questions are sent to a large number of respondents (monitors) to conduct a questionnaire. By aggregating and analyzing the answers obtained from the questionnaire, it is possible to obtain useful information for decision support, preparation of statistical data, academic research, etc. when conducting marketing and product development. In particular, in recent years, a survey method called net research, in which a questionnaire is conducted via a network, has been actively used. Internet research has the advantages that the burden on respondents is relatively light and that it is easy to aggregate the response results.

There are two types of questions in questionnaire surveys, including online research: multiple-choice questions and open-answer questions. The multiple-choice question is one in which the respondent selects an answer from the options prepared in advance, and there are an alternative-choice method and a multiple-answer method. On the other hand, the free answer type is also called a free description type or a free answer (FA) type, and is a method in which the respondent freely describes the answer to the question. Free-answer questions may be relatively simple, with numbers or words, or may require written questions. In addition, a combination of a selection type and a free answer type may be used, such as requesting text input only when "Other" is selected in the selection type question.

The open-ended questions have the following characteristics compared to the multiple-choice questions. The merit is that the free-answer type can obtain answers based on free ideas that the surveyor did not anticipate, while the multiple-choice type can only obtain the answers included in the existing options. It also has the advantage of providing detailed written answers. On the other hand, as a demerit, the selection formula is suitable for automatic aggregation, whereas the free response formula is difficult to quantitatively analyze and classify the response contents. There is also a risk of clerical errors by the respondents. Also, even when referring to the same thing, the terms used may differ depending on the respondent.

Therefore, in order to analyze the result of the free answer, a work called coding is performed. This is the work of classifying the free responses sent from the respondents by category, and assigns a code corresponding to the category as classification information. This makes it possible to make free answers optional and to quantify and analyze qualitative information. Coding is also called "after-coding (AC)" because it is done after free answer.

When coding is done by a natural person, the investigator visually checks the free answers and assigns classifications one by one. However, such manual coding increases time and cost. In addition, the judgment method tends to be personalized. Therefore, when there are a plurality of investigators, the judgment result varies depending on the skill level and the difference in the way of thinking. In addition, if the investigator is changed, it takes time to become proficient in the work, and there is a risk that the judgment results will not be consistent. Therefore, attempts have been made to perform coding automatically.

Patent Document 1 describes a classification device that obtains answers to a free-answer questionnaire and determines the contents using a machine learning model.
In addition, Patent Document 2 also describes a potential needs deriving device that applies a machine learning model to the answers to open-ended questionnaires.

Japanese Unexamined Patent Publication No. 2005-092492 Japanese Unexamined Patent Publication No. 2016-10188

In Patent Document 1, a morphological analysis is performed on a free answer text in Japanese, a machine learning model using a support vector machine is created after detecting a specific expression, and the machine learning model is used for classification.
Further, also in Patent Document 2, the sentences in the free answer column are morphologically analyzed to extract words, and the answer results are divided into a plurality of categories.

However, in these conventional techniques, since it is necessary to perform information processing such as morphological analysis on the free answer text as preprocessing, man-hours increase and appropriate computational resources and processing time are required. In addition, when performing morphological analysis on open-ended text, it is difficult to separate words and phrases in Japanese without word-separation. In addition, morphological analysis requires specialized skills and knowledge, and has problems such as an increase in human cost and the time required for preprocessing, which makes it impossible to concentrate on analysis. In addition, if there are words that are not included in the dictionary (such as when unregistered words appear or when errors are included), the analysis accuracy will decrease, so it is necessary to keep updating the dictionary manually, which is continuous. It becomes a factor of cost increase. In addition, the method of morphological analysis differs depending on the language. Therefore, it becomes necessary to create a learning model for each language.

The present invention has been made in view of the above problems, and an object of the present invention is to provide a technique for easily and accurately classifying a monitor's response to a free-answer questionnaire in online research. ..

In order to achieve the above object, the following configuration is adopted in the present invention. That is,
It is an information processing method that classifies the answers to the open-ended questions included in the questionnaire into one of multiple categories.
The steps to obtain the answer, which consists of textual information for the question, created by the respondent,
The step of dividing the character string included in the answer into characters and encoding each character,
A step of applying a trained model generated by a character-level convolutional neural network to the encoded answer to determine which of the plurality of categories the answer by the respondent falls into.
It is an information processing method characterized by having.

The present invention also employs the following configurations. That is,
It is an information processing method that classifies the answers to the questionnaire into multiple categories.
The survey includes a photo and a question requesting respondents to send comments about the photo.
A step of acquiring the image information of the photograph and the text information of the comment from the answer sent from the respondent, and
A step of dividing the character string included in the text information into characters and encoding each character, and
The step of encoding the image information for each pixel and
A step of applying a trained model generated by a convolutional neural network to the encoded text information and the image information to determine which of the plurality of categories the answer by the respondent falls into. When,
It is an information processing method characterized by having.

The present invention also employs the following configurations. That is,
An information processing device that classifies the answers to free-answer questions included in a questionnaire into one of multiple categories.
An answer receiver that acquires the answer, which consists of text information for the question, created by the respondent.
An answer content acquisition unit that divides the character string included in the answer into characters and encodes each character.
A classification determination unit that applies a trained model generated by a character-level convolutional neural network to the encoded answer to determine which of the plurality of categories the answer by the respondent falls into. ,
It is an information processing apparatus characterized by having.

The present invention also employs the following configurations. That is,
An information processing device that classifies the answers to questionnaires into multiple categories.
The survey includes a photo and a question requesting respondents to send comments about the photo.
An answer receiving unit that acquires the image information of the photograph and the text information of the comment from the answer sent from the respondent.
An answer content acquisition unit that divides the character string included in the text information into characters and encodes each character, and encodes the image information pixel by pixel.
By applying a trained model generated by a convolutional neural network to the coded text information and the image information, the text information and the image information can be subjected to the plurality of answers by the respondent using the same algorithm. A classification judgment unit that determines which of the categories is classified into
It is an information processing apparatus characterized by having.

According to the present invention, it is possible to provide a technique for easily and accurately classifying a monitor's response to an open-ended questionnaire in net research.

It is a figure explaining the overall functional structure of an information processing system. It is a figure explaining the structure of an information processing apparatus and a respondent terminal. It is a flowchart explaining the process of classifying free answer. It is a figure explaining the state of answering using a respondent terminal. It is a figure explaining setting of classification information. It is a flowchart explaining the process of Embodiment 2. It is a figure explaining the state of the answer in Embodiment 2.

A preferred embodiment of the present invention will be described below with reference to the drawings. However, the constituent blocks described below and their relative arrangements should be appropriately changed depending on various conditions of the system to which the invention is applied, and the scope of the invention is limited to the following description. It's not a thing.

The present invention relates to a technique for classifying answers to open-ended questionnaires. The present invention can be regarded as an information processing device or an information processing method using such a technique. The present invention can also be regarded as a program that operates by utilizing the arithmetic resources of the above-mentioned information processing apparatus and executes each process of the information processing method, or a storage medium in which such a program is stored. The storage medium may be a non-temporary storage medium that can be read by a computer. The present invention can also be regarded as an information processing system including the above-mentioned information processing device and a terminal in which a user answers a questionnaire.

In the following embodiments, the surveyor is a person who conducts a questionnaire to respondents (monitors) via the Internet, acquires answers, and analyzes them. The investigator can be any person who surveys a large number of monitors, such as companies, research institutes, or individuals. The researcher may be a company or the like itself, or may be a research company that has received a request from the company or the like.
Respondents are those who respond to questionnaires via the Internet. The research company organizes a large number of monitors as members, and when asked to conduct a survey, it contacts the monitors and conducts a questionnaire. By having such a research company also serve as a researcher, it is possible to smoothly carry out the process from the survey to the implementation of the questionnaire. Internet research companies have the advantage of being able to conduct surveys of a large number of monitors at a relatively low cost and in a short period of time.

[First Embodiment]
<Overall configuration>
The overall configuration of the system of the present embodiment will be described with reference to FIGS. 1 and 2.
The information processing system 1 generally includes an information processing device 10 used by the surveyor 15 and respondent terminals (respondent terminals 20a, 20b, ...) Used by each of a plurality of monitors (respondents 25a, b, ...). doing. The information processing device 10 and each of the respondent terminals 20 are connected to each other so as to be able to communicate with each other via a network such as the Web or a dedicated line.

The information processing device 10 is a device used by a researcher, and has a function of creating a questionnaire, acquiring answers from respondents 25 to the questionnaire, and classifying free responses.
The functional block of the information processing device 10 is shown at the bottom of the page of FIG. As will be described in detail later, the information processing apparatus 10 includes a questionnaire creation unit 101 in which the surveyor 15 creates a questionnaire, a questionnaire transmission unit 102 that transmits a questionnaire to the respondent terminal 20, and an answer reception unit that receives an answer from the respondent terminal 20. Unit 111, answer content acquisition unit 112 that acquires information (text or image) necessary for discrimination from the answers, feature extraction unit 113 that acquires the feature amount from the text, classification judgment unit 114 that gives classification information to the answer, judgment It includes a result output unit 115 that presents the result to the investigator 15, a data creation unit 121 that the investigator 15 creates training data, and a model creation unit 122 that creates a trained model from the training data. The detailed operation of each functional block will be described later. Each functional block does not necessarily have to be a physical entity, and each may be virtually realized as a program module. Further, the division of the functional blocks does not necessarily have to be as shown in the drawing, and it is sufficient that the information processing apparatus 10 can realize the processing of the present embodiment as a whole.

FIG. 2A shows a configuration example of the information processing device 10.
As the information processing device 10, a PC, a server, a workstation, or the like, which has computational resources and operates according to a program instruction or an instruction via an interface, can be used. As the information processing device 10, a cloud server that uses computing resources on the cloud may be used, or a net line or a plurality of directly connected PCs may be combined.

In the illustrated example, the information processing device 10 includes a CPU 1001 that performs information processing, a GPU 1002 that performs image processing, a memory 1003 that temporarily or permanently holds data, and a communication module 1004 that communicates with the outside via the Web or the like. Includes mouse 1005, keyboard 1006, display 1007. An external storage device such as an HDD or SSD may be connected to the information processing device 10 in addition to the memory 1003 or instead of the memory 1003.

Questionnaire creation unit 101, questionnaire transmission unit 102, response reception unit 111, response content acquisition unit 112, feature extraction unit 113, classification determination unit 114, result output unit 115, data creation unit 121, model creation unit 122 shown in FIG. Such a functional block is realized by operating each component of the information processing apparatus 10 shown in FIG. 2A in cooperation with each other.

The respondent terminal 20 is a terminal device for the respondent 25 to answer a question in a questionnaire.
The functional block of the respondent terminal 20 is shown in the upper part of the page of FIG. As will be described in detail later, the respondent terminal 20 includes a questionnaire receiving unit 201 that receives a questionnaire from the information processing device 10, an answer input unit 202 that the respondent 25 inputs an answer, and an answer transmission that transmits the answer to the information processing device 10. A unit 211 is provided. The detailed operation of each functional block will be described later. Each functional block does not necessarily have to be a physical entity, and each may be virtually realized as a program module. Further, the division of the functional blocks does not necessarily have to be as shown in the drawing, and it is sufficient that the respondent terminal 20 can realize the processing of the present embodiment as a whole.

FIG. 2B shows a configuration example of the respondent terminal 20.
As the respondent terminal 20, a PC, a server, a workstation, or the like, which has computational resources and operates according to a program instruction or an instruction via an interface, can be used. In addition, with the development of mobile devices in recent years, smartphones and tablet terminals can also be suitably used.
The figure shows a case where the respondent terminal 20 is a smartphone. The respondent terminal 20 includes a System (System on Chip) 2001 that performs information processing and image processing, a memory 2002, a communication module 2003, a touch panel 2004, and a microphone 2005.

The functional blocks such as the questionnaire receiving unit 201, the answer input unit 202, and the answer transmitting unit 211 shown in FIG. 1 are realized by the cooperative operation of each component of the respondent terminal 20 shown in FIG. 2 (b). Will be done.

<Processing flow>
(Premise)
The processing flow of the present embodiment will be described with reference to FIG. This flow starts when the investigator 15 decides to conduct the questionnaire. In addition, although the questionnaire is conducted in Japanese here, it can be processed in the same way even in other languages. This flow will be described mainly using the functional blocks of FIG.

In the present embodiment, a questionnaire (also referred to as this questionnaire) according to the purpose of the survey is immediately sent to the respondent 25, and the responses are collected. However, a questionnaire for screening (also referred to as a preliminary questionnaire) may be conducted before the main questionnaire to extract the respondents 25 who satisfy the attribute conditions and the history conditions. The technique of the present invention can be applied to both the answer to the preliminary questionnaire and the answer to the present questionnaire.

In this flow, the number of questions included in the questionnaire is one, and the content is to have the respondent 25 answer the name of the car he / she wants to purchase. However, there is no particular limit to the number of questions. In addition, not only open-ended questions but also multiple-choice questions may be included. In that case, the answers to the open-ended questions are processed.
In this flow, the vehicle name freely answered by the respondent 25 is automatically determined by the trained model as to which of the preset classifications.
In this flow, a word-based answer is coded corresponding to that word, but the type of answer is not limited to this, and any answer that may be coded, such as when a sentence or number is included, is used. Anything is the subject of the present invention.

(Start flow)
This flow is roughly divided into a preparatory process for creating a trained model and an implementation process for classifying and discriminating free answers using the created trained model.
Steps S101 to S103 are preparatory steps for creating a trained model using the coded sample data as training data.

In step S101, the investigator 15 codes the sample data using the data creation unit 121 of the information processing apparatus 10 to create the learning data. FIG. 5A is a table showing the classifications, and shows the correspondence between the vehicle name and the code code when coding. This correspondence is stored as a database table. For convenience, only three vehicle names are used here, but the category is not limited to this. In addition, the same model name may be subdivided according to the model, year, and the like. Further, in the case of an automobile, a plurality of categories having different viewpoints may be set for one survey target, such as "vehicle name", "color", and "model year". This makes it possible to search for tags from various perspectives on the survey results, and to analyze from various perspectives.

FIG. 5B is learning data created by the researcher 15 by manually coding the sample data (sample response). The method of acquiring the sample data is arbitrary, and answers to the same type of questions in the past may be acquired, or the investigator 15 may newly create the data. The number of samples depends on the required accuracy and the like, but is set to several hundreds or more, for example. Subsequently, the investigator 15 determines which classification the individual sample data corresponds to, and assigns appropriate classification information. Since the investigator 15 classifies vehicles based on manuals, dictionaries, know-how, etc., as shown in FIG. 5B, appropriate classification information is given even to vehicle names that are not always accurate.
The investigator 15 may enter the correct vehicle name as it is, as shown in FIG. 5D, when assigning the classification information.

In step S102, the investigator 15 digitizes the text information for the convolutional neural network (CNN) by using the model creation unit 122 of the information processing apparatus 10. Then, in step S103, the model creation unit 122 creates a trained model from the digitized text information.

Here, an example of CNN for text information will be described. CNN is a kind of machine learning using a hierarchical neural network, and enables highly accurate information processing by using a laminated structure in which a convolutional layer and a pooling layer are repeated multiple times. A feature of the present invention is that the text information used for learning in CNN is not separated by words or phrases, but is regarded as continuous data of characters and coded for each character. As a coding method, for example, a code point assigned to a character included in a character set of the Unicode standard can be used. However, the character set and the encoding method are not limited as long as the target text information can be encoded. For example, when the number of types of characters included in the character set is relatively small as in English, a method that requires a smaller number of bytes at the time of encoding may be used. If the amount of text differs for each answer, a predetermined number of characters may be obtained from the beginning of the sentence or zero padding may be performed in order to make the text a fixed length and facilitate processing by CNN. By encoding the Unicode character set in this way, various languages including Japanese can be processed in the same way.

Further, as described above, a character-level convolutional neural network (Character-level Convolutional neural network, hereinafter also referred to as "character-level CNN") in which coding is performed for each character is adopted. By adopting the character level CNN, the sentence can be treated as a simple character string. In other words, it is not necessary to consider the meaning and grammatical structure of the sentence, and only the arrangement of characters needs to be considered. Therefore, preprocessing such as morphological analysis performed in Patent Document 1 becomes unnecessary, and the characters are divided one by one. Pre-processing that only encodes is sufficient. Further, since it is not necessary to use a dictionary for sentence analysis, the cost of creating and maintaining the dictionary is not required. Further, as shown in FIG. 5C, the determination accuracy is improved even when typographical errors, typographical errors, and notational fluctuations are included.

The size and number of channels of input data in CNN, the size of filters, the number of filters, the filter movement interval, the number of convolutions and poolings, and various other parameters required for learning are appropriately determined according to the characteristics of the input data and the required discrimination accuracy. be able to. When creating the trained model, the coded sample data may be divided into training data and validation data, or the accuracy may be improved by a cross-validation method.

After step S104, it is a questionnaire implementation process in which a questionnaire is conducted and the trained model is applied to the answers.
In step S104, the investigator 15 creates a questionnaire using the questionnaire creation unit 101 of the information processing device 10. Here, in addition to creating the above-mentioned question text, various conditions such as the flow of questions included in the questionnaire and the display format are set. Then, the questionnaire is transmitted to the respondent terminal 20 using the questionnaire transmission unit 102.

In step S105, the respondent 25 responds to the questionnaire received by the questionnaire receiving unit 201 of the respondent terminal 20 by using the answer input unit 202. FIG. 4 shows a questionnaire and an answer displayed on the touch panel 2004 of the respondent terminal 20. FIG. 4A shows the respondent terminal 20 of this flow, and a smartphone is used here. The respondent 25 uses the touch panel 2004 as the answer input unit 202.

FIG. 4B shows a screen when the respondent 25 inputs an answer. On the touch panel 2004, a question display field 2041, an answer display field 2042, and an answer transmission button 2043 are displayed. The question in this example is to have respondent 25 answer the name of the car he wants to buy. When the respondent 25 inputs the answer in the answer display field 2042 using the IME installed on the smartphone, the respondent 25 presses the answer transmission button 2043 as shown in FIG. 5 (c). As a result, the response transmission unit 211 of the respondent terminal 20 transmits the response to the response reception unit 111 of the information processing device 10 via the network.

In step S106, the investigator 15 uses the response content acquisition unit 112 of the information processing device 10 to acquire information necessary for discrimination from the acquired response. Here, the text information of the free answer is acquired. Then, processing such as digitization by UTF-8 is performed. Subsequently, the feature extraction unit 113 applies the trained model to the text to extract the feature amount. Subsequently, the classification determination unit 114 determines which classification corresponds to the free response, and adds classification information. The classification determination unit 114 corresponds to the fully connected layer to the output layer in the CNN, and since the number of classifications is three in this embodiment, the output destination is also set to three nodes. Then, based on the extracted features, weighting in the hidden layer and calculation using the softmax function in the output layer are performed to calculate the probability of belonging to each classification.

In step S107, the result output unit 115 of the information processing device 10 presents the determination result to the investigator 15. FIG. 5C is a diagram for explaining the relationship between the text and the code in the free answer indicating the vehicle name output as a result of the classification determination and the result presented to the investigator 15. Each open-ended response data contains an inaccurate vehicle name as shown. However, accurate coding can be provided by applying the trained model of the present invention to the investigator. As a method of providing the determination result, the determined data may be saved in the memory or provided online. Further, the determined data and software having an analysis function of the data may be provided in the form of SaaS operating on the cloud.

As described above, in the flow of the present embodiment, the information processing device 10 classifies the text of the free answer based on the result of machine learning. In particular, free answers can be processed uniformly without the need for preprocessing such as morphological analysis. Therefore, even when various free responses including notation fluctuations and typographical errors are targeted, the determination result can be presented in a format that is easy for the investigator 15 to understand while reducing the cost related to time and manpower. Therefore, regardless of the investigator 15's coding proficiency and machine learning knowledge, he / she can concentrate on the analysis of the questionnaire results. Furthermore, not only Japanese but also multiple languages can be processed in the same way.

[Second Embodiment]
The second embodiment will be described focusing on the differences from the first embodiment, and the description of the common parts will be simplified.
The configuration of the information processing system 1 according to the present embodiment is basically the same as that of the first embodiment. The respondent terminal 20 of the present embodiment has a function of transmitting an image to the information processing device 10 when answering a questionnaire, in addition to the function of the first embodiment. When the respondent terminal 20 is a smartphone, it is preferable to enable the respondent 25 to transmit an image taken by the camera of the smartphone. The image may be a moving image.

<Processing flow>
(Premise)
The processing flow of the present embodiment will be described with reference to FIG. 6, focusing on a portion different from the flow of FIG. The questionnaire in this flow is a photo survey questionnaire in which respondents 25 send photos. Such a photo survey has various advantages. For example, it is possible to intuitively grasp the living conditions and behavioral conditions of the respondents 25 as compared with the text-only answers. In addition, it is possible to check the actual usage status of products in the home. In addition, the contents that the respondent 25 cannot answer in words can be understood by looking at the photograph.

In the present embodiment, as shown in FIG. 7A, a questionnaire is conducted in which the furniture (sofa) used at the home of the respondent 25 is photographed and a photograph is requested to be transmitted. In the present embodiment, the sent sofa photograph is classified into either "upholstered" or "leathered" according to the trained model.

(Start flow)
The flow of this embodiment is also roughly divided into a preparatory process for creating a trained model and a questionnaire implementation process for classifying and discriminating photographs. Steps S101 to S103 are preparatory steps for creating a trained model using the coded sample data as training data.

In step S201, the investigator 15 codes the sample data using the data creation unit 121 of the information processing apparatus 10 to create learning data, and saves it as data. As sample data, image information of photographs of various sofas is used. The investigator determines whether the sofa in each photo is upholstered or leather upholstered and gives classification information as a code. The number of categories is not limited to two and may be larger. In addition, various cut categories such as color, shape, furniture manufacturer name, etc. may be provided. Appropriate classification information is given to each photograph by the know-how possessed by the investigator 15.

In step S202, the investigator 15 digitizes the image information of the photograph for the convolutional neural network (CNN) by using the model creation unit 122 of the information processing apparatus 10. Then, in step S203, the model creation unit 122 creates a trained model from the digitized image information.

Here, an example of CNN targeting image information will be described. The input image here is image data encoded by a method in which the pixel value is quantified for each pixel. When processing such image data, the convolution layer performs a convolution operation on the input image to the input layer using a plurality of filters that detect features different from each other. As a result, the feature amount can be calculated by using the locality of the image. In CNN, the filter parameters are automatically learned according to the image recognition target.
The pooling layer subsamples the output of the convolution layer by a method such as max pooling that takes the maximum value within a predetermined region in the image to reduce the image resolution and compress the feature amount. The processing in the pooling layer has the effect of reducing the dependence on the position in the image and reducing the calculation cost. Finally, in the fully connected layer including the output layer, the features are combined and pattern classification is performed, and the result is output. In the generation of such a trained model, the filter weight is calculated and optimized so that the classification result of CNN approaches the correct answer. In that case, any known method such as backpropagation method may be used.

After step S204, it is a questionnaire implementation process in which a questionnaire is conducted and the trained model is applied to the answers.
In step S204, the investigator 15 creates a questionnaire using the questionnaire creation unit 101 of the information processing device 10. Then, the questionnaire is transmitted to the respondent terminal 20 using the questionnaire transmission unit 102.

In step S205, the respondent 25 answers the questionnaire using the answer input unit 202. FIG. 7A is a question displayed on the touch panel 2004 of the smartphone which is the respondent terminal 20. The respondent 25 selects a photo stored in the smartphone and displays it in the answer display field 2042. Then, as shown in FIG. 7B, a photograph is transmitted as an answer.

In step S206, the surveyor 15 uses the response content acquisition unit 112 to acquire image information as information necessary for discrimination from the acquired response. Then, the image data is encoded as described above. Subsequently, the feature extraction unit 113 extracts the feature amount using the trained model, and the classification determination unit 114 performs the classification determination and the category determination. Then, in step S207, the determination result is presented from the result output unit 115.

As described above, in the present embodiment, the image information obtained in the photo survey questionnaire is also given an appropriate classification based on the result of machine learning. As a result, the determination result can be presented in a format that is easy for the investigator 15 to understand while reducing the time and labor costs.

<Modification example>
In the first and second embodiments, the text information of the free answer and the image information of the photo survey were processed and automatically coded. In the first embodiment, the text information is encoded at the character level to process the character string as numerical information, and in the second embodiment, the image information is processed as numerical information based on the pixel value for each pixel. Therefore, in the present invention, text information and image information can be processed by the same algorithm. Therefore, the present invention can also be applied to a multimodal answer that combines text information and image information, such as when the respondent 25 responds by adding a comment to a photograph. In this case, the text information of the comment portion may be encoded at the character level as in the first embodiment, and the image information of the photographic portion may be pixel-processed as in the second embodiment. Then, the code based on the text information and the code based on the image information are input to the neural network and classified.
By performing multimodalization in this way, it is possible to improve the classification accuracy based on the comments of the respondents even when it is difficult to distinguish only by the feature amount of the image information.

10: Information processing device, 101: Questionnaire creation unit, 102: Questionnaire transmission unit, 111: Answer reception unit, 114: Classification judgment unit, 121: Data creation unit, 122: Model creation unit 20: Respondent terminal, 201: Questionnaire Receiver, 202: Answer input, 211: Answer transmitter

Claims (8)

Information that classifies the answers to the open-ended questions included in the questionnaire and the photos sent to the questions included in the questionnaire that require the respondents to send photos into one of multiple categories. It is an information processing method executed by the processing device .
A step of acquiring the reply consisting of text information regarding which have been, the free reply type interrogator created by the respondent,
The step of encoding by dividing the character string included in the answer into characters one by one and digitizing each character, and
A step of acquiring image information of the photograph transmitted from the respondent and digitizing it for each pixel to encode it.
A step of applying a trained model generated by a character-level convolutional neural network to the quantified answer to determine which of the plurality of categories the answer by the respondent falls into.
A step of applying a trained model generated by a convolutional neural network to the digitized image information to determine which of the plurality of categories the photograph transmitted from the respondent falls into. ,
An information processing method characterized by having. And creating a plurality of the learning data consisting of the open-ended type classification information assigned to the sample answers and the sample answers to questions,
A step of dividing the character strings included in the sample answers of the plurality of learning data into characters and encoding them, respectively.
A step of creating the trained model for classifying the answers of the respondents using the plurality of training data, and
The information processing method according to claim 1, further comprising a preparatory step. The information processing method according to claim 1 or 2, wherein the character string included in the answer is encoded by a coding method corresponding to the Unicode standard. The image information is encoded into a numerical value based on the pixel value for each pixel.
The information processing method according to any one of claims 1 to 3, wherein the information processing method is characterized. It is an information processing method executed by an information processing device that classifies the answers to the questionnaire into one of multiple categories.
The survey includes a photo and a question requesting respondents to send comments about the photo.
A step of acquiring the image information of the photograph and the text information of the comment from the answer sent from the respondent, and
A step of dividing the character string included in the text information into characters and digitizing each character, and a step of encoding the characters.
A step of encoding the image information by digitizing it for each pixel,
A step of applying a trained model generated by a convolutional neural network to the digitized text information and the image information to determine which of the plurality of categories the answer by the respondent falls into. When,
An information processing method characterized by having. A program for causing an information processing device to execute the information processing method according to any one of claims 1 to 5. Information that classifies the answers to the open-ended questions included in the questionnaire and the photos sent to the questions included in the questionnaire that require the respondents to send photos into one of multiple categories. It ’s a processing device,
The respondents created by the response receiver that acquires the reply consisting of text information regarding the open-ended type interrogator,
The first answer content acquisition unit, which encodes the character string included in the answer by dividing it into characters one by one and digitizing each character,
A second response content acquisition unit that acquires image information of the photograph transmitted from the respondent and encodes it by digitizing each pixel, and
The digitized said answer, by applying the learned model generated by the character level convolutional neural network, the first classification determines the answer by the respondent is classified into the plurality of categories Judgment unit and
A second method of applying a trained model generated by a convolutional neural network to the digitized image information to determine which of the plurality of categories the photograph transmitted from the respondent falls into. Classification judgment unit and
An information processing device characterized by having. An information processing device that classifies the answers to questionnaires into multiple categories.
The survey includes a photo and a question requesting respondents to send comments about the photo.
An answer receiving unit that acquires the image information of the photograph and the text information of the comment from the answer sent from the respondent.
The answer content acquisition unit, which encodes the character string included in the text information by dividing it into characters one by one and digitizing each character, and digitizes the image information for each pixel.
The digitized said text information and the image information, by applying the learned model generated by the neural network convolution, the same algorithm from the image information and the text information, the reply by the respondents said plurality A classification judgment unit that determines which of the categories is classified into
An information processing device characterized by having.

Images