JP2023123247A - Apparatus and method for classification and program - Google Patents

Abstract

To provide an apparatus and method for classification and a program to generate an index so as to retrieve also a similar case different in the domain to which information belongs.SOLUTION: An apparatus for classification, for classifying object-of-classification information into any of a plurality of classification results, comprises a classification-model learning section 31 and a classification executing section. The classification-model learning section learns a classification model by performing machine learning based on a domain code representative of a domain to which the object-of-classification information belongs, attribute information about the object-of-classification information and classification result information representative of a classification result of a correct answer to which the object-of-classification information should be classified, thereby creating a learnt model. The classification executing section classifies the object-of-classification information using the learnt model based on the domain code and the attribute information.SELECTED DRAWING: Figure 8

Description

The present invention relates to a classification device, a classification method and a program.

As a method for classifying information, a method based on deep learning using a neural network has been used in recent years.

Among them, the prototype-based method shown in Non-Patent Document 1 can perform not only classification but also retrieval of similar information by using the index that is the output, and can find similar cases that serve as the basis for AI judgment. It is attracting attention as a method that can show In the prototype-based method, the classification model converts the input data into an index in the encoder layer, calculates the distance between the index and multiple prototypes in the prototype layer, and classifies in the output layer based on the distance. be. A prototype here is a typical index value in training data, and is learned as a parameter of a classification model during training. With this configuration, information with similar input data content and the same classification result is learned so that the index values are close to each other.

In addition, the technology described in Patent Document 1 is an improvement of the prototype-based method so that column data such as text data and time-series data can be used as input data, and is applied to the classification of information expressed in text. be able to.

These techniques can be used, for example, to classify causes of defects in on-site repair services. In this case, the input data includes, for example, product items representing rough classifications of products, product product numbers for identifying products, codes such as major and middle classifications representing rough classifications of defects, and codes requested by customers. Text data representing the content of the defect, text data representing the results of on-site confirmation by a repair engineer, etc. are used. By applying a prototype-based method using these as input data, we classify the causes of failures, search for similar cases using an index, and refer to the repair reports of those cases to take appropriate measures. be able to.

Searching for similar cases using an index is expected to be advantageous not only for use in repair services, but also for quality analysis of accumulated repair service data. A housing equipment manufacturer that provides repair services needs to cross-sectionally analyze whether there are similar cases in other domains (for example, different product items) for defects that occurred in one domain (for example, a specific product item). There is

U.S. Patent Application Publication No. 2020/0364504

Oscar Li, et al., "Deep Learning for Case-Based Reasoning Through Prototypes: A Neural Network That Explains Its Predictions," Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 32, No. 1, 2018.

However, with the conventional technology, there is a problem that it is difficult to generate an index so that similar cases to which information belongs to different domains can also be retrieved.

For example, the technology described in Non-Patent Document 1 and Patent Document 1 converts all input data, including the product item code, into an index, so the index includes product item information, resulting in the content of the defect. Even if the factors and causes are similar, if the product items are different, the index values will not be close. Therefore, it is difficult to retrieve similar cases with different domains (merchandise items) with the index generated by the conventional technique.

The present invention has been made to solve the above problems, and provides a classification device, a classification method, and a program capable of generating an index so that similar cases with different domains to which information belongs can also be searched. aim.

An example of the classification device according to the present invention is
A classification device that classifies classification target information into one of a plurality of classification results,
The classification device comprises a classification model learning unit and a classification execution unit,
The classification model learning unit is based on a domain code representing a domain to which the classification target information belongs, attribute information related to the classification target information, and classification result information representing a correct classification result to which the classification target information should be classified. , trains a classification model by performing machine learning to generate a trained model,
The classification model includes:
an encoder layer that receives an input regarding part or all of the attribute information and outputs an index that is a multi-dimensional numerical vector;
a prototype layer that outputs the similarity between the prototype learned by the machine learning and the index;
an output layer that receives inputs regarding the domain code and the similarity and outputs a classification result;
a domain code prediction unit that predicts the domain code based on the index;
with
The classification model learning unit performs machine learning by applying adversarial learning in the machine learning so that the domain code prediction error by the domain code prediction unit is maximized,
The classification execution unit classifies the classification target information using the trained model based on the domain code and the attribute information.

In one example, the classifier further comprises:
an index creation unit that stores, in an index table, a relationship between the index output by the encoder layer and the classification target information corresponding to the index;
Based on a search target index, which is an index for the classification target information to be searched, and the index table, an index whose value is similar to the search target index is obtained as a similar index, and the classification target information associated with the similar index is output. a similar information search unit;
Prepare.

In one example, the similar information search unit
the Euclidean distance is less than or equal to a specified threshold, or
that the cosine similarity is greater than or equal to a specified threshold;
is used as a reference to determine whether the index values are similar.

In one example, the index is a vector of three or more dimensions,
The similarity information search unit dimensionally compresses the similarity index into a two-dimensional vector using either principal component analysis or t-SNE, and displays the two-dimensional vector as a scatter diagram.

In one example, the similar information search unit clusters the similar indexes by cluster analysis,
The scatter chart includes a bubble chart, and the bubble chart represents the center point of a cluster and the number of similar indices belonging to the cluster by the center point of a circle and the size of the circle, respectively.

A classification method according to the present invention is a classification method for classifying classification target information into one of a plurality of classification results, and is executed by the classification device described above.

A program according to the present invention causes a computer to function as the above-described classification device.

INDUSTRIAL APPLICABILITY According to the classification device, classification method, and program of the present invention, an index can be generated so that similar cases to which information belongs to different domains can also be retrieved.

1 is a configuration of a classification device 10 according to Embodiment 1 of the present invention; A configuration example of a classification target information table 21 . An example of the configuration of a classification result table 22. FIG. A configuration example of a training data table 23 . 3 is a configuration example of an index table 25; 4 is a flowchart showing an example of the operation of the classification device 10; A detailed example of step S1. A configuration example of a classification model. A detailed example of step S2. A detailed example of step S3. Detailed example of step S4.

An embodiment of the present invention will be described below with reference to the accompanying drawings.
[Embodiment 1]
FIG. 1 shows the configuration of a classification device 10 according to Embodiment 1 of the present invention. The classification device 10 is a device that classifies information to be classified into one of a plurality of classification results by executing the information classification method described herein. In this embodiment, the classification target information is product defect reports, but as a modification, various information can be used. As a specific example, it can be applied to time-series data such as images and electrocardiograms, text data, and the like.

The classification device 10 has a hardware configuration as a known computer, and includes a storage unit 20 and a processing unit 30, for example. Storage unit 20 includes a storage medium such as a semiconductor memory device and a magnetic disk device. Some or all of the storage media may be non-transitory storage media. Processing unit 30 includes, for example, a processor.

The classification device 10 may also include input/output means (not shown). The input/output means includes, for example, input devices such as a keyboard and mouse, output devices such as a display and printer, and communication devices such as a network interface.

The storage unit 20 stores a classification target information table 21 , a classification result table 22 , a training data table 23 , a learned classification model 24 and an index table 25 . The processing unit 30 functions as a classification model learning unit 31 , a classification executing unit 32 , an index creating unit 33 and a similar information searching unit 34 .

The storage unit 20 may store a program (not shown). Execution of this program by the processor may cause the classifier 10 to perform the functions described herein. That is, this program causes the computer to function as the classification device 10 (more specifically, the processor as the classification model learning unit 31, the classification execution unit 32, the index creation unit 33, and the similar information search unit 34). may

FIG. 2 shows a configuration example of the classification target information table 21. As shown in FIG. The classification target information table 21 stores, for each defect report, which is classification target information, associated information (information ID) for identifying the defect report with information related to the defect report.

Information related to bug reports includes domain codes and attribute information. The domain code represents a domain that is a rough classification to which the bug report belongs. In this embodiment, the domain is the product item related to the defect report, but in a modification, the domain can be appropriately defined by those skilled in the art.

The attribute information is not particularly limited as long as it is information related to the defect report and is different from the domain code. It contains text data representing the content of the offer and text data representing the results of confirmation of the defect by the worker.

FIG. 3 shows a configuration example of the classification result table 22. As shown in FIG. The classification result table 22 stores each trouble report in association with the information ID of the trouble report and the classification result of the trouble report. The classification result is a code representing the cause of the problem in this embodiment, but can be defined as appropriate by those skilled in the art in the modification. This classification result is, for example, an appropriate one specified by a human, and can be used as classification result information representing a correct classification result to which the classification target information should be classified in machine learning. Further, as will be described later, after the classification is performed by the classification model learning unit 31, the classification result may be included.

FIG. 4 shows a configuration example of the training data table 23. As shown in FIG. The training data table 23 stores, for each defect, the information ID of the defect report in association with the domain code of the defect report, the attribute information of the defect report, the classification result of the defect report, and the like. The domain code and attribute information can be the same as those in the classification target information table 21 in FIG. 2, and the classification results can be the same as in the classification result table 22 in FIG.

A classification model learning unit 31 of the classification device 10 learns a classification model by performing machine learning based on the training data table 23, and generates a trained classification model 24 as a trained model. Specific processing at this time will be described later with reference to FIG. 7 and the like.

FIG. 5 shows a configuration example of the index table 25. As shown in FIG. The index table 25 stores each defect report in association with the information ID of the defect report and the index of the defect report. The index is a multi-dimensional (preferably multi-dimensional) numeric vector that is output by the encoder layer 42, which will be described later, with a part or all of the attribute information of the defect report as an input. , index element 3, and so on.

FIG. 6 is a flow chart showing an example of the operation of the classification device 10. As shown in FIG. This flow chart represents the information classification method performed by the classification device 10 .

First, the classification model learning unit 31 of the classification device 10 learns a classification model (step S1). It is assumed that the classification target information table 21 and the classification result table 22 are stored in the storage unit 20 at this point.

FIG. 7 shows a detailed example of step S1. In step S1, the classification model learning unit 31 first connects the classification target information table 21 and the classification result table 22 by internal join using the information ID as a key, and stores them in the training data table 23 (step S11). The training data table 23 is thus generated.

Next, the classification model learning unit 31 learns a classification model based on the data in the training data table 23 (step S12). This will be specifically described below.

FIG. 8 shows a configuration example of the classification model learning unit 31. As shown in FIG. The classification model learning unit 31 learns a prototype-based neural network by machine learning. In particular, adversarial learning learns the index to maximize the prediction error of the domain code and avoids containing domain-specific information in the index. Then, the classification target information is classified by adding (concatenating) a vector corresponding to the domain code to the similarity vector output from the prototype layer and inputting it to the output layer.

The classification model learned by the classification model learning unit 31 includes an embedding layer 41 , an encoder layer 42 , a prototype layer 43 , an output layer 44 and a domain code prediction unit 45 .

The embedding layer 41 receives domain codes and attribute information included in the classification target information as inputs, and outputs a plurality of numerical vectors. In this embodiment, a vector v ₁ corresponding to the domain code (product item), a vector v ₂ corresponding to the product number, a vector v ₃ corresponding to the major defect classification, a vector v ₄ corresponding to the minor defect classification, the customer and vectors v _{n+5 to} v _{n+m+4 corresponding} to words 1 to n included in the confirmation result.

Specific processing for converting a code such as a domain code into a vector can be appropriately designed by those skilled in the art based on known techniques. Further, a specific process of dividing text data into words and a specific process of converting each word into a vector can also be appropriately designed by those skilled in the art based on known techniques.

The encoder layer 42 generates an index based on part or all of the vector corresponding to the attribute information among the vectors generated by the embedding layer 41 . In this embodiment, the vectors corresponding to the customer's offer and confirmation result are used, but which vectors to use can be designed as appropriate by those skilled in the art. For example, all vectors may be used. However, the vector corresponding to the domain code is not used, and only the vector corresponding to the attribute information is used. The encoder layer 42 can be configured using, for example, a recurrent neural network (RNN) including multiple layers 42a, but is not limited to this.

Thus, the encoder layer 42 accepts input regarding some or all of the attribute information and outputs an index.

The prototype layer 43 holds prototypes p ₁ to p _k (where k is 1 or more) as parameters. Each prototype is a vector of the same dimension as the index. The prototype is learned as part of the parameters of the classification model by machine learning performed by the classification model learning unit 31 based on the data in the training data table 23 . As a result, each prototype is learned to have typical index values for failure reports contained in the training data table 23 . Specific processing of machine learning at this time can be appropriately designed by those skilled in the art based on known techniques including Patent Document 1 and Non-Patent Document 1.

The prototype layer 43 calculates and outputs the similarity between each of the prototypes p ₁ to p _k and the index output from the encoder layer 42 .

The definition of similarity can be appropriately designed by those skilled in the art. For example, Euclidean distance between vectors can be used as similarity, or cosine similarity between vectors can also be used as similarity. Using such an approach, the similarity between the index and one prototype is computed as a single scalar value, so all (k) similarities are concatenated to generate a k-dimensional similarity vector. can be done.

The output layer 44 receives input including a vector _v1 corresponding to the domain code (product item) output from the embedding layer 41, a part of the vector corresponding to the attribute information, and a similarity vector. Output the cause of the problem. The output layer 44 can be configured using, for example, an RNN including multiple layers 44a, but is not limited to this. When RNN is used, vectors v ₁ to v ₄ corresponding to domain codes and the like can be connected to the similarity vector and input.

The cause of failure is represented by a code as shown in the classification result table 22 of FIG. 3, for example. A person skilled in the art can appropriately design a specific process for converting the output (for example, vector) of the output layer 44 into a code, but one example will be described below. First, a number is assigned to each of a plurality of classification results. Then, for each of the plurality of classification results, the output layer 44 generates a vector whose elements are values representing the probability that the classification target information is classified into that classification result. That is, when the code representing the classification result can take q values, the output layer 44 generates a q-dimensional vector. Then, the classification result corresponding to the element having the maximum value among the elements of this q-dimensional vector is output by the output layer 44 .

Here, the input of the output layer 44 includes the vector _v1 corresponding to the domain code and the similarity vector. In this way, it can be said that the output layer 44 receives inputs regarding domain codes and similarities, and outputs classification results.

A domain code prediction unit 45 predicts a domain code based on the index. The domain code prediction unit 45 can be configured using, for example, a neural network including a plurality of fully connected layers, but is not limited to this.

The classification model learning unit 31 learns a classification model by performing machine learning based on the classification result output by the output layer 44 and the domain code output by the domain code prediction unit 45 . Here, the classification model learning unit 31 minimizes the error between the output classification result and the classification result associated with the classification target information in the training data table 23, and sets the output domain code Then, the parameters of the classification model are updated so that the error with the domain code associated with the classification target information in the training data table 23 is maximized.

That is, the classification model learning unit 31 applies adversarial learning to perform machine learning so that the domain code prediction error by the domain code prediction unit 45 is maximized. This adversarial learning makes it possible to prevent domain-specific information from being included in the index, and to generate an index that allows retrieval of similar cases to which the classification target information belongs to different domains.

Further, in parallel with such hostile learning, the classification model learning unit 31 performs machine learning based on the domain code, the attribute information, and the classification result information representing the correct classification result to generate the classification model. learn. The classification model learning unit 31 generates the trained classification model 24 by performing such two types of machine learning in parallel.

A person skilled in the art can appropriately design specific error propagation calculations during learning. For example, for the learning of the encoder layer 42, a GRL (Gradient Reversal Layer) is arranged before the input layer of the domain code prediction unit 45, and the error backpropagated from the output layer 44 and the error from the domain code prediction unit 45 It may be learned to minimize the back-propagated error (gradient reversed by GRL), respectively.

As described above, the classification model learning unit 31 generates the trained classification model 24 . The trained classification model 24 that has completed learning can have the same layer structure as the classification model before learning, but the domain code prediction unit 45 may be omitted.

Thus, step S12 (FIG. 7) ends. Next, the classification model learning unit 31 stores the trained classification model 24 in the storage unit 20 (step S13). This completes step S1 (FIG. 6).

After step S1, the classification executing section 32 executes classification (step S2). That is, the classification executing unit 32 classifies the classification target information by predicting the classification result using the trained classification model 24 based on the domain code and attribute information included in the newly input classification target information.

FIG. 9 shows a detailed example of step S2. In step S2, first, the classification execution unit 32 selects classification target information specified by the user of the classification device 10 (for example, it may be newly input classification target information, or a record in the new classification target information table 21). ) is input, classification is executed using the trained classification model 24 (step S21).

Next, the classification execution unit 32 associates the classification result output by the output layer of the trained classification model 24 with the information ID of the input classification target information and stores it in the classification result table 22 (step S22). Thus, step S2 ends.

As shown in FIG. 6, after step S2, the index creation unit 33 creates the index table 25 (step S3).

FIG. 10 shows a detailed example of step S3. In step S3, the index creation unit 33 first receives all classification target information (for example, all records in the new classification target information table 21) and performs classification using the learned classification model 24 (step S31).

Next, the index creation unit 33 stores the relationship between the index output by the encoder layer 42 of the trained classification model 24 and the original classification target information corresponding to the index in the index table 25 (step S32). . Thus, step S3 ends.

As shown in FIG. 6, after step S3, the similar information search unit 34 searches for similar information (step S4).

FIG. 11 shows a detailed example of step S4. In step S4, first, the similar information search unit 34 acquires a specific index (hereinafter referred to as "search target index") (step S41). The search target index is obtained, for example, by referring to the index table 25 using the information ID as a key for the classification target information specified by the user, but it may be specified or obtained by other methods. good too.

Next, based on the search target index and the index table 25, the similar information search unit 34 identifies an index whose value is similar to the search target index from the index table 25 (step S42). The index specified here is hereinafter referred to as a "similar index". A plurality of similarity indices may be used.

Next, the similar information search unit 34 refers to the classification target information table 21 using the information ID associated with the acquired similarity index as a key, acquires the corresponding classification target information, and outputs it as a search result ( step S43).

Here, the above-described adversarial learning makes it possible to prevent domain-specific information from being included in the index, and it is possible to search for similar cases belonging to a domain different from the information to be searched.

The similarity determination process by the similar information search unit 34 can be appropriately designed by a person skilled in the art. , it may be determined whether the index values are similar. Alternatively, the similar information search unit 34 determines whether or not the index values are similar based on whether the cosine similarity between the search target index and each index in the index table 25 is equal to or greater than a specified threshold. may

Although not shown, the similarity information search unit 34 may display the similarity index as a scatter diagram. A scatter diagram is represented, for example, in two dimensions. In particular, when the index is a three-dimensional vector or more, the similarity information search unit 34 may dimensionally compress the similarity index into a two-dimensional vector and display the two-dimensional vector as a scatter diagram. A person skilled in the art can appropriately design the specific process of dimensionality reduction. For example, a technique of principal component analysis or a technique of t-SNE (t-distributed stochastic neighbor embedding) may be used.

By using such a two-dimensional scatter diagram, the user can more easily understand the similarity search results.

Moreover, although not shown, the similarity information search unit 34 may cluster similarity indexes by cluster analysis. Specific processing of cluster analysis can be appropriately designed by those skilled in the art. The clustering results can be displayed as a scatter plot as described above. For example, a scatter plot may include a bubble chart. The bubble chart represents the center point of a cluster and the number of similarity indices belonging to the cluster, respectively, by the center point of a circle (eg, represented by two-dimensional coordinates) and the size of a circle (eg, represented by a radius or area). may

By using such a bubble chart, the user of the classification device 10 can more easily understand the similarity search results.

The specific structure of the classification model is not limited to that shown in FIG. Those skilled in the art can make appropriate modifications based on known techniques including Patent Document 1 and Non-Patent Document 1.

REFERENCE SIGNS LIST 10 classification device 20 storage unit 21 classification target information table 22 classification result table 23 training data table 24 learned classification model 25 index table 30 processing unit 31 classification model learning unit 32 classification execution unit 33 Index creating unit 34 Similar information searching unit 41 Embedding layer 42 Encoder layer 43 Prototype layer 44 Output layer 45 Domain code prediction unit

Claims (7)

A classification device that classifies classification target information into one of a plurality of classification results,
The classification device comprises a classification model learning unit and a classification execution unit,
The classification model learning unit is based on a domain code representing a domain to which the classification target information belongs, attribute information related to the classification target information, and classification result information representing a correct classification result to which the classification target information should be classified. , trains a classification model by performing machine learning to generate a trained model,
The classification model includes:
an encoder layer that receives an input regarding part or all of the attribute information and outputs an index that is a multi-dimensional numerical vector;
a prototype layer that outputs the similarity between the prototype learned by the machine learning and the index;
an output layer that receives input regarding the domain code and the similarity and outputs a classification result;
a domain code prediction unit that predicts the domain code based on the index;
with
The classification model learning unit performs machine learning by applying adversarial learning in the machine learning so that the domain code prediction error by the domain code prediction unit is maximized,
The classification execution unit classifies the classification target information using the trained model based on the domain code and the attribute information.
Classifier. The classification device further comprises:
an index creation unit that stores, in an index table, a relationship between the index output by the encoder layer and the classification target information corresponding to the index;
Based on a search target index, which is an index for the classification target information to be searched, and the index table, an index whose value is similar to the search target index is obtained as a similar index, and the classification target information associated with the similar index is output. a similar information search unit;
2. The classifier of claim 1, comprising: The similar information search unit is
the Euclidean distance is less than or equal to a specified threshold, or
that the cosine similarity is greater than or equal to a specified threshold;
3. The classification device according to claim 2, wherein it is determined whether or not the index values are similar based on any of . the index is a vector of three or more dimensions;
The similarity information search unit uses either principal component analysis or t-SNE to dimensionally compress the similarity index into a two-dimensional vector, and displays the two-dimensional vector as a scatter diagram.
4. A sorting device according to claim 2 or 3. The similar information search unit clusters the similar indexes by cluster analysis,
The scatter diagram includes a bubble chart, wherein the bubble chart represents the center point of a cluster and the number of similar indices belonging to the cluster by the center point of a circle and the size of the circle, respectively.
5. A sorting device according to claim 4. A classification method for classifying information to be classified into one of a plurality of classification results, the classification method being executed by the classification device according to claim 1 . A program that causes a computer to function as the classification device according to claim 1 .

Images