JP7633015B2 - METHOD FOR PROPOSING SOLUTION MEANS, METHOD FOR GENERATING CLASSIFICATION MODEL, AND SYSTEM FOR PROPOSING SOLUTION MEANS - Google Patents

Description

The embodiments relate to a method for proposing a solution, a method for generating a classification model, and a system for proposing a solution.

Conventionally, attempts have been made to accumulate set data linking past problems with the solutions, and when considering solutions to new problems, to search the accumulated set data for past problems similar to the new problem and utilize the solutions of the similar past problems.
However, in some cases, the characteristics of a task are not fully expressed in past tasks, and it may not be possible to find an appropriate past task similar to a new task from the accumulated set data.

JP 2000-276487 A

The purpose of the embodiment is to provide a solution proposal method and proposal system that can detect appropriate past problems that are similar to a new problem, and a method for generating a classification model used in such a solution proposal method.

The method for proposing a solution according to the embodiment includes the steps of: acquiring a new task; estimating a category into which the new task will be classified from among a plurality of categories using a classification model trained with teacher data in which task tags corresponding to each of a plurality of past tasks are classified into a plurality of categories according to their mutual relevance; searching for a past task similar to the new task from among the plurality of past tasks within the estimated category; and outputting a solution associated with the searched similar past task, in which each of the past tasks is input to one of the task items in at least one hierarchical data in which task items and means items are arranged in a hierarchical structure; and each of the task tags is determined based on information input to each of the past tasks and to an item in the hierarchical data that has a parent-child relationship with the task item into which the each of the past tasks is input.

FIG. 2 is a block diagram showing a proposal system of a solution according to the first embodiment and a generation system of a classification model used in the proposal system. FIG. 2 is a diagram illustrating an example of hierarchical data. 4 is a flowchart showing a method for generating a classification model according to the first embodiment. FIG. 1 illustrates a method for determining issue tags. FIG. 1 illustrates a method for determining issue tags. FIG. 13 illustrates a method for classifying multiple issue tags into multiple categories. 4 is a flowchart showing a method for proposing a solution according to the first embodiment. FIG. 1 illustrates a method for inferring a category into which a new task falls. 13 is a graph showing the maximum relevance value for each category. FIG. 13 is a diagram showing a method for searching for similar past assignments. FIG. FIG. 12(a) is a diagram showing a method for estimating a category into which a new assignment will be classified in the system for proposing a solution relating to the second embodiment, and FIG. 12(b) is a diagram showing a method for extracting a past assignment similar to a new assignment from among multiple past assignments in the system for proposing a solution relating to the second embodiment. FIG. 11 is a diagram showing a setting screen for search conditions of the solution proposal system according to the second embodiment; FIG. 13 is a diagram showing an output of the solution proposal system according to the second embodiment. FIG. 13 is a diagram showing a method for determining an issue tag in the solution proposal system according to the third embodiment. FIG. 13 is a diagram showing past data according to the fourth embodiment;

First Embodiment
First, the first embodiment will be described.
(Outline of the proposed solution system)
FIG. 1 is a block diagram showing a system for proposing a solution according to the present embodiment and a system for generating a classification model used in the system.
A proposal system 1 of a solution according to the present embodiment (hereinafter simply referred to as the “proposal system 1”) is used, for example, in an organization consisting of multiple domains. A “domain” is a business area in which an organization such as a company develops its business.

When a new problem i1 is input from a terminal in any domain, the proposed system 1 searches for a past problem a11 similar to the new problem i1 from among multiple past problems stored in the organization, and outputs a solution a12 for the similar past problem a11. This makes it possible to utilize past knowledge regardless of the domain when considering a solution to the new problem i1 in each domain. Note that the proposed system 1 may be used within one domain.

The proposed system 1 is based on at least one hierarchical data D1 (past data) provided from each domain, a classification model D5 generated by a generation system 2 using the hierarchical data D1, and set data D2. Before describing the configuration of the proposed system 1 in detail, the hierarchical data D1 and the generation system 2 will be described in detail first.

(Hierarchical Data)
FIG. 2 is a diagram illustrating hierarchical data.
The hierarchical data D1 is data that breaks down knowledge considered in the past when realizing an object such as a technology or product into problems that occurred when realizing the object and means that can solve the problems, and links the means to the problems. By creating the hierarchical data D1, it is possible to organize past knowledge considered when realizing an object.

In the hierarchical data D1, the task item a1 and the means item a2 are arranged to form a hierarchical structure. A "hierarchical structure" refers to a structure having multiple hierarchies arranged from the top to the bottom, in which an item Y that is a prerequisite for a certain item X is arranged in a hierarchical structure immediately below the hierarchical structure in which the higher-level item X is arranged, and the lower-level item Y is linked to the higher-level item X. Hereinafter, such a relationship between the X item and the Y item is referred to as a "parent-child relationship." In addition, the X item is referred to as a "parent item" from the viewpoint of the Y item, and the Y item is referred to as a "child item" from the viewpoint of the X item. Furthermore, when a Z item is arranged in a hierarchical structure immediately below the hierarchical structure in which the Y item is arranged, and the Z item is linked to the Y item, the Z item is referred to as a "grandchild item" from the viewpoint of the X item. The hierarchical data D1 may be configured as a hierarchical structure that includes other items in addition to the task item a1 and the means item a2. In such a case, in the hierarchical data D1, multiple items including the task item a1 and the means item a2 are arranged to form a hierarchical structure.

In the example of Fig. 2, a target item a0 is placed at the top (root) of the hierarchical data D1. For example, a target technology, product, etc. are input to the target item a0. A child item of the target item a0 is a problem item a1 that arises when realizing the target item a0. A child item of the problem item a1 is a means item a2 that can solve the problem item a1, or another problem item a1 that is obtained by subdividing the problem item a1 into multiple subordinate problems. A child item of the means item a2 is a problem item a1 that arises when the means item a2 is adopted, or another means item a2 that is obtained by subdividing the means item a2 into multiple subordinate means.
In the hierarchical data D1, when a plurality of child items are linked to one parent item, the child items are of the same type.

The proposal system 1 and the generation system 2 treat an assignment inputted into any assignment item a1 of at least one piece of hierarchical data D1 as a "past assignment a11."
Furthermore, the proposal system 1 and the generation system 2 treat, for example, the means adopted in actually solving the past problem a11 among the means input to the means item a2 of the descendants such as the child or grandchild of the problem item a1 in which the past problem a11 was input, as the "solving means a12". Note that the proposal system 1 and the generation system 2 may treat, for example, all the means including the means not adopted among the means input to the means item a2 of the descendants of the problem item a1 in which the past problem a11 was input, as the "solving means a12".

(Configuration of classification model generation system)
Next, the configuration of the generation system 2 will be described.
As shown in FIG. 1, the generation system 2 includes a tag determination unit 21, a data creation unit 22, and a model generation unit 23.

The tag determination unit 21 determines the assignment tag c4 corresponding to each past assignment a11 based on the information input to each past assignment a11 and to the items in the hierarchical data D1 that have a parent-child relationship with the assignment item a1 to which the past assignment a11 was input. The "items in a parent-child relationship" are the parent item b1 and child item b2 of the assignment item a1 to which the past assignment a11 was input, as shown in FIG. 2.

The assignment tag c4 is determined, for example, based on words extracted from the sentences of each past assignment a11 and the sentences entered in the parent item b1 and child item b2, and on synonyms of the extracted words. As shown in FIG. 1, the tag determination unit 21 uses, for example, an ontology D3 to search for synonyms and determine the assignment tag c4.

Ontology D3 is a systematization of the relationships between concepts used in multiple domains. Even if a task has a similar meaning, it may be expressed using different words when the domain is different. Ontology D3 is a concept dictionary that incorporates, for example, the relationships between words used in each domain, and systematizes the hierarchical relationships and synonymous relationships between words.

The data creation unit 22 creates teacher data D4 in which multiple task tags c4 are classified into multiple categories according to their mutual relevance. Specifically, the data creation unit 22 uses the ontology D3 to divide the multiple task tags c4 into task tag groups that can be grouped under a unified concept across domains. The data creation unit 22 treats one task tag group as one category. In other words, each category contains multiple task tags c4 that can be grouped under a unified concept across domains. Thus, a "domain" is the business area of an organization, such as "train air conditioning business" or "nuclear power business," while a "category" is a task that is commonly performed in multiple domains, such as "procurement" or "installation."

The data creation unit 22 also creates set data D2 that links past issues a11 and solutions a12 extracted from the hierarchical data D1 to the issue tag c4 and organizes them by category. The set data D2 is provided to the search unit 13 and output unit 14 of the proposal system 1, which will be described later.

The model generation unit 23 generates a classification model D5 using the teacher data D4 created by the data creation unit 22. The generated classification model D5 is provided to the estimation unit 12 of the proposal system 1 described below, and is used by the estimation unit 12 to estimate the category into which the new task i1 is classified.

The classification model D5 is a text classification model. For example, the classification model D5 is a multi-class classification model that allows one new task i1 to be classified into multiple categories. The method used in the classification model D5 is, for example, a neural network. An example of a text classification model using a neural network is fastText, an open source from Facebook, Inc.

However, the method used in the classification model D5 is not limited to the above, and may be, for example, a support vector machine (SVM), a k-nearest neighbor method, logistic regression, etc. In addition, the classification model may be an exclusive classification model that classifies one new task i1 into only one category.

The classification model D5 created in this way can, for example, take any term as input and output the relevance of the input term in each category.

(Configuration of the proposed solution system)
Next, the configuration of the proposed system 1 will be described.
The proposed system 1 includes an input unit 11, an estimation unit 12, a search unit 13, and an output unit 14.

The input unit 11 accepts input to the proposal system 1. A user in each domain uses a terminal to input a new task i1, set search conditions, and give search instructions to the proposal system 1. The input unit 11 receives search instructions from the terminal and acquires the new task i1, search conditions, etc.

The estimation unit 12 uses the classification model D5 provided by the generation system 2 to estimate the category into which the new task i1 is classified among multiple categories.

The search unit 13 refers to the set data D2 in which multiple past assignments a11 are organized by category, and searches for past assignments a11 that are similar to the new assignment i1 among the multiple past assignments a11 within the estimated category. In this way, past assignments a11 similar to the new assignment i1 are searched for within a category to which highly related past assignments a11 belong. Therefore, it is possible to detect an appropriate past assignment a11 that is similar to the new assignment i1.

When searching for a past assignment a11 similar to the new assignment i1, the search unit 13 may use a vectorization tool D6 that vectorizes text. Examples of the vectorization tool D6 include Word2Vec and Doc2Vec. The vectorization tool D6 has already learned from the text entered in each item in the hierarchical data D1 as training data.

The search unit 13 uses a vectorization tool D6 to vectorize the text of the new assignment i1 and the text of each past assignment a11 belonging to the estimated category. For example, for each category, the search unit 13 determines that the past assignment a11 represented by a vector that is closest to the vector representing the new assignment i1 is the past assignment a11 similar to the new assignment i1. Note that the search unit 13 may detect past assignments a11 represented by vectors whose distance from the vector representing the new assignment i1 is equal to or less than a threshold, and determine the detected past assignments a11 as the past assignments a11 similar to the new assignment i1.

The output unit 14 references the set data D2 provided by the generation system 2 and outputs a solution a12 for a past assignment a11 that is similar to the new assignment i1.

The proposal system 1 and the generation system 2 are realized, for example, by a server having a CPU (Central Processing Unit) and memory and connected to terminals in the respective domains. The server may be composed of one computer or multiple computers. The proposal system 1 and the generation system 2 may be realized by separate computers. Furthermore, the hardware that realizes the proposal system and the generation system is not limited to a server. For example, the proposal system and the generation system may be realized by a computer that performs processing using various data such as new tasks and past data uploaded to the cloud. Furthermore, the hardware that realizes the proposal system and the hardware that realizes the generation system may be different.

(How to generate a classification model)
FIG. 3 is a flowchart showing a method for generating a classification model according to this embodiment.
4 and 5 are diagrams illustrating a method for determining issue tags.
Next, a method for generating a classification model according to this embodiment will be described using a specific example.
First, as shown in Figures 3 to 5, the tag determination unit 21 of the generation system 2 determines an assignment tag c4 corresponding to each past assignment a11 based on the information input to each past assignment a11 and to items b1 and b2 that have a parent-child relationship with the assignment item a1 to which each past assignment a11 was input within the hierarchical data D1 (step S11).

More specifically, as shown in FIG. 4, the tag determination unit 21 performs morphological analysis on the sentence of past assignment a11, the sentence entered in parent item b1, and the sentence entered in child item b2, and extracts words (independent words) from each sentence, excluding particles and auxiliary verbs (step S11a). Hereinafter, the extracted words are referred to as "extracted words c1." That is, in the example of FIG. 4, the sentence of past assignment a11 is "product durability." The sentence entered in parent item b1 is "procurement work for semiconductor application products." The sentences entered in the five child items b2 are "appropriate design," "appropriate material selection," "appropriate manufacturing," "appropriate inspection," and "appropriate shipping and transportation." The independent words in these sentences are extracted words c1.

Next, the tag determination unit 21 refers to the ontology D3 and adds synonyms c2 for each extracted word c1 (step S11b).

Next, as shown in FIG. 5, the tag determination unit 21 determines a representative word c3 of the extracted words c1 and the synonyms c2 by using the ontology D3 (step S11c).
Specifically, first, the tag determination unit 21 divides the extracted words c1 and the synonyms c2 into word groups g that can be grouped under a unified concept using the ontology D3. Next, the tag determination unit 21 selects some word groups g from the word groups g. The selection method is not particularly limited, but for example, the tag determination unit 21 may select the top word groups g that contain the largest number of words after removing word groups g with low priority from the viewpoint of expressing the problem.

Next, the tag determination unit 21 determines a representative word c3 that aggregates the extracted words c1 and synonyms c2 contained in each selected word group g. Each representative word c3 may be, for example, a superordinate word of the extracted words c1 and synonyms c2 contained in each word group g, or any word among the extracted words c1 and synonyms c2 contained in each word group g. Note that in the example of FIG. 5, the number of representative words c3 is six, but the number of representative words c3 is not particularly limited.

Next, the tag determination unit 21 determines a task tag c4 based on the determined representative words c3 (step S11d). The task tag c4 may be a combination of all the representative words c3, or may be a word or combination of words that aggregates the determined representative words c3. In the figure, for the sake of convenience, each task tag c4 is represented by a character string that combines "Y" with a number.

The tag determination unit 21 thus determines the task tag c4 based on the information entered not only in past task a11 but also in items b1 and b2 that are in a parent-child relationship. For this reason, using Figures 4 and 5 as an example, task tag c4 includes features such as "procurement," "design," "manufacturing," and "transportation" that are not expressed in the sentence "product durability" in past task a11. In this way, task tag c4 includes features of information from higher and lower hierarchical levels of past task a11.

The tag determination unit 21 also determines the assignment tag c4 based on the extracted word c1 extracted from the sentences of each past assignment a11 and the sentences entered in the parent-child items b1 and b2, and the synonyms c2 of the extracted word c1. This allows an appropriate assignment tag c4 to be assigned to each past assignment a11, without being limited to the expressions in the sentences of the past assignment a11 and the sentences entered in the parent-child items b1 and b2.

The method of determining the task tag c4 is not limited to the above. For example, the task tag c4 may not be mechanically determined by the tag determination unit 21 using the ontology D3 or the like, but may be manually determined by an expert or the like, or may be determined by a method that combines a mechanical method and an artificial method.

FIG. 6 is a diagram showing a method for classifying multiple issue tags into multiple categories.
As shown in FIG. 6, in step S11 described above, an assignment tag c4 is assigned to each past assignment a11 in each domain.

Next, the data creation unit 22 creates teacher data D4 in which a plurality of task tags c4 are classified into a plurality of categories according to their mutual relevance (step S12).
In step S12, for example, the data creation unit 22 uses the ontology D3 to divide multiple task tags c4 into task tag groups h that can be grouped under a unified concept across multiple domains. The data creation unit 22 treats one task tag group h as one category. In this case, for example, one task tag c4 may belong to multiple categories. In the example shown in FIG. 6, the categories include "procurement system", "installation system", "product system", "maintenance system", "design system", "production system", and "transportation system".

The method of creating the teacher data D4 is not limited to the above. For example, the task tags c4 may not be mechanically categorized by the data creation unit 22 using the ontology D3 or the like, but may be manually categorized by an expert or the like, or may be categorized using a method that combines mechanical and manual methods.

The data creation unit 22 also creates set data D2 (see FIG. 10) in which the assignment tag c4 is linked to the past assignment a11 and the solution a12 and organized by category. The data creation unit 22 provides the set data D2 to the search unit 13.

Next, the model generation unit 23 generates a classification model D5 using the teacher data D4 (step S13). The classification model D5 generated by the model generation unit 23 is provided to the estimation unit 12 of the proposed system 1.

When generating the classification model D5 using the teacher data D4, all of the teacher data D4 created by the data creation unit 22 may be used to train the classification model D5, or only a portion of the teacher data D4 created by the data creation unit 22 may be used to train the classification model D5.

(Method of Proposing a Solution)
FIG. 7 is a flowchart showing a method for proposing a solution according to this embodiment.
Next, a method for proposing a solution according to this embodiment will be described.
First, a user uses a terminal to input a new assignment i1, set search conditions, and instruct the proposal system 1 to start a search. The input unit 11 receives the instruction to start a search and acquires the new assignment i1 and the search conditions (step S21).

Next, the estimation unit 12 uses the classification model D5 trained in step S13 to estimate the category into which the new task i1 is classified among multiple categories (step S22).

FIG. 8 illustrates a method for inferring a category into which a new task falls.
For example, the estimation unit 12 extracts a plurality of partial words i2 by shifting the position from the sentence of the new task i1. For example, the estimation unit 12 sequentially extracts partial words i2 each consisting of three characters while shifting the start position of extraction by one character at a time. However, the number of characters of the partial word i2 and the number of characters to be shifted are not limited to the above.

Next, the estimation unit 12 inputs the extracted multiple partial words i2 to the classification model D5, and causes the classification model D5 to output the relevance i3 between each partial word i2 and each category. For example, the classification model D5 calculates a vertical vector V for each partial word i2, the components of which are the relevance i3 with each category. Hereinafter, the vertical vectors V of each partial word i2 arranged horizontally are referred to as the relevance matrix M.

Next, the estimation unit 12 calculates the maximum value of the relevance i3 of each category in the relevance matrix M, and determines whether the maximum value calculated for each category is equal to or greater than a threshold value. In the example of FIG. 8, the maximum value of the relevance i3 of the "procurement" category is 0.77. The maximum value of the relevance i3 of the "installation" category is 0.99. The maximum value of the relevance i3 of the "product" category is 0.11. The maximum value of the relevance i3 of the "maintenance" category is 0.73.

FIG. 9 is a graph showing the maximum relevance value for each category.
As shown in Fig. 9, the estimation unit 12 estimates that a category in which the maximum value of the relevance degree i3 is equal to or greater than the threshold value is the category in which the new issue i1 is classified. In the example of Fig. 9, the maximum values of the relevance degrees i3 of "installation", "maintenance", and "procurement" exceed the threshold value. Therefore, the estimation unit 12 estimates that "installation", "maintenance", and "procurement" are the categories in which the new issue i1 is classified. Note that the threshold value may be set to an arbitrary value by the user in step S21, for example.

In this way, multiple partial words i2 are extracted from the sentence of new task i1 at different positions. Then, the classification model D5 receives the multiple extracted partial words i2 as input, and outputs the relevance i3 between each partial word i2 and each category. Then, based on the relevance i3, the category into which new task i1 will be classified is estimated. At least one of the partial words i2 contains part of a group of words that are located in a nearby position in the sentence. As a result, it is possible to estimate a contextually related category for the sentence of new task i1. Note that the method of estimating the category into which new task i1 will be classified is not limited to the above.

FIG. 10 is a diagram showing a method for searching for similar past assignments.
Next, as shown in FIG. 10, the search unit 13 refers to the set data D2 and searches for past assignments a11 similar to the new assignment i1 among the multiple past assignments a11 in each of the estimated categories (step S23).

In step S23, the search unit 13 vectorizes the text of the new assignment i1 and the text of the past assignment a11 in each estimated category, for example, using a vectorization tool D6.

Next, the search unit 13 detects, for example, in each category, a past assignment a11 represented by a vector that is closest to the vector representing the new assignment i1, and determines that the detected past assignment a11 is a past assignment a11 similar to the new assignment i1.

In the example shown in FIG. 10, a past issue a11 with an issue tag c4 of "Y22" for the "installation" category is detected as a past issue a11 similar to the new issue i1. A past issue a11 with an issue tag c4 of "Y21" for the "maintenance" category is detected as a past issue a11 similar to the new issue i1. A past issue a11 with an issue tag c4 of "Y33" for the "procurement" category is detected as a past issue a11 similar to the new issue i1. For example, the search unit 13 may search for a past issue a11 similar to the new issue i1 among multiple past issues a11 in a category based on the issue tag c4. In this case, the issue tag c4 corresponding to the new issue i1 and the issue tag c4 corresponding to the past issue a11 in each estimated category may be vectorized using the vectorization tool D6.

In this way, past assignments a11 similar to the new assignment i1 are searched for within a category to which highly related past assignments a11 belong. This makes it possible to detect an appropriate past assignment a11 similar to the new assignment i1. As mentioned above, the assignment tag c4 contains information about the higher and lower hierarchical levels of the past assignment a11. Therefore, each past assignment a11 is classified into an appropriate category using the assignment tag c4. The method of searching for past assignments a11 similar to the new assignment i1 is not limited to the above.

FIG. 11 is a diagram illustrating an example of the output.
Next, the output unit 14 refers to the set data D2 and outputs solutions a12 of similar past assignments a11 for each estimated category (step S24).
The output unit 14 outputs, for example, the estimated categories, similar past assignments A11 in each category, and their solutions A12. There may be one or more similar past assignments A11 for each category. The user can consider the solution for the new assignment i1 by referring to the output solutions A12.

The classification model D5 may be generated at least once before the proposed system 1 is used for the first time. For example, the classification model D5 may be generated only once, and the proposed system 1 may use the same classification model D5 each time it estimates the category into which the new task i1 is classified. In addition, the classification model D5 may be regenerated, for example, when new hierarchical data D1 is provided or when the contents of the existing hierarchical data D1 are updated.

Next, the effects of this embodiment will be described.
In this embodiment, a past assignment a11 similar to the new assignment i1 is searched for within the category into which the new assignment i1 is estimated to be classified. Therefore, a past assignment a11 similar to the new assignment i1 is searched for within a category to which highly related past assignments a11 belong. As a result, an appropriate past assignment a11 similar to the new assignment i1 can be detected.

In this embodiment, each past assignment a11 is entered into one of the assignment items a1 in at least one hierarchical data D1 in which assignment items a1 and means items a2 are arranged in a hierarchical structure. The assignment tag c4 is determined based on the information entered into each past assignment a11 and into items b1 and b2 that have a parent-child relationship with the assignment item a1 to which each past assignment a11 is entered in the hierarchical data D1. Therefore, the assignment tag c4 contains characteristics of the information in the higher and lower hierarchical levels of the past assignment a11. Therefore, each past assignment a11 can be classified into an appropriate category using the assignment tag c4.

In addition, in this embodiment, each assignment tag c4 is determined based on an extracted word c1 extracted from the sentences of each past assignment a11 and the sentences entered in the parent-child items b1 and b2, and a synonym c2 of the extracted word c1. This allows an appropriate assignment tag c4 to be assigned to each past assignment a11, without being limited to the expression of the sentences of the past assignment a11.

In addition, in this embodiment, multiple partial words i2 are extracted from the sentence of new task i1 by shifting their positions. Then, the classification model D5 receives the multiple extracted partial words i2 as input, and outputs the relevance i3 between each partial word i2 and each category. Then, based on the output relevance i3, a category into which new task i1 is classified is estimated from among multiple categories. At least one of the partial words i2 contains a part of a group of words that are located in a nearby position in the sentence of new task i1. As a result, it is possible to estimate a contextually related category for the sentence of new task i1.

In addition, in this embodiment, each category includes multiple assignment tags c4 that can be grouped together under a unified concept across multiple domains. As a result, it is possible to search for an appropriate past assignment a11 that is similar to the new assignment i1, regardless of the domain.

According to the embodiment described above, it is possible to realize a method and system for proposing a solution means a12 that can detect an appropriate past problem a11 that is similar to a new problem i1, and a method for generating a classification model D5 used in the method for proposing such a solution means a12.

Second Embodiment
Next, a second embodiment will be described.
FIG. 12(a) is a diagram showing a method for estimating a category into which a new assignment will be classified in the system for proposing a solution according to this embodiment, and FIG. 12(b) is a diagram showing a method for extracting a past assignment similar to a new assignment from among multiple past assignments in the system for proposing a solution according to this embodiment.
FIG. 13 is a diagram showing a setting screen for search conditions of the solution proposal system according to this embodiment.
FIG. 14 is a diagram showing the output of the solution proposal system according to this embodiment.
In the following description, in principle, only the differences from the first embodiment will be described, and the matters other than those described below are the same as those in the first embodiment.

As shown in FIG. 12(a), the estimation unit 12 can calculate the relevance between the new task i1 and multiple categories, as in the first embodiment. Then, the estimation unit 12 estimates a category among the multiple categories whose relevance is equal to or greater than a first threshold TH1 as the category into which the new task i1 is classified.

The search unit 13 can also calculate the similarity between the new assignment i1 and a similar past assignment a11, as shown in FIG. 12(b). The search unit 13 then extracts, as similar past assignments a11, past assignments a11 that are classified into the estimated category from among the multiple past assignments a11 and have a similarity to the new assignment i1 equal to or greater than a second threshold value TH2. The similarity is not particularly limited as long as it represents the degree of similarity between the new assignment i1 and the past assignment a11, but examples of the similarity include cosine similarity or the distance between the new assignment i1 and the past assignment a11 in the semantic space.

As shown in FIG. 13, when inputting new task i1, the input unit 11 also accepts the setting of a first threshold TH1 and a second threshold TH2 by the user. For example, before a search, a screen such as that shown in FIG. 13 is displayed on a computer operated by the user. The user inputs new task i1 via an operation unit such as a keyboard or mouse of the computer, sets the first threshold TH1 and the second threshold TH2, and presses the search button 210. This starts a search using the set conditions.

As shown in FIG. 14, when outputting the solution a12 of a similar past assignment a11, the output unit 14 also outputs the relevance k1 between the estimated category and the new assignment i1. Furthermore, the output unit 14 outputs, as reference information, among the multiple categories, multiple categories with the highest relevance with the new assignment i1 that are less than the first threshold TH1 and their relevance k2. This allows the user to check the relevance k2 of the category into which the new assignment i1 was not estimated to be classified, and to determine whether the first threshold TH1 is appropriate. Note that, as reference information, only the category with the highest relevance and its relevance may be output among the categories with relevance less than the first threshold.

When outputting the solution a12 of the similar past assignment a11, the output unit 14 also outputs the similarity k3 between the new assignment i1 and the similar past assignment a11. Furthermore, the output unit 14 outputs, as reference information, the past assignment a11 with the highest similarity k3 to the new assignment i1 whose similarity k3 is less than the second threshold TH2 and its similarity k4 among the multiple past assignments a11. This allows the user to check the similarity k4 of the past assignment a11 that was not extracted and determine whether the second threshold TH2 is appropriate. Note that, as reference information, only the past assignment with the highest similarity and its relevance among the past assignments with similarity less than the second threshold may be output.

The input unit 11 is configured to be able to accept changes to the first threshold TH1 and the second threshold TH2 even after outputting the solution a12 of the similar past problem a11. For example, when a screen such as that shown in FIG. 14 is displayed on a computer operated by a user, the user re-sets the first threshold TH1 and the second threshold TH2 via an operation unit such as the computer's keyboard or mouse, and presses the search again button 220. This starts a search again using the set conditions.

As described above, the output unit 14 outputs the relevance k1 between the estimated category and the new task i1. Therefore, the user can understand the relevance k1 between the estimated category and the new task i1.

The output unit 14 also outputs the relevance k2 of the category with the highest relevance among the multiple categories whose relevance with the new assignment i1 is less than the first threshold TH1. The input unit 11 is also configured to be able to accept changes to the first threshold TH1. Therefore, the user can determine whether the first threshold TH1 is an appropriate value, and change the value of the first threshold TH1 as necessary to increase or decrease the number of categories into which the new assignment i1 is estimated to be classified.

The search unit 13 also outputs the similarity k3 between the extracted similar past assignment a11 and the new assignment i1. Therefore, the user can understand the similarity k3 between the extracted past assignment a11 and the new assignment i1.

The output unit 14 also outputs the similarity k4 of the past assignment a11 with the greatest similarity among the past assignments a11 whose similarity to the new assignment i1 is less than the second threshold TH2 among the multiple past assignments. The input unit 11 is also capable of accepting changes to the second threshold TH2. Therefore, the user can determine whether the second threshold TH2 is an appropriate value, and change the value of the second threshold TH2 as necessary to increase or decrease the number of similar past assignments a11 to be extracted.

Third Embodiment
Next, a third embodiment will be described.
FIG. 15 is a diagram showing a method for determining a problem tag in the solution proposal system according to this embodiment.
In the first embodiment, the assignment tag c4 of the past assignment a11 is determined based on the past assignment a11 and information entered in the item of the item in the hierarchical data D1 that has a parent-child relationship with the past assignment a11. In contrast, in the third embodiment, the assignment tag c4 of the past assignment a11 is determined based on the past assignment a11 and information related to the past assignment a11 in the domain 300 to which the past assignment a11 belongs in the hierarchical data D1.

As in the first embodiment, the multiple hierarchical data D1 are provided from multiple domains 300. Therefore, in the multiple hierarchical data D1, each past task a11 belongs to one of the domains 300. As shown in FIG. 15, when determining the task tag c4 of the past task a11 "check the quality of incoming goods" belonging to a certain domain 300, the tag determination unit 21 first extracts information related to the past task a11 in the same domain 300 from the multiple hierarchical data D1. The extracted information is not particularly limited, but may include, for example, a sentence entered in the task item a1, a sentence entered in the means item a2, or both of the task items a1 and a means item a2. In FIG. 15, for example, a sentence entered in the task item a1 "suppress variation by changing the production process or recipe" and a sentence entered in the task item a1 "introduce sampling inspection for redundant inspection and reliability" are extracted as related information.

The method of extracting information related to past assignments A11 is not particularly limited, but examples include a mechanical method using a text vectorization tool D6 such as Word2Vec, a manual method in which an expert extracts related information, or a combination of a mechanical method and a manual method. When the vectorization tool D6 is used, information that belongs to the same domain 300 and is close in distance in semantic space is extracted from multiple hierarchical data D1.

The procedure thereafter is the same as in the first embodiment, where words are extracted from the sentences "Check the quality of incoming goods", "Reduce variation by changing the production process or recipe", and "Implement sampling inspections for redundant inspections and reliability", synonyms of the extracted words are added, and the final issue tag c4 is determined.

As explained above, the information used to determine the assignment tag c4 of past assignment a11 in the hierarchical data D1 is not particularly limited to the information entered in an item that has a parent-child relationship with the assignment item a1 to which the past assignment a11 was entered, but is not particularly limited as long as it belongs to the same domain as the domain to which the past assignment a11 belongs in the hierarchical data D1 and is related to the past assignment a11.

Fourth Embodiment
Next, a fourth embodiment will be described.
FIG. 16 is a diagram showing past data according to this embodiment.
In the first, second, and third embodiments, a past assignment a11 is input, and the past data used to determine the assignment tag c4 is the hierarchical data D1. In contrast, in this embodiment, a past assignment a11 is input, and the past data used to determine the assignment tag c4 is the table data D41.

The table data D41 is data in a table format in which multiple past issues a11 are linked to multiple solution methods a12. In this case, the problem tag c4 of the past issue a11 is determined based on the past issue a11 and information in the table data D41 that belongs to the same domain 300 as the past issue a11 and is related to the past issue a11, as in the third embodiment. For example, when determining the problem tag c4 of the past issue a11 "check the quality of incoming goods" that belongs to a certain domain 300, the past issue a11 "suppressing variation by changing the production process or recipe" and the past issue a11 "incorporating redundant inspections and sampling inspections for reliability" are extracted as related information, as in the third embodiment. However, as in the third embodiment, the solution method a12 may be extracted as related information.

As explained above, the past assignment a11 is input, and the past data used to determine the assignment tag c4 is not limited to hierarchical data D1, as long as the solution method a12 is linked to the past assignment a11, but may be data in another format, such as table data D41.

Although the embodiments of the present invention have been described above, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the gist of the invention. These embodiments and their modifications are included within the scope and gist of the invention, and are included in the scope of the invention and its equivalents as set forth in the claims.

1: Proposal system 2: Generation system 11: Input unit 12: Estimation unit 13: Search unit 14: Output unit 21: Tag determination unit 22: Data creation unit 23: Model generation unit D1: Hierarchical data (past data)
D2: Set data D3: Ontology D4: Teacher data D5: Classification model D6: Vectorization tool D41: Table data M: Relevance matrix V: Vertical vector a0: Target item a1: Problem item a11: Past problem a2: Means item a12: Solution means b1: Parent item b2: Child item c1: Extracted term c2: Synonym c3: Representative term c4: Problem tag g: Word group h: Problem tag group i1: New problem i2: Partial word i3, k1, k2: Relevance k3, k4: Similarity TH1: First threshold TH2: Second threshold

Claims (7)

It runs on a computer.
acquiring new assignments;
a step of estimating a category into which the new assignment will be classified from among a plurality of categories using a classification model trained with training data in which assignment tags corresponding to each of a plurality of past assignments are classified into a plurality of categories according to their mutual relevance;
A step of searching for a past assignment similar to the new assignment among a plurality of past assignments within the estimated category;
outputting solutions associated with the searched similar past problems;
Equipped with
Each of the past issues is input to one of the issue items in at least one hierarchical data in which issue items and means items are arranged to form a hierarchical structure,
A method of proposing solutions in which each of the problem tags is determined based on information entered into each of the past problems and into items within the hierarchical data that have a parent-child relationship with the problem item into which each of the past problems was entered. The method for proposing the solution according to claim 1, wherein each of the assignment tags is determined based on words extracted from the text of each of the past assignments and the text entered in the parent-child item, and on synonyms of the words. The method for proposing a solution according to claim 1 or 2, wherein the multiple problem tags in the training data are classified into multiple categories using an ontology. The step of estimating the category includes:
extracting a plurality of subwords from the new subject sentence at different positions;
inputting a plurality of the partial words into the classification model and causing the classification model to output a degree of association between each of the partial words and each of the categories;
estimating a category into which the new task is classified among the plurality of categories based on the relevance;
The method for proposing the solution according to any one of claims 1 to 3, comprising: The plurality of hierarchical data are provided from a plurality of domains;
A method for proposing a solution described in any one of claims 1 to 4, wherein in the teacher data, each category includes multiple problem tags that can be grouped together under a unified concept that spans multiple domains. It runs on a computer.
determining assignment tags corresponding to each of the plurality of past assignments;
creating training data in which the multiple task tags are classified into multiple categories according to their mutual relevance;
generating a classification model using the training data to estimate a category into which a new task is classified among the multiple categories;
Equipped with
Each of the past issues is input to the issue item in at least one hierarchical data in which issue items and means items are arranged to form a hierarchical structure,
A method for generating a classification model in which each of the assignment tags is determined based on information entered into each of the past assignments and into items within the hierarchical data that have a parent-child relationship with the assignment item into which each of the past assignments was entered. an input unit for accepting input of a new task;
an estimation unit that estimates a category into which the new assignment will be classified from among a plurality of categories using a classification model trained with teacher data in which assignment tags corresponding to each of a plurality of past assignments are classified into a plurality of categories according to their mutual relevance;
a search unit that searches for a past assignment similar to the new assignment among the plurality of past assignments within the estimated category;
an output unit that outputs solutions associated with the searched similar past problems;
Equipped with
Each of the past issues is input to one of the issue items in at least one hierarchical data in which issue items and means items are arranged to form a hierarchical structure,
A solution proposal system in which each of the problem tags is determined based on information entered into each of the past problems and into items in the hierarchical data that have a parent-child relationship with the problem item into which each of the past problems was entered.

Images