JP7442246B1 - Program, information processing method, and information processing device - Google Patents

Abstract

[Problem] To provide programs that support the work of administrative agencies or facilities related to child welfare. [Solution] The program acquires text data related to child consultation, and when the text data related to child consultation is input, the first learning model is trained to output judgment information regarding the necessity of temporary protection. A computer is made to execute a process of inputting text data and outputting judgment information. Preferably, the first learning model outputs whether temporary protection is necessary. More preferably, the program causes the computer to execute a process of extracting sentence data related to child consultation with a high degree of similarity from the acquired sentence data, and outputting a plurality of extracted sentence data having a high degree of similarity. [Selection diagram] Figure 6

Description

The present invention relates to a program, an information processing method, an information processing device, and a learning model generation method.

In FY2020, child guidance centers received more than 200,000 consultations regarding child abuse, and the number of consultations at facilities such as child guidance centers, child care facilities, and family and child consultation rooms is increasing year by year. Child Guidance Centers receive a wide range of inquiries every day, not only regarding child abuse, but also regarding child rearing and disabilities. As a result, the workload of facility staff is increasing.

In light of this situation, there is a need to reduce the workload on staff and free up more time for interpersonal work. Therefore, technology has been disclosed to support child abuse counseling services. For example, a technology for supporting decision-making by staff at a child guidance center has been disclosed (Patent Document 1).

Japanese Patent Application Publication No. 2020-184185

Based on the Child Welfare Act, child guidance centers provide temporary protection by removing children suspected of abuse from their families. However, the technique disclosed in Patent Document 1 has room to improve the estimation accuracy for estimating whether or not to perform temporary protection.

One aspect is the purpose of providing programs that support the work of administrative agencies or facilities related to child welfare.

According to one aspect, the program acquires text data related to child consultation, and when the text data related to child consultation is input, the first learning model is trained to output judgment information regarding the necessity of temporary protection. The computer is made to execute a process of inputting the written text data and outputting judgment information.

In one aspect, it is possible to provide programs that support the work of administrative agencies or facilities related to child welfare.

FIG. 1 is a block diagram illustrating a configuration example of an information processing device. It is an explanatory diagram showing a data layout of child guidance center DB. FIG. 3 is a schematic diagram showing a configuration example of a first learning model. It is a flowchart which shows the generation process procedure of a 1st learning model. It is an example of the result screen of the determination information output by an information processing apparatus. 3 is a flowchart showing a processing procedure of a program executed by the information processing device. 3 is a flowchart showing a processing procedure in a subroutine for extracting similar cases. FIG. 2 is a schematic diagram showing the degree of similarity for each sentence included in a progress record. It is an example of a screen displayed by the display unit of the information processing device. 3 is a flowchart showing a processing procedure of a program executed by the information processing device. 3 is a flowchart showing a processing procedure of a program executed by the information processing device. FIG. 7 is a schematic diagram showing determination information for each division progress record. 3 is a flowchart showing a processing procedure of a program executed by the information processing device. It is a flowchart which shows the generation process procedure of a 2nd learning model. 3 is a flowchart showing a processing procedure of a program executed by the information processing device. 3 is a flowchart showing a processing procedure of a program executed by the information processing device. 3 is a flowchart showing a processing procedure of a program executed by the information processing device.

[Embodiment 1]
In this embodiment, a child consultation system that supports the work of administrative agencies or facilities related to child welfare will be described.

This system uses AI (Artificial Intelligence) to determine whether temporary protection is necessary based on text data related to child consultation. The text data is, for example, a progress record created by staff members in their daily work. This progress record is a document that records the child's information, and is created by staff members after telephone calls, interviews, or reports from each agency. Progress records include the child's family structure, home environment, living situation, personality, behavior at school, and chief complaint. This system supports staff in their work by inputting progress records into a learning model, which will be described later, and outputting predetermined information. The text data will be described below as a progress record.

Hereinafter, this embodiment will be described based on the drawings.

FIG. 1 is a block diagram showing a configuration example of an information processing device 10. As shown in FIG. The information processing device 10 is an information device such as a server computer or a personal computer that provides this system. In this embodiment, the explanation will be given assuming that one computer performs the processing, but the processing may be performed in a distributed manner by a plurality of computers. Further, a server computer on the cloud may execute some of the processing executed by the information processing device 10.

The information processing device 10 includes a control section 11 , a main storage section 12 , a communication section 13 , an auxiliary storage section 14 , a display section 15 , and an input section 16 .

The control unit 11 is one or more processors such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), a GPU (Graphics Processing Unit), or a quantum processor, and executes various information processing.

The main storage unit 12 is a temporary storage area such as SRAM (Static Random Access Memory) or DRAM (Dynamic Random Access Memory), and temporarily stores data necessary for the control unit 11 to execute processing.

The communication unit 13 is a communication interface for connecting to a communication network such as the Internet or a LAN (Local Area Network).

The auxiliary storage unit 14 is a memory such as an SSD (Solid State Drive) or an HDD (Hard Disk Drive).

The display unit 15 is a display screen such as a liquid crystal display or an organic EL (Electro Luminescence) display, and displays images. The display unit 15 displays a result screen that includes a risk determination format and determination information, which will be described later.

The input unit 16 is an input interface such as a touch panel or mechanical operation buttons, and receives operation inputs from the staff. The input unit 16 may be a microphone that collects voice commands from the staff member.

The auxiliary storage unit 14 stores a program (program product) 140 that causes the information processing device 10 to execute processing, a first learning model 141, a second learning model 142, a third learning model 143, a fourth learning model 144, and a fifth learning model. 145, child guidance center DB (data base) 150, and other data.

Note that the information processing device 10 may include a reading unit that reads the portable storage medium 10a, and may read the program 140 from the portable storage medium 10a. Further, the information processing device 10 may download the program 140 from another computer via a communication network.

Program 140 may be deployed to run on a single computer, located at one site, or distributed across multiple sites and executed on multiple computers interconnected by a communications network. I can do it.

In this embodiment, a case will be described in which the information processing device 10 learns each learning model, but another computer may learn the learning models. After that, the learning model is deployed to the information processing device 10.

FIG. 2 is an explanatory diagram showing the data layout of the child guidance center DB 150. The child guidance center DB 150 is a DB that stores actual progress records and child information based on consultation cases submitted to the child guidance center in the past. The child guidance center DB 150 is a DB that stores fields such as child ID (Identifier), child information, name, gender, grade, age, consultation case ID, name registrant, progress record, and presence/absence of temporary protection.

The child ID field stores identification information for identifying the child. The child information field stores the name, gender, grade, and age of the corresponding child. The name field stores the name of the corresponding child. The gender field stores the gender (eg, male or female) of the corresponding child. The grade field stores the grade indicating the age group of the corresponding child. The age field stores the age of the corresponding child.

The consultation case ID field stores identification information for identifying cases related to child consultation. The name field stores the name of the staff member who created the corresponding consultation case. The progress record field stores progress records regarding child consultation. The presence/absence of temporary protection field stores the presence or absence of temporary protection for the corresponding consultation case.

FIG. 3 is a schematic diagram showing a configuration example of the first learning model 141.

The first learning model 141 is a learning model that has been trained to output determination information regarding the necessity of temporary protection when a progress record regarding child consultation is input. The first learning model 141 is, for example, BERT (Bidirectional Encoder Representations from Transformers).

BERT is a natural language processing model that uses bidirectional encoded expression using Transformers, and is configured by connecting Transformer Encoders in a network. BERT receives as input a progress record divided into tokens (eg, words) and outputs a vector corresponding to each token.

In addition to BERT, the first learning model 141 includes LSTM (Long Short Term Memory), RNN (Recurrent Neural Network), GPT (Generative Pre-trained Transformer), SVM (Support Vector Machine), and Seq2Seq (Sequence to Sequence). Inference processing may be performed using a model such as , random forest, or decision tree.

The determination information is, for example, output data such as whether temporary protection is necessary, the probability of temporary protection being necessary, or the level of temporary protection necessity.

FIG. 4 is a flowchart showing the procedure for generating the first learning model 141. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 acquires progress records from the child guidance center DB 150 (step S101). The staff member inputs the judgment information in association with the obtained progress record. The control unit 11 receives input of determination information (step S102).

The control unit 11 stores training data in which the progress record and the determination information are associated with each other in the child guidance center DB 150 (step S103). Here, in step S101 and step S102, the progress record is input information, and the determination information is output information. The control unit 11 performs similar processing for a large number of consultation cases, and stores a plurality of training data sets in the child guidance center DB 150 (step S104).

The control unit 11 generates the first learning model 141 using the training data set (step S105). Specifically, the information processing device 10 uses, for example, error backpropagation to adjust the parameters so that there is no difference between the value output from the first learning model 141 and the actual presence or absence of temporary protection. Optimization is performed to generate a final first learning model 141. The control unit 11 ends the series of processing.

Next, various functions when this system is used at a child guidance center will be explained. Various functions include determining temporary protection, extracting similar cases, and extracting sentences from progress records.

[Determination of temporary protection]
The employee inputs a new progress record through the information processing device 10. The information processing device 10 inputs the input new progress record to the first learning model 141, and obtains whether temporary protection is necessary as determination information.

The information processing device 10 assigns a child ID and a consultation case ID to the new progress record. The information processing device 10 stores the child ID, the consultation case ID, and the necessity of temporary protection in the child guidance center DB 150 in association with each other. The information processing device 10 searches for the child ID of the target child and displays determination information, which will be described later with reference to FIG.

FIG. 5 is an example of a result screen of the determination information output by the information processing device 10.

The result screen displayed by the display unit 15 of the information processing device 10 includes child information, whether temporary protection is necessary, past similar cases, etc. for the corresponding consultation case.

The child information includes, for example, a child ID and name. The information processing device 10 reads child information from the child guidance center DB 150 and displays it.

Whether or not temporary protection is necessary is information used by staff when making the final decision as to whether temporary protection should be implemented for the child being consulted. The first learning model 141 outputs, for example, a value between 0 and 1 regarding the necessity of temporary protection. Each value is associated with a degree of risk. Specifically, the information processing device 10 displays risk 0 if the value is 0, risk 5 if the value is 0.5, and risk 10 if the value is 1.

Further, the information processing device 10 uses the attention mechanism to identify a portion of the progress record (for example, a sentence or a word) that contributed to the output of the determination information. Specifically, the information processing device 10 extracts the portion of the progress record that received attention based on the inner product of the portions of the progress record. When displaying a portion of the progress record on the display unit 15, the information processing device 10 performs an emphasized display such as making the relevant part bold or underlining the relevant part. This increases visibility for staff and supports decision-making. In addition, the information processing device 10 may specify a portion of the progress record by determining the SHAP value or the like.

The information processing device 10 identifies whether temporary protection is necessary based on a preset threshold value. The information processing device 10 determines that temporary protection is necessary when the threshold value is equal to or greater than the set value (for example, 0.5 or more), and performs temporary protection when the threshold value is less than the set value (for example, less than 0.5). It is determined that there is no need for

The staff member can set the threshold value for the temporary protection determination output by the first learning model 141 through the input unit 16 of the information processing device 10. The information processing device 10 updates the threshold value by accepting input of the threshold value from the staff member. For example, if the threshold value is changed from 0.5 to 0.7, the number of children determined to require temporary protection will be reduced. This makes it possible to adjust the number of temporary shelters according to their capacity, leading to a reduction in the workload of staff.

A past similar case is a progress record that is similar to a newly acquired progress record. Past similar cases include the consultation case ID, progress record, and whether or not temporary protection was actually obtained. The information processing device 10 extracts past similar cases from the child guidance center DB 150 and displays them by a similar case extraction process described later. By referring to similar past cases, officials can comprehensively judge the necessity of temporary protection.

In FIG. 5, a progress record of child ID: A220, child name: aaa, consultation case ID: C250, and an indicator representing the "degree of risk" are shown. This indicator is an "11-step" number line that indicates the degree of risk in correspondence with the value output by the first learning model 141. If the output value is less than 0.5, the degree of risk is low and an arrow is displayed to the left of the halfway point. On the other hand, if the output value is 0.5 or more, the degree of risk is high and an arrow is displayed to the right of the halfway point. Note that the output value and the corresponding risk level are merely examples, and are not limited to 11 levels.

In this progress record, the information processing device 10 determines that "temporary protection is necessary" and displays "O" in the corresponding direction. The information processing device 10 underlines and highlights the sentence "The parent caused serious injury to the child" which contributed to the output of the determination information. The first learning model 141 calculates the value as 0.8, and displays the degree of risk corresponding to the value as a numerical value and an arrow ("Risk 8" in FIG. 5) on the indicator. Furthermore, the information processing device 10 displays three past similar cases (C102, C042, and C008).

Further, the information processing device 10 displays “Requires close examination” in the range of Risk 5 and Risk 6 on the indicator. This is a display to alert the staff because if the value output by the first learning model 141 is between 0.5 and 0.6, certainty regarding the necessity of temporary protection cannot be guaranteed. be. If the arrow on the indicator is within the "requires close examination" range, staff members will need to conduct further scrutiny to determine whether temporary protection is necessary.

Note that the range of "requiring close examination" is not limited to between 0.5 and 0.6, and other value ranges may be set as "necessary of close examination."

FIG. 6 is a flowchart showing the processing procedure of the program 140 executed by the information processing device 10. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 acquires progress records regarding child consultation (step S201). The control unit 11 reads out the first learning model 141 (step S202). The control unit 11 inputs the acquired progress record to the first learning model 141 (step S203). The control unit 11 acquires the value output by the first learning model 141 (step S204).

The control unit 11 compares the acquired value with a preset threshold and determines whether the acquired value is greater than or equal to the threshold (step S205). If it is equal to or greater than the threshold (step S205: YES), the control unit 11 specifies that there is a need for temporary protection, and stores it in the child guidance center DB 150 in association with the consultation case ID (step S206). On the other hand, if it is not above the threshold (step S205: NO), the control unit 11 specifies that there is no need for temporary protection, and stores it in the child guidance center DB 150 in association with the consultation case ID (step S207).

The control unit 11 starts a subroutine for extracting similar cases (step S208). The similar case extraction subroutine is a subroutine that extracts a progress record similar to the present case from the child guidance center DB 150, and displays the progress record and whether temporary protection is available for the progress record. The flow of this process will be described later with reference to FIG.

The control unit 11 causes the display unit 15 to display whether or not temporary protection is necessary and an indicator corresponding to the consultation case ID (step S209). The control unit 11 uses the attention mechanism to identify and highlight a portion of the progress record that contributed to the output of the determination information (step S210). The control unit 11 ends the series of processing.

[Extraction of similar cases]
The information processing device 10 vectorizes a new progress record (hereinafter referred to as a new case) and a plurality of past progress records (hereinafter referred to as past cases), and then compares the vectors with each other to calculate a cosine similarity. The information processing device 10 extracts past cases similar to the new case based on the calculated cosine similarity. Cosine similarity is expressed between 0 and 1, and the closer the value is to 1, the higher the similarity between sentences.

Specifically, considering the case where five past cases are stored in the child guidance center DB 150, the information processing device 10 compares the "new case vector" and the five "past case vectors", respectively. , calculate five cosine similarities for each past case. The information processing device 10 extracts a predetermined number of past cases from the five past cases based on the calculated cosine similarity.

The information processing device 10 extracts a predetermined number of past cases from the child guidance center DB 150 and displays them on the display unit 15 , and also reads out the presence or absence of temporary protection in the extracted past cases and displays them on the display unit 15 .

In addition to the process of extracting past cases using the cosine similarity described above, past cases similar to the new case may be extracted using a trained learning model. In that case, when the information processing device 10 inputs two progress records, it uses a learning model trained to output the presence or absence of similarity between the progress records, Identify cases.

Furthermore, the information processing apparatus 10 is not limited to the process of extracting the predetermined number of past cases described above, and may extract past cases having a cosine similarity that is equal to or greater than a predetermined threshold.

In addition, the information processing device 10 identifies the degree of similarity based on the number of matches of a predetermined keyword (for example, violence) that appears in the two progress records, and then extracts some progress records with a high degree of similarity. Good too.

FIG. 7 is a flowchart showing the processing procedure in the similar case extraction subroutine. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 reads past cases from the child guidance center DB 150 (step S301). The control unit 11 compares the new case and a plurality of past cases to calculate cosine similarity (step S302). Specifically, the control unit 11 calculates cosine similarity for each sentence, and identifies a plurality of sentences in descending order of cosine similarity. The control unit 11 extracts a plurality of past cases with high similarity based on the calculated cosine similarity (step S303).

The control unit 11 reads out the presence or absence of temporary protection corresponding to each past case (step S304). The control unit 11 causes the display unit 15 to display the plurality of extracted progress records as well as the presence or absence of actual temporary protection corresponding to the past case (step S305). The control unit 11 ends the series of processing.

[Extraction of a sentence from the progress record]
In order to prevent child abuse, support from a variety of related organizations is required not only for early detection and early response, but also at each stage from ``prevention'' to ``helping abused children become independent.'' Furthermore, appropriate collaboration between child guidance centers and municipalities is desired, depending on the circumstances of each child consultation case. For this reason, a risk assessment format is used to facilitate information sharing regarding consultation cases of "child abuse" or "suspected child abuse" that child guidance centers and municipalities are aware of. This embodiment will be described assuming that the risk determination format is a risk assessment sheet.

The risk assessment sheet has judgment items for determining the necessity of temporary protection, such as abuse status, child situation, household situation, and guardian situation, and staff members can It is structured so that the situation understood for each item can be recorded. Staff use this risk assessment sheet to organize the facts of the case and determine the need for temporary protection.

Currently, staff read a huge amount of progress records and then enter the necessary information on the risk assessment sheet, but with this system, sentences that are highly relevant to the judgment items on the risk assessment sheet are automatically extracted from the progress records. Extracted. During automatic extraction, when a staff member inputs a new progress record through the information processing device 10, the information processing device 10 extracts a sentence from the progress record, and selects a sentence that is highly relevant to the judgment item as a judgment item. Classify and display by category.

Specifically, the information processing device 10 calculates a plurality of degrees of similarity between the extracted sentence and a plurality of model sentences associated with the determination items. The information processing device 10 stores in advance a plurality of model sentences in the auxiliary storage unit 14 for each of the 27 determination items. For example, if the judgment item is physical abuse, the model sentences are used to describe cases in which physical abuse is recognized, such as ``the child has been repeatedly assaulted by his parents'' and ``there are bruises on the child's body from being hit.'' This is a text that describes a typical situation and is prepared in advance by staff.

The information processing device 10 determines a predetermined number of sentences related to the determination item based on the plurality of similarities calculated for each sentence. Preferably, the similarity calculation is performed by specifying "item vectors" of a plurality of model sentences associated with the determination items, but this process will be described later.

FIG. 8 is a schematic diagram showing the degree of similarity for each sentence included in the progress record. An example will be shown in which four sentences in the progress record are extracted when the judgment item is "physical abuse."

The information processing device 10 extracts the following sentences from the progress record.
Sentence 1: A single-parent household with children and mothers.
Sentence 2: The child is on the verge of not going to school.
Sentence 3: There is a possibility that she has been subjected to violence from her father.
Sentence 4: The child has a bruise on his leg.

The information processing device 10 reads two model sentences associated with "physical abuse" from the auxiliary storage unit 14. Note that the number of model sentences may be any number, but is preferably about 4 to 7.
Model sentence 1: I have been repeatedly assaulted by my parents.
Model sentence 2: There are bruises on the child's body from being hit.

As shown in FIG. 8, the information processing device 10 calculates a plurality of degrees of similarity by comparing sentences 1 to 4 in the progress record with model sentences 1 and 2, respectively. The information processing device 10 determines a predetermined number of sentences related to "physical abuse" based on a plurality of degrees of similarity calculated for each sentence in the progress record. For example, the two sentences with the highest similarity are "Sentence 4: There is a bruise on the child's leg" with a similarity of 0.9, and "Sentence 3: The child was assaulted by his father" with a similarity of 0.8. The following two sentences have been decided:

The information processing device 10 performs the above-described process for each determination item, and classifies a plurality of sentences that are highly related to the relevant determination item.

FIG. 9 is an example of a screen displayed by the display unit 15 of the information processing device 10.

FIG. 9A is an example of a classification result screen for each determination item. The classification result screen in FIG. 9A includes child information (for example, child ID and name), judgment items, similarity, a sentence in the progress record, a check box 200 for inputting the risk level, and a drop-down 201. .

As shown in FIG. 9A, the information processing device 10 arranges the sentences in the progress record determined by the above-described process for each determination item, and displays the degree of similarity calculated between the sentences and the model sentences. The information processing device 10 displays the sentences in the progress record in order from the top in descending order of the calculated degree of similarity.

The employee performs the task of inputting the risk level based on a sentence in the progress record displayed by the information processing device 10. When the employee selects the check box 200 through the information processing device 10, a drop-down 201 for inputting a risk level is displayed on the display unit 15.

A drop-down 201 displays multiple risk levels that can be selected. For example, the risk level is expressed in six stages: "Risk S", which requires urgent intervention, "Risk I", which is high risk, "Risk II", which is medium risk, "Risk III", which is low risk, and "Problem", where there is no risk. There are "none" and "unknown" where risk assessment is not possible.

When inputting the risk level, the staff will set the risk level to ``Risk S'' for the situation ``The child has a bruise on his leg,'' and the situation ``The child may be experiencing violence from his father.'' However, the risk level is determined to be "Risk III". A staff member inputs such a risk level through the input unit 16 of the information processing device 10. The information processing device 10 stores the received risk level in association with the child ID and the determination item.

As described above, by checking the automatically extracted sentences (for example, sentence 4 and sentence 3), the employee can efficiently input the risk level without reading a huge amount of progress records. This allows you to efficiently create risk assessment sheets and free up time for interpersonal tasks.

FIG. 9B is an example of a risk assessment sheet.

As shown in FIG. 9B, the risk assessment sheet is displayed with the input contents in FIG. 9A reflected. Specifically, the information processing device 10 displays the risk assessment sheet including the determination items, and highlights the relevant portion based on the risk level input by the staff member in FIG. 9A. This can reduce the possibility that staff will overlook important points.

The information processing device 10 stores in advance a display format related to highlighting corresponding to the risk level in the auxiliary storage unit 14. For example, the information processing device 10 displays the judgment items of "Risk S", which requires urgency, in red (dark hatching in the example of FIG. 9B), and the judgment items of "Risk III", which has a low risk level, in yellow (in the example of FIG. 9B). In this example, displaying with light hatching improves visibility.

In addition to manually inputting the risk level by staff, AI may also be used to specify the risk level. Specifically, a learning model is prepared that outputs a risk level when multiple sentences from a progress record are input. The information processing device 10 inputs a plurality of sentences into a learning model and identifies a risk level.

FIG. 10 is a flowchart showing the processing procedure of the program 140 executed by the information processing device 10. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 acquires progress records regarding child consultation (step S401). The control unit 11 extracts one sentence from the acquired progress record (step S402). The control unit 11 calculates a plurality of degrees of similarity by comparing the plurality of sentences in the progress record and the model sentence, respectively (step S403). The control unit 11 determines a predetermined number of sentences related to the relevant determination item based on the plurality of similarities calculated for each sentence in the progress record (step S404). The control unit 11 performs this process using a plurality of determination items.

The control unit 11 classifies the determined sentence into a plurality of determination items, and outputs the classification results (step S405). The control unit 11 receives input of the risk level for each determination item from the staff through the input unit 16 (step S406). The control unit 11 highlights the corresponding location on the risk assessment sheet based on the input content of the risk level corresponding to the item (step S407). The control unit 11 ends the series of processing.

As described above, the work burden on staff can be reduced by determining temporary protection, extracting similar cases, and extracting sentences from progress records.

Note that the effects of creating a model sentence are, for example, as follows.

Training data can be created by creating multiple model sentences for each judgment item. Natural language processing models such as SentenceBERT can improve their accuracy by learning from triplet training data. Triplet-type training data is a set of three sentences: sentence A (left side), sentence B (middle), and sentence C (right side), where sentences A and B are a combination of sentences with high similarity, and sentence A is a combination of sentences with high similarity. and sentence C are created so that they are a combination of sentences with a low degree of similarity.

Such triplet-type learning data can be created by creating a plurality of model sentences for each judgment item and implementing the following algorithm.

Assuming that the i-th model sentence of judgment item j is expressed as sentence i,j, generalizing it will result in the following notation.
Left side: sentence i=p,j=q
Center: Sentence any i≠p,j=q
Right side: sentence any i, any j≠q

1) The p-th sentence of the judgment item q is acquired as the left side, and the sentences other than the p-th sentence of the same judgment item q are acquired as the center. If there are five model sentences each, there will be 10 combinations.
2) On the right side, sentences included in judgment items different from q are obtained. When there are 27 judgment items, there are 26 judgment items that are different from q, so by multiplying this by 5, there are 130 different model sentences.
3) By adding the sentences obtained in 2) to the left and center combinations created in 1), 10 x 130 = 1300 triplets can be created.
4) By doing this for all the determination items, 1300×27=35100 triplets can be created, which can be used for learning.

Next, processing using the "item vector" will be explained. In the similarity calculation described with reference to FIG. 8, since the similarity is calculated only using the vector of the model sentence, a sentence that is not semantically similar to the determination item may be extracted.

Specifically, when the model sentence is ``I have been repeatedly assaulted by my parents,'' a sentence that does not match the semantics of ``I have been repeatedly made fun of by my older brother'' may be inappropriately extracted. This is an error caused by the similarity of the words "repeated" which have nothing to do with physical abuse. In other words, if the degree of similarity is simply determined between the model sentence and the extracted sentence, the desired degree of similarity will not be obtained, and a sentence that is not semantically similar will be extracted. To avoid this error, perform the following processing.

The information processing device 10 identifies item vectors of a plurality of model sentences associated with the determination item. The item vector is a vector common to a plurality of model sentences, and refers to a vector common to model sentences 1 and 2 used in FIG. 8, for example. The item vector is specified by principal component scores, average values, or the like.

The information processing device 10 specifies the composite vector of each model sentence based on the vector of the model sentence of the judgment item and the item vector using the principal component score or the average value. The number of composite vectors calculated at this time is equal to the number of registered model sentences. Specifically, the information processing device 10 specifies an item vector based on model sentence 1 and model sentence 2, and then combines model sentence 1 and item vector into composite vector 1, model sentence 2, and item vector. A composite vector 2 obtained by combining the vectors is specified. After that, the information processing device 10 calculates a plurality of degrees of similarity between the vector of one sentence extracted from the progress record and the composite vector 1 and composite vector 2.

For example, when calculating the similarity between composite vector 1 and sentence 1, similarity 1-1 is calculated, and when calculating the similarity between composite vector 2 and sentence 1, similarity 2-1 is calculated. Similarly, the information processing device 10 calculates a plurality of degrees of similarity with the composite vector for the number of extracted sentences. In other words, if there are four sentences included in the progress record, 2×4=8 similarities are calculated.

Finally, the information processing device 10 compares the degrees of similarity (e.g. eight) and identifies sentences in descending order of degree of similarity. The information processing device 10 displays the sentences in the progress record in order of similarity from top to bottom.

FIG. 11 is a flowchart showing the processing procedure of the program 140 executed by the information processing device 10. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 identifies item vectors of a plurality of model sentences associated with the determination item (step S501). The control unit 11 specifies a composite vector of each model sentence based on the vector of the model sentence of the determination item and the item vector (step S502). The control unit 11 vectorizes the sentence extracted from the progress record (step S503). The control unit 11 calculates a plurality of degrees of similarity between the extracted one-sentence vector and a plurality of composite vectors (step S504).

The control unit 11 compares the degrees of similarity and identifies sentences extracted from the progress record in descending order of degree of similarity (step S505). The control unit 11 arranges the sentences in the progress record from the top and displays them on the display unit 15 in descending order of similarity (step S506). The control unit 11 ends the series of processing.

As described above, according to the first embodiment, the three functions of this system can support the work of a child guidance center.

[Embodiment 2]
In the second embodiment, a mode will be described in which a plurality of sentences are input to the first learning model 141, a plurality of pieces of judgment information are acquired, and whether or not temporary protection is necessary is obtained based on each piece of judgment information.

The information processing device 10 divides a sentence included in the progress record into a plurality of sentences. This is called a division progress record. The information processing device 10 acquires the division progress record and inputs it to the first learning model 141. The information processing device 10 acquires the determination information output by the first learning model 141. This division process is performed because up to 512 tokens can be input to BERT.

At this time, a plurality of pieces of determination information are acquired for each divided progress record, but the determination information may differ depending on the position at which the progress record is divided. In other words, even though the progress records are the same, the output results vary, making it difficult for staff to decide which results to use. Therefore, the information processing device 10 performs a process of acquiring final determination information based on a plurality of pieces of determination information.

An example in which a progress record is divided into four divided progress records will be described below.

For example, the information processing device 10 divides the progress record into division progress record a, division progress record b, division progress record c, and division progress record d so that it fits within 512 tokens. The information processing device 10 inputs each division progress record to the first learning model 141 and acquires determination information corresponding to each division progress record. FIG. 12 illustrates an example in which the first learning model 141 outputs determination information (whether temporary protection is necessary) corresponding to the division progress record as a value between 0 and 1.

FIG. 12 is a schematic diagram showing determination information for each division progress record. When division progress record a is input, the value is 0.3, when division progress record b is input, the value is 0.5, when division progress record c is input, the value is 0.9, and when division progress record d is input, the value is 0.3. When input, the value is output as 0.9. At this time, since the judgment information differs depending on the division progress record input to the first learning model 141, the information processing device 10 calculates a predetermined statistic such as an average value or a maximum value in order to obtain the final judgment information. calculate. Thereby, even if there are variations in the judgment information, it is possible to appropriately support the staff in making judgments.

In the above example, when the information processing device 10 calculates the average value,
(Average value) = (0.3 + 0.5 + 0.9 + 0.9) / 4 = 0.65
becomes. Based on this calculation formula, the information processing device 10 specifies the necessity of temporary protection in this progress record as 0.65 (risk 6.5) as final determination information. The information processing device 10 displays the specified final determination information on the display unit 15. Employees check the judgment information displayed on the information processing device 10 and utilize it for actual business decisions.

FIG. 13 is a flowchart showing the processing procedure of the program 140 executed by the information processing device 10. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 acquires a plurality of division progress records (step S601). The control unit 11 reads out the first learning model 141 (step S602). The control unit 11 inputs the plurality of acquired division progress records to the first learning model 141 (step S603). The control unit 11 acquires the determination information output by the first learning model 141 (step S604).

The control unit 11 calculates a predetermined statistic based on the acquired determination information (step S605). The control unit 11 outputs final determination information based on predetermined statistics (step S606). The control unit 11 causes the display unit 15 to display the final determination information (step S607). The control unit 11 ends the series of processing.

[Embodiment 3]
In the second embodiment, a mode has been described in which a predetermined statistical value is calculated to obtain final determination information. In the third embodiment, a plurality of division progress records are input to the first learning model 141 to obtain judgment information, and the judgment information is input to the second learning model 142 to obtain whether temporary protection is necessary. The form will be explained.

The second learning model 142 is a learning model that has been trained to output whether or not temporary protection is necessary when judgment information is input. The information processing device 10 acquires the determination information for each division progress record described in Embodiment 2, and creates a second learning model based on a training dataset consisting of a combination of the determination information and the necessity of temporary protection. 142 is generated.

The information processing device 10 inputs the determination information for each division progress record output by the first learning model 141 to the second learning model 142. The information processing device 10 acquires the necessity of temporary protection outputted by the second learning model 142. The second learning model 142 at this time is, for example, a decision tree model such as XGBoost. In addition, the second learning model 142 may be a neural network (NN), an SVM, or the like.

FIG. 14 is a flowchart showing the procedure for generating the second learning model 142. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 acquires the determination information output from the first learning model 141 (step S701). The staff member inputs whether there is a need for temporary protection in association with the obtained judgment information. The control unit 11 receives an input indicating whether or not temporary protection is necessary (step S702).

The control unit 11 stores the determination information and "the necessity of temporary protection" in the child guidance center DB 150 in association with each other (step S703). Here, in step S701 and step S702, the determination information is input information, and "whether or not temporary protection is necessary" is output information.

The control unit 11 performs similar processing for a large number of consultation cases, and stores a plurality of training data sets in the child guidance center DB 150 (step S704). The control unit 11 generates the second learning model 142 using the training data set (step S705). The control unit 11 ends the series of processing.

FIG. 15 is a flowchart showing the processing procedure of the program 140 executed by the information processing device 10. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 acquires the determination information output by the first learning model 141 for each division progress record (step S801). The control unit 11 reads out the second learning model 142 (step S802). The control unit 11 inputs determination information for each division progress record to the second learning model 142 (step S803). The control unit 11 acquires the necessity of temporary protection outputted by the second learning model 142 (step S804). The control unit 11 ends the series of processing.

As described above, according to the third embodiment, the information processing device 10 examines and outputs whether there is a need for temporary protection, which makes it easier for the staff to take appropriate measures and appropriately supports the work of the child guidance center. can do.

[Embodiment 4]
In Embodiment 4, a mode will be described in which learning models are annotated according to the attributes of employees and an alert is output when the results output by the respective learning models are different.

The third learning model 143 is a learning model that is trained to output judgment information regarding the necessity of temporary protection when a progress record regarding child consultation is input, and is annotated by an expert. The fourth learning model 144 is a learning model that is trained to output judgment information regarding the necessity of temporary protection when a progress record regarding child consultation is input, and is annotated by an unskilled person.

The information processing device 10 generates a third learning model 143 annotated by an expert among the staff of the child guidance center, and a fourth learning model 144 annotated by an unskilled person. A skilled person is, for example, a staff member who has a certain number of years of experience in child counseling services, and an unskilled person is, for example, a staff member who has less than a certain number of years of experience in child counseling services. Like the first learning model 141 described above, the third learning model 143 and the fourth learning model 144 are trained to output judgment information regarding the necessity of temporary protection when a progress record regarding child consultation is input. .

The information processing device 10 stores in the auxiliary storage unit 14 a third learning model 143 annotated by an expert and a fourth learning model 144 annotated by an unskilled person. The information processing device 10 reads the third learning model 143 and the fourth learning model 144, and inputs the new progress record to each learning model. The information processing device 10 acquires determination information output from the third learning model 143 and the fourth learning model 144.

Generally, it is thought that the third learning model 143 annotated by an expert will output more accurate determination information. Therefore, when the determination information output by each learning model is different, the information processing device 10 outputs an alert and uses it as a reference when considering whether temporary protection is necessary. This allows staff to judge the child guidance center's response policy from a more diverse perspective. On the other hand, if the determination information output by each learning model matches, the information processing device 10 does not output an alert.

In addition to the alert output processing described above, the following processing may be performed. For example, the value output from the third learning model 143 is 0.8 (risk 8), the value output from the fourth learning model 144 is 0.7 (risk 7), and so on. If the difference between the values is within a predetermined value, the information processing device 10 does not need to output an alert. This is because there is no significant difference in the content of judgments between experts and non-experts.

FIG. 16 is a flowchart showing the processing procedure of the program 140 executed by the information processing device 10. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 generates the third learning model 143 based on the training data annotated by the expert (step S901). The control unit 11 generates the fourth learning model 144 based on the training data annotated by an unskilled person (step S902). The control unit 11 stores the third learning model 143 and the fourth learning model 144 (step S903).

The control unit 11 reads the third learning model 143 and the fourth learning model (step S904). The control unit 11 acquires progress records regarding child consultation (step S905). The control unit 11 inputs the acquired progress record to the third learning model 143 and the fourth learning model 144 (step S906).

The control unit 11 acquires the determination information output from each learning model (step S907). The control unit 11 determines whether the pieces of determination information are different (step S908).

If the pieces of determination information are different (step S908: YES), the control unit 11 outputs an alert (step S909). If the pieces of judgment information are not different (step S908: NO), the control unit 11 outputs the judgment information without outputting an alert (step S910). The control unit 11 ends the series of processing.

As described above, according to the fourth embodiment, by annotating the learning model for each different group, the staff can comprehensively judge whether temporary protection is necessary.

[Embodiment 5]
In Embodiment 5, a fifth learning model 145 that determines whether or not to return home will be described.

After the child is placed in temporary custody, staff members provide guidance to the parents on how to interact with the child. Staff carefully consider whether to return the child to the home, taking into consideration the improvement status of the family that received the guidance and the child's condition at the temporary shelter. The purpose of temporary protection is to protect the child's life, but if parents and children live apart, the child's healthy development may be affected. Under these circumstances, staff must weigh the child's life against the child's healthy development at home, and are forced to make difficult decisions.

Therefore, in order to support the staff's decision after temporary protection, the fifth learning model 145 is used to determine whether or not to return to the home. The details of the fifth learning model 145 will be explained below.

In this embodiment, the input information is the progress record regarding the child under temporary protection, and the output information is the determination information regarding the necessity of returning to the home. In addition to the information obtained when the child received consultation, the progress record includes information about the family situation and the child's situation after the child was temporarily taken into custody. The determination information includes output data such as whether there is a need to return to the home, the probability of the need to return to the home, or the level of the need to return to the home.

The information processing device 10 generates the fifth learning model 145 using a plurality of training data sets consisting of a combination of "progress records regarding children in temporary care" and "whether the children have returned to their homes or not." Specifically, the information processing device 10 optimizes the parameters using, for example, an error backpropagation method so that there is no difference between the value output from the fifth learning model 145 and the actual result of returning home, A final fifth learning model 145 is generated.

The information processing device 10 uses the fifth learning model 145 described above to perform a determination process for returning to the home. The information processing device 10 also extracts past cases similar to the new case based on cosine similarity. These detailed processes overlap with those described in Embodiment 1 and will therefore be omitted.

FIG. 17 is a flowchart showing the processing procedure of the program 140 executed by the information processing device 10. The control unit 11 of the information processing device 10 executes the following processing based on the program 140.

The control unit 11 acquires progress records regarding the child under temporary protection (step S1001). The control unit 11 reads out the fifth learning model 145 (step S1002). The control unit 11 inputs the acquired progress record to the fifth learning model 145 (step S1003). The control unit 11 acquires the value output by the fifth learning model 145 (step S1004).

The control unit 11 compares the acquired value with a preset threshold and determines whether the acquired value is greater than or equal to the threshold (step S1005). If it is equal to or greater than the threshold (step S1005: YES), the control unit 11 specifies that there is a need to return to the home, and stores it in the child guidance center DB 150 in association with the consultation case ID (step S1006). On the other hand, if it is not above the threshold (step S1005: NO), the control unit 11 specifies that there is no need to return to the home, and stores it in the child guidance center DB 150 in association with the consultation case ID (step S1007).

The control unit 11 extracts progress records similar to this case from the child guidance center DB 150 based on cosine similarity (step S1008). The control unit 11 causes the display unit 15 to display whether or not there is a need for returning home and an indicator corresponding to the consultation case ID (step S1009). The control unit 11 uses the attention mechanism to identify and highlight a portion of the progress record that contributed to the output of the determination information (step S1010). The control unit 11 ends the series of processing.

As described above, according to Embodiments 1 to 5, the information processing device 10 outputs judgment information obtained by inputting progress records into a learning model to support staff in their work related to child consultation. can do.

The embodiments disclosed herein are illustrative in all respects and are not restrictive. The technical scope of the present invention is not defined in the meaning disclosed above, but is determined based on the description of the claims, and includes all changes within the meaning and range equivalent to the claims.

Items described in each embodiment can be combined with each other. Moreover, the independent claims and dependent claims recited in the appended claims may be combined with each other in any and all combinations, regardless of the form in which they are cited. Further, although the scope of claims uses a format in which claims refer to two or more other claims (multi-claim format), the invention is not limited to this format. It may be written using a multi-claim format that cites at least one multi-claim.

10 Information processing device 10a Portable storage medium 11 Control unit 12 Main storage unit 13 Communication unit 14 Auxiliary storage unit 15 Display unit 16 Input unit 140 Program (program product)
141 1st learning model 142 2nd learning model 143 3rd learning model 144 4th learning model 145 5th learning model 150 Child Guidance Center DB
200 Checkbox 201 Dropdown

Claims (10)

Obtain text data consisting of multiple sentences,
Extract multiple sentences from the acquired text data,
Calculate the degree of similarity between each of the multiple extracted sentences and each of the multiple model sentences associated with each of the multiple judgment items,
In each of the plurality of judgment items, based on the degree of similarity with the plurality of model sentences calculated for each of the extracted sentences, a sentence related to the judgment item is selected from among the plurality of sentences extracted from the text data. determine a predetermined number of sentences,
A program that causes a computer to execute a process of outputting the predetermined number of sentences determined for each of the determination items. Obtain text data consisting of multiple sentences,
Extract multiple sentences from the acquired text data,
obtaining an item vector that is common to a plurality of model sentences associated with each of a plurality of judgment items and that is identified based on a vector obtained by vectorizing the plurality of model sentences;
Identifying a composite vector of each of the model sentences based on the item vector common to the model sentences of the judgment item and a vector obtained by vectorizing each model sentence of the judgment item,
Calculating the degree of similarity between a vector obtained by vectorizing each of the extracted sentences and a composite vector of each of the plurality of model sentences associated with the judgment item,
In each of the plurality of judgment items, the judgment item is selected from among the plurality of sentences extracted from the text data based on the degree of similarity with the composite vector of the plurality of model sentences calculated for each of the extracted sentences. Determine a predetermined number of sentences related to the item,
A program that causes a computer to execute a process of outputting the predetermined number of sentences determined for each of the determination items. Claim 1 or Claim 2, wherein the computer executes a process of displaying the predetermined number of sentences determined for each of the plurality of judgment items in descending order of similarity calculated for each of the predetermined number of sentences. Programs listed. 4. The program according to claim 3, which causes the computer to execute a process of accepting evaluation values for each of the determination items while displaying the predetermined number of sentences in association with each of the determination items. For the assessment sheet in which evaluation contents for each of the plurality of evaluation values are displayed for each of the evaluation items, the evaluation contents corresponding to the evaluation values received for each of the evaluation items are added to the evaluation values. 5. The program according to claim 4, which causes the computer to perform a process of highlighting the display in accordance with the display. Extract text data with a high degree of similarity from the obtained text data,
The program according to claim 1 or 2, which causes the computer to execute a process of outputting a plurality of extracted text data having a high degree of similarity. Obtain text data consisting of multiple sentences,
Extract multiple sentences from the acquired sentence data,
Calculate the degree of similarity between each of the multiple extracted sentences and each of the multiple model sentences associated with each of the multiple judgment items,
In each of the plurality of judgment items, based on the degree of similarity with the plurality of model sentences calculated for each of the extracted sentences, a sentence related to the judgment item is selected from among the plurality of sentences extracted from the text data. determine a predetermined number of sentences,
An information processing method in which a computer executes a process of outputting the predetermined number of sentences determined for each of the determination items. In an information processing device having a control unit,
The control unit includes:
Obtain text data consisting of multiple sentences,
Extract multiple sentences from the acquired text data,
Calculate the degree of similarity between each of the multiple extracted sentences and each of the multiple model sentences associated with each of the multiple judgment items,
In each of the plurality of judgment items, based on the degree of similarity with the plurality of model sentences calculated for each of the extracted sentences, a sentence related to the judgment item is selected from among the plurality of sentences extracted from the text data. determine a predetermined number of sentences,
An information processing device that outputs the predetermined number of sentences determined for each of the determination items. Obtain text data consisting of multiple sentences,
Extract multiple sentences from the acquired text data,
obtaining an item vector that is common to a plurality of model sentences associated with each of a plurality of judgment items and that is identified based on a vector obtained by vectorizing the plurality of model sentences;
Identifying a composite vector of each of the model sentences based on the item vector common to the model sentences of the judgment item and a vector obtained by vectorizing each model sentence of the judgment item,
Calculating the degree of similarity between a vector obtained by vectorizing each of the extracted sentences and a composite vector of each of the plurality of model sentences associated with the judgment item,
In each of the plurality of judgment items, the judgment item is selected from among the plurality of sentences extracted from the text data based on the degree of similarity with the composite vector of the plurality of model sentences calculated for each of the extracted sentences. Determine a predetermined number of sentences related to the item,
An information processing method in which a computer executes a process of outputting the predetermined number of sentences determined for each of the determination items. In an information processing device having a control unit,
The control unit includes:
Obtain text data consisting of multiple sentences,
Extract multiple sentences from the acquired text data,
obtaining an item vector that is common to a plurality of model sentences associated with each of a plurality of judgment items and that is identified based on a vector obtained by vectorizing the plurality of model sentences;
Identifying a composite vector of each of the model sentences based on the item vector common to the model sentences of the judgment item and a vector obtained by vectorizing each model sentence of the judgment item,
Calculating the degree of similarity between a vector obtained by vectorizing each of the extracted sentences and a composite vector of each of the plurality of model sentences associated with the judgment item,
In each of the plurality of judgment items, the judgment item is selected from among the plurality of sentences extracted from the text data based on the degree of similarity with the composite vector of the plurality of model sentences calculated for each of the extracted sentences. Determine a predetermined number of sentences related to the item,
An information processing device that outputs the predetermined number of sentences determined for each of the determination items.

Images