Information retrieval system and information retrieval method

Abstract

To provide an information retrieval system, and the like, configured to make it possible to efficiently and easily decide on a property for reference.SOLUTION: An information retrieval system includes: a learning model generation unit which learns, by machine learning, data on architectural structures classified into multiple classes to generate a learning model configured to, when receiving data on an architectural structure, output a class that the architectural structure belongs to; a prediction unit which predicts a class that an architectural structure to be designed belongs to by inputting data on the architectural structure to be designed to the learning model; a similarity calculation unit which calculates similarities between the architectural structure to be designed and architectural structures belonging to the predicted class, on the basis of data on the architectural structures belonging to the predicted class and the data on the architectural structure to be designed; and a retrieval result output unit which outputs the architectural structures belonging to the predicted class in descending order of similarity to the architectural structure to be designed.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to an information retrieval system and an information retrieval method.

Background Art

[0002] When a structural designer conducts a structural design of a building to be designed, it is often the case that the structural drawings, calculation sheets, estimated cost materials, etc. of past buildings similar to the building to be designed are investigated and used as references. However, it is difficult for a structural designer with a short number of years of experience to find a structurally similar property from past properties based on their own knowledge. Therefore, it is necessary for a structural designer with a long number of years of experience to determine the properties to be referred to based on their knowledge and experience.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] When investigating past similar properties manually, it is not realistic to determine the properties to be referred to after investigating all properties, so the number of properties to be investigated is limited and it takes time to find the properties to be referred to. Also, when a structural designer with a long number of years of experience is absent, or even when a structural designer with a long number of years of experience cannot make a judgment, it is difficult to determine the properties to be referred to. The present invention has been made in view of the above problems, and its main object is to provide an information retrieval system or the like that can efficiently and easily determine the properties to be referred to.

Means for Solving the Problems

[0005] The present invention relates to an information retrieval system, which includes a learning model generation unit that performs machine learning on building data classified into a plurality of classes to generate a learning model that outputs the class to which a building belongs when the building data is input, a prediction unit that inputs the data of a building to be designed into the learning model to predict the class to which the building to be designed belongs, a similarity calculation unit that calculates the similarity between the building to be designed and the buildings belonging to the predicted class based on the data of the buildings belonging to the predicted class and the data of the building to be designed, and a search result output unit that ranks and outputs the buildings belonging to the predicted class in descending order of similarity to the building to be designed.

[0006] The present invention also relates to an information retrieval method, which includes a learning model generation step of performing machine learning on building data classified into a plurality of classes to generate a learning model that outputs the class to which a building belongs when the building data is input, a prediction step of inputting the data of a building to be designed into the learning model to predict the class to which the building to be designed belongs, a similarity calculation step of calculating the similarity between the building to be designed and the buildings belonging to the predicted class based on the data of the buildings belonging to the predicted class and the data of the building to be designed, and a search result output step of ranking and outputting the buildings belonging to the predicted class in descending order of similarity to the building to be designed.

Brief Description of the Drawings

[0007]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Embodiments for Carrying Out the Invention

[0008] Hereinafter, embodiments of the present invention will be described. Note that the embodiments described below do not unduly limit the content of the present invention described in the claims. Also, not all of the configurations described in this embodiment are necessarily essential components of the present invention.

[0009] 1. Configuration FIG. 1 is a diagram showing an example of the functional blocks of the information retrieval system (information retrieval device) according to this embodiment. The information retrieval system 1 includes a processing unit 100, an input unit 110, a storage unit 120, and a display unit 130.

[0010] The input unit 110 is a device for inputting (detecting) input information from the user, and outputs the input information (operation information) of the user to the processing unit 100. The functions of the input unit 110 can be realized by input devices such as a keyboard, a mouse, and a touch panel.

[0011] The storage unit 120 stores programs and various data for causing a computer to function as each part of the processing unit 100, and functions as a work area for the processing unit 100. The functions of the storage unit 120 can be realized by a hard disk, a memory (RAM), etc.

[0012] The display unit 130 displays the image generated by the processing unit 100, and its functions can be realized by a display such as an LCD or a touch panel that also functions as the input unit 110.

[0013] The processing unit 100 performs various processes using the storage unit 120 as a work area. The functions of the processing unit 100 can be realized by hardware such as various processors (CPU, DSP, etc.) and programs. The processing unit 100 includes a learning model generation unit 101, a prediction unit 102, a similarity calculation unit 103, and a search result output unit 104. Note that the device including the learning model generation unit 101 (the first device) and the device including the prediction unit 102, the similarity calculation unit 103, and the search result output unit 104 (the second device) may be made different, and the information retrieval system may be configured by the first device and the second device.

[0014] The learning model generation unit 101 performs machine learning on the data of buildings classified into a plurality of classes, and generates a learning model that outputs the class to which the building belongs when the data of the building is input. The generated (learned) learning model is stored in the storage unit 120. The learning model is, for example, a random forest composed of a plurality of decision trees that output the class to which the building belongs when the data of the building is input.

[0015] The prediction unit 102 inputs the data of the building to be designed (data input from the input unit 110) into the learning model stored in the storage unit 120, and predicts the class to which the building to be designed belongs. When the learning model is a random forest, the prediction unit 102 predicts the class to which the building to be designed belongs based on the classes output from the plurality of decision trees.

[0016] The similarity calculation unit 103 calculates the similarity between the building data belonging to the class predicted by the prediction unit 102 and the building data to be designed based on the building data belonging to the predicted class. The similarity calculation unit 103 obtains the similarity for each item included in the building data, and calculates the similarity between the building data belonging to the predicted class and the building data to be designed by a function (such as the weighted sum method) that weights the similarities of each item. It may be.

[0017] The search result output unit 104 ranks at least some of the predicted buildings in descending order of similarity to the building to be designed and outputs (displays on the display unit 130) them to the display unit 130.

[0018] Further, the prediction unit 102 predicts a plurality of classes to which the building to be designed belongs, and the similarity calculation unit 103 calculates the similarity between the building to be designed and the buildings belonging to each of the plurality of predicted classes based on the data of the buildings belonging to each of the plurality of predicted classes and the data of the building to be designed.

[0019] 2. Method of this embodiment Next, the method of this embodiment will be described with reference to the drawings.

[0020] FIG. 2 is a diagram showing the overall flow of the method of this embodiment. First, past buildings (past properties) are classified in advance into a plurality of groups so that buildings with structurally similar features belong to the same group. The classified groups are called classes. Then, using the data of the past buildings (past property data) classified into a plurality of classes as learning data, a learning model LM is generated that is learned to output, as a prediction result, the class (predicted class) to which the building belongs when the data of the building is input. Here, the data of the building is data consisting of a plurality of items indicating the structural features of the building, for example, data including the type of structure (such as S structure, RC structure, SRC structure, etc.), the number of floors, the span (column spacing), the presence or absence of a seismic isolation device (whether it is a seismic isolation building or not), the presence or absence of seismic resistance elements, the presence or absence of a void, etc. FIG. 3 shows an example of past property data. The past property data includes the data of each building (property) and the information of the class to which the building belongs. In the past property data shown in FIG. 3, the property "a" has an RC structure, is equipped with a seismic isolation device, and is classified into the class "1", and the property "b" has an S structure, a span of 15 m, and is classified into the class "4".

[0021] The learning model LM includes a plurality (N) of decision trees DT1 to DT NA random forest composed of..., and the prediction results are output by majority voting of a plurality of decision trees DT1 to DT N Examples of the decision trees that make up the learning model LM are shown in FIGS. 4 and 5. The decision tree shown in FIG. 4 is learned to output class "1" as the class to which the building belongs, for example, when the input building data indicates that the building is a "seismic isolation building" and the structural type is "RC construction". Also, the decision tree shown in FIG. 5 is learned to output class "4" as the class to which the building belongs, for example, when the input building data indicates that the structural type is not "SRC construction" and the span is "15 m or more".

[0022] The user (structural designer) inputs the data of the building to be designed (new property data) into the learning model LM. Then, the most frequent class among the classes output from the plurality of decision trees DT1 to DT N is output as the prediction result (predicted class) of the class to which the building to be designed belongs.

[0023] Next, from the past property data, the data of buildings belonging to the same class as the predicted class (same-class data) is extracted, and the similarity between the data of each extracted building and the data of the building to be designed (new property data) is calculated. The calculation of the similarity is performed by a scoring function that weights the similarity of each item (structural type, number of floors, span, etc.) of the building data. The weighting of the scoring function is set for each class. Specifically, when the similarity between the data of an arbitrarily selected building and the data of other buildings is calculated, the weighting of the scoring function for that class is adjusted so that other buildings belonging to the same class as the selected building are ranked higher in terms of similarity. This process is repeated to create a scoring function for each class. This process is repeated to create a scoring function for each class.

[0024] Next, among the buildings belonging to the same class as the predicted class, the top predetermined number of buildings with a high similarity to the building to be designed are ranked in descending order of similarity and displayed (ranked display) on the display unit 130 as the search results for similar properties. Note that all buildings belonging to the same class as the predicted class may be ranked and displayed. Also, the similarity values of each building may be displayed together with the ranking based on the similarity. The structural designer can select the similar properties (past properties for reference) that they require based on the ranking of similar properties output in response to the input of data for a new property.

[0025] According to the information retrieval system of the present embodiment, when data for a new property is input by a structural designer, the class to which the new property belongs is predicted using a learning model, and past properties belonging to the predicted class are ranked in descending order of similarity to the new property and output as search results. Thus, when the structural designer is working on the structural design of a new property, the working time for selecting past properties for reference can be significantly shortened. Also, even a structural designer with a short number of years of experience can determine the past properties for reference based on the search results. Further, by searching for properties similar to the new property from among all past properties, the evaluation of similarity can be performed comprehensively. Also, by displaying the search results in a ranking and displaying the similarity values of each property, while referring to the ranking and similarity, ultimately the structural designer can make a judgment and determine the past properties for reference.

[0026] In the above example, the case where one class determined by a majority vote of a plurality of decision trees DT1 to DT in the learning model LM is output as a prediction result has been described. However, for a plurality of decision trees DT1 to DT N The case where one class determined by a majority vote is output as a prediction result has been described, but for a plurality of decision trees DT1 to DT NAmong the classes output from [[ID=]], it is also possible to output, as the prediction result, the top predetermined number of classes with a large number of output (predicted number). (Output multiple predicted classes). In this case, the similarity between the data of past properties belonging to each of the multiple predicted classes and the new property data is calculated. For example, if the top two classes with a large predicted number are output as the prediction result, the class with the largest predicted number (ranked first in the prediction) is class "4", and the class with the next largest predicted number (ranked second in the prediction) is class "2", then the similarity between the past properties belonging to class "4" and the new property is calculated, and the similarity between the past properties belonging to class "2" and the new property is calculated, and the top predetermined number of past properties with a high similarity to the new property is displayed in a ranking.

[0027] Here, the similarity may be weighted according to the prediction rank, such that the higher the prediction rank of the class, the larger the coefficient multiplied by the similarity between the past property belonging to the class and the new property. For example, if the class ranked first in the prediction is class "4" and the class ranked second in the prediction is class "2", then the coefficient multiplied by the similarity between the past property belonging to class "4" and the new property is made larger than the coefficient multiplied by the similarity between the past property belonging to class "2" and the new property. Also, the similarity may be weighted according to the predicted number, such that the larger the predicted number of the class, the larger the coefficient multiplied by the similarity between the past property belonging to the class and the new property. For example, when there are 9 decision trees that output class "4" and 1 decision tree that outputs class "2" (the difference in the predicted numbers is large), the coefficient multiplied by the similarity between the past property belonging to class "4" and the new property is made larger than when there are 6 decision trees that output class "4" and 4 decision trees that output class "2" (the difference in the predicted numbers is small).

[0028] In the case of a building, there may be multiple structures that are structurally similar and span multiple uses, and they are not necessarily classified into the same class. Also, when conducting the structural design of a building, the reference building may be changed depending on the part. For example, while referring to Building A with a similar number of floors as the design target, Building B may be referred to for the beam structure of the reference floor. As described above, predicting multiple classes and outputting the results, and searching for similar buildings among the past buildings belonging to each of the multiple classes and displaying them in a ranking can handle cases where similar buildings span multiple classes or where the reference buildings differ for each part.

[0029] Also, in the above example, the case of using a random forest as the learning model was described, but the present invention is not limited to this, and algorithms such as the k-nearest neighbor method and logistic regression may be adopted as the machine learning algorithm.

[0030] 3. Processing Next, an example of the processing of the information retrieval system of the present embodiment will be described using the flowchart of FIG. 6. It is assumed that the learning model LM is generated by the learning model generation unit 101 and stored in the storage unit 120.

[0031] First, the prediction unit 102 inputs the data of the building to be designed input from the input unit 110 into the learning model LM stored in the storage unit 120, and obtains the prediction result of the class to which the building to be designed belongs (step S10).

[0032] Next, the similarity calculation unit 103 obtains the data of the buildings belonging to the same class as the class predicted by the prediction unit 102 from the past building data stored in the storage unit 120 (step S11), and based on the obtained data and the data of the building to be designed, calculates the similarity between the building to be designed belonging to the same class as the predicted class and the building to be designed using the scoring function corresponding to the class (step S12).

[0033] Next, the search result output unit 104 ranks the buildings belonging to the same class as the predicted class (information for specifying the building, such as the name of the building) in descending order of similarity to the building to be designed and outputs the ranked results to the display unit 130 (step S13).

[0034] The present invention is not limited to those described in the above embodiments, and various modifications can be made. For example, terms cited as broad or synonymous terms in the description of the specification or drawings can be replaced with broad or synonymous terms in other descriptions in the specification or drawings.

Description of Reference Numerals

[0035] 1... Information search system, 100... Processing unit, 101... Learning model generation unit, 102... Prediction unit, 103... Similarity calculation unit, 104... Search result output unit, 110... Input unit, 120... Storage unit, 130... Display unit

Claims

1. A learning model generation unit that performs machine learning on building data classified into a plurality of classes and generates a learning model that outputs the class to which a building belongs when the building data is input; A prediction unit that inputs the data of the building to be designed into the learning model and predicts the class to which the building to be designed belongs; A similarity calculation unit that calculates the similarity between the building belonging to the predicted class and the building to be designed based on the data of the building belonging to the predicted class and the data of the building to be designed; An information retrieval system comprising a search result output unit that ranks and outputs the buildings belonging to the predicted class in descending order of similarity to the building to be designed.

2. In Claim 1, The prediction unit Predicts a plurality of classes to which the building to be designed belongs, The similarity calculation unit Calculates the similarity between the building belonging to each of the plurality of predicted classes and the building to be designed based on the data of the building belonging to each of the plurality of predicted classes and the data of the building to be designed. An information retrieval system characterized by that.

3. In Claim 1, The similarity calculation unit An information retrieval system characterized in that the similarity between the building belonging to the predicted class and the building to be designed is calculated by a function that weights the similarity of each item of the building data.

4. A learning model generation step of performing machine learning on building data classified into a plurality of classes and generating a learning model that outputs the class to which a building belongs when the building data is input; A prediction step of inputting the data of the building to be designed into the learning model and predicting the class to which the building to be designed belongs; A similarity calculation step of calculating the similarity between the building belonging to the predicted class and the building to be designed based on the data of the building belonging to the predicted class and the data of the building to be designed; An information retrieval method comprising a search result output step of ranking and outputting the buildings belonging to the predicted class in descending order of similarity to the building to be designed.

Images