Information processing program, information processing device, and information processing method

Abstract

In the present invention, a machine learning model is trained to select a phrase (104) corresponding to a specific part of speech from among phrases extracted from atypical text data (101) in the manufacturing process of a product, and use training data in which an input feature amount including the selected phrase (104) and product configuration information (105) is associated with label information (106) indicating a determination result of a product, to output a determination result (107) corresponding to the input feature amount and a feature amount (108) contributory to the determination.

Description

Information processing program, information processing device, and information processing method

[0001] The present invention relates to an information processing program, an information processing device, and an information processing method.

[0002] To extract important words and phrases from unstructured text data such as notes, general machine learning, explainable AI (XAI), and natural language processing (in other words, morphological analysis) are sometimes used.

[0003] Explainable AI (XAI) is AI that can explain the process leading to prediction and classification results. Natural language processing is the process of dividing notes into parts of speech and extracting words.

[0004] Japanese Patent Application Laid-Open No. 2022-70766

[0005] However, in general machine learning, the output is only a prediction, which is often unconvincing and difficult to accept in the field. In the field, data such as notes that cannot be used for learning is generally accumulated, and there is a risk that the input content cannot be formalized.

[0006] It is anticipated that the accuracy of on-site notes could be improved by breaking them down using natural language processing and incorporating them into features, thereby providing more detailed evidence for predictions. However, there is a risk that unrelated keywords may be used as evidence, which could reduce the readability of the explanations.

[0007] In one aspect, the present invention aims to improve the readability of output data when unstructured text data is input into a machine learning model.

[0008] In one aspect, the information processing program causes a computer to execute a process of selecting a term corresponding to a specific part of speech from among terms extracted from non-standard text data in a product manufacturing process, and using training data that associates the selected term with input features including product configuration information and label information indicating a judgment result regarding the product, training a machine learning model that outputs a judgment result according to the input features and the features that contributed to the judgment.

[0009] In one aspect, the readability of output data can be improved when unstructured text data is input into a machine learning model.

[0010] 4 and 5. FIG. 5 is a graph illustrating the contribution of words and phrases to defect identification in related examples and examples. FIG. 6 is a block diagram schematically illustrating an example of a hardware configuration of an information processing device in an embodiment. FIG. 7 is a block diagram schematically illustrating an example of a software configuration of the information processing device shown in FIG. 2. FIG. 8 is a flowchart illustrating machine learning processing in the learning phase of the information processing device shown in FIG. 2. FIG. 9 is a flowchart illustrating machine learning processing in the prediction phase of the information processing device shown in FIG. 2. FIG. 10 is a table illustrating domain knowledge in part-of-speech selection shown in FIGS. 4 and 5. FIG. 11 is a table illustrating domain knowledge in keyword selection shown in FIGS. 4 and 5. FIG. 12 is a table illustrating data other than notes shown in FIGS. 4 and 5. FIG. 13 is a table illustrating data other than notes including the keywords shown in FIGS. 4 and 5. FIG. 14 is a diagram illustrating output data in the information processing device shown in FIG. 2. FIG. 15 is a diagram illustrating an explainable AI model that directly outputs an explanation in addition to a prediction result in an embodiment. FIG. 16 is a diagram illustrating a model that gives an explanation to a general AI in an embodiment. FIG. 17 is a diagram illustrating a model that gives an explanation to a general AI about image data in an embodiment.

[0011] [A] Embodiments An embodiment will be described below with reference to the drawings. However, the embodiments shown below are merely examples, and are not intended to exclude various modifications or application of techniques not explicitly stated in the embodiments. In other words, the present embodiment can be implemented with various modifications within the scope of its purpose. Furthermore, each figure does not intend to include only the components shown in the figure, but may include other functions, etc.

[0012] FIG. 1 is a graph illustrating the contribution of words to defect identification in the related examples and examples.

[0013] At production lines and other work sites, non-standard text data such as memos is created, and the data format required for AI predictions is often not consistent. Often, text data like memos are created that cannot be formatted, such as what types of defects are occurring on the production line. This can lead to the accumulation of data that cannot be used for learning unless some ingenuity is put into place.

[0014] For example, suppose a memo is entered saying, "Body 2 is damaged." When natural language processing (in other words, morphological analysis) is applied to this memo, all parts of speech contained in the memo are output as the result: "'body' (noun), '2' (noun), 'to' (particle), 'damaged' (noun), 'there' (verb)."

[0015] Furthermore, as shown in the related example of symbol A1, even if all parts of speech contained in the memo are input into the AI ​​model and the output shows that "2" and "ni" are important in terms of the contribution of each part of speech to identifying defects, this would not be convincing.

[0016] Therefore, in the embodiment, by utilizing the domain knowledge of using only nouns for the parts of speech included in the natural language processing result, "'Body' (noun), '2' (noun), 'damaged' (noun)" may be output.

[0017] Furthermore, in an embodiment, useful keywords for the explanation may be narrowed down by, for example, excluding parts of speech that contain only one letter or only numbers, thereby outputting "'body' (noun), 'damaged' (noun)."

[0018] As a result, as shown in the example indicated by symbol A2, only keywords that are effective in explanation are input into the AI ​​model, and output data is generated that indicates that "body" and "damage" are effective in identifying certain defects in terms of the contribution of each part of speech to identifying defects, thereby increasing the persuasiveness of the output data.

[0019] Also, suppose that a memo is entered saying, "The deviation is larger than the standard." In this case, similar to the example shown in Figure 1, when natural language processing is applied to this memo, all parts of speech included in the memo are output as a result: "'Deviation' (noun), 'ga' (particle), 'kiyoku' (noun), 'yori' (particle), 'ooi' (adjective)."

[0020] By utilizing the domain knowledge of using adjectives in addition to the nouns described above in Figure 1 for the parts of speech included in the natural language processing results, "'deviation' (noun), 'criterion' (noun), 'big' (adjective)" may be output.

[0021] Furthermore, keywords that are effective for explanation may be narrowed down by excluding parts of speech that consist of only one character or only numbers. In this example, the output result utilizing the above domain knowledge does not include parts of speech that consist of only one character or only numbers, so "'deviation' (noun), 'criterion' (noun), 'big' (adjective)" are output as they are.

[0022] This allows only keywords that are effective in explanation to be input into the AI ​​model, and output data is generated that indicates that "deviation," "standard," and "large" are effective in identifying certain defects in terms of the contribution of each part of speech to identifying defects, making the output data more convincing.

[0023] FIG. 2 is a block diagram schematically illustrating an example of the hardware configuration of the information processing device 1 according to the embodiment.

[0024] As shown in FIG. 1, the information processing device 1 includes a CPU (Central Processing Unit) 11, a memory unit 12, a display control unit 13, a storage device 14, an input interface (IF) 15, an external recording medium processing unit 16, and a communication IF 17.

[0025] The memory unit 12 is an example of a storage unit, and is, for example, a read-only memory (ROM) or a random access memory (RAM). A program such as a basic input / output system (BIOS) may be written to the ROM of the memory unit 12. The software program in the memory unit 12 may be read and executed by the CPU 11 as appropriate. The RAM of the memory unit 12 may be used as a temporary storage memory or a working memory.

[0026] The display control unit 13 is connected to and controls the display device 131. The display device 131 is a liquid crystal display, an organic light-emitting diode (OLED) display, a cathode ray tube (CRT), an electronic paper display, or the like, and displays various information to an operator, etc. The display device 131 may be combined with an input device, such as a touch panel. The display device 131 displays various information to a user of the information processing device 1.

[0027] The storage device 14 is a storage device with high IO performance, and may be, for example, a dynamic random access memory (DRAM), a solid state drive (SSD), a storage class memory (SCM), or a hard disk drive (HDD).

[0028] The input IF 15 may be connected to input devices such as a mouse 151 and a keyboard 152, and may control the input devices such as the mouse 151 and the keyboard 152. The mouse 151 and the keyboard 152 are examples of input devices, and an operator performs various input operations via these input devices.

[0029] The external recording medium processing unit 16 is configured to allow a recording medium 160 to be attached. The external recording medium processing unit 16 is configured to be able to read information recorded on the recording medium 160 when the recording medium 160 is attached. In this example, the recording medium 160 is portable. For example, the recording medium 160 is a flexible disk, an optical disk, a magnetic disk, a magneto-optical disk, a semiconductor memory, or the like.

[0030] The communication IF 17 is an interface that enables communication with an external device.

[0031] The CPU 11 is an example of a processor, and is a processing device that performs various controls and calculations. The CPU 11 executes an operating system (OS) and programs loaded into the memory unit 12, thereby realizing various functions as will be described later with reference to Fig. 3. The CPU 11 may be a multiprocessor including multiple CPUs, a multi-core processor having multiple CPU cores, or a configuration having multiple multi-core processors.

[0032] The device for controlling the overall operation of the information processing device 1 is not limited to the CPU 11, and may be, for example, any one of an MPU, DSP, ASIC, PLD, and FPGA. The device for controlling the overall operation of the information processing device 1 may also be a combination of two or more of the CPU, MPU, DSP, ASIC, PLD, and FPGA. Note that MPU is an abbreviation for Micro Processing Unit, DSP is an abbreviation for Digital Signal Processor, and ASIC is an abbreviation for Application Specific Integrated Circuit. Furthermore, PLD is an abbreviation for Programmable Logic Device, and FPGA is an abbreviation for Field Programmable Gate Array.

[0033] FIG. 3 is a block diagram schematically illustrating an example of the software configuration of the information processing device 1 according to the embodiment.

[0034] The CPU 11 of the information processing device 1 may function as an analysis unit 111, a part-of-speech selection unit 112, a part-of-speech word / phrase extraction unit 113, a valid part-of-speech extraction unit 114, and an AI processing unit 115.

[0035] The analysis unit 111 uses an existing open source algorithm such as MeCab to perform morphological analysis on on-site memo data 101 (described later using Figures 4 and 5, etc.), and outputs sentence segmentation results 102 (described later using Figures 4 and 5, etc.) divided into parts of speech.

[0036] The part-of-speech selection unit 112 selects parts of speech, such as nouns and adjectives, necessary for explainable AI processing based on domain knowledge 103 (described later using Figures 4 and 5, etc.).

[0037] The part-of-speech word / phrase extraction unit 113 extracts the parts of speech selected by the part-of-speech selection unit 112 from the sentence segmentation result 102 .

[0038] The effective part of speech extraction unit 114 extracts parts of speech that are effective for explanation from the parts of speech extracted by the part of speech phrase extraction unit 113 based on domain knowledge 103, such as excluding parts of speech that consist of only one character or only numbers, and outputs them as a list of keywords.

[0039] In other words, the part-of-speech word extraction unit 113 and the effective part-of-speech extraction unit 114 select words corresponding to specific parts of speech from words extracted from unstructured text data in the product manufacturing process. In the process of selecting words, the effective part-of-speech extraction unit 114 may further exclude at least one of words whose character length is equal to or shorter than a predetermined length and words consisting only of numbers. In the process of selecting words, the part-of-speech word extraction unit 113 may select noun phrases and adjective phrases.

[0040] The AI ​​processing unit 115 trains the explainable AI model based on the list of keywords 104 output by the effective part-of-speech extraction unit 114, data other than memos 105 (described later using Figures 4, 5, etc.), and correct answer labels 106 (described later using Figures 4, 5, etc.). Furthermore, the AI ​​processing unit 115 uses the trained explainable AI model to output prediction results 107 of defect causes (described later using Figures 4, 5, etc.) and explanations 108 (described later using Figures 4, 5, etc.).

[0041] In other words, the AI ​​processing unit 115 uses training data in which input features including the selected phrase and product configuration information and label information indicating a judgment result for the product are associated with each other to train a machine learning model that outputs a judgment result corresponding to the input features and features that contributed to the judgment. In the process of training the machine learning model, the AI ​​processing unit 115 may output the cause of the product defect as the judgment result, and may output the contribution of each selected phrase to identifying the defect as a contributing feature.

[0042] The machine learning process in the learning phase of the information processing device 1 shown in FIG. 2 will be described with reference to the flowchart (steps S1 to S5) shown in FIG.

[0043] The analysis unit 111 performs morphological analysis on the on-site memo data 101 using an existing algorithm such as MeCab, and outputs the sentence segmentation result 102 divided into parts of speech (step S1).

[0044] The part-of-speech selection unit 112 selects parts of speech, such as nouns and adjectives, necessary for explainable AI processing based on the domain knowledge 103 (step S2).

[0045] The part-of-speech word / phrase extraction unit 113 extracts the parts of speech selected by the part-of-speech selection unit 112 from the sentence segmentation result 102 (step S3).

[0046] Based on domain knowledge 103, such as excluding parts of speech consisting of only one character or only numbers, the effective part of speech extraction unit 114 extracts parts of speech that are effective for explanation from the parts of speech extracted by the part of speech phrase extraction unit 113 and outputs them as a list of keywords (step S4).

[0047] The AI ​​processing unit 115 trains the explainable AI model based on the keyword list 104 output by the effective part-of-speech extraction unit 114, the data other than notes 105, and the correct answer label 106 (step S5). Then, the machine learning process in the training phase ends. Note that the data other than notes 105 may be a table including the test date, test line, and specifications of the product for which the cause of the defect is to be identified, as will be described later with reference to FIG. 8 etc.

[0048] Next, the machine learning process in the prediction phase of the information processing device 1 shown in FIG. 2 will be described with reference to the flowchart (steps S1 to S4, S51) shown in FIG.

[0049] The processes in steps S1 to S4 are the same as those described with reference to FIG.

[0050] The AI ​​processing unit 115 uses the learned explainable AI model to output the defect cause prediction result 107 and explanation 108 (step S51). Then, the machine learning process in the prediction phase ends. Details of the defect cause prediction result 107 and explanation 108 will be described later using FIG. 10 etc.

[0051] FIG. 6 is a table illustrating the domain knowledge 103a in the part of speech selection shown in FIGS.

[0052] The part-of-speech selection unit 112 may select the part of speech necessary for the explainable AI processing based on the domain knowledge 103a in part-of-speech selection illustrated in FIG. 6 .

[0053] In the domain knowledge 103a for part-of-speech selection shown in FIG. 6, nouns and adjectives are conjugated (see the "◯" marks), but verbs are not conjugated (see the "X" marks).

[0054] Therefore, the part-of-speech selection unit 112 selects nouns and adjectives as parts of speech necessary for explainable AI processing.

[0055] FIG. 7 is a table illustrating domain knowledge in the keyword selection shown in FIGS.

[0056] As shown in FIG. 7, domain knowledge for keyword selection may include a condition list 103b and a no-no word list 103c.

[0057] The effective part of speech extraction unit 114 may extract effective parts of speech for explanation based on at least one of the condition list 103b and the NG word list 103c.

[0058] In the condition list 103b illustrated in FIG. 7, the condition "length(x) > 1" at index 1 indicates that the length of the character is greater than 1, and the condition "x.isdecimal(x) == False" at index 2 indicates that the character does not consist of only digits.

[0059] The NG word list 103c shown in FIG. 7 includes the NG word "tap" at index 1 and the NG word "place" at index 2.

[0060] Therefore, the valid part of speech extraction unit 114 excludes parts of speech that are one character or less in length, that consist of only numbers, or that correspond to the NG words "tap" or "place" from the parts of speech extracted by the part of speech and phrase extraction unit 113, and outputs the results as a list of keywords 104.

[0061] FIG. 8 is a table illustrating the non-memo data 105 shown in FIGS.

[0062] The data 105 other than memos shown in FIG. 8 includes the test date, test line, CPU type, memory capacity, exterior color, and display type for the product for which the cause of the defect is to be identified.

[0063] The AI ​​processing unit 115 may use data 105 other than notes to train and predict the explainable AI model.

[0064] FIG. 9 is a table illustrating data 105a other than memos that includes the keywords shown in FIGS.

[0065] As shown in FIG. 9, the data other than memos 105 shown in FIG. 8 may have keywords extracted by the effective part of speech extracting unit 114 registered as data other than memos 105a containing keywords.

[0066] 9, for data No. 0, "1" is registered in the flags of the keywords "body" and "damage", indicating that "body" and "damage" have been extracted by the effective part of speech extraction unit 114. Also, for data No. 1, "1" is registered in the flags of the keywords "deviation", "standard", and "large", indicating that "deviation", "standard", and "large" have been extracted by the effective part of speech extraction unit 114. Furthermore, for data No. 2, "0" is registered in the flags of all keywords, indicating that no keywords have been extracted by the effective part of speech extraction unit 114.

[0067] FIG. 10 is a diagram illustrating output data in the information processing device 1 shown in FIG.

[0068] As shown in FIG. 10, the output data may include data 105a other than memos containing keywords related to the identification of the defect cause (in other words, input data), a prediction result 107, and an explanation 108.

[0069] In the example shown in FIG. 10, "body" and "damage" are registered as data 105a other than memos containing keywords, "damage" is shown as a prediction result 107, and the contribution of each of "body," "damage," and "exterior=red" to defect identification is shown in bar graph form as explanation 108.

[0070] FIG. 11 is a diagram illustrating an example of an explainable AI model that directly outputs an explanation in addition to a prediction result in an embodiment.

[0071] The AI ​​processing unit 115 shown in FIG. 3 may function as an explainable AI model 211 as shown in FIG.

[0072] The explainable AI model 211 presents a prediction result 107 and its basis (in other words, an explanation 108) for a certain instance based on data other than notes 105 and a list of keywords 104.

[0073] As the explainable AI model 211, for example, Wide Learning (WL) may be applied.

[0074] FIG. 12 is a diagram illustrating a model for providing explanations to a general AI in an embodiment.

[0075] The AI ​​processing unit 115 shown in FIG. 3 may function as a general AI model 221 and an explanation providing model 222 as shown in FIG.

[0076] A typical AI model 221 outputs a prediction result 107 for a given instance based on non-memo data 105 and a list of keywords 104 .

[0077] The explanation assignment model 222 assigns an explanation 108 to an instance based on the non-memo data 105 and the list of keywords 104 .

[0078] For example, Deep Learning (DL) or Random Forest may be applied as the general AI model 221, and for example, LIME, SHAP, or WL may be applied as the explanation-giving model 222.

[0079] FIG. 13 is a diagram illustrating a model for providing explanations to a general AI about image data in an embodiment.

[0080] The AI ​​processing unit 115 shown in FIG. 3 may function as a general AI model 231 and an explanation providing model 232 as shown in FIG.

[0081] A typical AI model 231 outputs a prediction result 107 for a certain instance based on data 105 other than memos as image data.

[0082] The explanation model 232 provides an explanation 108 for an instance based on the list of keywords 104 .

[0083] For example, DL may be applied as the general AI model 231, and WL may be applied as the explanation-giving model 222.

[0084] In this way, if notes are attached to an image model, it can be used as input for an AI model.

[0085] [B] Effects According to the information processing program, information processing device, and information processing method in the above-described embodiments, the following effects can be achieved, for example.

[0086] The part-of-speech word extraction unit 113 and the effective part-of-speech extraction unit 114 select words corresponding to specific parts of speech from words extracted from unstructured text data in the product manufacturing process. The AI ​​processing unit 115 uses training data in which input features including the selected words and product configuration information are associated with label information indicating a judgment result related to the product to train a machine learning model that outputs a judgment result according to the input features and the features that contributed to the judgment.

[0087] This improves the readability of output data that incorporates non-standard text data into a machine learning model. Predictions made with explainable AI also increase persuasiveness. Furthermore, by effectively and automatically extracting necessary keywords based on domain knowledge, such as on-site memos on the production line, the readability and persuasiveness of explanations can be improved, automating the on-site learning and prediction cycle.

[0088] In the process of selecting words and phrases, the effective part-of-speech extracting unit 114 further excludes at least one of words and phrases whose character length is equal to or less than a predetermined length and words and phrases that are composed only of numbers.

[0089] This makes it possible to prevent unimportant words and phrases such as single-character words and numbers from being output as features that contributed to the determination.

[0090] In the process of selecting words and phrases, the part-of-speech extracting unit 113 selects noun phrases and adjective phrases.

[0091] This makes it possible to output effective parts of speech as features that contribute to the determination of nouns, adjectives, and the like.

[0092] In the process of training the machine learning model, the AI ​​processing unit 115 outputs the cause of the product defect as a judgment result, and outputs the contribution of each selected word to the identification of the defect as a contributing feature. This makes it possible to output appropriate data for identifying the cause of the product defect in the product manufacturing process.

[0093] [C] Others The disclosed technology is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present embodiment. The configurations and processes of the present embodiment can be selected or combined as needed.

[0094] In the above-described embodiment, an example of extracting important words and phrases from unstructured text data in Japanese has been described, but the present invention is not limited to this. Important words and phrases may also be extracted from unstructured text data in languages ​​other than Japanese, such as English.

[0095] For example, for a note that reads "Deviation is greater than the standard," the analysis unit 111 applies natural language processing to output "'deviation' (noun), 'is' (verb), 'greater' (adjective), 'than' (conjunction), 'the' (article), 'standard' (noun)."

[0096] The part-of-speech word / phrase extraction unit 113 utilizes the domain knowledge 103 of using nouns and adjectives to output "'deviation' (noun), 'greater' (adjective), 'standard' (noun)."

[0097] Furthermore, keywords that are effective for explanation may be narrowed down by excluding parts of speech that consist of only one character or only numbers. In this example, the effective part of speech extraction unit 114 outputs "deviation" (noun), "greater" (adjective), "standard" (noun)" as is, because parts of speech that consist of only one character or only numbers are not included in the output result that utilizes the above domain knowledge.

[0098] Also, for example, for a memo "Body 2 damaged," the analysis unit 111 applies natural language processing to output "'body' (noun), '2' (noun), 'damaged' (verb)."

[0099] The part-of-speech word extraction unit 113 utilizes the domain knowledge 103 that uses the past participle form of verbs in addition to nouns and adjectives, and outputs "'body' (noun), '2' (noun), 'damaged' (verb)" as is.

[0100] Furthermore, keywords that are effective for explanation may be narrowed down by excluding parts of speech that consist of only one character or only numbers. In this example, the effective part of speech extraction unit 114 outputs "'body' (noun), 'damaged' (verb)."

[0101] 1: Information processing device 11: CPU 12: Memory unit 13: Display control unit 14: Storage device 16: External recording medium processing unit 101: Memo data 102: Sentence segmentation result 103, 103a: Domain knowledge 103b: Condition list 103c: NG word list 104: Keyword list 105, 105a: Data other than memo 106: Correct answer label 107: Prediction result 108: Explanation 111: Analysis unit 112: Part of speech selection unit 113: Part of speech word / phrase extraction unit 114: Effective part of speech extraction unit 115: AI processing unit 131: Display device 151: Mouse 152: Keyboard 160: Recording medium 211: Explainable AI model 221, 231: General AI model 222, 232: Explanation model 15: Input IF 17: Communication IF

Claims

1. Selecting words that correspond to specific parts of speech from words extracted from non-standard text data in the manufacturing process of a product, training a machine learning model that outputs a judgment result according to the input feature amount and the feature amount that contributed to the judgment, using training data that associates input feature amounts including the selected phrase and product configuration information with label information that indicates a judgment result regarding the product; An information processing program that causes a computer to execute processing.

2. In the process of selecting the words and phrases, the selection is further performed by excluding at least one of words and phrases whose character length is equal to or shorter than a predetermined length and words and phrases consisting only of numbers. The information processing program according to claim 1 , which causes the computer to execute a process.

3. In the process of selecting a phrase, a noun phrase and an adjective phrase are selected. The information processing program according to claim 1 or 2, which causes the computer to execute processing.

4. In the process of training the machine learning model, the cause of the defect of the product is output as the judgment result, and the contribution of each of the selected words to the identification of the defect is output as the contributing feature. The information processing program according to claim 1 or 2, which causes the computer to execute processing.

5. Selecting words that correspond to specific parts of speech from words extracted from non-standard text data in the manufacturing process of a product, training a machine learning model that outputs a judgment result according to the input feature amount and the feature amount that contributed to the judgment, using training data that associates input feature amounts including the selected phrase and product configuration information with label information that indicates a judgment result regarding the product; An information processing device comprising a processor.

6. Selecting words that correspond to specific parts of speech from words extracted from non-standard text data in the manufacturing process of a product, training a machine learning model that outputs a judgment result according to the input feature amount and the feature amount that contributed to the judgment, using training data that associates input feature amounts including the selected phrase and product configuration information with label information that indicates a judgment result regarding the product; An information processing method in which processing is performed by a computer.