Text generation device, text generation method, and program

Abstract

To provide a sentence generation device, a method for generating a sentence, and a program which can unify writings of sentences with sites and states.SOLUTION: A sentence generation device 1 includes: an acquisition unit 100 for acquiring a target sentence including a site wording showing a site and a state wording showing the state of a site; and a FT diagram generation unit 102 for extracting the site wording and the state wording from the target sentence, performing a morphological analysis on the state wording, selecting a morpheme which shows a state from a morpheme group forming the state wording, and connecting the site wording and the morpheme showing a state to each other, thereby generating a first prediction sentence for the target sentence.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a text generation device, a text generation method, and a program.

Background Art

[0002] In response to problems such as a shortage of manpower and personalization in the after-service (maintenance) of equipment, a method of analyzing countermeasures by fault tree analysis is used.

[0003] In fault tree analysis, information necessary for maintenance scattered in the manuals of each model related to the equipment is aggregated, and a database is created using structured data consisting of three layers of event, cause, and treatment, that is, a fault tree diagram (FT diagram). As a result, even maintenance personnel with little experience can grasp the information necessary for maintenance for all models related to the equipment using the FT diagram.

[0004] In the process of aggregating information necessary for maintenance, it is necessary to create a document with unified terms in the text extracted from each manual. Patent Document 1 discloses a technique for unifying terms used in a document. Patent Document 1 describes a technique for analyzing trouble information such as accidents and failures that have occurred in an object in a top-down manner, and creating a tree-structured FT diagram from the causes and results of the trouble.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, while the technology described in Patent Document 1 allows for the linking of cause and effect in a problem, it does not allow for standardization of the wording. In the FT diagram, a series of structured data consisting of three layers—event, cause, and action—is created using a faulty part (component) that is the target of maintenance and a sentence describing the state of that part, for example, "The motor (component) is hot (state)." When extracting such sentences consisting of a part and its state from a manual, multiple sentences with inconsistent wording may be extracted, such as "The motor temperature is high," "The motor temperature is rising," "The temperature is high (motor)," and "The motor temperature is rising." If multiple sentences with inconsistent wording are listed in the FT diagram, it becomes troublesome to check whether these multiple sentences refer to the same thing or different things, making it difficult to obtain the information necessary for maintenance. For this reason, it is desirable to be able to standardize the wording of the sentences describing the parts and their states.

[0007] The present invention has been made to solve the above problems, and its objective is to provide a text generation device, a text generation method, and a program that can standardize the notation of texts that have parts and states. [Means for solving the problem]

[0008] To solve the above problem, one aspect of the present invention includes an acquisition unit that acquires a target document containing part-specific terminology and state-specific terminology indicating the state of the part; and a unit that extracts the part-specific terminology and state-specific terminology from the target document, performs morphological analysis on the state-specific terminology, selects a morpheme indicating the state from the group of morphemes constituting the state-specific terminology, and generates a first predicted document corresponding to the target document by linking the part-specific terminology and the morpheme indicating the state. Based on the degree of similarity between the combination of the part word and the morpheme indicating the state and the combination of the part word and the state word extracted from an existing text, a second predicted text corresponding to the target text is generated, and based on the correspondence between the part word in the first predicted text, the second predicted text, and the target text, one of the first predicted text and the second predicted text is selected, and the selected text is used in conjunction with the text generated to correspond to the target text. This is a text generation device that includes a generation unit.

[0009] Furthermore, one aspect of the present invention is a text generation method performed by a computer text generation device, wherein an acquisition unit acquires a target text that includes part terms indicating a part and state terms indicating the state of the part; a generation unit extracts the part terms and state terms from the target text, performs morphological analysis on the state terms, selects morphemes indicating the state from the group of morphemes constituting the state terms, and generates a first predicted text corresponding to the target text by linking the part terms and the morphemes indicating the state. Based on the degree of similarity between the combination of the part word and the morpheme indicating the state and the combination of the part word and the state word extracted from an existing text, a second predicted text corresponding to the target text is generated, and based on the correspondence between the part word in the first predicted text, the second predicted text, and the target text, one of the first predicted text and the second predicted text is selected, and the selected text is used in conjunction with the text generated to correspond to the target text. This is a method for generating text.

[0010] Furthermore, in one aspect of the present invention, a computer-based text generation device acquires a target text containing part terms indicating a part and state terms indicating the state of the part; extracts the part terms and state terms from the target text; performs morphological analysis on the state terms; selects morphemes indicating the state from the group of morphemes constituting the state terms; and generates a first predicted text corresponding to the target text by linking the part terms and the morphemes indicating the state. Based on the degree of similarity between the combination of the part word and the morpheme indicating the state and the combination of the part word and the state word extracted from an existing text, a second predicted text corresponding to the target text is generated, and based on the correspondence between the part word in the first predicted text, the second predicted text, and the target text, one of the first predicted text and the second predicted text is selected, and the selected text is used in conjunction with the text generated to correspond to the target text. This is a program that makes it happen. [Effects of the Invention]

[0011] According to this invention, it is possible to standardize the notation of sentences that describe parts and states. [Brief explanation of the drawing]

[0012] [Figure 1] This is a block diagram showing the configuration of the text generation device 1 according to the embodiment. [Figure 2] This diagram shows the overall processing flow performed by the text generation device 1 of the embodiment. [Figure 3] This flowchart shows the process flow for generating the initial FT diagram performed by the text generation device 1 of the embodiment. [Figure 4] This flowchart shows the process flow for generating the FT diagram performed by the text generation device 1 of the embodiment. [Figure 5] This flowchart shows the processing flow of the text generation device 1 of the embodiment, which extracts parts and states from text. [Figure 6] This is a diagram for explaining the process of extracting a part in FIG. 5. [Figure 7] This is a diagram for explaining the process shown in FIG. 5. [Figure 8] This is a flowchart showing the flow of the process of generating a sentence based on the matching prediction performed by the sentence generation device 1 of the embodiment. [Figure 9] This is a diagram for explaining the process shown in FIG. 8. [Figure 10] This is a flowchart showing the flow of the process of generating a sentence based on the AI prediction performed by the sentence generation device 1 of the embodiment. [Figure 11] This is a diagram for explaining the process shown in FIG. 10. [Figure 12] This is a flowchart for explaining the process of selecting a sentence generated by the matching prediction and the AI prediction performed by the sentence generation device 1 of the embodiment. [Figure 13] This is a diagram for explaining the process shown in FIG. 12. [Figure 14] This is a diagram for explaining the process of a human confirming the sentence generated by the sentence generation device 1 of the embodiment.

Mode for Carrying Out the Invention

[0013] Hereinafter, the sentence generation device in the embodiment of the present invention will be described with reference to the drawings. Hereinafter, the case where the sentence generation device generates a sentence described in the FT diagram used for the after - service (maintenance) of the device will be illustrated and described, but it is not limited thereto. The sentence generation device of the present embodiment can be applied when generating any sentence.

[0014] (Regarding the configuration of the sentence generation device 1) FIG. 1 is a block diagram showing the configuration of the sentence generation device 1 of the embodiment. The sentence generation device 1 is a computer managed by an operator who provides a service for generating sentences. As the sentence generation device 1, a PC, a server device, a cloud server, etc. can be applied.

[0015] The document generation device 1 comprises, for example, a control unit 10, a storage unit 12, and a display unit 14. The control unit 10 is implemented by causing the CPU (Central Processing Unit) provided as hardware in the document generation device 1 to execute a program. The control unit 10 comprises, for example, an acquisition unit 100, an initial FT diagram generation unit 101, and an FT diagram generation unit 102.

[0016] The acquisition unit 100 acquires various types of information. For example, the acquisition unit 100 acquires various manuals for the equipment to be maintained. The acquisition unit 100 acquires manuals, for example, that have been transmitted from an external device or stored on a storage medium such as a USB (Universal Serial Bus) memory. The acquisition unit 100 stores the acquired manuals in the storage unit 12 as manual information 120, which will be described later. The acquisition unit 100 also outputs information related to the acquired manuals to the initial FT diagram generation unit 101.

[0017] The initial FT diagram generation unit 101 generates an initial FT diagram. The initial FT diagram is a collection of text containing information that can be used for fault tree analysis (hereinafter referred to as FTA information) extracted from the manual. FTA information is a series of structured data consisting of three layers: event, cause, and action. Here, an event is an event that a worker must address, such as equipment failure or malfunction. The cause is the reason why the event occurs. The action is the action taken to resolve the event and is determined according to the cause. By generating an FT diagram that associates the cause and action corresponding to the event, it becomes possible to grasp the information necessary for equipment maintenance using the FT diagram. In other words, when a trouble event occurs, it is possible to grasp the cause and action for each model and take action to resolve the trouble event.

[0018] Each of the sentences corresponding to the event, cause, and action in the FTA information includes wording that indicates the faulty part (component) to be maintained, and wording that indicates the condition of that part. For example, suppose the sentence corresponding to the event is "motor malfunction," the sentence corresponding to the cause is "motor overheating," and the sentence corresponding to the action is "motor replacement." In this case, the wording that indicates the part is "motor." The wording that indicates the condition is "malfunction" for the event, "overheating" for the cause, and "replacement" for the action.

[0019] The initial FT diagram generation unit 101 generates the initial FT diagram by taking the text itself extracted from various manuals and using it as a collection of texts to be included in the initial FT diagram. In other words, the notation of the text included in the initial FT diagram is not standardized. For example, to indicate that "the motor is hot," multiple sentences with inconsistent notation are included, such as "the motor temperature is getting high," "the motor temperature is rising," "the temperature is high (motor)," and "the motor temperature is rising."

[0020] The FT diagram generation unit 102 generates an FT diagram. The FT diagram is a standardized version of the text in the initial FT diagram. For example, the FT diagram generation unit 102 uses phrases such as "the motor temperature is high," "the motor temperature is rising," "the temperature is high (motor)," and "the motor temperature is rising" to generate a standardized phrase such as "the motor is hot," and then generates the FT diagram as a collection of these standardized phrases. By generating an FT diagram with standardized notation, the FT diagram generation unit 102 makes it possible to present maintenance information to workers in an easy-to-understand manner.

[0021] The FT diagram generated by the FT diagram generation unit 102 is reviewed by a human (for example, an expert), and the reviewed FT diagram is stored in the storage unit 12 as reviewed FT diagram information 121, as described later.

[0022] The storage unit 12 is composed of storage media such as an HDD, flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), RAM (Random Access read / write Memory), ROM (Read Only Memory), or a combination thereof. The storage unit 12 stores programs for executing various processes of the document generation device 1, and temporary data used when performing these processes.

[0023] The memory unit 12 stores manual information 120 and reviewed FT diagram information 121. The manual information 120 is information about various manuals for the equipment to be maintained. The manual information 120 includes various information about the manuals, such as title, publication date, contents, and revision history. The reviewed FT diagram information 121 is information relating to reviewed FT diagrams. For example, the reviewed FT diagram information 121 includes information showing text corresponding to the event, cause, and action for each piece of equipment to be maintained. The text included in the reviewed FT diagram information 121 is text whose notation has been standardized by the FT diagram generation unit 102, and whose content has been reviewed by experts, etc.

[0024] The display unit 14 includes a display device such as a liquid crystal display and displays images corresponding to the control of the control unit 10, such as the initial FT diagram generated by the initial FT diagram generation unit 101, the FT diagram generated by the FT diagram generation unit 102, and the reviewed FT diagram stored in the storage unit 12 as reviewed FT diagram information 121.

[0025] (Regarding the processing performed by text generation device 1) Figure 2 is a diagram showing the overall flow of processing performed by the text generation device 1 of the embodiment. First, the text generation device 1 selects the model to be maintained (step S1). The text generation device 1 displays an input field on the display unit 14 for entering the model to be maintained, along with a message such as "Please select the model to be maintained." The text generation device 1 selects the model to be maintained based on the information entered into the input field using an input device such as a mouse or keyboard.

[0026] Next, the document generation device 1 acquires a manual to be used for generating the initial FT diagram (step S2). The document generation device 1 acquires various documents that are likely to contain FTA information, such as manuals for the model to be maintained, such as inspection and maintenance manuals and troubleshooting manuals. The document generation device 1 stores the information about the acquired manuals as manual information 120 in the storage unit 12.

[0027] Next, the text generation device 1 generates an initial FT diagram (step S3). The specific process by which the text generation device 1 generates the initial FT diagram will be explained in detail later. As a result, the text generation device 1 outputs the generated initial FT diagram (step S4).

[0028] The document generation device 1 responds to the generated initial FT diagram depending on whether or not a reviewed FT diagram already exists. Specifically, if a reviewed FT diagram already exists, the document generation device 1 generates an FT diagram using the initial FT diagram and the reviewed FT diagram (step S5). The specific process by which the document generation device 1 generates the FT diagram will be explained in detail later. As a result, the document generation device 1 outputs the generated FT diagram (step S6).

[0029] On the other hand, if there is no reviewed FT diagram, or if an FT diagram is generated in step S5, the document generation device 1 performs a content review (step S7). The content review is the process of having experts or others confirm whether the content described in the initial FT diagram or the text included in the FT diagram is correct. As a result, the document generation device 1 outputs a reviewed FT diagram (step S8). Then, the reviewed FT diagram is registered in the system (step S9). Specifically, the document generation device 1 stores the FT diagram and the initial FT diagram whose content, as determined by experts or others, is correct, in the storage unit 12 as reviewed FT diagram information 121.

[0030] (Regarding the process of generating the initial FT diagram) Figure 3 is a flowchart showing the process flow for step S3 in Figure 2, which is the process of generating the initial FT diagram.

[0031] The initial FT diagram generation unit 101 obtains manual information 120 by referring to the storage unit 12 (step S30). The initial FT diagram generation unit 101 uses the obtained manual information 120 to extract tabular pages contained in the obtained manual (step S31). In many manuals, FTA information is presented in tabular format. Specifically, in many cases, most of the text in the tables related to failures included in the manual can be used as FTA information, and tabular pages tend to be the most frequent source of FTA information. For this reason, the initial FT diagram generation unit 101 determines whether or not the manual contains tables, and if it does, it extracts the pages containing those tables as pages to be used to extract FTA information in the next step S32. For example, the initial FT diagram generation unit 101 extracts sections enclosed in frames in the manual as tables.

[0032] The initial FT diagram generation unit 101 extracts FTA information from the manual (step S32). For example, the initial FT diagram generation unit 101 extracts tables, flowcharts, diagrams, text, etc., from pages that describe equipment failures as FTA information. The initial FT diagram generation unit 101 extracts tables from the manual if they are included, and then extracts tables whose titles contain specific keywords, such as "failure," "trouble," or "error." This allows the initial FT diagram generation unit 101 to extract tables related to equipment failures, etc., as FTA information.

[0033] The initial FT diagram generation unit 101 displays a manual along with a message such as "Please select FTA information," and experts may visually review the displayed manual and manually select parts of the manual that can be used as FTA information, such as flowcharts, diagrams, and text. The text generation device 1 acquires the flowcharts, etc., selected by the experts and uses the acquired flowcharts, etc., as FTA information. In this case, the text generation device 1 may extract pages from the manual that frequently contain specific keywords such as "malfunction," "trouble," and "error," display the extracted pages, and present them to experts as pages that are highly likely to be extracted as FTA information.

[0034] The initial FT diagram generation unit 101 has an expert confirm the content of the FTA information extracted from the manual (step S33). For example, the initial FT diagram generation unit 101 displays the FTA information extracted in step S32 along with options such as OK / NG. The expert judges whether the displayed FTA information is correct to use as FTA information and inputs the judgment result by selecting an option such as OK / NG. In this way, the initial FT diagram generation unit 101 obtains the judgment result of the expert regarding whether the content of the FTA information extracted from the manual is correct or not.

[0035] The initial FT diagram generation unit 101 structures the FTA information, which has been confirmed to be correct by experts, etc. (step S34). The initial FT diagram generation unit 101 classifies the FTA information into events, causes, and actions. For example, the initial FT diagram generation unit 101 predicts sentences corresponding to events in the FTA information using a trained model that predicts sentences corresponding to events. The trained model that predicts sentences corresponding to events is a model that learns the correspondence between sentences and whether or not a sentence corresponds to an event by learning training data that indicates whether or not a sentence corresponds to an event for sentences included in the FTA information. By learning such correspondences, the trained model that predicts sentences corresponding to events will, for example, classify sentences indicating failures or malfunctions, such as "engine overheating" or "insufficient engine output," as "events." Furthermore, the initial FT diagram generation unit 101 uses a trained model that predicts cause and action sentences to predict sentences corresponding to events in the FTA information. The trained model that predicts cause and action sentences is a model that learns the correspondence between sentences and whether or not a sentence corresponds to a cause or action by learning training data that indicates whether or not a sentence in the FTA information corresponds to a cause or action associated with a certain event. By learning such correspondences, the trained model that predicts cause and action sentences classifies sentences that indicate causes corresponding to the FTA information classified as events, for example, sentences such as "deterioration of engine oil" and "unsuitable engine oil" as causes of engine overheating, as "causes" corresponding to the "engine overheating" event. The initial FT diagram generation unit 101 also classifies sentences that indicate actions corresponding to the FTA information classified as causes, for example, sentences such as "change the engine oil" as a countermeasure when the engine oil deteriorates, as "action" corresponding to engine oil deterioration. The initial FT diagram generation unit 101 generates structured FTA information as an initial FT diagram (step S35).

[0036] (Regarding the process of generating the FT diagram) Figure 4 is a flowchart showing the process flow for step S5 in Figure 2, that is, the process for generating the FT diagram.

[0037] The FT diagram generation unit 102 obtains the initial FT diagram generated in step S3 of Figure 2 and the reviewed FT diagram (step S50). The FT diagram generation unit 102 obtains the reviewed FT diagram by, for example, referring to the storage unit 12 to obtain the reviewed FT diagram information 121.

[0038] The FT diagram generation unit 102 deletes symbols such as item numbers included in the initial FT diagram (step S51). Since symbols such as item numbers are assigned in the manual, they become unnecessary when extracted as text to be included in the FT diagram. For this reason, the FT diagram generation unit 102 deletes symbols such as item numbers included in the initial FT diagram.

[0039] The FT diagram generation unit 102 unifies variations in wording in the text included in the initial FT diagram (step S52). Manuals often contain a mix of different notations due to differences in equipment manufacturers, manual creation dates, authors, etc. In order to generate an FT diagram with consistent notation, the FT diagram generation unit 102 unifies variations in wording in the text included in the initial FT diagram. For example, if the text in the initial FT diagram contains multiple notations for the same word, such as "air cleaner" and "air cleaner," the FT diagram generation unit 102 unifies it to one of them, for example, "air cleaner."

[0040] The FT diagram generation unit 102 completes the subject of sentences that lack one (step S53). In manuals, the subject may be omitted in some sentences. For example, on a page describing an engine, since it is clear that the sentence is about the "engine," the subject "engine" may be omitted, and only "overheating" or "insufficient power" may be written. Alternatively, following a sentence that corresponds to an event or cause, such as "the engine oil is degraded," the subject of the sentence corresponding to the action may be omitted because it is clear that the subject is "engine oil," and only "replace" may be written. In cases where the initial FT diagram includes sentences with omitted subjects, the FT diagram generation unit 102 generates a structured FT diagram of sentences with subjects (not omitted). To achieve this, it predicts the subject in the sentence with the omitted subject and uses the predicted subject as the subject of that sentence. The FT diagram generation unit 102 predicts the subject using, for example, AI (a trained model). The trained model that predicts the subject is a model that has learned the correspondence between documents such as manuals that contain training sentences in which the subject is omitted, and the subjects that are omitted in those training sentences. By learning such correspondences, the trained model that predicts the subject can, for example, predict that the subject of a sentence with an omitted subject on a page describing engines is "engine." Furthermore, by learning such correspondences, the trained model that predicts the subject can predict that the subject of the sentence "replace" following a sentence such as "the engine oil is deteriorating" is "engine oil."

[0041] The FT diagram generation unit 102 determines whether or not an error code is displayed in the top event (event) of the initial FT diagram (step S54). In the manual, some sentences corresponding to events are assigned codes (error codes). In this embodiment, the group of sentences in which events are managed by error codes are excluded from the processing shown in steps S55-S58 that are performed later. For this reason, the FT diagram generation unit 102 determines whether or not an error code is assigned to an event in the initial FT diagram and classifies the group of sentences in the initial FT diagram into sentences linked by error codes and sentences not linked by error codes. The FT diagram generation unit 102 then targets the sentences not linked by error codes for the processing shown in steps S55-S58.

[0042] The FT diagram generation unit 102 extracts the words corresponding to "part" and "condition" from the text in the initial FT diagram (step S55). Here, "part" refers to the location in the equipment subject to maintenance where events such as failures or malfunctions may occur. "Condition" refers to the words indicating the state of the part. The specific method by which the FT diagram generation unit 102 extracts the words corresponding to "part" and "condition" from the text will be explained in detail later.

[0043] The FT diagram generation unit 102 generates sentences corresponding to the "parts" and "states" extracted from the sentences in the initial FT diagram using two methods: matching prediction as shown in step S56 and AI prediction as shown in step S57.

[0044] In step S56, the FT diagram generation unit 102 generates text (matching prediction) based on the similarity between the combinations of "parts" and "states" extracted from the text in the reviewed FT diagram and the combinations of "parts" and "states" extracted from the text in the initial FT diagram. The specific method by which the FT diagram generation unit 102 generates text through matching prediction will be explained in detail later.

[0045] In step S57, the FT diagram generation unit 102 generates a text with consistent notation by linking the "parts" and "states" extracted from the text in the initial FT diagram with specific words (AI prediction). The specific method by which the FT diagram generation unit 102 generates text using AI prediction will be explained in detail later.

[0046] Then, the FT diagram generation unit 102 selects sentences generated by either the matching prediction method shown in step S56 or the AI ​​prediction method shown in step S57, and generates an FT diagram by replacing the sentences in the initial FT diagram with the selected sentences (step S58).

[0047] (Process to extract body parts and conditions from text) Figure 5 is a flowchart showing the process shown in step S55 of Figure 4, that is, the process of extracting parts and states from text. The FT diagram generation unit 102 obtains the target text from the initial FT diagram (step S550). The target text is the text from which the parts and states will be extracted. The FT diagram generation unit 102 predicts the location (or location phrase) in the acquired target text (step S551). For example, the FT diagram generation unit 102 uses AI (trained model) to predict the location (or location phrase) in the target text. The trained model here is a model that has learned the correspondence between training texts and the locations in those training texts.

[0048] Figure 6 shows an example of training data used to train a pre-trained model. The left side of Figure 6 shows sentences (training sentences), and the right side shows parts of the sentence. For example, the sentence "Damage to the cylinder packing" is associated with the part "cylinder packing". By training with training data that has been associated in this way, the pre-trained model learns the correspondence between sentences and parts of the sentence, and becomes able to predict parts of the sentence.

[0049] Returning to Figure 5, the FT diagram generation unit 102 determines whether to combine multiple predicted parts from the text into a single part phrase or not. For example, the FT diagram generation unit 102 extracts a word between two parts in the text, and if the extracted word is not a word indicating a state, it combines the two parts into a single part phrase.

[0050] The FT diagram generation unit 102 determines whether the words between the two parts are words that indicate a state, based, for example, on whether they belong to a specific part of speech or a specific word (stop word). The specific part of speech here refers to parts of speech that can indicate a state, such as nouns, adjectives, adverbs, ideograms, verbs, and pronouns. The specific word (stop word) here refers to words that may belong to a part of speech that can indicate a state, but do not become words that indicate a state, such as words like "become" and "be," or punctuation marks and symbols such as commas. In this case, the FT diagram generation unit 102 determines that if the word between the two parts is a part of speech other than a specific part of speech, such as a conjunction like "and" or "or", that word is not a word indicating a state, and predicts that the two parts be combined into a single part phrase. Alternatively, if the word between the two parts is a comma, the FT diagram generation unit 102 determines that the word is not a word indicating a state and predicts that the two parts are combined into a single part phrase. Furthermore, if the FT diagram generation unit 102 has prefixes or suffixes in the parts of the text it predicts, it includes those prefixes or suffixes in those parts.

[0051] The FT diagram generation unit 102 extracts the state portion from the text (step S552). The FT diagram generation unit 102 extracts the text from the beginning to the end of the text, section by section (or section phrase), as the state portion, from that section to the next section or the end of the text.

[0052] The FT diagram generation unit 102 performs morphological analysis on the state portion extracted in step S552 and uses specific parts of speech as state keywords (step S553). The specific parts of speech here are parts of speech that can indicate a state, such as nouns, adjectives, adverbs, ideograms, verbs, and pronouns. For example, the FT diagram generation unit 102 uses parts of speech that correspond to specific parts of speech, such as nouns, from among the morphemes that constitute the state portion as state keywords.

[0053] The FT diagram generation unit 102 removes specific words (stop words) from the state keywords obtained in step S553 (step S554). These specific words (stop words) are parts of speech that can indicate a state, but do not indicate a state themselves, such as words like "is," "become," and "are." The FT diagram generation unit 102 uses the state keywords with the stop words removed as the words indicating the "state" extracted from the text.

[0054] Figure 7 illustrates the process shown in Figure 5, that is, the process of extracting parts and states from text. Here, we will explain using the example where the target text is "The temperature inside the cooling sensor and the tank is high."

[0055] Code SA551 shows a specific example of the process shown in step S551 of Figure 5, that is, the process of predicting parts (or part phrases) from a text. As shown in this example, suppose the trained model predicts two parts from the target text: part 1 "cooling sensor" and part 2 "tank". In this case, the FT diagram generation unit 102 extracts the phrase "and," which is between part 1 and part 2 in the target text. The FT diagram generation unit 102 determines whether the extracted phrase corresponds to a specific part of speech such as a noun, and whether it corresponds to a specific word (stop word). Of the phrase "and," the conjunction "and" does not correspond to a specific part of speech. Also, of the phrase "and," the comma "," corresponds to a specific word (stop word). In this case, the FT diagram generation unit 102 determines that the phrase "and," which is between part 1 and part 2, is not a phrase that can indicate a state, and predicts a single part phrase "cooling sensor and, tank" which combines part 1 and part 2. The FT diagram generation unit 102 determines whether the predicted part phrase has a prefix or suffix. In this case, the FT diagram generation unit 102 determines that the part "tank" predicted from the text has the suffix "inside", and therefore includes the suffix, making the phrase "inside the tank" a single part phrase. As a result, the FT diagram generation unit 102 predicts that the part phrase in the target text is "cooling sensor and, inside the tank".

[0056] Code SA552 shows a specific example of the process shown in step S552 of Figure 5, that is, the process of extracting the state portion from the text. The FT diagram generation unit 102 extracts the wording from the predicted part phrase "cooling sensor and inside the tank" from the target text, up to the next part (or part phrase) or the end, in this example, the wording "the temperature is high", as the state portion.

[0057] Code SA553 shows a specific example of the process shown in step S553 of Figure 5, that is, the process of performing morphological analysis on the state portion and extracting state keywords. The FT diagram generation unit 102 divides the state portion "the temperature is getting high" into morphemes "of, temperature, is, high, become, be". The FT diagram generation unit 102 selects a specific part of speech, such as a noun, from the morphemes that make up the state portion and uses it as the state keyword "temperature, high, become, be".

[0058] Code SA554 shows a specific example of the process shown in step S554 of Figure 5, that is, the process of extracting words that indicate a state from a state keyword. The FT diagram generation unit 102 removes the specific words (stop words) "become" and "become" from the state keyword "temperature, high, become, be," and uses the remaining words "temperature, high" as words that indicate a state. As a result, the FT diagram generation unit 102 extracts the phrase "cooling sensor and inside the tank" from the target text as a phrase indicating a part. It also extracts the phrase "temperature, high" as a phrase indicating a state.

[0059] (Regarding the process of generating text through matching prediction) Figure 8 is a flowchart showing the process shown in step S56 of Figure 4, that is, the process of generating text through matching prediction. The FT diagram generation unit 102 acquires target data consisting of "parts" and "states" extracted in step S55 (step S560). The FT diagram generation unit 102 acquires comparison data consisting of "parts" and "states" from the text in the reviewed FT diagram (step S561). The FT diagram generation unit 102 calculates the similarity between the target data and the comparison data (step S562). The FT chart generation unit 102 calculates, for example, the degree to which the words indicating "parts" in the target data and the comparison data are similar to each other by using a database such as a thesaurus. The FT chart generation unit 102 also calculates the degree to which the words indicating "states" in the target data and the comparison data are similar to each other by using a database such as a thesaurus. The FT chart generation unit 102 uses the sum of the degree to which the words indicating "parts" are similar to each other and the degree to which the words indicating "states" are similar to each other as the similarity score between the target data and the comparison data. The FT diagram generation unit 102 calculates the similarity between the target data and the comparison data for each sentence in the reviewed FT diagram, and identifies the sentence corresponding to the comparison data with the highest calculated similarity as the sentence corresponding to the target data.

[0060] Figure 9 is a diagram illustrating the process shown in Figure 8. On the left side of Figure 9, the words representing the "part" and "state" extracted from the target data, i.e., the text in the initial FT diagram, are shown. In this example, the target text is "The cylinder is greatly deformed." The word representing the part extracted from the target text is "cylinder." The words representing the state are "large, deformed." On the right side of Figure 9, comparative data is shown, which consists of phrases indicating the "part" and "condition" extracted from the text in the reviewed FT diagram. In this example, four phrases are given as examples: "Severe cylinder wear," "Cylinder deformation," "Large pin play," and "Mast and carriage deformation." In the center of Figure 9, the similarity of the comparison data for each of the four sentences in the reviewed FT diagram to the target data is shown. The sentence with the greatest similarity is the second one shown, "deformation of the cylinder." In this case, the FT diagram generation unit 102 uses the sentence "deformation of the cylinder" as the sentence corresponding to the target data in the target sentence, "the cylinder is greatly deformed," which was generated by matching prediction.

[0061] (Regarding the process of generating text using AI prediction) Figure 10 is a flowchart showing the process shown in step S57 of Figure 4, that is, the process of generating text through AI prediction. The FT diagram generation unit 102 acquires target data consisting of "parts" and "states" extracted in step S55 (step S570). The FT diagram generation unit 102 uses a database such as a thesaurus to standardize the notation of the words indicating "parts" and "states" respectively (step S571). The FT diagram generation unit 102 generates sentences corresponding to the target data by AI prediction by linking the standardized words with specific words such as "of", "from", etc.

[0062] Figure 11 is a diagram illustrating the process shown in Figure 10. Here, we will explain using the example where the target sentence is "There is an unusual noise coming from the engine."

[0063] Code SA570 shows a specific example of the process shown in step S570 of Figure 10, that is, the process of acquiring target data. The FT diagram generation unit 102 acquires the words "engine" (part) and "abnormal, sound" (state) as target data corresponding to the target text. Code SA571 shows a specific example of the process shown in step S571 of Figure 10, that is, the process of unifying the notation. The FT diagram generation unit 102 unifies the notation by, for example, converting each of the words "abnormal" and "sound" to the word "unusual sound". Code SA572 shows a specific example of the process shown in step S572 of Figure 10, that is, the process of generating a sentence by concatenating the words for part and state. The FT diagram generation unit 102, for example, concatenates the words for part "engine" and state "abnormal, sound" with the specific word "from" to generate the sentence "abnormal sound from the engine," which is then used as the sentence generated by AI prediction.

[0064] (Regarding the process of selecting either matching prediction or AI-predicted text) Figure 12 is a flowchart showing the process of selecting either matching prediction or AI-predicted text, as shown in step S58 of Figure 4. The FT diagram generation unit 102 determines whether both sentences generated by matching prediction and sentences generated by AI prediction exist (step S580). If neither exists, the FT diagram generation unit 102 selects the existing sentence to be included in the FT diagram (step S581). In step S580, if both exist, the FT diagram generation unit 102 determines whether the respective parts of both sentences match (step S582). If the parts match, the FT diagram generation unit 102 selects the sentence generated by matching prediction as the sentence to include in the FT diagram (step S583). In step S582, if the parts do not match, the FT diagram generation unit 102 selects the text generated by AI prediction as the text to be included in the FT diagram (step S584).

[0065] Figure 13 is a diagram illustrating the process shown in Figure 12. Code SA581 shows a specific example of the process shown in step S581 of Figure 12, that is, the process when neither a sentence generated by matching prediction nor a sentence generated by AI prediction exists. For example, if the target sentence "Inspect the cylinder" is generated by matching prediction but not by AI prediction, the FT diagram generation unit 102 selects the sentence created by matching prediction, "Inspect and repair the cylinder". If the target sentence "Inspect or repair the filter" is generated by matching prediction but not by AI prediction, the FT diagram generation unit 102 selects the sentence created by matching prediction, "Inspect and repair the cylinder". If the target sentence "Replace the wheel" is not generated by matching prediction but is generated by AI prediction, the FT diagram generation unit 102 selects the sentence created by AI prediction, "Replace the wheel".

[0066] Code SA583 shows a specific example of the process shown in step S583 of Figure 12, that is, the process of selecting a sentence generated by matching prediction. The FT diagram generation unit 102, when given the target sentence "Inspect the engine", generates the sentence "Inspect and repair the engine" by matching prediction and generates the sentence "Inspect the engine" by AI prediction, and both the target sentence and the matching prediction match with the word "engine", selects the sentence "Inspect and repair the engine" generated by matching prediction.

[0067] Code SA584 shows a specific example of the process shown in step S584 of Figure 12, that is, the process of selecting a sentence generated by AI prediction. The FT diagram generation unit 102, when given the target sentence "replace the filter", generates the sentence "replace the cylinder" by matching prediction and the sentence "replace the filter" by AI prediction, and the parts "filter" and "cylinder" do not match between the target sentence and the matching prediction, selects the sentence "replace the filter" generated by AI prediction.

[0068] (Regarding content review by experts, etc.) Figure 14 illustrates the process shown in step S7 of Figure 2, that is, the process in which experts review the content of the text generated by the text generation device 1. As shown on the left side of Figure 14, each of the texts corresponding to events 1 through 30 will be reviewed by experts. The text generation device 1 sets a review level for each text. The review level is the level at which experts will review the content, and is a flag that indicates, for example, whether the text requires review or whether it does not necessarily require review. By setting a review level for each text, experts can identify which texts require priority review, thereby improving the efficiency of the review process. For example, text generation device 1 sets a refinement level depending on whether the text was generated using matching prediction or AI prediction. Text generated by matching prediction is already refined because it is in the refined FT diagram. Therefore, it is considered that there is little need to refine the content again. For this reason, text generation device 1 assigns a refinement level (refinement level 2) to text generated by matching prediction, for example, to indicate that the text does not necessarily need to be refined. On the other hand, text generated by AI prediction is newly generated text, so its content needs to be refined. For this reason, text generation device 1 assigns a refinement level (refinement level 1) to text generated by AI prediction, for example, to indicate that the text must be refined.

[0069] As described above, the text generation device 1 of the embodiment comprises an acquisition unit 100 and an FT diagram generation unit 102. The acquisition unit 100 acquires a target text. The target text is a text that includes words indicating a part (part words) and words indicating the state of a part (state words). The FT diagram generation unit 102 extracts the part words and the state words from the target text. The FT diagram generation unit 102 performs morphological analysis on the state words and selects morphemes indicating the state from the group of morphemes that constitute the state words. The FT diagram generation unit 102 generates a text using AI prediction (first predicted text) as a text corresponding to the target text. The text using AI prediction is a text generated by concatenating part words and morphemes indicating the state. As a result, the text generation device 1 of the embodiment can generate text by linking the parts and states extracted from the target text. For example, multiple texts with inconsistent notation, such as "inspect the cylinder" and "inspect the cylinder," can be unified by linking the part "cylinder" and the state "inspect," for example, to generate the text "inspect the cylinder." In other words, the notation of texts containing parts and states can be unified.

[0070] Furthermore, in the document generation device 1 of the embodiment, the FT diagram generation unit 102 generates a document (second predicted document) using matching prediction. The FT diagram generation unit 102 generates a document using matching prediction based on the similarity (degree of similarity) between the combination of part terms and morphemes indicating states and the combination of part terms and state terms extracted from the reviewed FT diagram (existing document). The FT diagram generation unit 102 selects one of the documents based on the correspondence between the document generated using AI prediction, the document generated using matching prediction, and the part terms in the target document. The FT diagram generation unit 102 uses the selected document as the document generated to correspond to the target document. As a result, the text generation device 1 of this embodiment can use either matching prediction or AI prediction to generate text that corresponds to the target text. Therefore, if an inconsistency occurs in the text generated by AI prediction, or if it is not possible to generate text by AI prediction, a similar text from the reviewed FT diagram can be selected as the text that corresponds to the target text.

[0071] Furthermore, in the document generation device 1 of the embodiment, if the part-specific words in the document generated using matching prediction do not match the part-specific words in the target document, the FT diagram generation unit 102 selects the document generated using AI prediction as the document generated corresponding to the target document. On the other hand, if the part-specific words in the document generated using matching prediction match the part-specific words in the target document, the FT diagram generation unit 102 selects the document generated using matching prediction as the document generated corresponding to the target document. As a result, in the document generation device 1 of this embodiment, documents where the parts of the target document and the generated documents match can be considered as documents generated corresponding to the target document. Therefore, documents where the parts of the target document do not match can be excluded as documents that do not correspond to the target document.

[0072] Furthermore, in the document generation device 1 of the embodiment, the FT diagram generation unit 102 assigns a review level (an indicator showing the degree to which the document has been reviewed) to the document generated in response to the target document. The FT diagram generation unit 102 assigns information as the review level that indicates whether the document was generated using AI prediction or matching prediction. As a result, the document generation device 1 of this embodiment can present priorities for the document review process (content scrutiny) by experts, thereby enabling efficient content scrutiny.

[0073] Furthermore, in the document generation device 1 of this embodiment, the FT diagram generation unit 102 generates documents corresponding to the target document (documents using AI prediction and / or documents using matching prediction) as candidates for documents to be used in the FT diagram (Fault Tree Diagram). As a result, in the document generation device 1 of this embodiment, the document generated in response to the target document can be used in the FT diagram. The FT diagram generation unit 102 can generate a document that indicates the parts and states and uses consistent notation as a document corresponding to the target document, so it can generate a document suitable for an FT diagram and use the generated document in the FT diagram. Therefore, an appropriate FT diagram can be generated.

[0074] Furthermore, in the document generation device 1 of this embodiment, the FT diagram generation unit 102 extracts particulate words from the target document based on the parts predicted by the trained model. The trained model is a model that has learned the correspondence between training documents and particulate words in those training documents. As a result, the text generation device 1 of this embodiment can easily extract particulate words from the target text using a trained model.

[0075] Furthermore, in the text generation device 1 of the embodiment, the FT diagram generation unit 102 selects the morphemes that remain after removing a specific part of speech or specific word from the group of morphemes that constitute the state word as morphemes that indicate the state. As a result, the text generation device 1 of the embodiment can extract words that indicate a state at the morpheme level (morphemes that indicate a state) from words that are written in various ways, and it becomes possible to generate text with a unified notation.

[0076] The document generation device 1 in the above-described embodiment may be implemented in whole or in part using a computer. In that case, the program for implementing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be loaded into a computer system and executed. Here, "computer system" includes hardware such as an OS and peripheral devices. Furthermore, "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, and storage devices such as hard disks built into a computer system. Moreover, "computer-readable recording medium" may also include those that dynamically hold programs for a short period of time, such as communication lines used when transmitting programs via networks such as the Internet or communication lines such as telephone lines, and those that hold programs for a certain period of time, such as volatile memory inside a computer system that acts as a server or client in such cases. Furthermore, the above-mentioned program may be for implementing a part of the above-mentioned function, or it may be a program that can implement the above-mentioned function in combination with a program already recorded in the computer system, or it may be implemented using a programmable logic device such as an FPGA (Field Programmable Gate Array).

[0077] Although embodiments of this invention have been described in detail above with reference to the drawings, the specific configuration is not limited to these embodiments and includes designs and the like that do not depart from the spirit of this invention. [Explanation of Symbols]

[0078] 1...Text generation device, 10...Control unit, 100...Acquisition unit, 101...Initial FT diagram generation unit, 102...FT diagram generation unit (generation unit)

Claims

1. An acquisition unit that acquires a target document that includes a part-specific term indicating a part and a state-specific term indicating the state of the part, A generation unit that extracts the part terms and state terms from the target text, performs morphological analysis on the state terms, selects morphemes indicating the state from the group of morphemes constituting the state terms, and generates a first predicted text corresponding to the target text by linking the part terms and the morphemes indicating the state; generates a second predicted text corresponding to the target text based on the degree of similarity between the combination of the part terms and the morphemes indicating the state and the combination of the part terms and state terms extracted from an existing text; and selects either the first or second predicted text based on the correspondence between the first predicted text, the second predicted text, and the part terms in the target text, and makes the selected text the text generated corresponding to the target text. A text generation device equipped with the following features.

2. The generation unit selects the first predicted sentence as the sentence generated in correspondence with the target sentence if the part word in the second predicted sentence does not match the part word in the target sentence, and selects the second predicted sentence as the sentence generated in correspondence with the target sentence if the part word in the second predicted sentence matches the part word in the target sentence. The text generation device according to claim 1.

3. The generation unit assigns to the text information indicating which of the first predicted text and the second predicted text it selected as the text generated in response to the target text, as an indicator of the degree to which the text has been reviewed. A text generation device according to claim 1 or claim 2.

4. The generation unit generates text corresponding to the target text to be used in the Fault Tree Diagram (FT). A text generation device according to any one of claims 1 to 3.

5. The generation unit extracts the part words from the target text based on the parts predicted by the trained model, The aforementioned trained model is a model that has learned the correspondence between training texts and the aforementioned part words within those training texts. A text generation device according to any one of claims 1 to 4.

6. The generation unit selects the morphemes remaining after removing a specific part of speech or specific word from the group of morphemes constituting the state word as morphemes that indicate the state. A text generation device according to any one of claims 1 to 5.

7. A method for generating text performed by a computer-based text generation device, The acquisition unit acquires a target document that includes a part term indicating a part and a state term indicating the state of the part. The generation unit extracts the part terms and state terms from the target text, performs morphological analysis on the state terms, selects morphemes indicating the state from the group of morphemes constituting the state terms, and generates a first predicted text corresponding to the target text by linking the part terms and the morphemes indicating the state. Based on the degree of similarity between the combination of the part terms and the morphemes indicating the state and the combination of the part terms and state terms extracted from the existing text, the generation unit generates a second predicted text corresponding to the target text. Based on the correspondence between the first predicted text, the second predicted text, and the part terms in the target text, the generation unit selects either the first or second predicted text and makes the selected text the text generated corresponding to the target text. Sentence generation method.

8. A computer-based text generation device, The system retrieves a target document that includes a part-specific term and a state-specific term indicating the state of the part. The system extracts the part terms and state terms from the target text, performs morphological analysis on the state terms, selects morphemes representing the state from the group of morphemes constituting the state terms, and generates a first predicted text corresponding to the target text by linking the part terms and the morphemes representing the state. Based on the degree of similarity between the combination of the part terms and the morphemes representing the state and the combination of the part terms and state terms extracted from the existing text, a second predicted text corresponding to the target text is generated. Based on the correspondence between the first predicted text, the second predicted text, and the part terms in the target text, either the first predicted text or the second predicted text is selected, and the selected text is made the text generated corresponding to the target text. program.

Images