Information extraction device, method, and program

Abstract

To support the efficient resolution of troublesome incidents. [Solution] The information extraction device according to the embodiment comprises a trouble event extraction unit, a countermeasure candidate extraction unit, and a data structure generation unit. The trouble event extraction unit extracts trouble events from text data. The countermeasure candidate extraction unit extracts from the text data a first countermeasure candidate for investigating the cause of the trouble event, and a plurality of second countermeasure candidates that differ according to the first effect of the first countermeasure candidate. The data structure generation unit associates the first countermeasure candidate with the trouble event and generates a data structure in which the plurality of second countermeasure candidates are associated with the first countermeasure candidate.

Description

Technical Field

[0001] Embodiments of the present invention relate to an information extraction device, method, and program.

Background Art

[0002] Generally, various trouble events can occur in the manufacture or operation of equipment. In order to quickly resolve a trouble event, it is important to present an effective countermeasure candidate for the trouble event. For this reason, the prior art extracts a trouble event and a plurality of countermeasure candidates from text data regarding various trouble events that have occurred in the past. Next, the prior art generates a data structure in which the plurality of countermeasure candidates are directly associated with the trouble event.

[0003] However, in the data structure generated by the prior art, a large number of countermeasure candidates can be associated with a trouble event. A user cannot determine which of the large number of countermeasure candidates should be preferentially implemented. Therefore, a priority is assigned to each of the large number of countermeasure candidates, and it is assumed that the user implements each countermeasure candidate in order of decreasing priority. In this case, depending on the work cost (e.g., cost, time) for implementing each countermeasure candidate, the user may require a great deal of cost to resolve the trouble event.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] The problem to be solved by the present invention is to support the efficient resolution of trouble events.

Means for Solving the Problems

[0006] The information extraction device according to the embodiment comprises a trouble event extraction unit, a countermeasure candidate extraction unit, and a data structure generation unit. The trouble event extraction unit extracts trouble events from text data. The countermeasure candidate extraction unit extracts from the text data a first countermeasure candidate for investigating the cause of the trouble event, and a plurality of second countermeasure candidates that differ according to the first effect of the first countermeasure candidate. The data structure generation unit associates the first countermeasure candidate with the trouble event and generates a data structure in which the plurality of second countermeasure candidates are associated with the first countermeasure candidate. [Brief explanation of the drawing]

[0007] [Figure 1] Figure 1 is a functional configuration diagram of the information extraction device according to the first embodiment. [Figure 2] Figure 2 is a flowchart of the information extraction device according to the first embodiment. [Figure 3] Figure 3 shows the text data related to the first example. [Figure 4] Figure 4 shows the text data related to the second example. [Figure 5] Figure 5 is a table showing the data structure for one example. [Figure 6] Figure 6 is a graph relating to the first example. [Figure 7] Figure 7 is a graph relating to the second example. [Figure 8] Figure 8 is a graph relating to the third example. [Figure 9] Figure 9 is a functional configuration diagram of the information extraction device according to the second embodiment. [Figure 10] Figure 10 is a flowchart of the information extraction device according to the second embodiment. [Figure 11] Figure 11 is a table showing the data structure for another example. [Figure 12] Figure 12 is a graph relating to another example. [Figure 13] Figure 13 is a hardware configuration diagram of the information extraction device according to each embodiment. [Modes for carrying out the invention]

[0008] The embodiments will be described below with reference to the drawings. Multiple parts assigned the same reference numeral are considered identical, and redundant explanations will be omitted as appropriate. "Trouble event" may mean "the nature of the trouble." "Candidate countermeasure" may mean "a candidate countermeasure for the trouble event." Examples of candidate countermeasures include countermeasures that have actually been implemented and countermeasures that may be implemented in the future. "Effect" may mean "the effect when a candidate countermeasure for the trouble event is implemented." "Effect" may mean "result." "Effective" and "Ineffective" may mean "effective" and "Ineffective," respectively.

[0009] (First embodiment) Figure 1 is a functional configuration diagram of the information extraction device 1A according to the first embodiment. The information extraction device 1A is a device for extracting information. The information extraction device 1A comprises a text data DB 11, a trouble event extraction unit 12, a countermeasure candidate extraction unit 13, an effect determination unit 14A, a data structure generation unit 15, a data structure DB 16, and a graph presentation unit 17.

[0010] Text data DB11 is a database that stores text data. Text data DB11 stores text data for each of multiple documents (e.g., reports, daily reports) that describe a trouble event. Text data DB11 is also called the text data storage unit. Text data DB11 may also be a data file.

[0011] The trouble event extraction unit 12 is a means for extracting trouble events. The trouble event extraction unit 12 extracts trouble events from text data stored in the text data DB 11. For example, the trouble event extraction unit 12 extracts trouble events based on manual input from the user to the text data (e.g., annotations). The trouble event extraction unit 12 may also extract trouble events using a machine learning model for sequential labeling or the like.

[0012] The troubleshooting candidate extraction unit 13 is a means for extracting troubleshooting candidates. Firstly, the troubleshooting candidate extraction unit 13 extracts troubleshooting candidates (hereinafter also referred to as "first troubleshooting candidates") from the text data stored in the text data DB 11 to investigate the cause of the trouble event (i.e., isolate the problem). The troubleshooting candidate extraction unit 13 may extract multiple first troubleshooting candidates. Secondly, the troubleshooting candidate extraction unit 13 extracts multiple different troubleshooting candidates (hereinafter also referred to as "second troubleshooting candidates") from the text data stored in the text data DB 11, depending on the effect of the first troubleshooting candidate (hereinafter also referred to as "first effect"). For example, the troubleshooting candidate extraction unit 13 extracts the first troubleshooting candidates and second troubleshooting candidates based on manual input from the user to the text data (e.g., annotation). The troubleshooting candidate extraction unit 13 may also extract the first troubleshooting candidates and second troubleshooting candidates using a machine learning model for sequential labeling or the like.

[0013] The effect determination unit 14A is a means for determining the effect. Firstly, the effect determination unit 14A determines the first effect of the first candidate for countermeasures from the text data stored in the text data DB 11. Secondly, the effect determination unit 14A determines the effect of each of the multiple second candidate for countermeasures (hereinafter also referred to as the "second effect") from the text data stored in the text data DB 11. For example, the effect determination unit 14A determines the first and second effects based on manual input from the user to the text data (e.g., annotation). The effect determination unit 14A may also determine the first and second effects using a machine learning model for sequence labeling, binary classification, etc.

[0014] The data structure generation unit 15 is a means for generating a data structure. The data structure generation unit 15 uses the trouble event extracted by the trouble event extraction unit 12, the countermeasure candidate extracted by the countermeasure candidate extraction unit 13, and the effect discriminated by the effect discrimination unit 14A. The data structure generation unit 15 generates a data structure in which a first countermeasure candidate is associated with a trouble event and a plurality of second countermeasure candidates are associated with the first countermeasure candidate. The data structure generation unit 15 may associate a first effect with the first countermeasure candidate, or may associate a second effect with each of the plurality of second countermeasure candidates. The data structure generation unit 15 stores the generated data structure in the data structure DB 16.

[0015] The data structure DB 16 is a database for storing a data structure. The data structure DB 16 stores the data structure generated by the data structure generation unit 15. The data structure DB 16 may store a data structure for each trouble event. The data structure DB 16 is also referred to as a storage unit for the data structure. The data structure DB 16 may be a data file.

[0016] The graph presentation unit 17 is a means for presenting a graph or the like. The graph presentation unit 17 presents a graph representing the data structure stored in the data structure DB 16. The graph presentation unit 17 may select, from the data structure DB 16, a trouble event identical to or similar to the trouble event specified by the user. For example, the graph presentation unit 17 may search for a trouble event in consideration of synonyms and similar words. Alternatively, the graph presentation unit 17 may search for a trouble event using a machine learning model that determines semantic similarity between words or sentences. The graph presentation unit 17 may select the trouble event identified by this search. The graph presentation unit 17 may present a graph representing the data structure for the selected trouble event. The graph presentation unit 17 may cause the graph to be displayed on a display device.

[0017] The graph display unit 17 may highlight first action candidates whose first effect (in particular, the degree of the first effect) is higher than a threshold using a mark or the like. Similarly, the graph display unit 17 may highlight second action candidates whose second effect (in particular, the degree of the second effect) is higher than a threshold using a mark or the like. The threshold may be set to a predetermined value or may be set arbitrarily by the user or the like. The graph display unit 17 may display a graph composed of a set of nodes and a set of edges.

[0018] The solution candidate extraction unit 13 and the data structure generation unit 15 may be implemented using a machine learning model (e.g., a neural network) trained by a machine learning method. For example, training data is created with "trouble events" as input data and "solution candidates" as correct output data. Alternatively, training data is created with a set consisting of "trouble events," "solution candidates," and "effects of the solution candidates" as input data and "solution candidates" as correct output data. Next, the machine learning model may be trained by supervised learning using the training data. The trained machine learning model can output solution candidates corresponding to arbitrary input data from a user or the like. The training data may be created from table 150A (see Figure 5).

[0019] Similarly, training data may be created using "trouble events" as input data and "data structures" as correct output data. Likewise, a machine learning model may be trained using supervised learning with this training data. The trained machine learning model can output a data structure corresponding to any trouble event input by a user or other party. The training data may be created from Table 150A (see Figure 5).

[0020] The troubleshooting candidate extraction unit 13 and the data structure generation unit 15 may be implemented using a large-scale language model (LLM). This large-scale language model may be optimized using known tuning techniques (e.g., prompt tuning, fine tuning, instruction tuning). This large-scale language model may also be optimized using text data stored in the text data DB 11, data structures stored in the data structure DB 16, etc. When an arbitrary trouble event is input by a user or the like, the optimized large-scale language model can output troubleshooting candidates, data structures, etc., corresponding to the trouble event.

[0021] Figure 2 is a flowchart of the information extraction device 1A according to the first embodiment. The information extraction device 1A performs the following steps S1 to S9. If steps S1 to S6 have already been performed, steps S8 and S9 may be performed.

[0022] (Step S1) First, the trouble event extraction unit 12 selects text data. Specifically, the trouble event extraction unit 12 selects text data related to the document to be processed from the text data DB 11. For example, the trouble event extraction unit 12 randomly selects text data related to any document. Alternatively, the trouble event extraction unit 12 may select text data related to a document specified by the user (see Figures 3 and 4).

[0023] (Step S2) Next, the trouble event extraction unit 12 extracts trouble events. Specifically, the trouble event extraction unit 12 extracts trouble events from the text data selected in step S1 (see Figures 3 and 4).

[0024] (Step S3) Next, the countermeasure candidate extraction unit 13 extracts countermeasure candidates. Specifically, the countermeasure candidate extraction unit 13 extracts a first countermeasure candidate and a second countermeasure candidate for the trouble event extracted in step S2 from the text data selected in step S1 (see Figures 3 and 4).

[0025] (Step S4A) Next, the effect determination unit 14A determines the effect. Specifically, the effect determination unit 14A determines the first effect based on the first countermeasure candidate extracted in step S3 from the text data selected in step S1, and determines the second effect based on the second countermeasure candidate extracted in step S3 (see Figures 3 and 4).

[0026] (Step S5) Next, the data structure generation unit 15 generates a data structure. Specifically, the data structure generation unit 15 associates the first countermeasure candidate extracted in step S3 with the trouble event extracted in step S2, and associates the multiple second countermeasure candidates extracted in step S3 with the first countermeasure candidate. The data structure generation unit 15 may associate the first effect determined in step S4A with the first countermeasure candidate, or it may associate the second effect determined in step S4A with the second countermeasure candidate. As a result, the data structure generation unit 15 generates a data structure that includes multiple second countermeasure candidates branching off from the first countermeasure candidate (see Figure 5).

[0027] (Step S6) Next, the data structure generation unit 15 saves the data structure. Specifically, the data structure generation unit 15 saves the data structure generated in step S5 to the data structure DB 16 (see Figure 5).

[0028] (Step S7) Next, the data structure generation unit 15 determines whether to complete the series of processes from Step S1 to Step S6. For example, the data structure generation unit 15 completes the series of processes when the number of times the series of processes has been executed reaches a predetermined value. Alternatively, the data structure generation unit 15 may complete the series of processes in response to instructions from the user. If the series of processes is completed (Step S7-YES), the process proceeds to Step S8. If the series of processes is not completed (Step S7-NO), the process returns to Step S1.

[0029] (Step S8) Next, the graph display unit 17 selects a trouble event. Specifically, the graph display unit 17 selects a trouble event to be processed from the data structure DB 16. For example, the graph display unit 17 selects a trouble event that is the same as or similar to the trouble event specified by the user. The graph display unit 17 may display a screen on the display device for the user to specify a trouble event. The graph display unit 17 may also select a trouble event in response to user input on the screen.

[0030] (Step S9) Finally, the graph display unit 17 displays a graph. Specifically, the graph display unit 17 displays a graph representing the data structure for the trouble event selected in step S8 (see Figures 6, 7, and 8).

[0031] Figure 3 shows text data 111A relating to the first example. Text data 111A includes a description of a trouble event TE that occurred in a certain device. The information extraction device 1A extracts (or identifies) various data, taking into account the context of the description.

[0032] For example, the trouble event extraction unit 12 extracts "current drops" as the trouble event TE from the text data 111A. The countermeasure candidate extraction unit 13 extracts "power-on test of A" as the first countermeasure candidate AC1 performed to investigate the cause of the trouble event TE from the text data 111A. The effect determination unit 14A determines that the first effect EF11 of the first countermeasure candidate AC1 is "no problem" because it is written as "no problem" as the effect of the first countermeasure candidate AC1. The effect determination unit 14A may also determine the first effect EF11 as "power on" or "YES".

[0033] Furthermore, the countermeasure candidate extraction unit 13 extracts "The looseness of the B cable was corrected" as the second countermeasure candidate AC21 implemented in response to the first effect EF11. The effect determination unit 14A determines that the second effect EF21 due to the second countermeasure candidate AC21 is a "recurrence report" because it is recorded as a "recurrence report" on the day following the day (3 / 9) when the second countermeasure candidate AC21 was implemented (3 / 9). The effect determination unit 14A may also determine that the second effect EF21 is "no effect".

[0034] Similarly, the countermeasure candidate extraction unit 13 extracts "C was replaced" as the second countermeasure candidate AC22 implemented in response to the first effect EF11. The effect determination unit 14A determines that the second effect EF22 due to the second countermeasure candidate AC22 is "no recurrence" because it states "no recurrence" two days after the day (3 / 11) when the second countermeasure candidate AC22 was implemented (3 / 13). The effect determination unit 14A may also determine that the second effect EF22 is "effective".

[0035] The effect determination unit 14A may associate the effect EF with a combination of multiple countermeasure candidates AC (for example, multiple second countermeasure candidates). For example, consider a case where the text data states, "The problem was resolved after fixing the looseness of cable B and replacing C." In this case, it is presumed that the problem was resolved by performing both countermeasure candidates (fixing the looseness of cable B and replacing C). Therefore, the effect determination unit 14A may determine that the effect of the combination of countermeasure candidate "fixing the looseness of cable B" and countermeasure candidate "replacing C" is "effective" and associate "effective" with that combination.

[0036] Furthermore, the effect determination unit 14A may associate the effect EF with the order of multiple countermeasure candidates AC (for example, multiple second countermeasure candidates). For example, consider a case where the text data states, "After restarting C, B was replaced, but the problem was not solved. However, after replacing B, C was restarted, and the problem was solved." In this case, it is presumed that the problem was solved when the two countermeasure candidates (replace B and restart C) were implemented in the order of "replace B" followed by "restart C". Therefore, the effect determination unit 14A may determine that the effect of the order of countermeasure candidate "replace B" followed by countermeasure candidate "restart C" is "effective," and associate "effective" with that order.

[0037] Figure 4 shows text data 111B relating to the second example. Text data 111B covers equipment that is the same as or similar to the equipment in text data 111A. Text data 111B is the same as text data 111A.

[0038] For example, the trouble event extraction unit 12 extracts "current drops" as the trouble event TE from the text data 111B. The countermeasure candidate extraction unit 13 extracts "power-on test of A" as the first countermeasure candidate AC1 performed to investigate the cause of trouble event TE from the text data 111B. The effect determination unit 14A determines that the first effect EF12 of the first countermeasure candidate AC1 is "no power," since the effect of the first countermeasure candidate AC1 is described as "no power." The effect determination unit 14A may also determine the first effect EF12 as "problematic" or "NO."

[0039] Furthermore, the countermeasure candidate extraction unit 13 extracts "The looseness of the D cable was corrected" as the second countermeasure candidate AC23 implemented in response to the first countermeasure candidate EF12. The effect determination unit 14A determines that the second effect EF23 resulting from the first countermeasure candidate AC23 is a "recurrence report" because it is recorded as a "recurrence report" on the day following the day (9 / 3) when the second countermeasure candidate AC23 was implemented (9 / 2). The effect determination unit 14A may also determine that the second effect EF23 has "no effect".

[0040] Similarly, the countermeasure candidate extraction unit 13 extracts "E was replaced" as the second countermeasure candidate AC24 implemented in response to the first effect EF12. The effect determination unit 14A determines that the second effect EF24 due to the second countermeasure candidate AC24 is "no recurrence" because it states "no recurrence" on the day following the day (9 / 4) when the second countermeasure candidate AC24 was implemented (9 / 5). The effect determination unit 14A may also determine that the second effect EF24 is "effective".

[0041] Figure 5 shows Table 150A, which illustrates a data structure for one example. Table 150A holds various data extracted from text data 111A and 111B. Table 150A has columns for "id", "document id", "trouble event", "candidate solution", "effect", and "edge".

[0042] For example, records from the first to the third line (id1 to id3) hold various data extracted from document id1 (i.e., text data 111A). In particular, record id1 in the first line holds (2,3) as an edge. This edge indicates the id of the next associated record. In this example, record id1 is associated with records id2 and id3. Consequently, the action candidate "Test the power supply of A" in record id1 is associated with the action candidate "Tighten the looseness of cable B" in record id2, and with the action candidate "Replace C" in record id3.

[0043] Similarly, records from the 4th to the 6th row (id4 to id6) hold various data extracted from document id2 (i.e., text data 111B). In particular, record id4 in the 4th row holds (5,6) as an edge. This edge indicates the id of the next associated record. In this example, record id4 is associated with records id5 and id6. Consequently, the action candidate "Test the power supply of A" in record id4 is associated with the action candidate "Tighten the looseness of cable D" in record id5, and with the action candidate "Replace E" in record id6.

[0044] The effect may be expressed qualitatively as "yes" or "no," or quantitatively as a probability value in the range of 0.0 to 1.0. The probability value may be assigned manually by the user or automatically by a machine learning model. The probability value may also be the confidence level of the extraction process performed by the machine learning model. In Table 150A, the probability value is shown as a numerical value inside parentheses.

[0045] Figure 6 is graph 170A relating to the first example. Graph 170A schematically shows a series of potential solutions implemented to resolve a trouble event and the effect of each solution. In graph 170A, each trouble event and solution candidate is represented as a node ND, and the relationship between the trouble event and the solution candidate is represented as an edge ED. Graph 170A represents the data structure corresponding to the records (id1 to id3) from the first to the third row in table 150A.

[0046] For example, node ND1, which relates to the trouble event "current drops," is associated with node ND2, which relates to the proposed solution "power test of A," via edge ED12. Node ND2 is associated with node ND3, which relates to the proposed solution "tighten loose cable B," via edge ED23. Node ND2 is associated with node ND4, which relates to the proposed solution "replace C," via edge ED24.

[0047] In particular, node ND4 is assigned the mark MK, indicating that the suggested solution "replace C" is effective in resolving the trouble event (i.e., it is effective). Probability values ​​indicating the degree of effectiveness of the suggested solution may be assigned to the vicinity of mark MK. This probability value may be "0.8" based on the effect of record id3 in the third row of table 150A. Similarly, similar probability values ​​may be assigned to the vicinity of nodes ND2 and ND3 related to other suggested solutions.

[0048] As can be seen from Graph 170A, some entity (especially a person) has confirmed that the results of the proposed solution "Power supply test of A" to investigate the cause of the trouble event "Current drops" are not problematic, and has then implemented the proposed solutions "Tighten loose cable B" and "Replace C". In particular, the proposed solution "Replace C" is marked with a "effective" mark MK. Therefore, the user can understand that replacing C is more effective than tightening loose cable B in addressing the trouble event.

[0049] Figure 7 is graph 170B relating to the second example. Graph 170B is similar to graph 170A. Graph 170B represents the data structure corresponding to records (id4 to id6) from row 4 to row 6 in table 150A.

[0050] For example, node ND1, which relates to the trouble event "current drops," is associated with node ND2, which relates to the proposed solution "power test of A," via edge ED12. Node ND2 is associated with node ND5, which relates to the proposed solution "tighten loose cable D," via edge ED25. Node ND2 is associated with node ND6, which relates to the proposed solution "replace E," via edge ED26.

[0051] In particular, node ND6 is assigned the mark MK, indicating that the suggested solution "Replace E" is effective in resolving the trouble event (i.e., it is effective). A probability value indicating the degree of effectiveness of the suggested solution may be assigned to the vicinity of the mark MK. This probability value may be "0.9" based on the effectiveness of record id6 in row 6 of table 150A. Similarly, similar probability values ​​may be assigned to the vicinity of nodes ND2 and ND5 related to other suggested solutions.

[0052] As can be seen from Graph 170B, some entity (especially a person) investigates the cause of the trouble event "current drops" by testing the continuity of A, confirming that there is no continuity, and then implements the following countermeasures: "tighten the looseness of cable D" and "replace E". In particular, the countermeasure "replace E" is marked with a "effective" mark MK. Therefore, the user can understand that replacing E is more effective than tightening the looseness of cable D in addressing this trouble event.

[0053] Figure 8 is graph 170M relating to the third example. Graph 170M is a combined graph of graphs 170A and 170B. Between graphs 170A and 170B, multiple identical or similar nodes ND are merged into a single node ND. Similarly, multiple identical or similar edges ED are merged into a single edge ED.

[0054] As can be seen from Graph 170M, an entity (especially a person) implements several different countermeasures depending on the result of the "power-on test of A" to investigate the cause of the trouble event "current drops". Firstly, if the result is "no problem", the entity implements countermeasures "tighten loose B cable" and "replace C". Secondly, if the result is "no power", the entity implements countermeasures "tighten loose D cable" and "replace E". Therefore, the user can understand that they should implement several different countermeasures for the trouble event depending on the result of the "power-on test of A".

[0055] In addition, in graph 170M, multiple potential solutions for investigating the cause of a trouble event may exist in parallel (or in series). These multiple potential solutions may exist in parallel (or in series) at the same hierarchical level where node ND2 is located. That is, these multiple potential solutions may constitute one hierarchical level or multiple hierarchical levels.

[0056] The graph display unit 17 may display partial graphs, such as graphs 170A and 170B, or an integrated graph, such as graph 170M. The graph display unit 17 may also interactively display graphs 170A, 170B, and 170M. For example, the graph display unit 17 may, after presenting one action candidate with a priority higher than a threshold for a trouble event, receive the result of implementing that action candidate from the user and then present the next action candidate to be implemented. The threshold may be set to a predetermined value or may be arbitrarily set by the user or others.

[0057] For example, the graph display unit 17 may calculate a priority for each of the multiple possible solutions (e.g., a second solution) based on each of the multiple effects (e.g., a second effect). The graph display unit 17 may calculate a higher priority value for each effect that is stronger.

[0058] Next, the graph display unit 17 may present multiple possible solutions (e.g., a second solution) in order of their calculated priority. After presenting the solution with the highest priority, if the graph display unit 17 receives information from the user regarding the results of implementing that solution (e.g., that the trouble issue was not resolved), it may present the solution with the second highest priority (interactive presentation). Alternatively, the graph display unit 17 may present multiple solution options simultaneously as a list.

[0059] According to the information extraction device 1A described above, the trouble event extraction unit 12 extracts trouble events from text data. The countermeasure candidate extraction unit 13 extracts from the text data a first countermeasure candidate for investigating the cause of the trouble event, and a plurality of second countermeasure candidates that differ depending on the first effect of the first countermeasure candidate. The data structure generation unit 15 associates the first countermeasure candidate with the trouble event and generates a data structure that associates the plurality of second countermeasure candidates with the first countermeasure candidate.

[0060] In the data structure generated by the information extraction device 1A, multiple potential solutions are associated with a single trouble event in a stepwise (or hierarchical) manner. In particular, this data structure includes a first potential solution for identifying the cause of the trouble event. Based on the results (or effects) of implementing the first potential solution, the user can narrow down the second potential solution to be implemented next. Furthermore, the user can narrow down the potential solution to be implemented next even if they do not possess the extensive knowledge or skills necessary to identify the cause of the trouble event (i.e., they are not a veteran).

[0061] From another perspective, the information extraction device 1A reduces the number of second countermeasures to be implemented by generating a data structure that includes the first countermeasure. Since the user does not need to implement numerous second countermeasures to resolve the trouble, the work cost required to resolve the trouble can be reduced. Therefore, the information extraction device 1A can support the efficient resolution of troubles.

[0062] (Second embodiment) Figure 9 is a functional configuration diagram of the information extraction device 1B according to the second embodiment. The information extraction device 1B has the same configuration as the information extraction device 1A. The information extraction device 1B includes a cost determination unit 14B instead of the effect determination unit 14A that is included in the information extraction device 1A.

[0063] The cost determination unit 14B is a means for determining costs (particularly labor costs). Firstly, the cost determination unit 14B determines the cost of a first candidate solution (hereinafter also referred to as the "first cost") from the text data stored in the text data DB 11. Secondly, the cost determination unit 14B determines the cost of each of a plurality of second candidate solutions (hereinafter also referred to as the "second cost") from the text data stored in the text data DB 11. For example, the cost determination unit 14B determines the first cost and the second cost based on manual input from the user to the text data (e.g., annotation). The cost determination unit 14B may also determine the first cost and the second cost using machine learning models such as extraction by series labeling or regression. The first cost and the second cost may be calculated from the price of replacement parts, man-hours, etc., or may be manually assigned by the user, etc.

[0064] The cost determination unit 14B may highlight first action candidates whose first cost is lower than the threshold by marking or the like. Similarly, the cost determination unit 14B may highlight second action candidates whose second cost is lower than the threshold by marking or the like. The threshold may be set to a predetermined value or may be set arbitrarily by the user or the like.

[0065] Figure 10 is a flowchart of the information extraction device 1B according to the second embodiment. The information extraction device 1B performs the same operations as the information extraction device 1A. The information extraction device 1B performs step S4B instead of step S4A performed by the information extraction device 1A.

[0066] (Step S4B) Following step S3, the cost determination unit 14B determines the cost. Specifically, the cost determination unit 14B determines the first cost based on the first countermeasure candidate extracted in step S3 from the text data selected in step S1, and determines the second cost based on the second countermeasure candidate extracted in step S3.

[0067] (Step S5) Following step S4B, the data structure generation unit 15 generates a data structure. Specifically, the data structure generation unit 15 associates the first countermeasure candidate extracted in step S3 with the trouble event extracted in step S2, and associates a plurality of second countermeasure candidates extracted in step S3 with the first countermeasure candidate. The data structure generation unit 15 may associate the first cost determined in step S4B with the first countermeasure candidate, or it may associate the second cost determined in step S4B with the second countermeasure candidate. As a result, the data structure generation unit 15 generates a data structure that includes a plurality of second countermeasure candidates branching off from the first countermeasure candidate (see Figure 11).

[0068] (Step S9) Finally, the graph display unit 17 displays a graph. Specifically, the graph display unit 17 displays a graph representing the data structure for the trouble event selected in step S8 (see Figure 12).

[0069] Figure 11 shows Table 150B, which illustrates a data structure for another example. Table 150B holds various types of data similar to Table 150A. Table 150B has a column called "Cost" instead of the column "Effect" that Table 150A has.

[0070] The "Cost" column holds the cost of implementing a proposed solution. For example, record id1 in the first row has a primary cost of "3". Record id2 in the second row has a secondary cost of "10". Record id3 in the third row has a secondary cost of "40". In table 150B, the larger the cost value, the greater the work cost (e.g., expenses, time) required to implement the proposed solution.

[0071] Figure 12 is graph 170N, which relates to another example. Graph 170N is an integrated graph, similar to graph 170M. Graph 170N has a cost in the vicinity of node ND related to each candidate for action. This cost may be derived from the "Cost" column in table 150B.

[0072] Similar to the first embodiment, the graph display unit 17 may interactively display the graph 170N. For example, the graph display unit 17 may, after presenting one action candidate with a priority higher than a threshold for a trouble event, receive the result of implementing that action candidate from the user and then present the next action candidate to be implemented. The threshold may be set to a predetermined value or may be arbitrarily set by the user or the like.

[0073] Similar to the first embodiment, the graph display unit 17 may calculate a priority for each of the multiple possible solutions (e.g., a second solution) based on each of the multiple costs (e.g., a second cost). The graph display unit 17 may calculate a higher priority value for lower costs. The graph display unit 17 may present the multiple possible solutions (e.g., a second solution) in descending order of their calculated priorities. Alternatively, the graph display unit 17 may present the multiple possible solutions as a list simultaneously.

[0074] According to the information extraction device 1B described above, the graph display unit 17 presents a graph 170N that associates the costs of each proposed solution. The user can compare the costs among multiple proposed solutions in the presented graph 170N. As a result, the user can determine which of the multiple proposed solutions should be prioritized. For example, the user can determine that the proposed solution with the lowest cost should be prioritized.

[0075] The first and second embodiments can be combined in any manner. For example, the information extraction device 1 may include both an effect determination unit 14A and a cost determination unit 14B. The information extraction device 1 may present a graph that integrates graph 170M and graph 170N. This graph may include both the mark MK in graph 170M and the cost in graph 170N. The user can compare the "effect" and "cost" among multiple countermeasure candidates in the presented graph. As a result, the user can determine which of the multiple countermeasure candidates should be prioritized. For example, the user can determine that the countermeasure candidate with a greater effect and lower cost should be prioritized.

[0076] Figure 13 is a hardware configuration diagram of the information extraction device 1 according to each embodiment. The information extraction device 1 comprises a CPU 101, RAM 102, ROM 103, storage 104, display device 105, input device 106, and communication device 107 as its components. Each component is connected to each other via an internal bus so as to be able to communicate with each other. The information extraction device 1 may include at least some of each component.

[0077] The CPU 101 is a processor that executes various processes according to a program. The CPU 101 uses a predetermined area of ​​the RAM 102 as a working area. The CPU 101 realizes each part (e.g., trouble event extraction unit 12, countermeasure candidate extraction unit 13, effect determination unit 14A, cost determination unit 14B, data structure generation unit 15, graph presentation unit 17) by reading and executing each program stored in the ROM 103 or storage 104. Each part may be realized by dedicated hardware circuits (e.g., ASIC, PLD, FPGA). Each part may be realized on-premise or in the cloud. The CPU 101 is an example of a processing unit.

[0078] RAM102 is a memory that stores various types of data in a rewritable format. For example, RAM102 is SDRAM (Synchronous Dynamic Random Access Memory). RAM102 is an example of a storage unit (e.g., text data DB11, data structure DB16).

[0079] ROM103 is a memory that stores various types of data in a non-rewritable format. ROM103 is an example of a storage unit (e.g., text data DB11, data structure DB16).

[0080] The storage 104 is various types of storage media. The storage 104 may also be a drive device that writes various types of data to the storage media or reads various types of data from the storage media. The storage 104 may be controlled by the CPU 101. The storage 104 is an example of a storage unit (e.g., text data DB 11, data structure DB 16).

[0081] The display device 105 is a device that displays various types of data. The display device 105 may be an LCD (Liquid Crystal Display). The display device 105 displays various types of data based on display signals from the CPU 101. The display device 105 is an example of a display unit.

[0082] Input device 106 is a device that receives various input operations from the user. Input device 106 may be a mouse or a keyboard. Input device 106 receives the operations entered by the user as instruction signals and transmits the instruction signals to the CPU 101. Input device 106 is an example of an input unit.

[0083] The communication device 107 communicates with external devices via a network in accordance with the control of the CPU 101. The communication device 107 is an example of a communication unit.

[0084] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be carried out in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents. [Explanation of Symbols]

[0085] 1,1A,1B...Information extraction device, 11...Text data DB, 12...Trouble event extraction unit, 13...Solution candidate extraction unit, 14A...Effect determination unit, 14B...Cost determination unit, 15...Data structure generation unit, 16...Data structure DB, 17...Graph presentation unit, 101...CPU, 102...RAM, 103...ROM, 104...Storage, 105...Display device, 106...Input device, 107...Communication device, 111A,111B...Text data, 150A,150B...Table, 170A,170B,170M,170N...Graph, AC...Solution candidate, ED...Edge, EF...Effect, MK...Mark, ND...Node, TE...Trouble event

Claims

1. A trouble event extraction unit that extracts trouble events from text data, A countermeasure candidate extraction unit extracts from the text data a first countermeasure candidate for investigating the cause of the trouble event, and a plurality of second countermeasure candidates that differ according to the first effect of the first countermeasure candidate. A data structure generation unit generates a data structure in which the first countermeasure candidate is associated with the trouble event, and the plurality of second countermeasure candidates are associated with the first countermeasure candidate. An information extraction device equipped with the following.

2. The system further comprises a storage unit that stores the data structure for each of the aforementioned trouble events, The information extraction device according to claim 1.

3. The system further comprises a graph display unit that selects from the storage unit a trouble event identical or similar to a trouble event specified by the user, and displays a graph representing the data structure for the selected trouble event. The information extraction device according to claim 2.

4. The system further comprises an effect determination unit that determines the second effect of each of the plurality of second countermeasure candidates from the text data, The information extraction device according to claim 1.

5. The effect determination unit determines the effect of the combination of the plurality of second countermeasure candidates and associates the determined effect with the combination. The information extraction device according to claim 4.

6. The effect determination unit determines the effect based on the order of the plurality of second countermeasure candidates, and associates the determined effect with respect to the order. The information extraction device according to claim 4.

7. The system further comprises a graph display unit that calculates a priority for each of the plurality of second countermeasure candidates based on each of the second effects, and presents the plurality of second countermeasure candidates in order of the highest calculated priority. The information extraction device according to claim 4.

8. The graph representing the data structure further comprises a graph presentation unit that highlights, among the plurality of second countermeasure candidates, the second countermeasure candidate whose second effect is higher than a threshold. The information extraction device according to claim 4.

9. The system further comprises a cost determination unit that determines the second cost for each of the plurality of second response candidates from the text data, The information extraction device according to claim 1.

10. The system further comprises a graph display unit that calculates a priority for each of the plurality of second countermeasure candidates based on each of the second costs, and presents the plurality of second countermeasure candidates in order of the highest calculated priority. The information extraction device according to claim 9.

11. The graph representing the data structure further comprises a graph presentation unit that highlights, among the plurality of second response candidates, the second response candidate whose second cost is lower than a threshold. The information extraction device according to claim 9.

12. Computers Extract trouble events from text data, From the text data, a first candidate for action to investigate the cause of the trouble event and a plurality of second candidates for action that differ depending on the first effect of the first candidate for action are extracted. A data structure is generated by associating the first candidate solution with the aforementioned trouble event, and associating the plurality of second candidates solution with the first candidate solution. Information extraction method.

13. On the computer, A trouble event extraction function that extracts trouble events from text data, A function for extracting countermeasure candidates from the text data, which extracts a first countermeasure candidate for investigating the cause of the trouble event, and a plurality of second countermeasure candidates that differ according to the first effect of the first countermeasure candidate, A data structure generation function that generates a data structure by associating the first countermeasure candidate with the trouble event and associating the plurality of second countermeasure candidates with the first countermeasure candidate, An information extraction program that achieves this.

Images