JP7077683B2 - Answer candidate creation program, answer candidate creation device and answer candidate creation method - Google Patents

Description

Regarding the answer candidate creation program, the answer candidate creation device, and the answer candidate creation method.

Conventionally, FAQ (Frequently Asked Questions) answer sentences have been generated by clustering incidents (see Patent Document 1 below). Specifically, the FAQ answer sentence was generated by adopting the elements that frequently appear in the incident and the elements that appear in the incident of the operator who has skill and experience.

Japanese Patent Application Laid-Open No. 2003-263443

However, when there are many operators who have no skill or experience, if a frequently used element is adopted, there is a problem that it is not possible to generate an FAQ answer sentence that enables an appropriate answer. Further, when the FAQ answer sentence is generated by adopting the elements appearing in the incident of the operator who has skill and experience, there is a problem that the personal data of the operator is required and extra processing increases.

Therefore, one aspect of the present invention is to generate high-quality FAQ answer candidates.

In one example of the embodiment, it is determined whether or not there is an element described in the incident before the arbitrary time, but described in the incident after the arbitrary time, and if there is the element, the element is described. Is added to generate answer candidates for the question and answer collection, and the computer of the answer candidate creation device is made to execute the process.

It is possible to generate high-quality FAQ answer candidates.

It is a figure which shows the functional structure of the answer candidate making apparatus of 1st Embodiment. It is a figure which shows an example of the incident data in 1st Embodiment. It is a figure which shows the result of extracting a word from an incident in 1st Embodiment. It is a figure which shows the result of having calculated the distance between incidents in 1st Embodiment. It is a figure which shows the result of having divided the incident in 1st Embodiment into a cluster. It is a figure which shows an example of the result of clustering incident data in 1st Embodiment. It is a figure which shows an example of the result of dividing the answer sentence of each answer of a cluster into an element in 1st Embodiment. It is a figure which shows an example of the result of totaling the description and non-description of the element in the 1st embodiment for each operator. It is a figure which shows the result of having aggregated whether the element in 1st Embodiment is described or not described in the incident for each correspondence day. It is a figure which shows the result of having separated the date and time of the incident which is described and the date and time of the incident which is not described for each element in the 1st Embodiment, and totaling the number of each incident. It is a figure to be used for the explanation of the determination method by the determination unit in 1st Embodiment. It is a figure which shows an example of the result of totaling the description and non-description of the element in the 1st embodiment for each operator. It is a figure which shows the result of totaling the ratio (description ratio) of the number of description incidents of the operator in 1st Embodiment for a predetermined period. It is a figure which compares the description ratio in the 1st Embodiment in the period before and after, and shows the result calculated by dividing the description ratio of the period (the later period) by the description ratio of the previous period. It is a figure which shows an example of the result of totaling the description and non-description of the element in the 1st embodiment for each operator. It is a figure which shows an example of the result of totaling the description and non-description of the element in the 1st embodiment for each operator. It is a figure which shows the hardware configuration of the answer candidate making apparatus in 1st Embodiment. It is a flowchart which shows the extraction processing flow of the FAQ answer candidate of the answer candidate creation apparatus in 1st Embodiment. It is a flowchart which shows the FAQ answer candidate determination processing flow of the answer candidate creation apparatus in 1st Embodiment. It is a flowchart which shows the consistent description determination and the additional processing flow of the answer candidate creation apparatus in 1st Embodiment. It is a flowchart which shows the other FAQ answer candidate determination processing flow of the answer candidate creation apparatus in 1st Embodiment. It is a flowchart which shows the description increase determination and the additional processing flow of the answer candidate creation apparatus in 1st Embodiment. It is a flowchart which shows the other FAQ answer candidate determination processing flow of the answer candidate creation apparatus in 1st Embodiment. It is a flowchart which shows the consistent description / non-description determination and additional processing flow of the answer candidate creation apparatus in 1st Embodiment. It is a flowchart which shows the extraction processing flow of the FAQ answer candidate at the time of combining the adoption method of the plurality of FAQ answer candidates of the answer candidate creation apparatus in 1st Embodiment. It is a figure which shows the functional structure of the answer candidate making apparatus in 2nd Embodiment. It is a figure which shows the method of generating the incident cluster from the FAQ cluster in the 2nd Embodiment. It is a figure which shows an example of the result of totaling the description and non-description of the element in the second embodiment for each operator. It is a figure which shows the other example of the result of summarizing the description and non-description of the element in the second embodiment for each operator. It is a figure which shows the description situation for each operator used for operator selection in 2nd Embodiment. It is a figure which shows the representative operator corresponding to the cluster in 2nd Embodiment. It is a figure which shows the other example of the result of summarizing the description and non-description of the element in the second embodiment for each operator. It is a flowchart which shows an example of the representative operator setting and selection process in 2nd Embodiment. It is a flowchart which shows the extraction process of the FAQ answer candidate at the time of combining the adoption method of the plurality of FAQ answer candidates of the answer candidate creation apparatus in 2nd Embodiment.

<First Embodiment>
Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings.

FIG. 1 shows an example of the configuration of the response candidate creating device of the first embodiment.
The answer candidate creating device (hereinafter, also referred to as a creating device) 1 outputs an FAQ answer candidate for generating an FAQ as a model answer from the incident data. By preparing FAQs based on the FAQ answer candidates that are output, it is possible to answer frequently asked questions efficiently and with high quality without depending on the knowledge range of the individual operator such as a contact center. The incident is a correspondence history exchanged between the questioner and the operator, and is recorded as, for example, a sentence.

Specifically, the creating device 1 reads the input incident data 2 and processes the read incident data 2 to output the FAQ answer candidate 3. The specific processing of the incident data 2 will be described later.

Although the incident data 2 is configured to be input from the outside of the creation device 1 here, the incident data 2 may be stored in the storage area in the creation device 1 and the incident data 2 may be read. .. Further, although the FAQ is generated based on the output FAQ answer candidates, the FAQ may be generated by, for example, an FAQ generator (not shown).

Here, an example of incident data will be described with reference to FIG.
Incident data is composed of incident ID, response date and time, response operator, question, and answer items. The items to be configured are not limited to this, and may include other items.

The incident ID is unique identification information for identifying an incident, such as i000001 and i000002.

The response date and time indicates the date and time when the operator responded to the question from the questioner. For example, an incident with an incident ID of i000001 corresponds to 9:10 on January 4, 2016. Shows.

The corresponding operator indicates an operator who responded to the question from the questioner. For example, an incident with an incident ID of i000001 is indicated to be handled by an operator named AAA. It should be noted that the operator's name is not required, and it is sufficient to distinguish which incident the operator responded to.

The question shows the content of the question from the questioner. For example, in the case of an incident with an incident ID of i000001, I was able to see the Internet, but I could not see it when I started up my PC (Personal Computer) for the first time in a week. Therefore, the question is what to do.

The answer indicates the answer by the operator to the question, for example, it is shown that the answer to the question of the incident whose incident ID is i000001 is to check the Internet settings.

Next, an example of the functional configuration of the creating device of the embodiment will be described with reference to FIG.
The creating device 1 is composed of a clustering unit 11, an element partitioning unit 12, a tabulation unit 13, a determination unit 14, a generation unit 15, and a storage unit 16. The configuration of the creating device 1 is not limited to this, and may include other configurations.

The clustering unit 11 clusters the input incident data 2. The clustering here is performed by a conventional method, such as the K-means method.

Here, incident clustering will be described. First, the words appearing in each incident are extracted by, for example, morphological analysis. The extracted results are shown in FIG. 3A. For example, in one incident (Incident-01), the words "Internet", "Connect", and "Disconnect" are extracted, and in another Incident (Incident-02), the words "Wireless LAN", "Connect", and "Disconnect" are extracted. Has been extracted.

Incident-01 and Incident-02 each have three words that appear, and the words "connect" and "disconnect" are common.

To perform incident clustering, the distance between incidents is calculated by the cosine similarity, and the calculated cosine similarity is used as the similarity between the incidents. The cosine similarity is used when calculating the similarity of sentences, and can be expressed by the cosine similarity cosθ (x, y) = x * y / || x || * || y ||. x * y is the inner product, || x || is the size of x, and || y || is the size of y.

For example, the cosine similarity between Incident-01 and Incident-02 is calculated by the following equation (1).

Cosine similarity (incident-01, incident-02) = number of common words / sqrt (number of incident-01 words) x sqrt (number of incident-02 words)
... (1)
Therefore, the cosine similarity in this case (incident-01, incident-02) = 2 / sqrt (3) × srqt (3) = 2/3 = 0.667. When calculated in the same manner, the cosine similarity between Incident-01 and Incident-03 is 1/3 (= 0.333).

The result of calculating the distance between each incident is shown in FIG. 3B. For example, the distance between Incident-01 and Incident-02 is 0.667, and the distance between Incident-01 and Incident-03 is 0.333.

The calculated cosine similarity is defined as the similarity between incidents. For example, the incident is divided into three clusters as shown in FIG. 3C by the K-means method, and each of them is used as the original data of FAQ. In this example, all incidents are divided into one of three clusters. As a result, Incident-01, Incident-02 and the like belong to the first cluster, Incident-03, Incident-03 and the like belong to the second cluster, and Incident-05 and the like belong to the third cluster. For the above clustering, the above-mentioned Patent Document 1 can be referred to.

FIG. 4 shows an example of the case where the incident data shown in FIG. 2 is clustered by the above-mentioned clustering method. In this clustering, an Internet and a wireless LAN (Local Area Network) system cluster (first cluster), a printer system cluster (second cluster), and the like are used.

The element division unit 12 divides the answer text of each answer of each cluster clustered by the clustering unit 11 into elements. Here, the element means a delimiter for each punctuation mark, or one bullet point when the answer sentence consists of a plurality of bullet points.

An example in which the answer sentence of each answer of each cluster is divided into elements will be described with reference to FIG.

Here, the first cluster will be described as an example. For example, for the answer that the incident ID of the first cluster shown in FIG. 4 is i000001, "Check Internet settings", "On the" Connection "tab of" Internet Properties "in Control Panel", "Dial-up and ..."・ "Is divided into elements such as.

For example, for the answer that the incident ID of the first cluster shown in FIG. 4 is i006002, "Confirm wireless LAN connection", "Click the wireless LAN icon on the taskbar", and "Current connection destination". It is divided into elements such as "・ ・" and "There is a display that is not connected at the" current connection destination "".

In this way, each answer sentence for each cluster is divided into elements.
The aggregation unit 13 aggregates for each operator whether each of the divided elements is described or not described (not described) in the incident by the operator. That is, the aggregation unit 13 aggregates for each operator whether or not each element obtained by dividing the response sentence of the incident is described in the incident, including the number of descriptions and the number of non-statements, and generates aggregated data. The non-stated number means the number of incidents in which the element is not described. The number of incidents at this time and the date and time of response are also totaled. FIG. 6 shows an example of the aggregated results.

Here, the aggregated results for the response elements of the first cluster are shown. It should be noted that aaa, bbb, and the like are described as the elements of the answer, but these indicate the contents of certain elements as well as the elements such as "Check the Internet settings".

For example, looking at the element "Confirm Internet settings", 102 correspondence dates and times are described in the description column of the operator AAA. The number in parentheses indicates the number of incidents. This indicates that the operator AAA describes the element "Check Internet settings" in all incidents (102 incidents). Therefore, there is no description of the corresponding date and time in the non-stated column, and the number of incidents is 0. That is, the sum of the number of incidents in the column described and the number of incidents in the column not described is the total number of incidents handled by the operator AAA. Similarly, for the operator BBB, it is shown that the element "Check the Internet settings" is described in all incidents (86 incidents).

Looking at the element "confirm wireless LAN connection", there is no description of the corresponding date and time in the column of the operator AAA, and the number of incidents is 0. On the other hand, in the non-stated column, 102 correspondence dates and times are described. This indicates that the operator AAA does not describe the element "Confirm wireless LAN connection" in all incidents (102 incidents). Therefore, there is no description of the corresponding date and time in the description column, and the number of incidents is 0.

Further, regarding the element "aaa", there is a description of the corresponding date and time in the description column of the operator AAA, and the number of incidents is 32. On the other hand, there is a description of the corresponding date and time in the non-stated column, and the number of incidents is 70. This indicates that the operator AAA describes the element "aaa" in 32 incidents and not in 70 incidents. Similarly, for the operator BBB, it is shown that the element "aaa" is described in 46 incidents and not in 40 incidents. The same can be considered for the element "bbb".

Here, the aggregation method by the aggregation unit 13 will be described with reference to FIGS. 7A and 7B. Here, the target cluster is set as the third cluster, and the number of incidents described and not described in each element and the response date and time are totaled by taking the element of the answer of the third cluster as an example.

First, as shown in FIG. 7A, whether or not each element is described in the incident is totaled for each corresponding date. For example, if the element "aaa" is described in an incident on all corresponding dates, it is indicated that it is described on all corresponding dates. Here, what is described is indicated by a circle.

Similarly aggregation shows that the element "bbb" is not listed in the incident on all corresponding dates. In addition, it is shown that the element "ccc" has been described in the incident since February 1, 2016. In this way, whether or not each element is described in the incident is totaled for each response date.

When aggregated for each corresponding date, as shown in FIG. 7B, the date and time of the incident described for each element and the date and time of the incident not described are separated, and the number of incidents is aggregated for each. For example, for the element "aaa", the number of incidents described is 10 and the number of incidents not described is 0. The opposite is true for the element "bbb". Based on FIG. 7B, the aggregation result as shown in FIG. 6 is generated.

The determination unit 14 determines whether or not to adopt each element as an FAQ answer candidate based on the aggregation result aggregated by the aggregation unit 13.

The determination unit 14 determines from the aggregation result shown in FIG. 6 whether or not a certain element has been described at a certain time, and if it is described, the generation unit 15 determines whether or not the element is described. The elements are adopted as FAQ answer candidates. Specifically, a certain element is not described in the incident of the initial time (early time), but if it comes to be described from the incident after a certain arbitrary time, that element is described as a FAQ answer candidate. adopt. That is, it is determined whether or not there is an element described in the incident before the arbitrary time (a certain time), but described in the incident after the arbitrary time, and if there is an element, the generation unit 15 Adds elements and generates (adopts) answer candidates for the question and answer collection (FAQ).

This is because even if the operator in the early stage (for example, in the new age) has little knowledge and cannot give a sufficient answer, he / she gains experience every day and the knowledge increases so that he / she can gradually give a sufficient answer. Therefore, the elements that have been described from an arbitrary time (for example, skillful time) with experience are preferable as answer information, and are adopted (output) as FAQ answer candidates.

A specific determination method will be described with reference to FIG. Here, the target cluster will be described as the first cluster.

The element "bbbb" of the first cluster in the operator BBB is described in all incidents (42 incidents) after August 1, 2016, and incidents before July 31, 2016 (44 incidents). Not listed in.

In addition, the element "ddd" of the first cluster in the operator BBB is described in all incidents (28 incidents) after June 1, 2016, and incidents before May 31, 2016 (58 incidents). Not listed in Incident).

In this way, the element "bbb" will be described from incidents after a certain time (for example, August 1, 2016), and before that, that is, before July 31, 2016, the element "bbb" will be described. If there is no description of "bbb", the element "bbb" is adopted as a FAQ answer candidate. The adopted elements are output as FAQ answer candidates.

The determination unit 14 described above determines whether or not a certain element has been described at a certain time, and if it is described, the generation unit 15 determines the element as a FAQ answer candidate. I try to adopt it. The method of adopting FAQ answer candidates other than this method will be described below.

The first is not the case where the description and non-statement are completely separated at a certain time as described above, but the FAQ about the element when the description is less before the certain time and the description is more after the certain time. This is a method to be adopted as an answer candidate.

As shown in FIG. 9, regarding the element “aaa”, the number of incidents described in the incidents by the operator AAA is 32 before January 9, 2016, and the number of incidents not described after January 10, 2016. Is 70. On the other hand, regarding the element "eee", the number of incidents described in the incidents by the operator AAA is 50 after January 26, 2016, and the number of incidents not described before January 25, 2016 is 52. There is.

From the table shown in FIG. 9, the ratio (description ratio) of the number of described incidents of each operator is totaled as shown in FIG. 10 for each predetermined period. The predetermined period here is 14 days, and the respective periods overlap, but the setting of the predetermined period is not limited to this. It should be noted that the values of the number of times and the ratio shown in FIG. 10 are for the purpose of explaining in an easy-to-understand manner, and are different from the numerical values aggregated from the table shown in FIG.

For example, in the element "aaa" in the operator AAA, the number of times of description is 2 times and the number of times of non-statement is 3 times between January 4, 2016 and January 18, 2016, and the description ratio is 2/5. There is. In addition, in the element "eee" in the operator AAA, the number of times of description is 0 and the number of times of non-statement is 5 times between January 4, 2016 and January 18, 2016, and the description ratio is 0/5. There is.

In addition, the number of times "aaa" in the operator AAA was described between January 12, 2016 and January 26, 2016 was 0 times, the number of times not described was 3 times, and the description ratio was 0/3. There is. Further, the number of times the element "eee" is described between January 12, 2016 and January 26, 2016 is once and the number of times not described is two, and the description ratio is 1/3. The same can be considered for other operators such as BBB.

After the description ratios are totaled, the description ratios are compared between the previous and next periods, and the description ratio of the previous period (later period) is divided by the description ratio of the previous period to calculate the ratio. The calculated result is shown in FIG. The ratio is calculated by the formula (2). If the description ratio of the period is 0, the ratio is 0, and if the description ratio of the previous period is 0, the ratio is 1.

For example, the ratio of the element "aaa" in the operator AAA in the period from January 12 to 26, 2016 is 0 because the description ratio in the period is 0/3 = 0. Further, the ratio of the element "eee" in the operator AAA in the period from January 12 to 26, 2016 is 1 because the description ratio in the period before the period is 0/5 = 0.

There is no value for the ratio from January 4th to 18th, 2016 in the first period because there is no description ratio in the previous period. The same applies to other operators.

When the calculated ratio is a predetermined value (for example, 0.8) or more, the corresponding element is adopted as a FAQ answer candidate. That is, the determination unit 14 calculates the description ratio of each element in the incident at regular intervals for each operator, and based on the calculated description ratio, it is not described in the incident before any time, but it is arbitrary. It is determined whether or not there is an element described in the incident after the time of. Here, if the value of the calculated ratio is 1 or more, it is always described from the previous period, but the predetermined value is set to 0.8 in consideration of fluctuations and the like.

Here, it is determined whether or not to adopt the element as an FAQ answer candidate based on a predetermined value, but the description ratio is approximated by a predetermined function (for example, a linear function), and the approximated function increases. In the case of a function, the element may be adopted as a FAQ answer candidate.

The second method is to adopt the elements that all operators always describe in the incident as FAQ answer candidates. That is, the determination unit 14 determines whether or not the elements are described in the incidents of all operators, and the generation unit 15 adds the elements described to generate answer candidates for the question and answer collection.

As shown in FIG. 12, for example, the elements "Check Internet settings" and "Control panel" Internet properties "..." are always described in the incident by all operators. The table shown in FIG. 12 omits the corresponding date and time of the aggregation result shown in FIG.

Therefore, "Check the element Internet settings described by all operators" and "..." of "Internet properties" in the control panel are adopted as FAQ answer candidates.

The third method is to adopt elements that are consistently described or not described as FAQ answer candidates, other than the elements that are always described in the incident by all operators. That is, the determination unit 14 determines for each element whether or not the element is consistently described or not described in the incident, and the generation unit 15 adds an element that is consistently described or not described. And generate answer candidates for the question and answer collection.

As shown in FIG. 13, regarding the element "confirm wireless LAN setting", the operator AAA is not consistently described in the incident, and the operator BBB and the operator CCC are consistently described in the incident. As described above, although it differs depending on the operator, the elements that are consistently described or not described are adopted as FAQ answer candidates.

Next, the hardware configuration of the creating device of the embodiment will be described with reference to FIG. The hardware of the creation device 1 is a CPU (Central Processor Unit) 140, an HDD (Hard Disk Drive) 141, a RAM (Random Access Memory) 142, a ROM (Read Only Memory) 143, a display device 144, an input device 145, and a communication interface. It is composed of 146 and 147 buses. The CPU 140, HDD 141, RAM 142, ROM 143, display device 144, input device 145, and communication interface 146 are connected to each other via, for example, a bus 147.

The CPU 140 reads a program for performing various processes (for example, FAQ answer candidate extraction process) stored in the HDD 141 or the like via the bus 147, temporarily stores the read program in the RAM 142, and then stores the program. Various processes are performed according to the above. The CPU 140 mainly functions as the clustering unit 11, the element division unit 12, the aggregation unit 13, the determination unit 14, and the generation unit 15 described above.

The HDD 141 stores an application program for performing various processes of the creating device 1, data necessary for the processing of the creating device 1, and the like. The HDD 141 mainly functions as a storage unit 16.

The RAM 142 is a volatile memory, and a part of an OS (Operating System) program or an application program for being executed by the CPU 140 is temporarily stored. Further, the RAM 142 stores various data necessary for processing by the CPU 140.

The ROM 143 is a non-volatile memory and stores programs such as a boot program and a BIOS (Basic Input / Output System).

The display device 144 displays information, such as a display.
The input device 145 inputs information, such as a keyboard.

The communication interface 146 transmits / receives data to / from the outside (such as a device having incident data) via a network.

The bus 147 is a path that mediates the transfer of control signals, data signals, and the like between the devices.
Next, the extraction processing flow of the FAQ answer candidate in the creating apparatus of the embodiment will be described with reference to FIG. First, the creation device 1 performs clustering on the read incident (step S1501). The creation device 1 divides the response portion of the clustered incident into elements (step S1502). The creating device 1 determines the FAQ answer candidate for each divided element (FAQ answer candidate determination process: step S1503). The details of the FAQ answer candidate determination process will be described later. As a result of the determination, the creation device 1 outputs if there is an element of the FAQ answer candidate (FAQ answer candidate output: step S1504).

Next, the details of the above-mentioned FAQ answer candidate determination process will be described with reference to FIG. The FAQ answer candidate determination process here determines whether or not a certain element has been described at a certain time, and if it is described, the element is adopted as an FAQ answer candidate. It is a process when doing. The other FAQ answer candidate determination processing will be described later.

The creation device 1 totals the number of incidents described and not described and the corresponding date and time for each divided element for each operator (step S1601). For example, as shown in FIG. 6 described above, they are tabulated and tabulated. Based on the aggregation result, the creation device 1 determines whether or not there is an operator who has consistently described the elements in the incident after a certain period (step S1602).

When there is an operator who has come to be described (Yes in step S1602), the creation device 1 adds (adopts) the element in the operator to the FAQ answer candidate (step S1603). On the other hand, if there is no operator to be described (No in step S1602), the FAQ answer candidate determination process is terminated.

Here, the detailed contents of the determination process of whether or not there is an operator who has consistently described the elements in the incident after a certain period and the FAQ answer candidate addition process (consistent description determination and additional process) are shown in the figure. 17 will be described.

For each of the divided elements, the creation device 1 aggregates the number of cases described and not described in the incident and the corresponding date and time for each operator, and then whether or not there is an unselected element that is not included in the FAQ answer candidates. Is determined (step S1701). When there is an unselected element that is not included in the FAQ answer candidates (Yes in step S1701), the creation device 1 selects one element (step S1702) and selects one arbitrary operator (step S1703).

The creation device 1 determines whether or not the earliest response date and time of the selected operator's description of the element in the incident is after the latest response date and time of not describing the element in the incident (step S1704). If the earliest response date and time described in the incident of the element is after the latest response date and time in which the element is not described in the incident (Yes in step S1704), the creation device 1 is selected as the FAQ answer candidate of the cluster. Add (adopt) the added elements (step S1705).

On the other hand, if the earliest response date and time described in the incident of the element is not after the latest response date and time in which the element is not described in the incident (No in step S1704), whether or not the creation device 1 has an unselected operator. (Step S1706). If there is an unselected operator (Yes in step S1706), the process returns to step S1703, and if there is no unselected operator (No in step S1706), the process returns to step S1701.

If there is no unselected element not included in the FAQ answer candidates (No in step S1701), the consistent description determination and additional processing are completed.

Next, the details of the other FAQ answer candidate determination processing will be described with reference to FIG. As described above, the FAQ answer candidate determination process here is a process in which the elements when the description is small before a certain period and the description is large after a certain period are adopted as the FAQ answer candidate.

The creation device 1 totals the number of incidents described and not described and the corresponding date and time for each divided element for each operator (step S1801). Based on the aggregation result, the creating device 1 determines whether or not there is an operator who has few descriptions in the incident with an early response date and time and many descriptions in the later incident (step S1802).

When there is an operator who has few descriptions in the incident with an early response date and many in the later incident (Yes in step S1802), the creation device 1 adds (adopts) the element in the operator to the FAQ answer candidate (step). S1803). On the other hand, when there are few descriptions in the incident with an early response date and time and there is no operator who has many descriptions in the later incident (No in step S1802), the FAQ answer candidate determination process is terminated.

Here, about the detailed contents of the judgment process of whether or not there is an operator who has a small number of descriptions in the incident with an early response date and the large number of descriptions in the later incident, and the FAQ answer candidate addition process (description increase judgment and additional process). This will be described with reference to FIG.

The creation device 1 aggregates the number of incidents described and not described and the corresponding date and time for each operator for each divided element, and then determines whether or not there is an unselected element that is not included in the FAQ answer candidates. Is determined (step S1901). When there is an unselected element that is not included in the FAQ answer candidates (Yes in step S1901), the creation device 1 selects one element (step S1902) and selects one arbitrary operator (step S1903).

As shown in FIG. 10 described above, the creating apparatus 1 calculates the description ratio of the element to the incident for each fixed period (predetermined period) of the corresponding operator (step S1904). The creating device 1 determines whether or not the description ratio in the incident at regular intervals tends to increase (step S1905). Specifically, as described above, whether or not the calculated ratio (the ratio described in the previous period divided by the ratio described in the period (later period)) is equal to or higher than a predetermined value. To judge.

When the creation device 1 tends to increase (Yes in step S1905), the creating device 1 adds (adopts) the selected element to the FAQ answer candidate of the cluster (step S1906).

On the other hand, when there is no increasing tendency (No in step S1905), the creating device 1 determines whether or not there is an unselected operator (step S1907). If there is an unselected operator (Yes in step S1907), the process returns to step S1903, and if there is no unselected operator (No in step S1907), the process returns to step S1901.

If there is no unselected element not included in the FAQ answer candidates (No in step S1901), the description increase determination and additional processing are completed.

Next, the details of the other FAQ answer candidate determination processing will be described with reference to FIG. As described above, the FAQ answer candidate determination process here is a process in which elements that are consistently described or not described in the incident are adopted as FAQ answer candidates.

The creation device 1 totals the number of incidents described and not described and the corresponding date and time for each divided element for each operator (step S2001). The creation device 1 determines whether or not there is an operator who consistently describes or does not describe the element in the incident based on the aggregation result (step S2002).

When there is an operator who consistently describes or does not describe the element in the incident (Yes in step S2002), the creating device 1 adds (adopts) the element in the operator to the FAQ answer candidate (step S2003). On the other hand, if there is no operator who consistently describes or does not describe the element in the incident (No in step S2002), the FAQ answer candidate determination process ends.

Here, FIG. 21 describes the detailed contents of the determination process of whether or not there is an operator who consistently describes or does not describe the element in the incident and the FAQ answer candidate addition process (consistent description / non-statement determination and additional process). Will be described using.

The creation device 1 aggregates the number of incidents described and not described and the corresponding date and time for each operator for each divided element, and then determines whether or not there is an unselected element that is not included in the FAQ answer candidates. Is determined (step S2101). When there is an unselected element that is not included in the FAQ answer candidates (Yes in step S2101), the creation device 1 selects one element (step S2102) and selects one arbitrary operator (step S2103).

The creation device 1 determines whether the number of incidents in which the element is described in the incident (number of described incidents) is 0 or the number of incidents in which the element is not described in the incident is 0 in the selected operator. Determine (step S2104).

If any of them is 0 (Yes in step S2104), the creating device 1 determines whether or not there is an unselected operator (Step S2105), and if not (No in step S2105), the FAQ answer candidate of the cluster is selected. The selected element is added (adopted) (step S2106). On the other hand, if there is an unselected operator (Yes in step S2105), the process returns to step S2103.

If any of them is not 0 (No in step S2104), the process returns to step S2101.

If there is no unselected element that is not included in the FAQ answer candidates (No in step S2101), the consistent description / non-statement determination and additional processing are completed.

It is also possible to combine the above-mentioned methods for adopting a plurality of FAQ answer candidates. There are multiple combinations, but an example will be described below. For example, as shown in FIG. 22, for each of the divided elements, the number of description and non-statement in the incident and the corresponding date and time are totaled for each operator (step S2201). After that, it is determined whether or not all the operators described above always describe the element in the incident (step S2202). If it is not described (No in step S2202), it is determined whether or not the element is consistently described in the incident described in FIG. 20 (step S2203).

If the element is consistently described or not described in the incident (No in step S2203), there is an operator who has consistently described the element in the incident since a certain time as described in FIG. It is determined whether or not to do so (step S2204). When such an operator does not exist (No in step S2204), it is determined whether or not there is an operator who has a small number of descriptions in the incident with an early response date and time described in FIG. 18 and a large number of descriptions in a later incident (step S2205).

If "Yes" is determined in each of the determination steps from steps S2202 to S2205, the creating device 1 adds the element to the FAQ answer candidate (step S2206).

According to one aspect of the production device of the embodiment, it is possible to generate high quality FAQ answer candidates. In addition, if it is possible to generate high-quality FAQ answer candidates, sufficient information is included in the FAQ, so that the operator does not have to browse other materials and data at the time of response, which leads to efficiency.

<Second embodiment>
Hereinafter, the second embodiment will be described with reference to the drawings. In the second embodiment, the description of the same configuration and processing as in the first embodiment will be omitted.

As shown in FIG. 23, the creating device 1 of the present embodiment includes a clustering unit 11, an element partitioning unit 12, an aggregation unit 13, a determination unit 14, a generation unit 15, a storage unit 16, a setting unit 17, and a selection unit 18. include. The creation device 1 of the present embodiment is different from the creation device 1 of the first embodiment in that the FAQ is input and the processing of the clustering 11 is different. Further, the creating device 1 of the second embodiment is different from the creating device 1 of the first embodiment in that the setting unit 17 and the selection unit 18 are included.

Incident data 2 and past FAQ data 4 are input to the creation device 1. The incident in the incident data 2 and the FAQ in the FAQ data 4 may be clustered in advance, in which case the clustering unit 11 does not have to execute the clustering process described later.

The clustering unit 11 clusters past FAQs. Further, the clustering unit 11 associates an incident with an FAQ, and classifies (clusters) the incident into a plurality of clusters according to the associated FAQ cluster. The details of the clustering process will be described later. Similar to the first embodiment, the clustering unit 11 may cluster incidents without using FAQ by using a method such as the K-means method.

The setting unit 17 corresponds to a cluster including an element that is not described in an incident before an arbitrary time of a predetermined operator but is described in an incident after an arbitrary time. The predetermined operator is set as an operator candidate.

If there is an element that the predetermined operator consistently describes in the incident, the setting unit 17 may set the predetermined operator as an operator candidate corresponding to the cluster including the element.

The selection unit 18 selects an operator (representative operator) corresponding to the cluster from the operator candidates set by the setting unit 17. The selection unit 18 calculates an evaluation value for each operator candidate based on the description status of each cluster of elements included in the incident described by the operator candidate, and among the operator candidates based on the magnitude of the evaluation value. Select the representative operator from. The selection unit 18 evaluates, for example, based on the number of elements described in all incidents described by the operator candidate and the number of elements described in all incidents after a predetermined time described by the operator candidate. Calculate the value. The details of the selection process of the selection unit 18 will be described later.

FIG. 24 is a diagram showing a method of generating an incident cluster from a FAQ cluster.

The clustering unit 11 acquires FAQs divided for each field (cluster). In the example shown in FIG. 24, it is assumed that the FAQ divided into three clusters of a browser, a meter, and integrated software is acquired. Note that there may be FAQs included in a plurality of clusters, as in the "FAQ of Mailer A" shown in FIG. 24. Further, the FAQ in one cluster may be further divided into a plurality of clusters.

The clustering unit 11 refers to each FAQ and each incident, and associates the FAQ and the incident to which the contents are related. The clustering unit 11 may directly associate the incident used when generating the FAQ in the past with the FAQ. Further, the clustering unit 11 may associate the FAQ with the incident based on the similarity between the FAQ and the description of the incident.

The clustering unit 11 clusters, for example, incidents associated with the FAQ of the same cluster into the same cluster. For example, in FIG. 24, the clustering unit 11 clusters the incidents “i00101”, “i00102”, and “i00231” corresponding to the cluster “FAQ-1 of the browser A” in the first cluster.

Note that the clustering unit 11 may cluster incidents without using FAQ by using a method such as the K-means method as in the first embodiment.

The clustering unit 11 of the present embodiment associates an incident with an FAQ, and clusters the incident according to the associated FAQ cluster. Therefore, the clustering unit 11 can cluster the incident into the cluster corresponding to the FAQ.

FIG. 25 is a diagram showing an example of the result of totaling the description and non-statement of the elements in the second embodiment for each operator. Similar to the first embodiment, the aggregation unit 13 aggregates for each operator whether each divided element is described or not described (not described) in the incident by the operator. That is, the aggregation unit 13 aggregates for each operator whether or not each element obtained by dividing the response sentence of the incident is described in the incident, including the number of descriptions and the number of non-statements, and generates aggregated data.

The setting unit 17 refers to the table shown in FIG. 25, and if there is an element that the predetermined operator consistently describes in the incident, the setting unit 17 sets the predetermined operator as the operator candidate corresponding to the cluster including the element. do.

For example, in FIG. 25, the operator AAA consistently describes the element "confirm the Internet settings" and the element "confirm the" Internet properties "". Therefore, the setting unit 17 sets the operator AAA as the operator candidate corresponding to the first cluster.

For example, in FIG. 25, the operator BBB consistently describes the element "check" Internet properties "". Therefore, the setting unit 17 sets the operator BBB as an operator candidate corresponding to the first cluster.

For example, when the number of elements consistently described by a predetermined operator in an incident is equal to or greater than a predetermined threshold value, the setting unit 17 sets a predetermined operator as an operator candidate corresponding to a cluster including the element. You may. For example, in the example shown in FIG. 25, when the predetermined threshold value is “2”, the setting unit 17 sets the operator AAA as the operator candidate corresponding to the first cluster, but does not set the operator BBB.

FIG. 26 is a diagram showing another example of the result of totaling the description and non-statement of the elements in the second embodiment for each operator. In the example shown in FIG. 26, in addition to the number of descriptions and the number of non-statements, the date and time of incident response are totaled.

The setting unit 17 is not described in the incident before the arbitrary time of the predetermined operator, and if the element described in the incident after the arbitrary time of the predetermined operator exists, the setting unit 17 is included in the cluster including the element. A predetermined operator is set as the corresponding operator candidate.

For example, the element "bbbb" is not consistently described in incidents prior to July 31, 2016 by Operator BBB, but consistently in incidents after August 1, 2016. In addition, the element "ddd" is not consistently described in the incidents of the operator BBB before May 31, 2016, and is consistently described in the incidents after June 1, 2016. Therefore, the setting unit 17 sets the operator BBB as an operator candidate corresponding to the first cluster including the elements “bbbb” and “dddd”.

FIG. 27 is a diagram showing a description status for each operator used for operator selection in the second embodiment. It is assumed that the operator included in FIG. 27 is an operator candidate corresponding to the cluster set by the setting unit 17 for each cluster. Although the incident cluster is used in FIG. 27, the FAQ cluster may be used.

As shown in FIG. 27, the aggregation unit 13 aggregates the description status of the incident elements included in each cluster for each operator. The description status is expressed in the form of "all description" or "statement from a predetermined time".

The notation "all described" means that the element is described in all incidents described by the operator. The notation "stated from a given time" means that the element is described in all incidents after the given time stated by the operator.

The selection unit 18 selects a representative operator from the operator candidates for each cluster set by the setting unit 17 using the description situation shown in the example of FIG. 27.

The selection method of the representative operator will be described. In order to select a representative operator, the selection unit 18 calculates, for example, the evaluation value shown in any of the following (a) to (d) for each operator.
(A) Evaluation value = number of elements corresponding to "all description" (b) Evaluation value = number of elements corresponding to "all description" or "statement from a predetermined time" (c) "all description" or "predetermined time" In the evaluation values of the weighted evaluation values (d) and (c) based on the elements corresponding to "described from", an example of calculating the evaluation value of the evaluation value (c) in which "described from a predetermined time" is further weighted according to the time will be described. For example, assume that the weight is set as follows.
Weight for the number of elements (A) corresponding to "all description": 2
Weight for the number of elements (B) corresponding to "stated from a predetermined time": 1
When the weight is set as described above, the selection unit 18 calculates the evaluation value as shown in the following formula.
Evaluation value = 2 x A + 1 x B

For example, when A = 2 and B = 3, the selection unit 18 calculates the evaluation value as shown in the following formula.
Evaluation value = 2 × 2 + 1 × 3 = 7
An example of calculating the evaluation value of (d) will be described. For example, assume that the weight is set as follows.
Weight for the number of elements (A) corresponding to "all description": 3
Weight for the number of elements (B2) that corresponds to "statement from the specified time" and the specified time is after January 2016: 2
Weight for the number of elements (B1) that corresponds to "stated from the specified time" and the specified time is after February 2016: 1

When the weight is set as described above, the selection unit 18 calculates the evaluation value as shown in the following formula.
Evaluation value = 3 x A + 2 x B2 + 1 x B1

For example, when A = 2, B2 = 3, and B1 = 4, the selection unit 18 calculates the evaluation value as shown in the following equation.
Evaluation value = 3 × 2 + 2 × 3 + 1 × 4 = 16

In the calculation of the evaluation values shown in (c) and (d) above, the weight may be a different value for each cluster.

Further, the selection unit 18 calculates any of the evaluation values shown in the above (a) to (d), and then applies any of the following processes (1) and (2) to among the operator candidates. Select the representative operator from.
(1) The selection unit 18 selects the operator having the largest evaluation value. When there are a plurality of operators having the same evaluation value, the selection unit 18 may select the plurality of operators as the representative operator.
(2) The selection unit 18 ranks operators in descending order of evaluation value, and selects a predetermined number of operators having higher ranks.

The thick frame in FIG. 27 shows the selection result using the above methods (a) and (1). For example, with respect to the first cluster, the number of elements corresponding to "all described" is that the operator AAA and the operator DDD are "3", so that the evaluation values of the operator AAA and the operator DDD are the largest. Therefore, the selection unit 18 selects the operator AAA and the operator DDD as the representative operator corresponding to the first cluster.

Further, regarding the second cluster, the number of FAQs corresponding to "all described" is "2" for the operator BBB, and the evaluation value of the operator BBB is the largest. Therefore, the selection unit 18 selects the operator BBB as the representative operator corresponding to the second cluster.

When an operator is selected by using the methods (C) and (1) above, the selection unit 18 calculates the evaluation value of each operator for the first cluster as follows.
Operator AAA evaluation value = 2 × 3 + 1 × 2 = 8
Operator BBB evaluation value = 2 × 2 + 1 × 2 = 6
Operator DDD evaluation value = 2 x 3 + 1 x 1 = 7
Therefore, when the operator is selected by using the methods (C) and (1), the selection unit 18 selects the operator AAA having the largest evaluation value, and does not select the operator BBB and the operator DDD.

FIG. 28 is a diagram showing a representative operator corresponding to the cluster. The correspondence table shown in FIG. 28 is stored in the storage unit 16. In the example shown in FIG. 28, the operator AAA, the operator DDD, and the operator FFF are associated with the first cluster. That is, the representative operators of the first cluster are the operator AAA, the operator DDD, and the operator FFF. Further, the operator BBB and the operator EEE are associated with the second cluster. That is, the representative operators of the second cluster are the operator BBB and the operator EEE.

When the representative operator corresponding to the cluster is selected from the operator candidates, the selection unit 18 records the cluster and the representative operator in association with each other in the correspondence table shown in FIG. 28. Although FIG. 28 shows an example in which a representative operator is set for each incident cluster, a representative operator may be set for each FAQ cluster.

FIG. 29 is a diagram showing another example of the result of totaling the description and non-statement of the elements in the second embodiment for each operator. For example, the aggregation unit 13 outputs the aggregation result shown in FIG. 29 based on the clustered incident data as shown in the example of FIG.

Note that FIG. 29 is an aggregation result showing the description and non-description of the elements of the second cluster, and as shown in FIG. 28, it is assumed that the operator BBB is selected as the representative operator of the second cluster.

The determination unit 14 determines whether the operator (operator BBB) corresponding to the second cluster describes each element included in the second cluster. In the example shown in FIG. 29, the operator BBB describes the element "aaa" and the element "bbbb". Therefore, the generation unit 15 adds the element "aaa" and the element "bbb" to the FAQ answer candidates, and generates the FAQ answer candidates corresponding to the second cluster.

FIG. 29 shows the aggregated results clustered in the cluster of incidents, but the determination unit 14 may generate FAQ answer candidates using the results of clustering in the clusters of FAQs corresponding to the incidents.

For example, an operator with high skill and knowledge in a given field (prescribed cluster) may find it appropriate to answer questions about a new version of the product in a given field. Therefore, the creation device 1 of the present embodiment selects a representative operator for each cluster, and uses the elements described in the incident of the representative operator for the FAQ of the cluster.

FIG. 30 is a flowchart showing an example of representative operator setting and selection processing in the second embodiment.

The clustering unit 11 classifies the incident into a plurality of clusters (step S2901). The creation device 1 starts iterative processing for each cluster (step S2902).

The setting unit 17 determines whether or not there is an unselected element in the target cluster (step S2903). If YES in step S2903, the setting unit 17 selects one incident element in the target cluster (step S2904).

The setting unit 17 determines whether or not there is an operator who consistently describes the selected element (step S2905). If YES in step S2905, the setting unit 17 sets an operator who consistently describes the element as an operator candidate corresponding to the cluster including the element (step S2907).

If NO in step S2905, the setting unit 17 does not consistently describe the selected element in the incident before the arbitrary time, but consistently describes the selected element in the incident after the arbitrary time. It is determined whether an operator exists (step S2906). If YES in step S2906, the setting unit 17 sets the operator as an operator candidate corresponding to the cluster including the element (step S2907).

If NO in step S2906, or after the process of step S2907, the process returns to step S2901.

If NO in step S2903, the selection unit 18 selects a representative operator corresponding to the target cluster from the operator candidates set by the setting unit 17 (step S2908). When the process of steps S2903 to S2908 is executed for all the clusters, the creation device 1 ends the iterative process.

FIG. 31 is a flowchart showing a FAQ answer candidate extraction process when a plurality of FAQ answer candidate adoption methods of the answer candidate creation device according to the second embodiment are combined. In FIG. 31, steps S3001 to S3005 are the same as steps S2201 to S2205 in FIG. 22. Further, step S3007 is the same as step S2206. The description of the same processing as in FIG. 22 will be omitted. The process of FIG. 31 is executed for each element.

The determination unit 14 refers to the table shown in the example of FIG. 28 and the aggregation result shown in the example of FIG. 29, and determines whether or not the representative operator corresponding to the cluster including the target element describes the target element (step S3006). .. When the representative operator is set for each FAQ cluster, the determination unit 14 determines which FAQ cluster the target element is included in, and the representative operator corresponding to the FAQ cluster including the target element determines the target element. If YES in step S3006 for determining whether or not the description is described, the generation unit 15 adds the target element to the FAQ answer candidates and generates answer candidates for the question and answer collection corresponding to the cluster including the target element (step). S3007).

As described above, in the present embodiment, even if all are NO in steps S3002 to S3005, FAQ answer candidates can be added. For example, when generating FAQ answer candidates for a new product or the like, the number of incidents described in the target element is small, and there is a high possibility that all of them will be NO in steps S3002 to S3005. However, even if the number of description of the target element is small, the response of the representative operator of the cluster including the target element is considered to be highly reliable. Therefore, the creation device 1 of the present embodiment can generate appropriate FAQ answer candidates for a new product or the like in a short period of time.

The following additional notes are further disclosed with respect to the above embodiments.
(Appendix 1)
A determination unit that determines whether or not there is an element that is not described in the incident before the arbitrary time but is described in the incident after the arbitrary time.
If there is the element, add the element and generate an answer candidate for the question and answer collection.
An answer candidate creation device characterized by having.
(Appendix 2)
It further has an aggregation unit that generates aggregated data by aggregating for each operator, including the number of descriptions and the number of non-descriptions, whether or not each element obtained by dividing the response sentence of the incident is described in the incident.
The answer candidate creating device according to Appendix 1, wherein the determination unit makes the determination based on the aggregated data.
(Appendix 3)
The determination unit calculates the description ratio of each element in the incident at regular intervals for each operator, and based on the calculated description ratio, it is not described in the incident before the arbitrary time. , The answer candidate creating apparatus according to Appendix 1 or 2, wherein it is determined whether or not there is an element described in the incident after the arbitrary time.
(Appendix 4)
The determination unit determines for each element whether or not the element is consistently described in the incident.
The answer according to any one of Supplementary note 1 to 3, wherein the generation unit generates an answer candidate of the question answer collection by adding elements that are consistently described or not described. Candidate creation device.
(Appendix 5)
The determination unit determines whether or not the element is described in the incident of all operators.
The answer candidate creating device according to any one of Supplementary note 1 to 4, wherein the generation unit adds the described elements to generate answer candidates for the question and answer collection.
(Appendix 6)
A clustering unit that classifies the incident into multiple clusters,
If there is an element that is not described in the incident before the arbitrary time of the predetermined operator, but is described in the incident after the arbitrary time of the predetermined operator, it corresponds to the cluster containing the element. A setting unit that sets the predetermined operator as an operator candidate,
A selection unit for selecting an operator corresponding to the cluster from the operator candidates is provided.
The generator adds the elements described in the incident of the operator corresponding to the predetermined cluster to generate the answer candidates of the question and answer collection corresponding to the predetermined cluster.
The answer candidate creating device according to Appendix 1, wherein the answer candidate is created.
(Appendix 7)
A clustering unit that classifies the incident into multiple clusters,
When there is an element that the predetermined operator consistently describes in the incident, a setting unit that sets the predetermined operator as an operator candidate corresponding to the cluster including the element, and a setting unit.
A selection unit for selecting an operator corresponding to the cluster from the operator candidates is provided.
The generator adds the elements described in the incident of the operator corresponding to the predetermined cluster to generate the answer candidates of the question and answer collection corresponding to the predetermined cluster.
The answer candidate creating device according to Appendix 1, wherein the answer candidate is created.
(Appendix 8)
The selection unit is based on the number of elements described in all incidents described by the operator candidate and the number of elements described in all incidents after a predetermined time described by the operator candidate. An evaluation value is calculated for each operator candidate, and an operator corresponding to the cluster is selected from the operator candidates based on the magnitude of the evaluation value.
The answer candidate creating device according to Appendix 6 or 7, characterized in that.
(Appendix 9)
The clustering unit clusters past questions and answers, associates the past questions and answers with the incidents, and classifies the incidents into a plurality of clusters according to the associated clusters of the past questions and answers.
The answer candidate creating device according to Appendix 8, wherein the answer candidate is created.
(Appendix 10)
It is determined whether or not there is an element described in the incident before the arbitrary time, but in the incident after the arbitrary time.
If there is the element, add the element to generate answer candidates for the question and answer collection.
A method of creating answer candidates, which is characterized in that the processing is executed by a computer.
(Appendix 11)
Whether or not each element obtained by dividing the response sentence of the incident is described in the incident is aggregated for each operator including the number of descriptions and the number of non-statements to generate aggregated data.
The method for creating answer candidates according to Appendix 10, wherein the computer executes a process of performing the determination based on the aggregated data.
(Appendix 12)
The description ratio of each element to the incident is calculated for each operator at regular intervals, and based on the calculated description ratio, the incidents before the arbitrary time are not described, but the arbitrary time. The method for creating answer candidates according to Appendix 10 or 11, wherein the computer executes a process of determining whether or not there is an element described in the subsequent incidents.
(Appendix 13)
It is determined for each element whether or not the elements are consistently described or not described in the incident, and the elements that are consistently described or not described are added to obtain the answer candidates of the question and answer collection. The method for creating an answer candidate according to any one of Supplementary note 10 to 12, wherein the computer executes the generated process.
(Appendix 14)
It is characterized in that the computer executes a process of determining whether or not an element is described in the incident of all operators, adding the described element, and generating an answer candidate of the question and answer collection. The method for creating answer candidates according to any one of Supplementary Notes 10 to 13.
(Appendix 15)
The incident was classified into multiple clusters and
If there is an element that is not described in the incident before the arbitrary time of the predetermined operator, but is described in the incident after the arbitrary time of the predetermined operator, it corresponds to the cluster containing the element. The predetermined operator is set as an operator candidate to be operated, and the operator is set.
From the operator candidates, select the operator corresponding to the cluster, and select
Add the elements described in the incident of the operator corresponding to the predetermined cluster to generate the answer candidates of the question and answer collection corresponding to the predetermined cluster.
The method for creating answer candidates according to Appendix 10, wherein the processing is executed by the computer.
(Appendix 16)
The incident was classified into multiple clusters and
If there is an element that the predetermined operator consistently describes in the incident, the predetermined operator is set as the operator candidate corresponding to the cluster including the element.
From the operator candidates, select the operator corresponding to the cluster, and select
Add the elements described in the incident of the operator corresponding to the predetermined cluster to generate the answer candidates of the question and answer collection corresponding to the predetermined cluster.
The method for creating answer candidates according to Appendix 10, wherein the processing is executed by the computer.
(Appendix 17)
Evaluation is made for each operator candidate based on the number of elements described in all incidents described by the operator candidate and the number of elements described in all incidents described by the operator candidate after a predetermined time. A value is calculated, and an operator corresponding to the cluster is selected from the operator candidates based on the magnitude of the evaluation value.
The method for creating an answer candidate according to Appendix 15 or 16, wherein the processing is executed by the computer.
(Appendix 18)
The past question and answer are clustered, the past question and answer are associated with the incident, and the incident is classified into a plurality of clusters according to the associated cluster of the past question and answer.
The method for creating answer candidates according to Appendix 17, wherein the processing is executed by the computer.

1 Answer candidate creation device 2 Incident data 3 FAQ answer candidate 4 FAQ data 11 Clustering section 12 Element division section 13 Aggregation section 14 Judgment section 15 Generation section 16 Storage section 17 Setting section 18 Selection section 140 CPU
141 HDD
142 RAM
143 ROM
144 Display device 145 Input device 146 Communication interface 147 Bus

Claims (11)

An answer candidate creation program executed by an answer candidate creation device that generates answer candidates for a question and answer collection.
It is determined whether or not there is an element described in the incident before the arbitrary time, but in the incident after the arbitrary time.
If there is the element, add the element to generate answer candidates for the question and answer collection.
An answer candidate creation program characterized in that the processing is executed by the computer of the answer candidate creation device. Whether or not each element obtained by dividing the response sentence of the incident is described in the incident is aggregated for each operator including the number of descriptions and the number of non-statements to generate aggregated data.
The response candidate creation program according to claim 1, wherein the determination is made based on the aggregated data. The description ratio of each element to the incident is calculated for each operator at regular intervals, and based on the calculated description ratio, it is not described in the incident before the arbitrary time, but after the arbitrary time. The response candidate creation program according to claim 1 or 2, wherein it is determined whether or not there is an element described in the incident. It is determined for each element whether or not the elements are consistently described or not described in the incident, and the elements that are consistently described or not described are added to obtain the answer candidates of the question and answer collection. The answer candidate creation program according to any one of claims 1 to 3, characterized in that it is generated. Any of claims 1 to 4, wherein it is determined whether or not the element is described in the incident of all operators, and the described element is added to generate an answer candidate of the question and answer collection. Answer candidate creation program described in one. The incident was classified into multiple clusters and
If there is an element that is not described in the incident before the arbitrary time of the predetermined operator, but is described in the incident after the arbitrary time of the predetermined operator, it corresponds to the cluster containing the element. The predetermined operator is set as an operator candidate to be operated, and the operator is set.
From the operator candidates, select the operator corresponding to the cluster, and select
Add the elements described in the incident of the operator corresponding to the predetermined cluster to generate the answer candidates of the question and answer collection corresponding to the predetermined cluster.
The answer candidate creation program according to claim 1. The incident was classified into multiple clusters and
If there is an element that the predetermined operator consistently describes in the incident, the predetermined operator is set as the operator candidate corresponding to the cluster including the element.
From the operator candidates, select the operator corresponding to the cluster, and select
Add the elements described in the incident of the operator corresponding to the predetermined cluster to generate the answer candidates of the question and answer collection corresponding to the predetermined cluster.
The answer candidate creation program according to claim 1. Evaluation is made for each operator candidate based on the number of elements described in all incidents described by the operator candidate and the number of elements described in all incidents described by the operator candidate after a predetermined time. A value is calculated, and an operator corresponding to the cluster is selected from the operator candidates based on the magnitude of the evaluation value.
The answer candidate creation program according to claim 6 or 7, wherein the answer candidate is created. The past question and answer are clustered, the past question and answer are associated with the incident, and the incident is classified into a plurality of clusters according to the associated cluster of the past question and answer.
The answer candidate creation program according to claim 8, wherein the answer candidate is created. A determination unit that determines whether or not there is an element that is not described in the incident before the arbitrary time but is described in the incident after the arbitrary time.
If there is the element, add the element and generate an answer candidate for the question and answer collection.
An answer candidate creation device characterized by having. It is determined whether or not there is an element described in the incident before the arbitrary time, but in the incident after the arbitrary time.
If there is the element, add the element to generate answer candidates for the question and answer collection.
A method of creating answer candidates, which is characterized in that the processing is executed by a computer .

Images