Information processing apparatus, model generation apparatus, information processing method, model generation method, and program

Abstract

To provide a device, a method, and a program for facilitating determination of a sentence.SOLUTION: The information processing apparatus 10 includes an importance level identifying unit 140 and an importance level judging unit 150. The importance degree specifying unit 140 specifies a representative sentence corresponding to a question sentence and an importance degree of the representative sentence by using the question sentence and one or more models prepared in advance. The importance determination unit 150 determines whether or not to extract the representative sentence using the importance of the representative sentence. In one example, the one or more models include a model trained by machine learning.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing device, a model generating device, an information processing method, a model generating method, and a program. [Background technology]

[0002] There are devices such as chatbots that respond to questions from customers, etc. To improve the performance of such devices, it is necessary to enrich the databases used by the devices.

[0003] Patent Document 1 describes a device for efficiently storing sets of unanswered questions and answers in a database. The device described in Patent Document 1 outputs a user's question based on a determination result such as the degree of similarity between the user's question and existing questions in an existing knowledge DB. An administrator then inputs an answer to the user's question, and the user's question and the input answer from the administrator are linked and registered in the existing knowledge DB. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2022-32278 Summary of the Invention [Problem to be solved by the invention]

[0005] However, if a person had to make the final decision on whether to register each user's question in the database, it would take time and effort because even if the user's question has the same meaning, it can be expressed in a variety of ways, and the content of each question would need to be checked and understood one by one.

[0006] The present invention provides an apparatus, a method, and a program that facilitate making judgments about sentences. [Means for solving the problem]

[0007] According to one aspect of the present invention, there are provided an information processing device, a model generating device, an information processing method, a model generating method, and a program as follows.

[0008] 1. an importance determination unit that uses a question sentence and one or more models prepared in advance to determine a representative sentence corresponding to the question sentence and an importance of the representative sentence; an importance determination unit that determines whether or not to extract the representative sentence using the importance of the representative sentence; Information processing device. 2. In the information processing device according to item 1, The question sentence and the representative sentence corresponding to the question sentence have a relationship in which an answer to the question sentence can be used as an answer to the representative sentence. Information processing device. 3. In the information processing device according to 1. or 2., The one or more models include a trained model by machine learning, The importance identification unit inputs the question sentence to the trained model and causes the trained model to output at least the representative sentence corresponding to the input question sentence. Information processing device. 4. In the information processing device according to 3., The importance identification unit further outputs the importance of the representative sentence corresponding to the input question sentence from the trained model. Information processing device. 5. In the information processing device according to 3., the one or more models further include a rule model; The rule model indicates a rule for specifying the importance of the representative sentence based on one or more words contained in the representative sentence specified for the question sentence. Information processing device. 6. In the information processing device according to any one of 3. to 5., The trained models are prepared in advance for each category, The importance identification unit identifies the representative sentence of the question sentence using the trained model corresponding to the category of the question sentence. Information processing device. 7. In the information processing device according to any one of 1. to 6., a target registration unit that determines whether or not information about the dialogue is a target based on the information about the dialogue; a question extraction unit that extracts the question from information about the dialogue determined to be a target by the target registration unit. Information processing device. 8. In the information processing device according to item 7., the target registration unit determines whether or not information related to the dialogue is a target based on a response time indicated in the information related to the dialogue; Information processing device. 9. In the information processing device according to any one of 1. to 8., The system further includes a similarity determination unit that calculates the similarity between the identified representative sentence and each of a plurality of registered representative sentences that have been registered in advance in an existing database, and determines whether or not to extract the identified representative sentence based on the calculated similarities. Information processing device. 10. In the information processing device according to any one of 1. to 9., the importance determination unit determines whether to extract the representative sentence based on a comparison result between the importance of the representative sentence and a predetermined threshold value; The threshold value is a value determined based on a plurality of importance levels determined using the one or more models for a plurality of predetermined example question sentences. Information processing device. 11. A learning unit is provided that performs machine learning on the model using learning data, The learning data includes a question sentence, a representative sentence corresponding to the question sentence, and information indicating the basis for identifying the representative sentence. Model generation device. 12. One or more computers Using a question sentence and one or more models prepared in advance, a representative sentence corresponding to the question sentence and an importance of the representative sentence are identified; The importance of the representative sentence is used to determine whether or not to extract the representative sentence. Information processing methods. 13. Computer, an importance determination unit that uses a question sentence and one or more models prepared in advance to determine a representative sentence corresponding to the question sentence and the importance of the representative sentence; and The importance determination unit determines whether or not to extract the representative sentence by using the importance of the representative sentence. program. 14. One or more computers Machine learning is performed on the model using the training data, The learning data includes a question sentence, a representative sentence corresponding to the question sentence, and information indicating the basis for identifying the representative sentence. Model generation method. 15. The computer functions as a learning unit that performs machine learning on the model using the learning data; The learning data includes a question sentence, a representative sentence corresponding to the question sentence, and information indicating the basis for identifying the representative sentence. program. [Effects of the Invention]

[0009] According to the present invention, it is possible to provide an apparatus, a method, and a program that make it easier to make judgments about sentences. [Brief explanation of the drawings]

[0010] [Figure 1]FIG. 2 is a diagram illustrating an example of a functional configuration of the information processing device according to the first embodiment. [Figure 2] FIG. 1 illustrates an example of a computer. [Figure 3] 4 is a flowchart illustrating the flow of an information processing method according to the first embodiment. [Figure 4] FIG. 2 is a diagram illustrating an example of a trained model that can be used by an importance identification unit according to the first embodiment. [Figure 5] FIG. 10 is a diagram illustrating an example of the functional configuration of a response device that uses a question and answer database to make a response. [Figure 6] 4 is a flowchart illustrating the flow of a process executed by the response device according to the first embodiment. [Figure 7] FIG. 1 is a diagram illustrating an example of a functional configuration of a model generation device according to a first embodiment. [Figure 8] 3 is a flowchart illustrating the flow of processing executed by the model generation device according to the first embodiment. [Figure 9] FIG. 10 is a diagram illustrating an example of the configuration of a trained model and a rule model used by an importance identification unit according to the second embodiment. [Figure 10] FIG. 10 is a diagram illustrating an example of a functional configuration of an information processing device according to a fourth embodiment. [Figure 11] 13 is a flowchart illustrating the flow of processing executed in step S10 by an importance level identification unit according to the fourth embodiment. [Figure 12] FIG. 10 is a block diagram illustrating a functional configuration of an information processing device according to a fifth embodiment. [Figure 13] 13 is a flowchart illustrating the flow of processing executed by an information processing apparatus according to the fifth embodiment. [Figure 14] FIG. 13 is a block diagram illustrating a functional configuration of an information processing device according to a sixth embodiment. [Figure 15] 13 is a flowchart illustrating the flow of processing executed by an information processing apparatus according to the sixth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In all the drawings, like components are designated by like reference numerals, and the description thereof will be omitted as appropriate.

[0012] (First embodiment) FIG. 1 is a diagram illustrating an example of the functional configuration of an information processing device 10 according to the first embodiment. The information processing device 10 according to this embodiment includes an importance identification unit 140 and an importance determination unit 150. The importance identification unit 140 uses a question sentence and one or more models prepared in advance to identify a representative sentence corresponding to the question sentence and the importance of the representative sentence. The importance determination unit 150 uses the importance of the representative sentence to determine whether or not to extract the representative sentence.

[0013] According to the information processing device 10 of this embodiment, by replacing a question sentence with a representative sentence, it is easier to understand the intent of the question than by checking the original question sentence. Therefore, it becomes easier to judge the sentence, and important question sentences can be extracted and selected efficiently. Furthermore, by replacing a question sentence with a representative sentence and aligning the sentence expressions, it becomes easier to perform aggregation by sorting, etc.

[0014] The hardware configuration of the information processing device 10 will be described below. Each functional component of the information processing device 10 (importance identification unit 140 and importance determination unit 150) may be realized by hardware that realizes each functional component (e.g., a hardwired electronic circuit, etc.), or may be realized by a combination of hardware and software (e.g., a combination of an electronic circuit and a program that controls it, etc.). Below, a case where each functional component of the information processing device 10 is realized by a combination of hardware and software will be further described.

[0015] FIG. 2 is a diagram illustrating a computer 1000 for realizing the information processing device 10. The computer 1000 is any computer. For example, the computer 1000 is a system on chip (SoC), a personal computer (PC), a server machine, a tablet terminal, or a smartphone. The computer 1000 may be a dedicated computer designed to realize the information processing device 10, or may be a general-purpose computer. Furthermore, the information processing device 10 may be realized by one computer 1000 or by a combination of multiple computers 1000.

[0016] The computer 1000 includes a bus 1020, a processor 1040, a memory 1060, a storage device 1080, an input / output interface 1100, and a network interface 1120. The bus 1020 is a data transmission path through which the processor 1040, the memory 1060, the storage device 1080, the input / output interface 1100, and the network interface 1120 transmit and receive data to and from each other. However, the method of interconnecting the processor 1040 and other components is not limited to bus connection. The processor 1040 may be any of various processors, such as a central processing unit (CPU), a graphics processing unit (GPU), or a field-programmable gate array (FPGA). The memory 1060 is a main storage device implemented using a random access memory (RAM) or the like. The storage device 1080 is an auxiliary storage device implemented using a hard disk, a solid state drive (SSD), a memory card, a read-only memory (ROM), or the like.

[0017] The input / output interface 1100 is an interface for connecting the computer 1000 to an input / output device. For example, an input device such as a keyboard and an output device such as a display are connected to the input / output interface 1100. The input / output interface 1100 may be connected to the input device or output device by wireless connection or by wired connection.

[0018] The network interface 1120 is an interface for connecting the computer 1000 to a network. This communication network may be, for example, a LAN (Local Area Network) or a WAN (Wide Area Network). The network interface 1120 may be connected to the network via a wireless connection or a wired connection.

[0019] The storage device 1080 stores program modules that realize each functional component of the information processing device 10. The processor 1040 reads each of these program modules into the memory 1060 and executes them to realize the function corresponding to each program module.

[0020] 3 is a flowchart illustrating the flow of an information processing method according to this embodiment. The information processing method according to this embodiment includes steps S10 and S11. In step S10, one or more computers use a question sentence and one or more models prepared in advance to identify a representative sentence corresponding to the question sentence and the importance of the representative sentence. Then, in step S11, the one or more computers use the importance of the representative sentence to determine whether or not to extract the representative sentence.

[0021] The information processing method according to this embodiment can be executed by the information processing device 10.

[0022] FIG. 4 is a diagram illustrating a trained model 20 that can be used by the importance identification unit 140 according to this embodiment. The one or more models that can be used by the importance identification unit 140 include a trained model 20 based on machine learning. The importance identification unit 140 inputs a question sentence to the trained model 20. The importance identification unit 140 then causes the trained model 20 to output at least a representative sentence corresponding to the input question sentence. The question sentence is, for example, text data, but the question sentence may also be voice data. The importance identification unit 140 may acquire the question sentence as voice data, convert it to text data, and input it to the trained model 20, or input it directly to the trained model 20. The representative sentence output from the trained model 20 is, for example, text data, but the representative sentence may also be voice data. If the representative sentence is voice data, the importance identification unit 140 may convert the voice data to text data for subsequent processing.

[0023] The trained model 20 may be a large-scale language model or a natural language processing AI model. The trained model 20 may be an interactive generative model or an interactive AI. Input to the trained model 20 can be performed using a prompt. The configuration of the output data of the trained model 20 may be specified by a prompt of the trained model 20 each time input is made.

[0024] The importance identification unit 140 can be said to have a function of converting a question sentence into a representative sentence. A question sentence and its corresponding representative sentence have a relationship in which an answer to the question sentence can be used as an answer to the representative sentence. For example, when considering a question sentence in a customer inquiry, there may be question sentences with various expressions for the same intention. The importance identification unit 140 can convert a question sentence into a representative sentence so as to unify the terminology and expressions used in the sentence. In other words, the same representative sentence can correspond to multiple question sentences with different expressions. Furthermore, by converting question sentences with different intentions into representative sentences, the terminology is unified, making them easier for people to understand and easier to handle as targets for sorting, searching, etc.

[0025] The relationship between a question sentence and its corresponding representative sentence will be explained in more detail. The representative sentence may contain words that are different from those contained in the question sentence but have a similar meaning to those contained in the question sentence. The question sentence and the representative sentence may differ in at least some of the word order. The question sentence and the representative sentence may differ in whether or not they use polite language. For example, the question sentence and the representative sentence have the same "topic," the same "situation" that forms the background of the sentence, and the same "request" indicated by the sentence.

[0026] The input data of the trained model 20 does not have to include an answer sentence to the question sentence. Also, the output data of the trained model 20 does not have to include an answer sentence to the representative sentence.

[0027] A storage unit accessible by the importance identification unit 140 stores, in advance, a prompt template including a command statement to be input to the trained model 20. This prompt template is prepared in advance through trials and adjustments so as to obtain a desired output. The importance identification unit 140 can generate a prompt by reading the prompt template from the storage unit and combining the template with a question sentence and other information as needed. The importance identification unit 140 then inputs the generated prompt into the trained model 20 to obtain desired output data. Note that the importance identification unit 140 may obtain the desired output through multiple exchanges with the trained model 20.

[0028] The storage unit that stores the prompt templates may be provided inside or outside the information processing device 10. When the storage unit is provided inside the information processing device 10, the storage unit is realized, for example, by using the storage device 1080 of the computer 1000 that realizes the information processing device 10.

[0029] In the example of FIG. 4, the importance identification unit 140 further outputs the importance of the representative sentence corresponding to the input question sentence from the trained model 20. As an example, the importance is a numerical value, and the larger this numerical value, the higher the importance, i.e., the more important it is. Note that the importance may be indicated in two stages: important or unimportant.

[0030] The trained model 20 is stored in advance in a storage unit accessible by the importance identification unit 140. The importance identification unit 140 can read and use the trained model 20 from this storage unit. This storage unit may be provided in the information processing device 10 or may be provided externally to the information processing device 10. When this storage unit is provided in the information processing device 10, this storage unit is realized, for example, by a storage device 1080 of a computer 1000 that realizes the information processing device 10. On the other hand, when the trained model 20 is stored in a storage unit provided externally to the information processing device 10, the importance identification unit 140 may input data to the trained model 20 and acquire output data from the trained model 20 via a communication network such as a WAN.

[0031] The importance determination unit 150 determines whether to extract a representative sentence based on the comparison result between the importance of the representative sentence and a predetermined threshold. For example, if the importance of the representative sentence is greater than the threshold, the importance determination unit 150 extracts the representative sentence. On the other hand, if the importance of the representative sentence is not greater than the threshold, the importance determination unit 150 does not extract the representative sentence.

[0032] As an example, the threshold value is a value determined based on a plurality of importance levels identified using one or more models for a plurality of predetermined example question sentences.

[0033] An example of a method for determining the threshold will be described in detail below. To determine the threshold, a plurality of example questions are prepared. Then, the plurality of example questions are input into the trained model 20, and the importance of each of the plurality of example questions is obtained. The threshold can be determined by referring to the relationship between the example questions thus obtained and the importance. For example, the threshold may be set to a value lower than the importance of example questions that are considered important, or may be set to a value higher than the importance of example questions that are considered unimportant.

[0034] Alternatively, a decision tree may be created based on a combination of example question sentences and importance levels, and the threshold value may be determined using the decision tree. For example, each node of the decision tree may be associated with a keyword, and a determination may be made as to whether or not the keyword is included in the question sentence. The endpoints of the decision tree indicate importance levels. By using such a decision tree, it is possible to identify the importance threshold value for when the representative sentence extracted by the importance level determination unit 150 includes a desired keyword.

[0035] The threshold is not limited to the above example, and may be determined taking into consideration, for example, the number or ratio of representative sentences to be extracted, an appropriate processing time of the information processing device 10, and the like.

[0036] The flow of processing performed by the information processing device 10 according to this embodiment will be described with reference to FIG.

[0037] The importance identification unit 140 according to this embodiment acquires a question sentence. The question sentence is stored in advance in a storage unit accessible from the importance identification unit 140. The importance identification unit 140 can, for example, acquire the question sentence by reading it from this storage unit. This storage unit may be provided in the information processing device 10 or may be provided outside the information processing device 10. When this storage unit is provided in the information processing device 10, this storage unit is realized, for example, by the storage device 1080 of the computer 1000 that realizes the information processing device 10. As another example, the importance identification unit 140 may acquire the question sentence from another functional configuration unit provided in the information processing device 10, or may acquire it from a device different from the information processing device 10.

[0038] The importance identifying unit 140 may acquire questions one by one and process them as they are acquired, or may acquire and process a plurality of questions in a batch format.

[0039] A question sentence may be associated with an answer sentence to the question sentence. The answer sentence is, for example, text data, but may also be audio data. The importance identification unit 140 may then acquire the answer sentence in addition to the question sentence in a similar manner. If the answer sentence is audio data, the importance identification unit 140 may convert the audio data into text data and perform subsequent processing. However, the importance identification unit 140 does not necessarily have to acquire the answer sentence.

[0040] The importance level identification unit 140 may acquire a table having a column for question sentences and a column for answer sentences. Such a table is prepared based on input by an operator responding to an inquiry or based on a record of a conversation by a chatbot.

[0041] The importance identification unit 140 inputs the acquired question sentence into the trained model 20 and causes the trained model 20 to output a representative sentence and importance. The importance identification unit 140 identifies the representative sentence output from the trained model 20 as a representative sentence for the question sentence, and identifies the importance output from the trained model 20 as the importance of the representative sentence for the question sentence (S10). The importance identification unit 140 associates the identified representative sentence with the importance. The importance identification unit 140 may further associate the identified representative sentence with the answer sentence that was associated with the question sentence.

[0042] When the importance identification unit 140 acquires a table having a column for question sentences and a column for answer sentences, an incomplete sentence may be entered in at least some of the multiple question sentence columns. For such an incomplete sentence, the importance identification unit 140 assigns the lowest importance. For example, an incomplete sentence may be entered when the send button is pressed while inputting to the chatbot.

[0043] Furthermore, for example, requests outside the specifications of the question subject may be assigned a lower importance level than questions within the specifications, or may be classified as "out of scope." Note that "out of scope" representative sentences may be excluded from subsequent processing, or may be determined as representative sentences not to be extracted in the processing of the importance level identification unit 140, which will be described next.

[0044] The importance determination unit 150 acquires the representative sentence and importance identified by the importance identification unit 140. The importance determination unit 150 may further acquire an answer sentence associated with the representative sentence from the importance identification unit 140. The importance determination unit 150 compares the acquired importance with a threshold value to determine whether or not to extract the representative sentence (S11). As a specific example, as described above, if the importance is greater than the threshold value, the importance determination unit 150 determines to extract the representative sentence associated with that importance. On the other hand, if the importance is not greater than the threshold value, the importance determination unit 150 determines not to extract the representative sentence associated with that importance.

[0045] The importance determination unit 150 outputs the representative sentence that it has determined to be extracted. The importance determination unit 150 does not output the representative sentence that it has determined not to extract. The method for outputting the representative sentence is not particularly limited. For example, the importance determination unit 150 may store the representative sentence in a storage unit accessible from the importance determination unit 150. This storage unit may be provided in the information processing device 10 or may be provided external to the information processing device 10. When this storage unit is provided in the information processing device 10, this storage unit is realized, for example, by the storage device 1080 of the computer 1000 that realizes the information processing device 10. As another example, the importance determination unit 150 may display the representative sentence on a display connected to the information processing device 10, or may output the representative sentence to another functional configuration unit provided in the information processing device 10, or may output the representative sentence to a device different from the information processing device 10.

[0046] Furthermore, in addition to the representative sentence that has been determined to be extracted, the importance determination unit 150 may also output at least one of the importance and the response sentence associated with the representative sentence in a similar manner.

[0047] The output result of the information processing device 10 is used, for example, as follows: When a representative sentence is output, the user of the information processing device 10 checks the representative sentence and determines whether or not to register it in the question and answer database 103 (see FIG. 5) used by the response device. At this time, the representative sentence is a sentence that is easier for a person to understand than the original question sentence. Therefore, a more efficient decision can be made than if the decision were made by checking the original question sentence.

[0048] Furthermore, if the importance of the representative sentence is also output from the information processing device 10, the user can further refer to this importance and decide whether or not to register the representative sentence.

[0049] The user registers the representative sentence that the user judges should be registered in the question and answer database 103. At this time, the user may modify the representative sentence as necessary before registering it in the question and answer database 103. On the other hand, the representative sentence that the user does not judge should be registered is not registered in the question and answer database 103.

[0050] In order to reduce the effort required to prepare answer sentences, it is preferable that further answer sentences be output from the information processing device 10. If an answer sentence associated with a representative sentence that has been determined to be registered has been further output from the information processing device 10, the user associates the answer sentence with the representative sentence and registers it in the question and answer database 103. At this time, the user may modify the answer sentence as necessary before registering it in the question and answer database 103. If an answer sentence associated with a representative sentence that has been determined to be registered has not been output from the information processing device 10, the user creates an answer sentence to the representative sentence and registers the created answer sentence in the question and answer database 103 in association with the representative sentence. In this way, pairs of representative sentences and answer sentences are stored in the question and answer database 103.

[0051] The information processing device 10 can process multiple question sentences. For example, the importance identification unit 140 acquires multiple question sentences and identifies a representative sentence and importance for each question sentence. The importance determination unit 150 determines whether or not to extract each of the multiple representative sentences. The importance determination unit 150 then outputs one or more representative sentences that it has determined should be extracted.

[0052] The user of the information processing device 10 determines whether or not each of the output representative sentences should be registered in the question and answer database 103. Even if the user must make a decision about multiple representative sentences, the user can easily understand the content of each representative sentence because the representative sentences use consistent terminology and expressions. Therefore, the user can make the decision with less time and effort than if they had to make a decision about the original multiple question sentences.

[0053] 5 is a diagram illustrating an example of the functional configuration of a response device 90 that makes a response using the question and answer database 103. The response device 90 will be described below. The response device 90 includes a response unit 910.

[0054] The hardware configuration of a computer that realizes the response device 90 according to this embodiment is shown in Fig. 2, for example, similar to the information processing device 10 according to this embodiment. However, a program module that realizes the function of the response unit 910 is stored in a storage device 1080 of the computer 1000 that realizes the response device 90 according to this embodiment. The computer 1000 that realizes the response device 90 may also serve as the computer 1000 that realizes the information processing device 10, or may be provided separately from the computer 1000 that realizes the information processing device 10.

[0055] The response unit 910 can access the question and answer database 103. In the example of Fig. 5, the response device 90 further includes the question and answer database 103. However, the question and answer database 103 may be provided outside the response device 90. When the question and answer database 103 is provided in the response device 90, the question and answer database 103 is realized, for example, by a storage device 1080 of the computer 1000 that realizes the response device 90.

[0056] The response device 90 can be used, for example, as a chatbot. The response device 90 is used, for example, to respond to inquiries from questioners (for example, customers).

[0057] 6 is a flowchart illustrating the flow of processing executed by the response device 90 according to this embodiment. In the example of FIG. 5, the questioner inputs one or more sentences as a query to the questioner terminal 92. The questioner terminal 92 is, for example, a computer or a mobile terminal. The sentences input by the questioner to the questioner terminal 92 may be input by voice or text. When voice input is performed, the voice is converted into text in the questioner terminal 92, for example.

[0058] The response unit 910 acquires one or more sentences input by the questioner to the questioner terminal 92 (S20). The response device 90 is connected to the questioner terminal 92 via a communication network via an input / output interface 1100 or a network interface 1120 of the computer 1000 that realizes the response device 90.

[0059] The response unit 910 performs preprocessing on the one or more acquired sentences to extract a question from the one or more sentences (S21). The preprocessing can be performed using an existing method. Specifically, the response unit 910 recognizes periods and question marks in the one or more acquired sentences and divides the one or more sentences into individual sentences. Then, based on the wording and other factors contained in each sentence, it determines whether or not each sentence corresponds to a question. In this way, the response unit 910 extracts a question from the one or more sentences.

[0060] The response unit 910 inputs the extracted question sentence into the trained model 20 and causes the trained model 20 to output a representative sentence corresponding to the question sentence. In this way, the response unit 910 converts the question sentence into a representative sentence (S22).

[0061] However, if the trained model 20 is capable of extracting a question sentence, step S21 does not need to be performed after step S20. In this case, the response unit 910 may simply input the acquired one or more sentences to the trained model 20 as is, and cause the trained model 20 to output a representative sentence corresponding to the question sentence contained in the acquired one or more sentences. The response unit 910 may also switch whether or not to perform step S21 depending on the length of the acquired one or more sentences. For example, step S21 may be performed if the length of the acquired one or more sentences is equal to or greater than a predetermined length, and step S21 may not be performed if the length of the acquired one or more sentences is shorter than the predetermined length.

[0062] The trained model 20 used here may be the same as the trained model 20 used by the importance identification unit 140 described above. However, the response unit 910 does not have to output the importance from the trained model 20. The trained model 20 is stored in advance in a storage unit accessible by the response unit 910. The response unit 910 can read and use the trained model 20 from this storage unit. This storage unit may be provided in the response device 90 or may be provided external to the response unit 910. If this storage unit is provided in the response device 90, this storage unit is realized, for example, by the storage device 1080 of the computer 1000 that realizes the response device 90. On the other hand, if the trained model 20 is stored in a storage unit provided external to the response device 90, the response unit 910 may input data to the trained model 20 and acquire output data from the trained model 20 via a communication network such as a WAN.

[0063] As with the importance identification unit 140, the response unit 910 can read and use a prompt template stored in advance in a storage unit accessible by the response unit 910. As with the importance identification unit 140, the response unit 910 may obtain a desired output through multiple exchanges with the trained model 20.

[0064] The storage unit that stores the prompt template may be provided inside or outside the response device 90. When this storage unit is provided inside the response device 90, this storage unit is realized, for example, using the storage device 1080 of the computer 1000 that realizes the response device 90.

[0065] The response unit 910 searches the representative sentence output from the trained model 20 among the representative sentences stored in the question and answer database 103 (S23). Then, the response unit 910 reads out and outputs the answer sentence associated with the representative sentence found in the question and answer database 103 (S24).

[0066] The response unit 910 can output the answer sentence to the questioner terminal 92. The answer sentence may be displayed in text format on the questioner terminal 92, or may be output as audio.

[0067] It is possible that the same representative sentence as the representative sentence output from the trained model 20 does not exist among the multiple representative sentences stored in the question-and-answer database 103. In this case, the response unit 910 may read and output from the question-and-answer database 103 an answer sentence associated with the representative sentence that has the highest similarity to the representative sentence output from the trained model 20 among the multiple representative sentences stored in the question-and-answer database 103. The similarity can be calculated using an existing method, such as calculating cosine similarity.

[0068] Furthermore, the response unit 910 may output information indicating that an answer is not possible if there is no representative sentence among the multiple representative sentences stored in the question and answer database 103 whose similarity to the representative sentence output from the trained model 20 is equal to or greater than a predetermined similarity. The response unit 910 may output information indicating that an answer is not possible to an operator terminal other than the questioner terminal 92. This makes it possible to have a person answer instead of the response device 90.

[0069] In this way, the response device 90 can respond to questions.

[0070] The person asking the question is not limited to a customer. The person asking the question may be, for example, an employee. That is, the response device 90 may be used, for example, within a company to respond to inquiries from employees or the like.

[0071] The generation of the trained model 20 used by the information processing device 10 will be described below. FIG. 7 is a diagram illustrating an example of the functional configuration of a model generation device 50 according to this embodiment. FIG. 8 is a flowchart illustrating an example of the flow of processing executed by the model generation device 50 according to this embodiment. The model generation device 50 according to this embodiment includes a learning unit 510. The learning unit 510 performs machine learning on the model using training data (S30). The training data includes a question sentence and a representative sentence corresponding to the question sentence. In this embodiment, the training data also includes the importance of the representative sentence corresponding to the question sentence.

[0072] The hardware configuration of the computer that realizes the model generation device 50 according to this embodiment is shown in, for example, FIG. 2 , similar to the information processing device 10 according to this embodiment. However, a program module that realizes the function of the learning unit 510 is stored in the storage device 1080 of the computer 1000 that realizes the model generation device 50 according to this embodiment. The computer 1000 that realizes the model generation device 50 may also function as the computer 1000 that realizes the information processing device 10, or may be provided separately from the computer 1000 that realizes the information processing device 10. Furthermore, the computer 1000 that realizes the model generation device 50 may also function as the computer 1000 that realizes the response device 90, or may be provided separately from the computer 1000 that realizes the response device 90.

[0073] Machine learning in the model generation device 50 can be performed using an existing method. In the model generation device 50, for example, machine learning is performed as fine tuning on a pre-trained model to obtain a trained model 20. As the pre-trained model, an AI model from a general-purpose service can be used. Note that examples of pre-trained models include trained models 20 that have already been fine-tuned to a certain extent. In this case, the model generation device 50 can generate a trained model 20 with higher accuracy by performing further machine learning on that model.

[0074] Machine learning may be performed by inputting training data to a prompt of a pre-trained model. That is, machine learning may be performed by inputting a combination of a question sentence and correct answer data for the question sentence to a prompt of the pre-trained model. In this embodiment, the correct answer data includes an appropriate representative sentence and importance for the question sentence.

[0075] Learning data for machine learning can be prepared, for example, in the following manner. A question sentence can be input to a pre-trained model, causing the pre-trained model to output a representative sentence. The user of the model generation device 50 checks whether the output representative sentence is appropriate in light of the input question sentence. If the output representative sentence is not appropriate, the user of the model generation device 50 corrects the representative sentence and includes the corrected representative sentence in the correct answer data for the input question sentence.

[0076] At this time, the representative sentences are modified to conform to predetermined rules, such as standardizing terminology, whether or not to use polite language, and word order, thereby achieving uniformity of expression and terminology among the multiple representative sentences generated by the trained model 20.

[0077] In this embodiment, the importance of a representative sentence corresponding to an input question sentence can be further output from the pre-trained model. The user of the model generation device 50 checks whether the output importance is appropriate in light of the input question sentence. If the output importance is not appropriate, the user of the model generation device 50 corrects the importance and includes the corrected importance in the correct answer data for the input question sentence.

[0078] The appropriate importance level can be determined by the judgment of one or more experts in question-and-answer sessions with questioners. In this way, the know-how for judging importance can be reflected in the trained model 20.

[0079] In this embodiment, the input question may be further output from a pre-trained model.

[0080] Note that the input of a question to the pre-trained model to prepare learning data may be performed manually, or may be configured such that, for example, when the importance identification unit 140 of the information processing device 10 acquires a question, the model generation device 50 acquires a copy of the question and inputs it to the pre-trained model. In this case, the information processing device 10 and the model generation device 50 are connected so as to be able to communicate with each other.

[0081] The learning unit 510 acquires the learning data prepared as described above and performs additional learning on the pre-trained model, thereby obtaining the trained model 20 that has undergone additional learning.

[0082] In addition, to confirm whether the learning has been performed appropriately, the question sentence included in the learning data may be input to the trained model 20 that has undergone additional training, and it may be confirmed whether an output similar to the correct answer data is obtained.

[0083] The learning unit 510 stores the obtained trained model 20 in a storage unit accessible from the learning unit 510. This storage unit may be provided in the model generation device 50 or may be provided external to the model generation device 50. If this storage unit is provided in the model generation device 50, this storage unit is realized, for example, by the storage device 1080 of the computer 1000 that realizes the model generation device 50. Also, this storage unit may be provided in the information processing device 10 or may be provided external to the information processing device 10. If this storage unit is provided in the information processing device 10, this storage unit is realized, for example, by the storage device 1080 of the computer 1000 that realizes the information processing device 10. Also, this storage unit may be provided in the response device 90 or may be provided external to the response device 90. If this storage unit is provided in the response device 90, this storage unit is realized, for example, by the storage device 1080 of the computer 1000 that realizes the response device 90.

[0084] The information processing device 10 according to this embodiment does not necessarily need to use a model trained by the above-described model generation device 50. For example, the information processing device 10 may use an interactive AI model available through a general-purpose service without additional training.

[0085] Next, the operation and effect of this embodiment will be described. According to this embodiment, the importance identification unit 140 uses a question sentence and one or more models prepared in advance to identify a representative sentence corresponding to the question sentence and the importance of the representative sentence. Therefore, by checking the representative sentence and the importance, the user of the information processing device 10 can more easily make a judgment about the sentence than if they checked the original question sentence.

[0086] (Second embodiment) The information processing device 10 according to the second embodiment is the same as the information processing device 10 according to the first embodiment, except for the points described below. The information processing method according to the second embodiment is the same as the information processing method according to the first embodiment, except for the points described below.

[0087] 9 is a diagram illustrating the configuration of the trained model 20 and the rule model 22 used by the importance identification unit 140 according to this embodiment. In the second embodiment, the one or more models used by the importance identification unit 140 further include a rule model 22. The rule model 22 is a model that indicates rules for identifying the importance of a representative sentence identified for a question sentence, based on one or more words (keywords) included in the representative sentence.

[0088] The rule model 22 may be, for example, a decision tree. Each node of the decision tree may be associated with a keyword, and a determination may be made as to whether or not the keyword is included in the representative sentence. The endpoints of the decision tree indicate importance. By using such a decision tree, importance can be identified.

[0089] The importance identifying unit 140 according to this embodiment does not need to output the importance from the trained model 20.

[0090] The rule model 22 can be prepared based on a plurality of example questions and the importance level assigned to each example question. The importance level for each example question can be determined by the judgment of one or more experts in question-answering with questioners. This allows the rule model 22 to reflect the know-how of importance judgment.

[0091] The rule model 22 is stored in advance in a storage unit accessible by the importance identification unit 140. The importance identification unit 140 can read out and use the rule model 22 from this storage unit. This storage unit may be provided in the information processing device 10 or may be provided outside the information processing device 10. When this storage unit is provided in the information processing device 10, this storage unit is realized by, for example, a storage device 1080 of the computer 1000 that realizes the information processing device 10.

[0092] The flow of processing executed by the information processing device 10 according to this embodiment will be described. The importance identification unit 140 acquires at least a question sentence, as in the first embodiment. Then, the importance identification unit 140 inputs the acquired question sentence into the trained model 20 and causes the trained model 20 to output a representative sentence. The importance identification unit 140 identifies the representative sentence output from the trained model 20 as a representative sentence for the question sentence. Furthermore, the importance identification unit 140 applies the representative sentence output from the trained model 20 to the rule model 22 to identify the importance. The importance identification unit 140 identifies the importance identified using the rule model 22 as the importance of the representative sentence.

[0093] The importance determination unit 150 determines whether or not to extract a representative sentence in the same manner as in the first embodiment, and then outputs the representative sentence that has been determined to be extracted.

[0094] The importance identification unit 140 according to this embodiment may further output one or more keywords contained in the representative sentence from the trained model 20. In this case, instead of applying the representative sentence to the rule model 22, the importance can be identified by determining whether or not the one or more keywords output from the trained model 20 include a keyword associated with each node in the decision tree.

[0095] As another example, the importance identification unit 140 may identify the importance using the rule model 22 as described above, and may output the importance from the trained model 20 as in the first embodiment. In this case, two importance levels are obtained: the importance level according to the rule model 22 and the importance level according to the trained model 20. The importance determination unit 150 makes a determination for each of the two importance levels, and if at least one of the determination results indicates that the representative sentence should be extracted (i.e., the representative sentence is important), the importance determination unit 150 extracts the representative sentence. On the other hand, if both determination results indicate that the representative sentence should not be extracted (i.e., the representative sentence is not important), the importance determination unit 150 does not extract the representative sentence.

[0096] The model generation device 50 according to the second embodiment is the same as the model generation device 50 according to the first embodiment, except for the points described below. The model generation method according to the second embodiment is the same as the model generation method according to the first embodiment, except for the points described below.

[0097] The learning data used for machine learning by the model generation device 50 according to this embodiment does not need to include importance.

[0098] On the other hand, when the importance identification unit 140 causes the trained model 20 to further output keywords contained in the representative sentence, one or more keywords are included as part of the supervised data in the training data used for machine learning by the model generation device 50. The one or more keywords to be included in the supervised data can be determined as follows.

[0099] In this embodiment, one or more keywords can be further output from the pre-trained model. The user of the model generation device 50 checks whether the one or more output keywords are appropriate in light of the input question. If the one or more output keywords are inappropriate, the user of the model generation device 50 corrects the one or more keywords and includes the corrected one or more keywords in the correct answer data for the input question.

[0100] For example, multiple keyword candidates are predetermined, and the user of the model generation device 50 identifies, among the multiple candidates, a candidate included in a representative sentence output from a pre-trained model as a keyword to be included in the correct answer data. This unifies the keywords output from the trained model 20.

[0101] Next, the operation and effect of this embodiment will be described. In this embodiment, the same operation and effect as in the first embodiment can be obtained.

[0102] (Third embodiment) The functional configuration of the model generation device 50 according to the third embodiment is illustrated in FIG. 7, as in the first embodiment. Also, the flow of processing executed by the model generation device 50 according to the third embodiment is illustrated in FIG. 8, as in the first embodiment. The model generation device 50 according to the third embodiment is the same as the model generation device 50 according to the first or second embodiment, except for the points described below. The model generation method according to the third embodiment is the same as the model generation method according to the first or second embodiment, except for the points described below.

[0103] In this embodiment, the learning unit 510 performs machine learning on the model using learning data. The learning data includes a question sentence, a representative sentence corresponding to the question sentence, and evidence information indicating the basis for identifying the representative sentence.

[0104] In the model generation method according to the present embodiment, one or more computers perform machine learning on a model using training data, which includes a question sentence, a representative sentence corresponding to the question sentence, and evidence information indicating the basis for identifying the representative sentence.

[0105] The model generation method according to this embodiment can be executed by the model generation device 50 according to this embodiment.

[0106] The model generation device 50 and model generation method according to this embodiment provide a trained model 20 that can appropriately output the basis for identifying a representative sentence. This makes it easier for a person to determine the validity of the representative sentence when obtaining training data for further training the trained model 20. This reduces the effort required for a person to evaluate the output representative sentence.

[0107] Examples of evidence information include text such as "Because XX," "The evidence is ZZ," "Because the question contains the word YY," and "The evidence is the statement AA in the question." Examples of evidence information include text that includes one or more of the words "from," "because," "ground," "translation," and "reason."

[0108] As another example, the evidence information includes text indicating at least one of the "topic," "situation," and "desired thing" that is common to the question sentence and the representative sentence. It is preferable that the evidence information includes text indicating the "topic," "situation," and "desired thing" that is common to the question sentence and the representative sentence.

[0109] In the method for generating training data according to this embodiment, a question sentence is input to a pre-trained model, and the model outputs at least a representative sentence and also outputs evidence information indicating the basis for identifying the representative sentence. This allows a person preparing training data to refer to the evidence information when checking whether the output representative sentence is appropriate in light of the input question sentence. This allows the task of determining whether the representative sentence is appropriate to proceed efficiently.

[0110] If the user of the model generating device 50 thinks that the basis information is inappropriate, the user performs additional learning by including the corrected basis information in the correct answer data.

[0111] Furthermore, the model that is the subject of machine learning may further output keywords contained in the representative sentence, similar to the trained model 20 according to the second embodiment. The keywords can also help the user to confirm whether the output representative sentence is appropriate.

[0112] The information processing device 10 according to the third embodiment is the same as the information processing device 10 according to the first or second embodiment. The information processing method according to the third embodiment is the same as the information processing method according to the first or second embodiment.

[0113] Next, the actions and effects of this embodiment will be described. In this embodiment, the same actions and effects as in the first embodiment can be obtained. In addition, in the model generation device 50 according to this embodiment, the output data of the model on which machine learning is performed includes basis information indicating the basis for identifying the representative sentence. Therefore, a trained model 20 can be obtained that can appropriately output the basis for identifying the representative sentence. Ultimately, it becomes easier to determine the validity of the representative sentence, reducing the effort required for a person to evaluate the output representative sentence.

[0114] (Fourth embodiment) 10 is a diagram illustrating an example of the functional configuration of an information processing device 10 according to a fourth embodiment. The information processing device 10 according to this embodiment is the same as the information processing device 10 according to at least one of the first to third embodiments, except for the points described below. The information processing method according to this embodiment is the same as the information processing method according to at least one of the first to third embodiments, except for the points described below.

[0115] In this embodiment, trained models 20 are prepared in advance for each category. Then, the importance identification unit 140 uses the trained model 20 corresponding to the category of the question sentence to identify a representative sentence of the question sentence.

[0116] 10, the information processing device 10 further includes a model database 100. The model database 100 may be provided in the information processing device 10, or may be provided outside the information processing device 10. When the model database 100 is provided in the information processing device 10, the model database 100 is realized by, for example, a storage device 1080 of a computer 1000 that realizes the information processing device 10.

[0117] The model database 100 stores trained models 20 by category. That is, the model database 100 stores trained models 20 of multiple categories (Model A, Model B, ...). The category relates to what the question is about. The multiple categories include, for example, one or more of "waybill," "shipping receipt," "printing or reprinting," "shipping," "land or air mail," "fee," and "unissued or issued." As another example, the category may be a category related to the content of the question. The multiple categories include, for example, one or more of "deleting or canceling waybills or shipping receipts," "creating and editing waybills," "printing or reprinting waybills or shipping receipts," "managing address books and customer information," and "data management and data output."

[0118] 11 is a flowchart illustrating the flow of processing executed by the importance identification unit 140 according to this embodiment in step S10. The importance identification unit 140 according to this embodiment acquires a question sentence in the same manner as in the first embodiment (S40). Next, the importance identification unit 140 identifies a category of the question sentence (S41). For example, the importance identification unit 140 may identify the category by determining whether or not a predetermined keyword is included in the question sentence. Alternatively, the importance identification unit 140 may identify the category using a trained model that has been machine-learned to identify a category.

[0119] The importance identification unit 140 then reads out the trained model 20 of the identified category from the model database 100 (S42). The importance identification unit 140 also inputs a question sentence into the read trained model 20, causing the trained model 20 to output at least a representative sentence (S43). The importance identification unit 140 then uses the output data from the trained model 20 to identify the representative sentence and the importance, in the same manner as in the first or second embodiment (S44).

[0120] The model generation device 50 according to this embodiment is the same as the model generation device 50 according to at least one of the first to third embodiments, except that it generates trained models 20 for each category. The model generation method according to this embodiment is the same as the model generation method according to at least one of the first to third embodiments, except that it generates trained models 20 for each category.

[0121] The model generation device 50 according to this embodiment uses training data related to questions in a specific category to generate one trained model 20. This makes it possible to particularly improve the accuracy of the trained model 20 for the specific category. In other words, a trained model 20 that can output data more appropriate for questions in that category is obtained. The model generation device 50 similarly generates trained models 20 for each of multiple categories.

[0122] As described above, by using a trained model 20 for each category and by using a trained model 20 that is suitable for the category of the question, it becomes easier to obtain appropriate output data.

[0123] Furthermore, when multiple questions are input into the trained model 20 at once, the number of questions that can be processed by the trained model 20 may be limited. Even in this case, the processing time can be reduced by processing in parallel by category.

[0124] As explained in the second embodiment, when the rule model 22 is used in combination, rule models 22 for each category may also be prepared in advance and used according to the category of the question sentence.

[0125] Next, the operation and effect of this embodiment will be described. In this embodiment, the same operation and effect as in the first embodiment can be obtained. In addition, the importance identification unit 140 identifies a representative sentence of a question sentence using a trained model 20 corresponding to the category of the question sentence. Therefore, the accuracy of identifying the representative sentence can be improved.

[0126] The information processing device 10 according to this embodiment may use the prompt templates described above in the first embodiment prepared for each question category. In this case, one or more models used by the information processing device 10 according to this embodiment may be the same for all categories. Even in this case, the accuracy of identifying a representative sentence can be improved by using an appropriate prompt corresponding to the question category. Furthermore, processing time can be reduced by using the same multiple models in parallel.

[0127] (Fifth embodiment) 12 is a block diagram illustrating the functional configuration of an information processing device 10 according to a fifth embodiment. The information processing device 10 according to the fifth embodiment is the same as the information processing device 10 according to at least one of the first to fourth embodiments, except for the points described below. The information processing method according to the fifth embodiment is the same as the information processing method according to at least one of the first to fourth embodiments, except for the points described below.

[0128] The information processing device 10 according to the fifth embodiment further includes an object registration unit 110 and a question extraction unit 120. The object registration unit 110 determines whether or not dialogue information is to be treated as a target, based on dialogue information relating to the dialogue. The question extraction unit 120 extracts a question from the dialogue information determined by the object registration unit 110 to be a target.

[0129] 12, the information processing device 10 further includes a target database 101. The target database 101 holds dialogue information to be processed by the question extraction unit 120. However, the information processing device 10 does not necessarily have to include the target database 101. The target database 101 may be provided outside the information processing device 10.

[0130] The hardware configuration of a computer that realizes the information processing device 10 according to this embodiment is shown in, for example, Fig. 2, similar to the information processing device 10 according to the first embodiment. However, the storage device 1080 of the computer 1000 that realizes the information processing device 10 according to this embodiment further stores program modules that realize the functions of the object registration unit 110 and the question sentence extraction unit 120 of this embodiment. When the object database 101 is provided inside the information processing device 10, the object database 101 is realized using the storage device 1080 of the information processing device 10 that realizes the information processing device 10 according to this embodiment.

[0131] 13 is a flowchart illustrating the flow of processing executed by the information processing device 10 according to this embodiment. The information processing device 10 according to this embodiment will be described in detail with reference to FIGS.

[0132] The dialogue information includes a series of dialogue contents between the questioner and the responder. The responder may be a person or a response device such as a chatbot (e.g., response device 90). The dialogue information is, for example, a dialogue log. In the dialogue information, the dialogue contents may be represented by text data or by voice data. When the dialogue contents are represented by voice data, the target registration unit 110 may convert the voice data into text data.

[0133] The dialogue information preferably includes at least a question from the questioner. The dialogue information preferably further includes a response from the responder to the question. This question is preferably a question in a category that should be handled by the response device 90. The dialogue information preferably does not include questions outside the category that should be handled by the response device 90. The dialogue information preferably does not include personal information such as the name of the questioner, amounts, etc. Personal information such as the name of the questioner, amounts, etc. may be masked in the dialogue information.

[0134] The dialogue information is preferably information about a dialogue that has been completed. The dialogue information preferably further includes information indicating the time it took the person to respond. Furthermore, if the person to respond was unable to answer the question from the questioner, the dialogue information preferably includes information indicating that the person was unable to answer.

[0135] The dialogue information may be a table having a column for a question and a column for a response. Such a table may be prepared based on input by an operator responding to an inquiry or a record of a dialogue from a chatbot. In this case, a blank column for the response in the table may indicate that the question could not be answered.

[0136] The object registration unit 110 can read and acquire the dialogue information from, for example, a storage unit that stores a dialogue log. As an example, dialogue information is generated each time a dialogue is held and stored in this storage unit. This storage unit may be provided in the information processing device 10 or may be provided outside the information processing device 10. When this storage unit is provided in the information processing device 10, this storage unit is realized by, for example, a storage device 1080 of a computer 1000 that realizes the information processing device 10.

[0137] However, the object registration unit 110 may acquire the dialogue information from another device, or from another functional configuration unit included in the information processing device 10.

[0138] The object registration unit 110 may acquire and process dialogue information every time it is generated, or may acquire and process a plurality of pieces of dialogue information together in a batch format.

[0139] When the object registration unit 110 acquires the dialogue information, it determines whether or not the dialogue information is to be treated as a target. Then, the object registration unit 110 registers the dialogue information determined to be a target in the object database 101 (S50). On the other hand, the dialogue information determined not to be a target is not registered in the object database 101.

[0140] The object registration unit 110 determines whether or not to include the dialogue information based on, for example, the response time indicated in the dialogue information. Specifically, if the response time is longer than a predetermined reference time, the object registration unit 110 determines that the dialogue information is to be included. On the other hand, if the response time is not longer than the predetermined reference time, the object registration unit 110 determines that the dialogue information is not to be included.

[0141] Alternatively, the object registration unit 110 may determine whether or not to target dialogue information based on whether or not the dialogue information includes information indicating that an answer could not be made. Specifically, for dialogue information including information indicating that an answer could not be made, the object registration unit 110 determines that the dialogue information is to be targeted. On the other hand, for dialogue information not including information indicating that an answer could not be made, the object registration unit 110 determines that the dialogue information is not to be targeted.

[0142] The target registration unit 110 may make a determination by combining the two examples described above. That is, for dialogue information that satisfies at least one of the conditions that the response time is longer than a predetermined reference time or that the dialogue information includes information indicating that a reply could not be made, the target registration unit 110 determines that the dialogue information is to be included in the target. On the other hand, for dialogue information that does not satisfy either of these conditions, the target registration unit 110 determines that the dialogue information is not to be included in the target.

[0143] By performing the above-described processing by the object registration unit 110, dialogue information relating to matters for which the information used by the responder for responding should be strengthened can be targeted for processing by the question extraction unit 120.

[0144] The question extraction unit 120 reads dialogue information from the target database 101. The question extraction unit 120 extracts a question from the dialogue content included in the acquired dialogue information (S51). The dialogue content may include one or more sentences. That is, the question extraction unit 120 extracts a question from a "string" of one or more sentences. The extraction of a question can be performed using an existing method. Specifically, the question extraction unit 120 recognizes periods and question marks in the text indicating the dialogue content, and divides the "string" of one or more sentences into individual sentences. Then, the question extraction unit 120 determines whether each sentence corresponds to a question. In this way, the question extraction unit 120 extracts a question from the dialogue information.

[0145] The determination of whether each sentence corresponds to a question can be made based on the wording of each sentence. Specifically, the question extraction unit 120 determines that a sentence that includes at least one of a plurality of predetermined wordings corresponds to a question. On the other hand, the question extraction unit 120 determines that a sentence that does not include any of the plurality of predetermined wordings does not correspond to a question.

[0146] The predetermined plurality of phrases may include phrases indicating questions, phrases indicating requests, and phrases indicating dissatisfaction, because phrases indicating requests and phrases indicating dissatisfaction may also indirectly indicate questions.

[0147] Examples of phrases expressing a question include "Can you do it...?", "Can you do it...?", "Is it possible?", "Is it possible?", and "Is it possible?". Examples of phrases expressing a request include "I want to do it..." and "I want to do it.". Examples of phrases expressing a dissatisfaction include "I can't do it...", "I can't do it", "It won't be done", and "It won't come out."

[0148] The question sentence extraction unit 120 may further identify an answer sentence to the extracted question sentence from one or more sentences. The answer sentence can be identified by an existing method.

[0149] If the dialogue information is a table having a column for question sentences and a column for answer sentences, the above-mentioned division or extraction of question sentences based on wording may not be performed.

[0150] The importance identification unit 140 according to this embodiment acquires the question sentence extracted by the question sentence extraction unit 120. Furthermore, as described in the first embodiment, the importance identification unit 140 according to this embodiment may further acquire the answer sentence extracted by the question sentence extraction unit 120. The importance identification unit 140 identifies a representative sentence and an importance level for the acquired question sentence, similar to the importance identification unit 140 according to at least one of the first to fourth embodiments (S52).

[0151] However, if the trained model 20 is capable of extracting a question sentence, step S51 does not need to be performed after step S50. In this case, the importance identification unit 140 may simply input the acquired dialogue information to the trained model 20 and cause the trained model 20 to output a representative sentence corresponding to the question sentence included in the dialogue information. The importance identification unit 140 may also switch whether or not to perform step S51 depending on the length of the dialogue indicated in the acquired dialogue information. For example, step S51 may be performed if the length of the dialogue indicated in the acquired dialogue information is equal to or greater than a predetermined length, and step S51 may not be performed if the length of the dialogue indicated in the acquired dialogue information is shorter than the predetermined length.

[0152] Then, the importance determination unit 150 according to this embodiment determines whether or not to extract a representative sentence, in the same manner as described in the first embodiment (S53).

[0153] Note that instead of registering the dialogue information that the object registration unit 110 has determined to be registered in the object database 101 in the object database 101, the question extraction unit 120 may obtain the dialogue information from the object registration unit 110 and extract a question. In this case, the object database 101 may not be provided.

[0154] The model generation device 50 according to this embodiment is the same as the model generation device 50 according to at least one of the first to fourth embodiments. The model generation method according to this embodiment is the same as the model generation method according to at least one of the first to fourth embodiments.

[0155] Next, the operation and effect of this embodiment will be described. In this embodiment, the same operation and effect as in the first embodiment can be obtained. In addition, according to this embodiment, the object registration unit 110 determines whether or not to register the dialogue information in the object database 101. Therefore, the processing load on the information processing device 10 is reduced compared to when the question sentence extraction unit 120 processes all the dialogue information.

[0156] (Sixth embodiment) 14 is a block diagram illustrating the functional configuration of an information processing device 10 according to a sixth embodiment. The information processing device 10 according to the sixth embodiment is the same as the information processing device 10 according to at least one of the first to fifth embodiments, except for the points described below. The information processing method according to the sixth embodiment is the same as the information processing method according to at least one of the first to fifth embodiments, except for the points described below.

[0157] The information processing device 10 according to the sixth embodiment further includes a similarity determination unit 170. The similarity determination unit 170 calculates the similarity between the representative sentence identified by the importance identification unit 140 and each of a plurality of registered representative sentences that have been registered in advance in an existing database (the question and answer database 103 in FIG. 14). Then, the similarity determination unit 170 determines whether or not to extract the identified representative sentence based on the calculated similarities.

[0158] 14, the information processing device 10 further includes a question and answer database 103 as an existing database. However, the information processing device 10 does not include the question and answer database 103, and the question and answer database 103 may be provided outside the information processing device 10. As described in the first embodiment, the question and answer database 103 holds pairs of representative sentences and answer sentences. The pairs of representative sentences and answer sentences held in the question and answer database 103 are used by the response device 90.

[0159] The hardware configuration of a computer that realizes the information processing device 10 according to this embodiment is shown in Fig. 2, for example, similar to the information processing device 10 according to the first embodiment. However, the storage device 1080 of the computer 1000 that realizes the information processing device 10 according to this embodiment further stores a program module that realizes the function of the similarity determination unit 170. When the question and answer database 103 is provided inside the information processing device 10, the question and answer database 103 is realized using the storage device 1080 of the information processing device 10 that realizes the information processing device 10 according to this embodiment.

[0160] 15 is a flowchart illustrating the flow of processing executed by the information processing device 10 according to this embodiment. The information processing device 10 according to this embodiment will be described in detail with reference to FIGS.

[0161] The importance identification unit 140 according to this embodiment identifies a representative sentence and its importance (S60), similarly to the importance identification unit 140 according to at least one of the first to fifth embodiments. Then, the importance determination unit 150 according to this embodiment determines whether or not to extract a representative sentence based on the importance, similarly to the first embodiment (S61).

[0162] Next, the similarity determination unit 170 acquires the representative sentence determined by the importance determination unit 150 to be extracted. The similarity determination unit 170 may further acquire, from the importance determination unit 150, answer sentences associated with the representative sentence. The similarity determination unit 170 calculates the similarity between the acquired representative sentence (hereinafter also referred to as the "target representative sentence") and each of the multiple registered representative sentences registered in the question and answer database 103. The similarity can be calculated using an existing method, such as calculating cosine similarity. The similarity determination unit 170 then determines whether or not to extract the target representative sentence based on the multiple calculated similarities (S62). Specifically, if the maximum similarity among the multiple calculated similarities is equal to or less than a predetermined threshold, the similarity determination unit 170 determines to extract the target representative sentence. Specifically, if the maximum similarity among the multiple calculated similarities is not equal to or less than a predetermined threshold, the similarity determination unit 170 determines not to extract the target representative sentence.

[0163] The threshold value used here may be determined, for example, taking into consideration the number or ratio of representative sentences to be extracted, an appropriate processing time, etc., or may be determined based on a desired degree of similarity.

[0164] The similarity determination unit 170 outputs the representative sentence that it has determined to be extracted. The similarity determination unit 170 does not output the representative sentence that it has determined not to extract. The method for outputting the representative sentence is not particularly limited. For example, the similarity determination unit 170 may store the representative sentence in a storage unit accessible from the similarity determination unit 170. This storage unit may be provided in the information processing device 10 or may be provided external to the information processing device 10. When this storage unit is provided in the information processing device 10, this storage unit is realized, for example, by the storage device 1080 of the computer 1000 that realizes the information processing device 10. As another example, the similarity determination unit 170 may display the representative sentence on a display connected to the information processing device 10, or may output the representative sentence to another functional configuration unit provided in the information processing device 10, or may output the representative sentence to a device different from the information processing device 10.

[0165] Furthermore, the similarity determination unit 170 may output, in addition to the representative sentence determined to be extracted, at least one of the importance and the answer sentence associated with the representative sentence in a similar manner.

[0166] Note that, although FIG. 15 shows an example in which a judgment based on importance (S61) is followed by a judgment based on similarity (S62), a judgment based on importance (S61) may also be performed after a judgment based on similarity (S62).

[0167] Moreover, the information processing device 10 according to this embodiment may further include an object registration unit 110 and a question sentence extraction unit 120, similar to the fifth embodiment.

[0168] The model generation device 50 according to this embodiment is the same as the model generation device 50 according to at least one of the first to fifth embodiments. The model generation method according to this embodiment is the same as the model generation method according to at least one of the first to fifth embodiments.

[0169] Next, the operation and effect of this embodiment will be described. In this embodiment, the same operation and effect as in the first embodiment can be obtained. In addition, according to this embodiment, the similarity determination unit 170 determines whether or not to extract the identified representative sentence based on the similarity. Therefore, new question sentences that are not registered in the question and answer database 103 can be efficiently added to the question and answer database 103.

[0170] Although the embodiments of the present invention have been described above with reference to the drawings, these are merely examples of the present invention, and various other configurations may be adopted. For example, in the sequence diagrams and flowcharts used in the above description, multiple steps (processes) are described in order, but the order of execution of the steps performed in each embodiment is not limited to the order described. In each embodiment, the order of the steps shown in the drawings may be changed to the extent that the content is not affected. Furthermore, the above-described embodiments may be combined to the extent that the content is not contradictory. [Explanation of symbols]

[0171] 10. Information processing equipment 20 Pre-trained models 22 Rule Model 50 Model generation device 90 Answering Machine 92 Questioner terminal 100 Model Database 101 Target Database 103 Question and Answer Database 110 Target Registration Department 120 Question sentence extraction part 140 Importance Identification Section 150 Importance judgment section 170 Similarity determination unit 510 Learning Department 910 Response Section 1000 calculator 1020 Bus 1040 processor 1060 memory 1080 storage device 1100 Input / Output Interface 1120 Network Interface

Claims

1. an importance determination unit that uses a question sentence and one or more models prepared in advance to determine a representative sentence corresponding to the question sentence and an importance of the representative sentence; an importance determination unit that determines whether or not to extract the representative sentence using the importance of the representative sentence; Information processing device.

2. 2. The information processing device according to claim 1, The question sentence and the representative sentence corresponding to the question sentence have a relationship in which an answer to the question sentence can be used as an answer to the representative sentence. Information processing device.

3. 3. The information processing device according to claim 1, The one or more models include a trained model by machine learning, The importance identification unit inputs the question sentence to the trained model and causes the trained model to output at least the representative sentence corresponding to the input question sentence. Information processing device.

4. 4. The information processing device according to claim 3, The importance identification unit further outputs the importance of the representative sentence corresponding to the input question sentence from the trained model. Information processing device.

5. 4. The information processing device according to claim 3, the one or more models further include a rule model; The rule model indicates a rule for specifying the importance of the representative sentence based on one or more words contained in the representative sentence specified for the question sentence. Information processing device.

6. 6. The information processing device according to claim 3, The trained models are prepared in advance for each category, The importance identification unit identifies the representative sentence of the question sentence using the trained model corresponding to the category of the question sentence. Information processing device.

7. 7. The information processing device according to claim 1, a target registration unit that determines whether or not information about the dialogue is a target based on the information about the dialogue; a question extraction unit that extracts the question from information about the dialogue determined to be a target by the target registration unit. Information processing device.

8. 8. The information processing device according to claim 7, the target registration unit determines whether or not information related to the dialogue is a target based on a response time indicated in the information related to the dialogue; Information processing device.

9. 9. The information processing device according to claim 1, The system further includes a similarity determination unit that calculates the similarity between the identified representative sentence and each of a plurality of registered representative sentences that have been registered in advance in an existing database, and determines whether or not to extract the identified representative sentence based on the calculated similarities. Information processing device.

10. 10. The information processing device according to claim 1, the importance determination unit determines whether to extract the representative sentence based on a comparison result between the importance of the representative sentence and a predetermined threshold value; The threshold value is a value determined based on a plurality of importance levels determined using the one or more models for a plurality of predetermined example question sentences. Information processing device.

11. A learning unit is provided that performs machine learning on the model using learning data, The learning data includes a question sentence, a representative sentence corresponding to the question sentence, and information indicating the basis for identifying the representative sentence. Model generation device.

12. One or more computers Using a question sentence and one or more models prepared in advance, a representative sentence corresponding to the question sentence and an importance of the representative sentence are identified; The importance of the representative sentence is used to determine whether or not to extract the representative sentence. Information processing methods.

13. Computer, an importance determination unit that uses a question sentence and one or more models prepared in advance to determine a representative sentence corresponding to the question sentence and the importance of the representative sentence; and The importance determination unit determines whether or not to extract the representative sentence by using the importance of the representative sentence. program.

14. One or more computers Machine learning is performed on the model using the training data, The learning data includes a question sentence, a representative sentence corresponding to the question sentence, and information indicating the basis for identifying the representative sentence. Model generation method.

15. The computer functions as a learning unit that performs machine learning on the model using the learning data; The learning data includes a question sentence, a representative sentence corresponding to the question sentence, and information indicating the basis for identifying the representative sentence. program.

Images