Information processing method, information processing system, and program

Abstract

A terminal device 10 is equipped with: an information generating unit 41 for generating instruction information P including question information Q representing a question from a user relating to a guidance target; an information acquiring unit 42 for acquiring answer information R corresponding to the result of processing of the instruction information P by a trained generative model M, the answer information R representing an answer for providing guidance relating to the guidance target to the user in relation to the question represented by the question information Q, and the answer corresponding to auxiliary information X relating to the user; and an information output unit 43 for outputting the answer information R.

Description

Information processing method, information processing system, and program

[0001] The present disclosure relates to a technique for providing information to a user.

[0002] Techniques have been proposed for providing users with information guiding them to various facilities, such as commercial facilities, transportation facilities, exhibition facilities, etc. For example, Patent Document 1 discloses a configuration in which, of multiple pieces of guide information registered in advance in a database for different identification information, guide information corresponding to identification information input by a user into a user terminal is transmitted to the user terminal.

[0003] Japanese Patent Application Laid-Open No. 2002-109112

[0004] However, there is a problem in that the diverse information desired by the user cannot necessarily be provided if the user is simply presented with information prepared in advance, as in Patent Document 1. In consideration of the above circumstances, one aspect of the present disclosure aims to appropriately provide the diverse information desired by the user.

[0005] In order to solve the above problems, an information processing method according to one aspect of the present disclosure generates instruction information including question information representing a question from a user regarding a guidance target, obtains answer information corresponding to the result of a trained generative model processing the instruction information, where the answer is an answer that guides the user to the guidance target in response to the question represented by the question information and is also an answer corresponding to auxiliary information regarding the user, and outputs the answer information.

[0006] An information processing system according to one aspect of the present disclosure includes an information generation unit that generates instruction information including question information that represents a question from a user regarding a guidance target, an information acquisition unit that acquires answer information corresponding to the result of a trained generative model processing the instruction information, the answer being an answer that guides the user to the guidance target in response to the question represented by the question information and representing an answer corresponding to auxiliary information regarding the user, and an information output unit that outputs the answer information.

[0007] A program according to one aspect of the present disclosure causes a computer system to function as an information generation unit that generates instruction information including question information representing a question from a user regarding a guidance target, an information acquisition unit that acquires answer information corresponding to the result of a trained generative model processing the instruction information, the answer being an answer that guides the user to the guidance target in response to the question represented by the question information and representing an answer corresponding to auxiliary information regarding the user, and an information output unit that outputs the answer information.

[0008] FIG. 1 is a block diagram of an information system in a first embodiment. FIG. 2 is a block diagram of a terminal device. FIG. 3 is a block diagram illustrating an example of the functional configuration of a terminal device. FIG. 4 is a flowchart of a guidance process. FIG. 5 is a block diagram illustrating an example of the functional configuration of a terminal device in a fourth embodiment. FIG. 6 is an explanatory diagram of a guidance character. FIG. 7 is a block diagram of an answer generation system in a sixth embodiment. FIG. 8 is a block diagram illustrating an example of the functional configuration of an answer system. FIG. 9 is a flowchart of a basic training process. FIG. 10 is a block diagram illustrating an example of the functional configuration of an answer generation system in a seventh embodiment. FIG. 11 is a block diagram illustrating an example of the functional configuration of an answer generation system in an eighth embodiment. FIG. 12 is a flowchart of a selection process. FIG. 13 is a block diagram illustrating an example of the functional configuration of a terminal device in a modified example.

[0009] A: First Embodiment FIG. 1 is a block diagram illustrating the configuration of an information system 100 in the first embodiment. The information system 100 is a communication system that provides information to a user U. The information system 100 includes a terminal device 10 and an answer generation system 20. The terminal device 10 is an information processing system used by the user U. Examples of the terminal device 10 include portable information devices such as smartphones, tablet terminals, and personal computers. The terminal device 10 can communicate with the answer generation system 20 via a communication network 200 such as the Internet.

[0010] A user U is located in an exhibition facility 30 such as a museum or art gallery. A plurality of exhibits 31 are installed in the exhibition facility 30. Each exhibit 31 is an item to be guided to by the user U. The exhibits 31 are, for example, works of art such as paintings or statues, or items related to history or science. The user U uses the terminal device 10 near a specific exhibit 31 in the exhibition facility 30 while viewing the exhibit 31.

[0011] The terminal device 10 transmits instruction information P to the answer generation system 20. The instruction information P is information (prompt) in response to a question from the user U regarding the exhibit 31, and is expressed in, for example, natural language. The answer generation system 20 is a server system that provides answer information R corresponding to the instruction information P to the terminal device 10. The answer information R is information representing an answer that guides the user U around the exhibit 31 in response to the question expressed by the instruction information P. The terminal device 10 outputs the answer information R provided by the answer generation system 20 to the user U. In other words, the information system 100 guides the user U around the exhibition facility 30.

[0012] The answer generation system 20 generates answer information R by processing instruction information P using a trained generative model M. That is, the answer information R is information corresponding to the result of the generative model M processing the instruction information P. The generative model M is a generative probabilistic model that generates answer information R corresponding to the instruction information P. The generative model M has learned the tendency of answer information R in response to instruction information P through prior machine learning (pre-training). Specifically, the generative model M is an interactive large-scale language model (LLM) trained specifically for natural language processing tasks such as generating answers to questions. For example, a natural language processing model realized by a Transformer model using a self-attention mechanism is an example of the generative model M.

[0013] 2 is a block diagram of the terminal device 10. The terminal device 10 includes a control device 11, a storage device 12, a communication device 13, an operation device 14, a sound collection device 15, a position detection device 16, and an output device 19. The terminal device 10 may be realized as a single device, or may be realized as a plurality of devices configured separately from each other.

[0014] The control device 11 is configured with one or more processors that control each element of the terminal device 10. For example, the control device 11 is configured with one or more types of processors such as a central processing unit (CPU), a sound processing unit (SPU), a digital signal processor (DSP), a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC).

[0015] The storage device 12 is one or more memories that store programs executed by the control device 11 and various data used by the control device 11. The storage device 12 is configured with a known storage medium such as a magnetic storage medium or a semiconductor storage medium. The storage device 12 may be configured with a combination of multiple types of storage media. Furthermore, a portable storage medium that is detachable from the terminal device 10, or a storage medium (e.g., cloud storage) that the control device 11 can write to or read from via the communication network 200 may be used as the storage device 12.

[0016] The communication device 13 communicates with the answer generation system 20 under the control of the control device 11. Specifically, the communication device 13 transmits instruction information P to the answer generation system 20 and receives answer information R transmitted from the answer generation system 20.

[0017] The operation device 14 is an input device that accepts instructions from the user U. The operation device 14 is, for example, an operator operated by the user U, or a touch panel that detects contact by the user U. The sound collection device 15 is a microphone that collects surrounding sounds. The sound collection device 15 generates a collected sound signal by collecting the sound uttered by the user U, for example.

[0018] The position detection device 16 generates position information X1 that indicates the position of the terminal device 10. Specifically, the position detection device 16 generates the position information X1 by analyzing radio waves received from multiple positioning satellites of a satellite positioning system such as the Global Positioning System (GPS). The position information X1 can also be said to be information that indicates the position of the user U who carries the terminal device 10.

[0019] The output device 19 outputs the answer information R. Specifically, the output device 19 outputs the answer information R in a manner that can be perceived by the user U. The output device 19 includes a display device 17 and a sound emitting device 18. The display device 17 displays an image under the control of the control device 11. For example, the display device 17 displays the answer information R as an image. The display device 17 is configured with a display panel such as a liquid crystal panel or an organic EL panel. The sound emitting device 18 emits sound waves under the control of the control device 11. For example, the sound emitting device 18 emits the answer information R as sound. The sound emitting device 18 is, for example, a speaker or headphones. As explained above, the output of the answer information R by the output device 19 includes the display of an image and the emission of sound.

[0020] 3 is a block diagram illustrating an example of the functional configuration of the terminal device 10. The control device 11 executes a program stored in the storage device 12 to realize multiple functions (an information generation unit 41, an information acquisition unit 42, and an information output unit 43).

[0021] The user U can input a question to the terminal device 10 regarding an exhibit 31 located near the user U within the exhibition facility 30. For example, the user U can input a question such as "Please tell me the year and creator of this exhibit 31" regarding the exhibit 31 located in front of the user U. Specifically, the user U can input a string of characters of a question by operating the operation device 14. The user U can also input a question by voice, for example, by speaking to the sound collection device 15.

[0022] 3 is information representing a question from the user U regarding the exhibit 31. In other words, the question represented by the question information Q includes information that can identify the matter that the user U wishes to confirm regarding the guidance target. Furthermore, the auxiliary information X in FIG. 3 is auxiliary information regarding the user U. The auxiliary information X in the first embodiment includes the position information X1 generated by the position detection device 16.

[0023] The information generation unit 41 generates instruction information P. Specifically, the information generation unit 41 generates instruction information P including question information Q and auxiliary information X (location information X1). For example, the instruction information P is a natural language prompt requesting an answer to a question indicated by the question information Q about the exhibit 31 located at the position indicated by the location information X1. In other words, the guidance target (i.e., the exhibit 31) that cannot be uniquely identified by the question indicated by the question information Q alone is auxiliary identified by the location information X1. For example, the instruction information P is a prompt representing the question, "Please tell me the year of production and the creator of the exhibit 31 located near (location information X1)."

[0024] The information generation unit 41 transmits the instruction information P exemplified above from the communication device 13 to the answer generation system 20. The answer generation system 20 generates answer information R in natural language by processing the instruction information P received from the terminal device 10 using the generation model M, and transmits the answer information R to the terminal device 10. The answer information R is information corresponding to the question information Q and the auxiliary information X (location information X1). Specifically, the answer information R is information representing an answer that guides the user U through the exhibit 31 in response to a question specified by the question information Q and the auxiliary information X. In other words, the answer information R is generated to represent an answer corresponding to the question represented by the question information Q and the location represented by the location information X1. For example, the answer information R represents an answer such as, "This work was created by Mr. ABC in 1990."

[0025] The information acquisition unit 42 acquires the answer information R. Specifically, the information acquisition unit 42 receives the answer information R transmitted from the answer generation system 20 via the communication device 13.

[0026] The information output unit 43 outputs the answer information R via the output device 19. Specifically, the information output unit 43 displays a character string of the answer represented by the answer information R on the display device 17. Furthermore, the information output unit 43 reproduces the voice of the answer represented by the answer information R (hereinafter referred to as "answer voice V") via the sound emitting device 18. For example, the information output unit 43 generates a voice signal A by performing a voice synthesis process on the answer information R, and reproduces the answer voice V by outputting the voice signal A to the sound emitting device 18. The voice synthesis process is, for example, a segment concatenative voice synthesis process that connects multiple speech segments, or a probabilistic model-type voice synthesis process that uses a probabilistic model such as a deep neural network or an HMM (Hidden Markov Model).

[0027] 4 is a flowchart of a process (hereinafter referred to as "guidance process") executed by the control device 11. The guidance process is started in response to an instruction from the user U via the operation device 14. When the guidance process is started, the control device 11 waits for input of a question by the user U (Sa1: NO).

[0028] When a question input by the user U is accepted (Sa1: YES), the control device 11 acquires auxiliary information X (Sa2). Specifically, the control device 11 acquires position information X1 from the position detection device 16. The control device 11 (information generation unit 41) generates instruction information P including the question information Q and the auxiliary information X (Sa3), and transmits the instruction information P to the answer generation system 20 via the communication device 13 (Sa4).

[0029] The control device 11 (information acquisition unit 42) receives the answer information R transmitted from the answer generation system 20 via the communication device 13 (Sa5). Then, the control device 11 (information output unit 43) outputs the answer information R via the output device 19 (Sa6).

[0030] The control device 11 determines whether a predetermined termination condition is met (Sa7). The termination condition is, for example, when the user U issues a termination command via an operation on the operation device 14. If the termination condition is not met (Sa7: NO), the control device 11 proceeds to step Sa1 and waits for the user U to input a new question. As described above, each time the user U inputs a question, the generation (Sa3) and transmission (Sa4) of instruction information P and the acquisition (Sa5) and output (Sa6) of answer information R are repeated. If the termination condition is met (Sa7: YES), the control device 11 terminates the guidance process.

[0031] As described above, in the first embodiment, in addition to question information Q representing the question posed by the user U, answer information R reflecting auxiliary information X related to the user U is output. Therefore, compared to a configuration in which a guidance voice corresponding to the identification information input by the user is simply played back from among a plurality of guidance voices prepared in advance, a variety of information desired by the user U can be appropriately provided.

[0032] For example, in a configuration in which pre-prepared guide information is presented to a user as in Patent Document 1, the guide information often contains information other than the information desired by the user, resulting in the problem of the user having to confirm unnecessarily redundant guide information. Therefore, for example, when guide information is output as audio, the user may miss the information they actually need. In contrast to the configuration of Patent Document 1, the first embodiment outputs answer information R according to the results of processing question information Q representing a question from a user U by a generation model M. That is, compared to the configuration of Patent Document 1, the first embodiment has the advantage of being able to generate appropriate and concise answer information R by directly processing the question from the user U. For example, the user U can easily understand the desired information from the answer audio V without missing it.

[0033] In the first embodiment, in particular, the instruction information P corresponding to the auxiliary information X including the position information X1 is processed by the generative model M. Therefore, it is possible to provide appropriate information according to the position of the user U. For example, as illustrated above, it is possible to provide the user U with information about the exhibit 31 located near the user U among the multiple exhibits 31 in the exhibition facility 30.

[0034] B: Second Embodiment A second embodiment will be described. Note that, for elements in the following exemplary aspects that have the same functions as those in the first embodiment, the same reference numerals as those in the first embodiment will be used, and detailed descriptions of each element will be omitted as appropriate.

[0035] The auxiliary information X in the second embodiment includes attribute information X2 in addition to the location information X1 similar to that in the first embodiment. The attribute information X2 is information representing the attributes of the user U. The attribute information X2 represents attributes such as the age, generation, gender, place of residence (country or region), and language used of the user U, for example.

[0036] As in the first embodiment, the information generating unit 41 generates instruction information P including question information Q and auxiliary information X. For example, the instruction information P is a prompt representing a question such as "Please answer the questioner of (attribute information X2) about the year of production and the creator of the exhibit 31 located near (location information X1)."

[0037] As in the first embodiment, the generation model M processes the instruction information P to generate answer information R. In the second embodiment, the auxiliary information X includes attribute information X2, so the answer information R represents an answer according to the attribute of the user U represented by the attribute information X2. For example, the answer information R expressed in the language used by the user U is generated.

[0038] Furthermore, for example, the answer information R is expressed in a natural language in a tone that corresponds to the attribute information X2. For example, the answer information R represents an answer in a tone that corresponds to the age of the user U. For example, if the age represented by the attribute information X2 is below a threshold (for example, a child's age), answer information R expressed in a tone that is suitable for children is generated. On the other hand, if the age represented by the attribute information X2 is above a threshold (for example, an adult's age), answer information R expressed in a tone that is suitable for adults is generated. Furthermore, if the attribute information X2 represents a language used, answer information R expressed in that language is generated. Acquisition and output of answer information R are the same as in the first embodiment.

[0039] The second embodiment also achieves the same effects as the first embodiment. Moreover, in the second embodiment, instruction information P corresponding to auxiliary information X including attribute information X2 is processed by the generative model M. Therefore, appropriate answer information R corresponding to the attributes of user U can be provided to user U. Moreover, in the second embodiment, the tone of the answer information R is controlled according to the auxiliary information X (e.g., attribute information X2) of user U, so that answer information R can be provided in a tone appropriate for user U (e.g., a tone of voice that is familiar to user U).

[0040] C: Third Embodiment The auxiliary information X in the third embodiment includes history information X3 in addition to the location information X1 similar to that in the first embodiment. The history information X3 is information representing the behavioral history of the user U. For example, the history information X3 is information such as places the user U has visited in the past (e.g., a museum or art gallery), websites the user U has viewed in the past, questions the user U has entered in the past, or products the user U has purchased in the past. The history information X3 can also be expressed as information representing the interests of the user U.

[0041] As in the first embodiment, the information generating unit 41 generates instruction information P including question information Q and auxiliary information X. For example, the instruction information P is a prompt representing a question such as "Please tell us the year of production and the creator of the exhibit 31 located near (location information X1). The questioner is interested in (history information X3)."

[0042] The generation model M generates answer information R by processing the instruction information P, as in the first embodiment. In the third embodiment, the auxiliary information X includes history information X3, so the answer information R represents an answer according to the behavioral history represented by the history information X3. Specifically, the answer information R includes supplementary information related to the behavioral history represented by the history information X3. For example, answer information R is generated that includes a guide to an exhibition facility 30 that is estimated to be of high interest to the user U based on the behavioral history represented by the auxiliary information X. Note that the acquisition and output of the answer information R are similar to those in the first embodiment.

[0043] The third embodiment also achieves the same effects as the first embodiment. Furthermore, in the third embodiment, instruction information P corresponding to auxiliary information X including history information X3 is processed by the generative model M. Therefore, appropriate answer information R corresponding to the behavioral history of the user U can be provided to the user U.

[0044] D: Fourth Embodiment Fig. 5 is a block diagram illustrating the functional configuration of the terminal device 10 in the fourth embodiment. In the fourth embodiment, the auxiliary information X (X1 to X3) is used to generate the instruction information P as in the above-described embodiments, and is also used to output the answer information R by the information output unit 43.

[0045] The information output unit 43 of the fourth embodiment generates a voice signal A by performing voice synthesis processing on the answer information R, as in the first embodiment. The information output unit 43 controls the voice quality of the answer voice V represented by the voice signal A in accordance with the auxiliary information X (X1 to X3). Voice quality is an auditory characteristic of the answer voice V. For example, voice quality includes various elements related to the answer voice V, such as the frequency characteristics (timbre) of the answer voice V, as well as the pitch, range, intonation, or rhythm of the answer voice V. For example, when the age represented by the attribute information X2 is below a threshold (e.g., when the user is a child), a clear answer voice V with abundant harmonic components and a high-pitched range is generated. On the other hand, when the age represented by the attribute information X2 is above a threshold (e.g., when the user is an adult), a low-pitched answer voice V with abundant non-harmonic components is generated.

[0046] The information output unit 43 may control the voice quality of the answer information R by any method. For example, the information output unit 43 controls the voice quality of the answer voice V by controlling parameters applied to the voice synthesis process in accordance with the auxiliary information X. Furthermore, the information output unit 43 performs various acoustic processes on the voice signal A with a standard voice quality generated by the voice synthesis process, thereby generating the voice signal A with a voice quality corresponding to the auxiliary information X. The acoustic processes are, for example, well-known signal processes that change the frequency characteristics of the voice. The voice quality of the answer voice V can be controlled by controlling parameters applied to the acoustic processes in accordance with the auxiliary information X. As can be understood from the above description, outputting the answer information R in the fourth embodiment includes playing back the answer voice V that pronounces the answer information R with a voice quality corresponding to the auxiliary information X.

[0047] The fourth embodiment also achieves the same effects as the first embodiment. Furthermore, in the fourth embodiment, the voice quality of the answer information R is controlled according to the auxiliary information X (X1 to X3) of the user U, so that the answer information R can be provided in a voice quality that is appropriate for the user U (for example, a voice quality that is familiar to the user U).

[0048] E: Fifth Embodiment As in the above-described embodiments, the information output unit 43 displays the answer information R on the display device 17 and reproduces the answer voice V on the sound emitting device 18. In addition to the above operations, the information output unit 43 of the fifth embodiment displays a guide character C exemplified in Fig. 6 on the display device 17. The guide character C is a virtual display object that is displayed in parallel with the output of the answer information R to notify the user U of the answer information R.

[0049] The information output unit 43 controls the guide character C in parallel with the output of the answer information R. Specifically, the information output unit 43 controls the facial expression (e.g., mouth shape) and movement of the guide character C in conjunction with the reproduction of the answer voice V.

[0050] For example, the information output unit 43 controls the mouth shape of the guide character C according to the phonemes (vowels / consonants) or volume of the answer voice V. For example, the degree to which the mouth of the guide character C opens increases in the pronunciation section of a vowel in the answer voice V compared to the pronunciation section of a consonant. Also, the degree to which the mouth of the guide character C opens increases as the volume of the answer voice V increases. Furthermore, the guide character C swings both hands in conjunction with the playback of the answer voice V. As illustrated above, outputting the answer information R in the fifth embodiment includes displaying the guide character C, which notifies the user U of the answer information R, on the display device 17.

[0051] The fifth embodiment also achieves the same effects as the first embodiment. Furthermore, according to the fifth embodiment, the guide character C is displayed to notify the user of the answer information R, so that the answer information R can be provided to the user U in a familiar and memorable manner.

[0052] F: Sixth Embodiment Fig. 7 is a block diagram illustrating the configuration of an answer generation system 20 in a sixth embodiment. As illustrated in Fig. 7, the answer generation system 20 includes a control device 21, a storage device 22, and a communication device 23. The answer generation system 20 may be realized as a single device, or may be realized as a plurality of devices configured separately from each other.

[0053] The control device 21 is composed of one or more processors that control each element of the answer generation system 20. For example, the control device 21 is composed of one or more types of processors such as a CPU, an SPU, a DSP, an FPGA, or an ASIC.

[0054] The storage device 22 is one or more memories that store programs executed by the control device 21 and various data used by the control device 21. The storage device 22 is configured with a known storage medium such as a magnetic storage medium or a semiconductor storage medium. The storage device 22 may be configured with a combination of multiple types of storage media. Furthermore, a portable storage medium that is detachable from the answer generation system 20, or a storage medium (e.g., cloud storage) that the control device 21 can write to or read from via the communication network 200 may be used as the storage device 22.

[0055] The communication device 23 communicates with the answer generation system 20 under the control of the control device 21. As described above, the communication device 23 receives instruction information P transmitted from the terminal device 10, and transmits answer information R generated by the generation model M through processing of the instruction information P to the terminal device 10.

[0056] 8 is a block diagram of the function of constructing the generation model M in the answer generation system 20. The control device 21 executes a program stored in the storage device 22 to realize the generation model M, a training data acquisition unit 51, and a training processing unit 52. The generation model M generates answer information R according to instruction information P, as exemplified in each of the above-mentioned forms.

[0057] A plurality of basic training data Za are stored in the storage device 22. Each of the plurality of basic training data Za is composed of a combination of training instruction information Pt and training answer information Rt. The instruction information Pt includes question information Qt and auxiliary information Xt. The basic training data Za generated by the control device 21 may be stored in the storage device 22, or the basic training data Za generated by an external device may be stored in the storage device 22.

[0058] A portion of the multiple basic training data Za is generated, for example, from existing stored data held by the exhibition facility 30. The stored data is a database that collects information on each exhibit 31 in the exhibition facility 30. Specifically, the stored data includes an explanation and a position (position information X1) for each exhibit 31 in the exhibition facility 30. The question information Qt and answer information Rt of the instruction information Pt are generated from the explanation of each exhibit 31 registered in the stored data. Furthermore, the auxiliary information X (position information X1) of the instruction information Pt is generated from the position of each exhibit 31 registered in the stored data.

[0059] Furthermore, as illustrated in FIG. 8 , some of the multiple basic training data Za are generated from a history of a real guide U1 guiding a visitor U2 within the exhibition facility 30. The guide U1 guides the visitor U2 while carrying an information device 32 (e.g., a smartphone) including a sound collection device 321 and a position detection device 322. The sound collection device 321 collects audio of questions posed by the visitor U2 to the guide U1 and audio of answers from the guide U1 to those questions. Basic training data Za is generated for each pair of a question from the visitor U2 and an answer from the guide U1. Specifically, the question information Qt of the instruction information Pt is generated from the audio of the question collected by the sound collection device 321, and the answer information Rt is generated from the audio of the answer collected by the sound collection device 321. The auxiliary information X (position information X1) of the instruction information Pt is information generated by the position detection device 322 of the information device 32 at the time of the question by the visitor U2 or the answer by the guide U1.

[0060] The training data acquisition unit 51 acquires multiple pieces of basic training data Za from the storage device 22. The training processing unit 52 establishes a generative model M through machine learning using the multiple pieces of basic training data Za. Specifically, the training processing unit 52 trains the generative model M so that answer information Rt of the basic training data Za is generated in response to instruction information Pt of the basic training data Za.

[0061] 9 is a flowchart of a process (hereinafter referred to as the "basic training process") in which the control device 21 establishes a generative model M through machine learning using a plurality of basic training data Za. For example, the basic training process is started in response to an instruction from the administrator of the answer generation system 20.

[0062] When the basic training process begins, the control device 21 (training data acquisition unit 51) acquires one of multiple basic training data Za (hereinafter referred to as "selected data Z") from the storage device 22 (Sb1). The control device 21 (training processing unit 52) ​​generates answer information R by processing instruction information Pt of the selected data Z using an initial or provisional generative model M (hereinafter referred to as "provisional model M"), as illustrated in FIG. 8 (Sb2). The control device 21 (training processing unit 52) ​​calculates a loss function representing the error between the answer information R generated by the provisional model M and the answer information Rt of the selected data Z (Sb3). The control device 21 (training processing unit 52) ​​updates multiple variables of the provisional model M (e.g., weights, biases, number of layers, etc.) so as to reduce (ideally minimize) the loss function (Sb4).

[0063] The control device 21 (training processing unit 52) ​​determines whether a predetermined termination condition is met (Sb5). The termination condition may be, for example, that a predetermined number of basic training data Za have been processed, that the loss function falls below a predetermined threshold, or that the amount of change in the loss function falls below a predetermined threshold. If the termination condition is not met (Sb5: NO), the control device 21 selects unselected basic training data Za stored in the storage device 22 as new selected data Z (Sb1). That is, the process of updating multiple variables of the provisional model M0 (Sb2 to Sb4) is repeated until the termination condition is met (Sb5: YES).

[0064] If the termination condition is met (Sb5: YES), the control device 21 terminates the basic training process. The provisional model M0 at the time the termination condition is met is confirmed as the trained generative model M. As can be understood from the above explanation, the generative model M learns the latent relationship between the instruction information Pt and the answer information Rt in multiple basic training data Za. Therefore, the trained generative model M outputs statistically valid answer information R for unknown instruction information P based on the above relationship.

[0065] G: Seventh Embodiment FIG. 10 is a block diagram illustrating the functional configuration of an answer generation system 20 in the seventh embodiment. The control device 21 functions as a training data acquisition unit 51 and a training processing unit 52, as in the sixth embodiment. The training data acquisition unit 51 and the training processing unit 52 establish a generative model M through basic training processing using multiple sets of basic training data Za, as in the sixth embodiment. In addition to the basic training processing described above, in the seventh embodiment, the trained generative model M is subsequently updated through additional training using multiple sets of additional training data Zb (hereinafter referred to as "additional training processing"). The additional training processing is additional learning (e.g., fine tuning) that additionally adjusts multiple variables of the trained generative model M.

[0066] Each additional training data Zb is composed of a combination of training instruction information Pt and training response information Rt, similar to the basic training data Za. The additional training data Zb is training data generated after training the generative model M through the basic training process. Specifically, multiple additional training data Zb are generated in parallel with the actual use of the trained generative model M.

[0067] For example, in an exhibition facility 30, a guide service is provided in which a real guide U1 guides a visitor U2, in parallel with the provision of a guide service that provides answer information R to a terminal device 10 using a trained generative model M. As described above in the sixth embodiment, the guide U1 guides the visitor U2 while carrying an information device 32 including a sound collection device 321 and a position detection device 322. The sound collection device 321 collects the audio of the question posed by the visitor U2 to the guide U1 and the audio of the answer from the guide U1 to the question. Question information Qt in the additional training data Zb is generated from the audio of the question posed by the visitor U2 collected by the sound collection device 321, and answer information Rt in the additional training data Zb is generated from the audio of the answer from the guide U1 collected by the sound collection device 321. Furthermore, the auxiliary information X (position information X1) of the instruction information Pt is information generated by the position detection device 322 of the information device 32 at the time of the question by the visitor U2 or the answer by the guide U1. As can be understood from the above explanation, the additional training data Zb represents the question by the visitor U2 and the answer from the guide U1 to the question. The visitor U2 is an example of a "questioner," and the guide U1 is an example of an "answerer."

[0068] The training data acquisition unit 51 of the seventh embodiment acquires the additional training data Zb described above. Specifically, the training data acquisition unit 51 generates the additional training data Zb from the sound collection results by the sound collection device 321 and the detection results by the position detection device 322, and stores the additional training data Zb in the storage device 22. The training processing unit 52 additionally trains the generative model M by additional training processing using the multiple pieces of additional training data Zb stored in the storage device 22.

[0069] The specific procedure of the additional training process is the same as that of the basic training process illustrated in Fig. 9. That is, multiple variables of the generative model M are iteratively updated so that the loss function between the answer information R generated by the generative model M from the instruction information Pt of each additional training data Zb and the answer information Rt of the additional training data Zb is reduced (ideally minimized).

[0070] As described above, according to the seventh embodiment, it is possible to make the generative model M learn the tendencies of the conversation that actually occurs between the guide U1 and the visitor U2 in the exhibition facility 30. Therefore, the generative model M can generate response information R that is statistically valid in response to instruction information P based on the underlying tendencies in the actual conversation.

[0071] H: Eighth Embodiment In the seventh embodiment, the additional training data Zb was generated from the results of recording the conversation between the guide U1 and the visitor U2. However, the conversation between the guide U1 and the visitor U2 also includes conversations (e.g., casual conversations) that are unrelated to the exhibits 31 in the exhibition facility 30. In the eighth embodiment, the additional training data Zb is generated from only the appropriate conversations related to the exhibits 31 among the conversations between the guide U1 and the visitor U2.

[0072] 11 is a block diagram illustrating the functional configuration of the answer generation system 20 according to the eighth embodiment. The training data acquisition unit 51 according to the eighth embodiment acquires a plurality of pieces of material data Y. Each piece of material data Y is composed of a combination of instruction information Pt and answer information Rt, similar to the additional training data Zb. The plurality of pieces of material data Y are generated in parallel with the actual use of the trained generative model M.

[0073] Specifically, question information Qt in the material data Y is generated from the voice of the question posed by visitor U2 picked up by the sound collection device 321, and answer information Rt in the material data Y is generated from the voice of the answer posed by guide U1 picked up by the sound collection device 321. In addition, auxiliary information X (position information X1) in the instruction information Pt is information generated by the position detection device 322. As described above, the material data Y represents the question posed by visitor U2 and the answer given by guide U1 to the question.

[0074] The training data acquisition unit 51 determines the appropriateness of each of the multiple pieces of material data Y and selects the material data Y determined to be appropriate from the multiple pieces of material data Y as additional training data Zb. For example, existing basic training data Za is guaranteed to be information related to the exhibits 31 in the exhibition facility 30. Therefore, material data Y correlated with the basic training data Za can be evaluated as appropriate information related to each exhibit 31 in the exhibition facility 30. Taking the above into consideration, the training data acquisition unit 51 of the eighth embodiment determines the appropriateness of the material data Y by comparing the material data Y with each piece of basic training data Za. That is, the training data acquisition unit 51 selects material data Y correlated with the basic training data Za as appropriate information and selects it as additional training data Zb, and determines material data Y not correlated with the basic training data Za as inappropriate information and discards it. As described above, additional training data Zb is selected by data cleansing of the multiple pieces of material data Y.

[0075] 12 is a flowchart of a process (hereinafter referred to as the "selection process") in which the control device 21 (training data acquisition unit 51) of the eighth embodiment selects material data Y. For example, the selection process is performed for each conversation between a guide U1 and a visitor U2.

[0076] When the selection process starts, the control device 21 (training data acquisition unit 51) acquires material data Y (Sc1). For each of the multiple basic training data Za stored in the storage device 22, the control device 21 (training data acquisition unit 51) calculates a correlation index between the material data Y and the basic training data Za (Sc2). That is, the control device 21 compares the material data Y with each basic training data Za. The correlation index is an index related to the degree of correlation between the material data Y and the basic training data Za. To calculate the correlation index, any known analysis technique for evaluating the correlation between sentences may be used, for example.

[0077] The control device 21 (training data acquisition unit 51) determines whether the correlation index between any of the multiple basic training data Za and the material data Y exceeds a predetermined threshold (Sc3). If the correlation index exceeds the threshold (Sc3: YES), the material data Y can be evaluated as being correlated with the basic training data Za. In other words, the material data Y is appropriate information related to the exhibit 31 in the exhibition facility 30. Therefore, the control device 21 (training data acquisition unit 51) stores the material data Y as additional training data Zb in the storage device 22 (Sc4).

[0078] On the other hand, if the correlation index is below the threshold (Sc3: NO), it can be determined that the material data Y is not correlated with the basic training data Za. In other words, the material data Y is information that has little relevance to the exhibits 31 in the exhibition facility 30, such as information about a conversation between the guide U1 and the visitor U2. Therefore, the control device 21 (training data acquisition unit 51) discards the material data Y without using it as additional training data Zb (Sc5).

[0079] The procedure for the additional training process using the additional training data Zb is the same as in the seventh embodiment. Therefore, the same effects as in the seventh embodiment are achieved in the eighth embodiment. Furthermore, in the eighth embodiment, the appropriateness of the material data Y is determined, and material data Y that is not appropriate for generating answers to questions from the user U is excluded from the additional training data Zb. Therefore, compared to a form in which the material data Y of all conversations that occur between the guide U1 and the visitor U2 are uniformly applied to the additional training process of the generation model M, the generation model M after the additional training process can generate appropriate answers to questions from the user U.

[0080] Furthermore, the plurality of basic training data Za applied to the basic training process of the generation model M is also used to determine the appropriateness of the material data Y. Therefore, the appropriateness of the material data Y can be determined with high accuracy without preparing special data dedicated to determining the appropriateness of the material data Y.

[0081] I: Modifications Specific modifications that can be added to the above-mentioned embodiments are exemplified below. Two or more embodiments arbitrarily selected from the following examples may be combined as appropriate within the scope of not mutually contradicting each other.

[0082] (1) In each of the above-described embodiments, the instruction information P including the auxiliary information X is processed by the generative model M to generate answer information R corresponding to the auxiliary information X. However, the configuration and method for reflecting the auxiliary information X in the answer information R are not limited to the above examples. The embodiment illustrated in FIG. 13 is a embodiment in which the instruction information P does not include the auxiliary information X. The information generator 41 generates instruction information P including question information Q. The generative model M generates answer information R regardless of the auxiliary information X.

[0083] The information acquisition unit 42 adjusts the answer information R in accordance with the auxiliary information X. For example, the information acquisition unit 42 generates answer information R in accordance with the attribute information X2 by executing a processing process to which the attribute information X2 is applied on the answer information R. Similarly, the information acquisition unit 42 may generate answer information R in accordance with the history information X3 by executing a processing process to which the history information X3 is applied on the answer information R. For example, the information acquisition unit 42 changes the content or tone of the answer information R in accordance with the attribute information X2 or the history information X3.

[0084] As can be understood from the above example, the answer information R is an answer that guides the user U to the guidance target (e.g., the exhibit 31) in response to the question expressed by the question information Q, and is comprehensively expressed as information that represents an answer corresponding to the auxiliary information X related to the user U. The instruction information P may or may not include the auxiliary information X. Furthermore, part of the auxiliary information X (X1 to X3) may be included in the instruction information P and applied to the generation of the answer information R by the generative model M, and another part of the auxiliary information X may be applied to the adjustment of the answer information R by the information acquisition unit 42. For example, the instruction information P may include the question information Q and the position information X1, and the information acquisition unit 42 may adjust the answer information R according to the attribute information X2 and the history information X3.

[0085] (2) The information included in the auxiliary information X is not limited to the information exemplified in the above-described embodiments (location information X1, attribute information X2, history information X3). For example, information specifying various conditions related to the answer information R (hereinafter referred to as "condition information") may be included in the auxiliary information X.

[0086] The condition information specifies, for example, the length of the answer information R. For example, the total number of sentences or the number of characters that make up the answer information R, the playback time of the audio when the answer information R is output by voice, etc. are included as condition information in the auxiliary information X. Further, information indicating the level of the length of the answer information R, such as long / normal / short, is also exemplified as condition information. The generation model M generates answer information R under the conditions specified by the condition information. For example, answer information R of the length specified by the condition information is generated.

[0087] The condition information may be set in response to an instruction from the user U via the operation device 14, or may be information that is fixedly set in advance. The condition information may also be included in the instruction information P as information separate from the auxiliary information X.

[0088] (3) In the second embodiment, answer information R with a tone corresponding to attribute information X2 is generated, but the auxiliary information X applied to control the tone of answer information R is not limited to attribute information X2. For example, answer information R with a tone corresponding to location information X1 or history information X3 may be generated by the generation model M.

[0089] (4) In the fifth embodiment, the guide character C may be selected from a plurality of candidates. For example, the guide character C selected by the user U through an operation on the operation device 14 from among a plurality of guide characters C prepared in advance is displayed on the display device 17. The sound quality of the response voice V may differ for each guide character C. That is, the information output unit 43 generates the audio signal A with a voice quality corresponding to the guide character C selected by the user U.

[0090] In a configuration in which one of a plurality of types of guide characters C is selected, auxiliary information X representing the characteristics (e.g., appearance or personality) of the guide character C may be included in the instruction information P. The generation model M processes the instruction information P to generate answer information R corresponding to the characteristics of the guide character C. For example, answer information R with a tone of voice corresponding to the characteristics of the guide character C is generated.

[0091] (5) In the fifth embodiment, the mouth shape of the guide character C is controlled in accordance with the phonemes or volume of the answer voice V. However, the method of controlling the guide character C is not limited to the above example. For example, the information output unit 43 may cause the guide character C to move in accordance with the content of the answer information R. Specifically, the information output unit 43 causes the guide character C to move in accordance with an impression estimated from the content of the answer information R. For example, the guide character C performs different actions when the answer information R is an answer with a happy impression and when the answer information R is an answer with a sad impression.

[0092] (6) In the above-described embodiments, the position detection device 16 generates the position information X1 using a satellite positioning system. However, the method by which the terminal device 10 acquires the position information X1 is not limited to the above examples. For example, the position information X1 may be generated from the location of a wireless base station with which the terminal device 10 communicates wirelessly. The wireless base station is, for example, a wireless device used for mobile communication or Wi-Fi (registered trademark). An identification code (e.g., a QR code (registered trademark)) representing the position information X1 may be installed in the exhibition facility 30, and the terminal device 10 may generate the position information X1 by reading the identification code. Alternatively, the terminal device 10 may receive the position information X1 transmitted by short-range wireless communication from a transmitter installed in the exhibition facility 30. An example of short-range wireless communication is infrared communication. Another example of short-range wireless communication is acoustic communication using sound waves emitted by a sound-emitting device as a transmission medium. In acoustic communication, the sound-emitting device functions as the transmitter. Furthermore, the exhibit 31 may be estimated from an image of the exhibit 31 captured by an imaging device (not shown) of the terminal device 10 .

[0093] (7) In each of the above-described embodiments, guidance to the exhibition facility 30 is assumed, but the guidance target is not limited to the exhibition facility 30. For example, a specific location or operation status of a transportation facility such as a railway, airline, or ship is guided to the user U using the same configuration and method as in each of the above-described embodiments. For example, ticket vending machines, boarding locations, entrances and exits at a railway facility are exemplified as guidance targets. Furthermore, for example, a sales location for a specific item at a commercial facility such as a retail store or entertainment facility may be guided to the user U using the same configuration and method as in each of the above-described embodiments. As shown in the above examples, the present disclosure is applicable to providing the user U with information regarding any guidance target.

[0094] (8) In the above-described embodiments, the answer generation system 20 separate from the terminal device 10 generates the answer information R. However, the terminal device 10 may generate the answer information R. For example, the terminal device 10 may be equipped with a generative model M. The information acquisition unit 42 generates the answer information R by processing the instruction information P generated by the information generation unit 41 using the generative model M. As can be understood from the above explanation, the information acquisition unit 42 is comprehensively expressed as an element that acquires the answer information R. The "acquisition" of the answer information R includes not only the "reception" of the answer information R but also the "generation" of the answer information R.

[0095] Furthermore, as can be understood from the above description, the “information processing system” in the present disclosure may be the terminal device 10 alone in each of the above-described forms, or the entire information system 100 including the terminal device 10 and the answer generation system 20. The answer generation system 20 may also be expressed as a machine learning system that trains the generation model M through basic training processing or additional training processing.

[0096] (9) The functions of the terminal device 10 according to each of the above-described embodiments are realized by cooperation between one or more processors constituting the control device 11 and a program stored in the storage device 12. Similarly, the functions of the answer generation system 20 according to each of the above-described embodiments are realized by cooperation between one or more processors constituting the control device 21 and a program stored in the storage device 22.

[0097] The programs according to the above embodiments can be provided in a form stored on a computer-readable recording medium and installed on a computer. The recording medium is, for example, a non-transitory recording medium, such as an optical recording medium (optical disk) such as a CD-ROM, but also includes any known type of recording medium, such as a semiconductor recording medium or a magnetic recording medium. Note that a non-transitory recording medium includes any recording medium other than a transitory, propagating signal, and does not exclude volatile recording media. In addition, in a configuration in which a distribution device distributes a program via a communication network, the recording medium storing the program in the distribution device corresponds to the non-transitory recording medium.

[0098] J: Supplementary Note From the above-described exemplary embodiments, the following configurations can be understood, for example.

[0099] An information processing method according to one aspect (aspect 1) of the present disclosure generates instruction information including question information representing a question from a user regarding a guidance target, obtains answer information corresponding to a result of processing the instruction information by a trained generative model, the answer being an answer that guides the user to the guidance target in response to the question represented by the question information and representing an answer corresponding to auxiliary information related to the user, and outputs the answer information. According to the above aspect, answer information that reflects not only the question information representing the question posed by the user but also the auxiliary information related to the user is output. Therefore, compared to a configuration in which a guidance voice corresponding to identification information input by the user is simply played, a variety of information desired by the user can be appropriately provided.

[0100] A "guide subject" is any thing (item or matter) that is the object of a user's question and the guidance provided in response to that question. Examples of "guide subjects" include exhibits at exhibition facilities such as museums or art galleries, specific locations or operational status of transportation facilities such as railways, airlines, or ships, and sales locations of specific items at commercial facilities. Note that a "guide subject" can also be expressed as a thing that is an element that constitutes a user's question.

[0101] "Question information" is information in any format that represents a question posed by a user. For example, the question information may be a character string of a question entered by a user. However, the question information may also include an image representing the question posed by the user or a voice of the user pronouncing the question.

[0102] "Auxiliary information" refers to various types of information related to a user. Specifically, "auxiliary information" is information other than the question that is added to the question information in order to obtain an appropriate answer to the question from the user. In other words, "auxiliary information" is expressed as information other than the question that supports the question information. "auxiliary information" can also be described as information that clarifies the content of the question from the user when referenced together with the question information. For example, location information indicating the location of the user, attribute information indicating the attributes of the user, or history information indicating the behavioral history of the user are examples of "auxiliary information." As can be understood from the above examples, "auxiliary information" is, for example, information that is acquired separately from the question information. "auxiliary information" can also be expressed as information that includes matters not included in the question information (i.e., the question from the user).

[0103] "Instruction information" is a prompt to be input to the trained generative model, and is generated in response to question information representing a question posed by a user. However, "instruction information" is not limited to information that directly includes question information. For example, information obtained by editing question information may be included in "instruction information." Furthermore, information obtained by converting the information format of instruction information or information obtained by converting the information format of auxiliary information may be included in "instruction information." The information format refers to the representation format of information, and includes, for example, formats such as character strings, images, or audio. Note that auxiliary information may be included in instruction information, or may be information separate from instruction information.

[0104] A "trained generative model" is a probabilistic model that generates output information in response to instruction information. The generative model learns the tendency of output information in response to instruction information through prior machine learning. Specifically, the "generative model" is a generative large-scale language model (LMM) that is trained specifically for natural language processing tasks such as answer generation. For example, a probabilistic model realized by a Transformer model that utilizes a self-attention mechanism is an example of a "generative model."

[0105] "Answer information" is information corresponding to the result (output information) of the generative model processing the instruction information. "Answer information" is expressed in any information format, such as a character string, an image, or audio. Specifically, "answer information" is an answer that guides the user to the guidance target in response to the question represented by the question information, and is information that represents an answer corresponding to the auxiliary information. Note that the generative model may generate output information corresponding to both the question information and the auxiliary information, or the answer information may be generated by editing the output information generated by the generative model from the instruction information corresponding to the question information according to the auxiliary information.

[0106] The generative model can generate answer information that directly answers a user's question (for example, information that contains little or no information unrelated to the question). That is, the possibility of generating a redundant answer that contains information unrelated to the question can be reduced. Therefore, the time required to output the answer information by voice can be shortened, or the display area can be reduced when outputting the answer information by image. However, the answer information may include items that are not directly related to the user's question, as long as the answer information does not become excessively long.

[0107] "Output of answer information" is, for example, an operation of notifying the user of the answer information. Specifically, examples of "output" include an operation of displaying a character string or an image represented by the answer information, or an operation of playing back a sound represented by the answer information.

[0108] In a specific example (Aspect 2) of Aspect 1, the auxiliary information includes location information indicating the location of the user, and the answer information indicates an answer corresponding to the location indicated by the location information. According to the above aspect, instruction information corresponding to the auxiliary information including the location information is processed by a generative model. Therefore, appropriate information corresponding to the user's location can be provided.

[0109] "Location information" refers to any type of information that can identify the location of a user. For example, "location information" includes information that directly indicates the user's location (e.g., latitude and longitude information), as well as information that indicates an object of interest to the user in the location (e.g., a museum exhibit).

[0110] Furthermore, information that can be obtained only from the location where the user is located may be used as “location information.” For example, information transmitted only to a specific location, or information that can be optically read from an image (e.g., a QR code (registered trademark)) placed in a specific location, may be used as “location information” that indicates the user’s location.

[0111] In a specific example (Aspect 3) of Aspect 1 or Aspect 2, the auxiliary information includes attribute information that represents an attribute of the user, and the answer information represents an answer corresponding to the attribute represented by the attribute information. According to the above aspect, instruction information corresponding to the auxiliary information that includes attribute information is processed by a generative model. Therefore, appropriate answer information corresponding to the user's attribute can be provided to the user.

[0112] "Attribute information" is any type of information related to a user's attributes. Examples of "attribute information" include the user's age, generation, gender, place of residence (country or region), and language used. Other examples of "attribute information" include the user's level of knowledge about the target of guidance or the user's interests.

[0113] In a specific example (Aspect 4) of any of Aspects 1 to 3, the auxiliary information includes history information representing the user's behavioral history, and the answer information represents an answer corresponding to the history represented by the history information. According to the above aspect, instruction information corresponding to the auxiliary information including the history information is processed by a generative model. Therefore, appropriate information corresponding to the user's past behavioral history can be provided.

[0114] "History information" is information that represents the user's past behavior history. Examples of "history information" include places the user has visited in the past (e.g., museums or art galleries), routes the user has taken in the past, and means of transportation the user has used in the past.

[0115] In a specific example (Aspect 5) of any of Aspects 1 to 4, the answer information is expressed in natural language in a tone that corresponds to the auxiliary information. According to the above aspects, the tone of the answer information is controlled in accordance with the auxiliary information of the user, so that the answer information can be provided in a tone that is appropriate for the user (for example, a tone that is familiar to the user).

[0116] The "tone" of the answer information refers to the tone of the words expressed in the answer information. When the answer information is expressed in speech, the tone of the speech is an example of the "tone." In addition, the form of endings such as "desu" (is) or "de aru" (is).

[0117] The method for controlling the tone of the answer information is arbitrary. For example, the generative model may process instruction information including auxiliary information to generate answer information with a tone corresponding to the auxiliary information, or the output information of the generative model may be edited in accordance with the auxiliary information to generate answer information with a tone corresponding to the auxiliary information.

[0118] In a specific example (Aspect 6) of any one of Aspects 1 to 5, outputting the answer information includes playing back a voice that pronounces the answer information in a voice quality that corresponds to the auxiliary information. According to the above aspects, the voice quality of the answer information is controlled in accordance with the auxiliary information of the user, so that the answer information can be provided in a voice quality that is appropriate for the user (for example, a voice quality that is familiar to the user).

[0119] The "voice quality" of the answer information refers to the characteristics of the voice represented by the answer information. For example, the "voice quality" is affected by various voice-related elements such as the frequency characteristics (timbre) of the voice represented by the answer information, as well as the pitch, range, intonation, or rhythm of the voice.

[0120] In a specific example (Aspect 7) of any one of Aspects 1 to 6, outputting the answer information includes displaying a guide character that notifies the user of the answer information. According to the above aspects, the guide character that notifies the user of the answer information is displayed, so the answer information can be provided in a way that is easy for the user to understand and memorable.

[0121] The "guide character" is a display object that is displayed on the display device in parallel with the output of answer information. For example, a display object representing a virtual living thing (e.g., a human or an animal) existing in a virtual space, or a display object representing an inanimate object such as a robot, is exemplified as the "guide character." The "guide character" may operate in conjunction with the output of answer information. For example, it is conceivable that the facial expression (e.g., mouth shape), gestures, or hand movements of the guide character change in parallel with the playback of a sound representing the answer information.

[0122] In a specific example (Aspect 8) of any of Aspects 1 to 7, additional training data representing questions posed by a questioner and answers posed by a respondent to the questions is further acquired, and the generative model is additionally trained using the additional training data. According to the above aspect, it is possible to make the generative model learn trends in actual dialogue between a questioner and a respondent. Therefore, the generative model can generate statistically valid response information for instruction information based on trends underlying in actual dialogue.

[0123] "Additional training data" is training data generated after training a generative model. For example, the additional training data is generated or acquired in parallel with the actual use of a trained generative model. "Additional training data" is, for example, audio data representing the voice of a questioner asking a question and the voice of a respondent answering, or text data representing a question by a questioner and an answer by a respondent.

[0124] Additional training of a generative model is a process (relearning) in which a generative model trained by machine learning using existing training data (basic training data) is updated to specialize in the trends of the additional training data. For example, transfer learning (fine-tuning) is used for additional training.

[0125] In a specific example of aspect 8 (aspect 9), the generative model has been trained using a plurality of basic training data representing questions and answers, and in acquiring the additional training data, material data representing questions posed by the questioner and answers posed by the answerer is compared with each of the plurality of basic training data to determine the suitability of the material data, and the material data determined to be appropriate is selected as the additional training data.

[0126] In the above-described embodiment, the appropriateness of material data representing a dialogue between a questioner and a respondent is determined, and material data that is inappropriate for generating an answer to a user's question, such as casual conversation between the questioner and the respondent, is excluded from the additional training data. Therefore, compared to an embodiment in which material data of all dialogues occurring between the questioner and the respondent is applied to the additional training of the generative model, the generative model can generate appropriate answers to questions from users.

[0127] Furthermore, the plurality of basic training data applied to the initial or basic training of the generative model M is used to determine the appropriateness of the material data. Therefore, the appropriateness of the material data can be determined with high accuracy without preparing special data dedicated to determining the appropriateness of the material data.

[0128] The method for determining the suitability of the material data is arbitrary. For example, a conceivable form is to calculate the correlation between a set of questions and answers represented by each basic training data and a set of questions and answers represented by the material data, and select material data whose correlation exceeds a predetermined threshold as additional training data, and discard material data whose correlation is below the threshold.

[0129] An information processing system according to one aspect (aspect 10) of the present disclosure includes an information generation unit that generates instruction information including question information representing a question from a user regarding a guidance target, an information acquisition unit that acquires answer information corresponding to the result of a trained generative model processing the instruction information, the answer being an answer that guides the user to the guidance target in response to the question represented by the question information and representing an answer corresponding to auxiliary information regarding the user, and an information output unit that outputs the answer information.

[0130] A program according to one aspect (aspect 11) of the present disclosure causes a computer system to function as an information generation unit that generates instruction information including question information representing a question from a user regarding a guidance target, an information acquisition unit that acquires answer information corresponding to the result of a trained generative model processing the instruction information, the answer being an answer that guides the user to the guidance target in response to the question represented by the question information and representing an answer corresponding to auxiliary information regarding the user, and an information output unit that outputs the answer information.

[0131] 100...information system, 10...terminal device, 11...control device, 12...storage device, 13...communication device, 14...operation device, 15...sound collection device, 16...position detection device, 17...display device, 18...sound emission device, 19...output device, 20...answer generation system, 21...control device, 22...storage device, 23...communication device, 30...exhibition facility, 31...exhibit, 32...information device, 321...sound collection device, 322...position detection device, 41...information generation unit, 42...information acquisition unit, 43...information output unit, 51...training data acquisition unit, 52...training processing unit.

Claims

1. An information processing method implemented by a computer system, which generates instruction information including question information representing a question from a user regarding a guidance target, obtains answer information according to the result of a trained generative model processing the instruction information, the answer being an answer that guides the user to the guidance target in response to the question represented by the question information and representing an answer according to auxiliary information regarding the user, and outputs the answer information.

2. The information processing method according to claim 1, wherein the auxiliary information includes location information indicating the location of the user, and the answer information indicates an answer corresponding to the location indicated by the location information.

3. The information processing method according to claim 1, wherein the auxiliary information includes attribute information that represents an attribute of the user, and the answer information represents an answer according to the attribute represented by the attribute information.

4. The information processing method according to claim 1, wherein the auxiliary information includes history information representing the user's behavioral history, and the answer information represents an answer corresponding to the history represented by the history information.

5. The information processing method according to claim 1, wherein the answer information is expressed in a natural language in a tone that corresponds to the auxiliary information.

6. The information processing method according to claim 1, wherein outputting the answer information includes reproducing a voice that pronounces the answer information with a voice quality corresponding to the auxiliary information.

7. The information processing method according to claim 1, wherein outputting the answer information includes displaying a guide character that notifies the user of the answer information.

8. The information processing method according to claim 1, further comprising: acquiring additional training data representing questions posed by questioners and answers to those questions posed by answerers; and additionally training the generative model using the additional training data.

9. The information processing method of claim 8, wherein the generative model has been trained using a plurality of basic training data representing questions and answers, and the acquisition of the additional training data involves comparing material data representing questions posed by the questioner and answers posed by the answerer with each of the plurality of basic training data to determine the suitability of the material data, and selecting material data determined to be appropriate as the additional training data.

10. An information processing system comprising: an information generation unit that generates instruction information including question information that represents a question from a user regarding a guidance target; an information acquisition unit that acquires answer information corresponding to the result of a trained generative model processing the instruction information, the answer being an answer that guides the user to the guidance target in response to the question represented by the question information and representing an answer corresponding to auxiliary information regarding the user; and an information output unit that outputs the answer information.

11. A program that causes a computer system to function as: an information generation unit that generates instruction information including question information that represents a question from a user regarding a guidance target; an information acquisition unit that acquires answer information that corresponds to the result of a trained generative model processing the instruction information, which is an answer that guides the user to the guidance target in response to the question represented by the question information and which represents an answer that corresponds to auxiliary information regarding the user; and an information output unit that outputs the answer information.

Images