Response generation learning device, response generation device, response generation learning method, response generation method and program

Abstract

To improve a control property and an interpolation property regarding generation of utterance of a specific speaker's characteristic.SOLUTION: A response generation learning device includes a learning unit that trains a speaker embedding model so that the speaker embedding model, when inputted with speaker information containing a token for identifying the speaker and a text describing the speaker's characteristics, can reproduce the speaker information, wherein the learning unit also trains a response generation model to reproduce the correct response corresponding to the speaker information and dialogue history, by inputting both a latent representation generated by the speaker embedding model based on the speaker information and the history of utterances in a dialogue.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a response generation learning device, a response generation device, a response generation learning method, a response generation method, and a program. [Background technology]

[0002] Conventionally, in response generation (a dialogue system that generates a system response using a user's speech as input), research has been conducted to learn, reproduce, and reflect the characteristics of a specific speaker (Non-Patent Document 1, Non-Patent Document 2).

[0003] In the method of Non-Patent Document 1 (hereinafter referred to as "Method 1"), dialogue data is associated with a speaker ID and learning is performed to generate responses that reflect the speaker's characteristics. Specifically, the speaker ID is converted into a vector called a speaker embedding, and the characteristics are reflected in the response by reading the vector into any part of the response generation model. Dialogue data with speaker IDs is available in large quantities on the Internet, making learning easy.

[0004] In the method of Non-Patent Document 2 (hereinafter referred to as "Method 2"), in order to generate a response that reflects the characteristics of the speaker, learning is performed by associating dialogue data with text (persona sentences) that describes the characteristics of the speaker. In Method 2, the persona sentences are input in the same way as the utterances. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Jiwei Li et al., "A Persona-Based Neural Conversation Model", arXiv:1603.06155v2 [cs.CL] 8 Jun 2016 [Non-Patent Document 2] Saizheng Zhang et al., "Personalizing Dialogue Agents: I have a dog, do you have pets too?", Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Long Papers), pages 2204-2213 Melbourne, Australia, July 15 - 20, 2018 Summary of the Invention [Problem to be solved by the invention]

[0006] Method 1 has the problem of low controllability (the ability of a user / human to use natural language to instruct the system on the speaker features reflected by the system) because it learns by associating the features of each speaker with a response sentence and a speaker ID. In other words, method 1 has the problem that it is difficult to fine-tune the speaker characteristics because the feature space of the speaker ID is learned so that the features of each speaker are separated. For example, method 1 does not accept instructions or operations in natural language such as "speak politely" for any speaker.

[0007] Method 2 has high controllability because it learns the features of each speaker by associating the natural language (persona sentences) that express them with response sentences, but the cost of collecting data given persona sentences is high, so it is not possible to collect a large amount of such data. For these reasons, method 2 does not have a speaker embedding vector and has a small amount of training data, so it is difficult to generate response sentences with expressions that have interpolation properties for speaker information (the ability to generate and indicate a speaker with the characteristics of the sum of two speakers, or the characteristics of an intermediate speaker, by making it possible to add and subtract speaker features). In other words, method 2 has high controllability, but has issues with interpolation properties.

[0008] The present invention has been made in consideration of the above-mentioned points, and has an object to improve the controllability and interpolation capability regarding the generation of utterances with characteristics of a specific speaker. [Means for solving the problem]

[0009] In order to solve the above problem, the response generation training device has a learning unit configured to train a speaker embedding model to which speaker information including a token that identifies a speaker and text indicating characteristics of the speaker is input so that the speaker embedding model reproduces the speaker information, and to train a response generation model to which latent expressions generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue are input so that the response generation model reproduces correct responses to the speaker information and the history. Effect of the Invention

[0010] Greater control and interpolation of speech production of specific speaker characteristics is possible. [Brief description of the drawings]

[0011] [Figure 1] 1 is a diagram illustrating an example of a hardware configuration of a response generating device 10 according to an embodiment of the present invention. [Diagram 2] 2 is a diagram illustrating an example of a functional configuration of a response generating device 10 according to an embodiment of the present invention. [Diagram 3] FIG. 2 is a diagram illustrating an example of the configuration of a machine learning model m1 according to an embodiment of the present invention. [Figure 4] FIG. 13 is a diagram showing an example of the configuration of a speaker embedding model m12. [Diagram 5] FIG. 13 is a diagram for explaining a first specific example of speaker embedding model m12. [Figure 6] FIG. 11 is a diagram for explaining a second specific example of speaker embedding model m12. [Figure 7] FIG. 11 is a diagram showing an example of the configuration of a response generation model m11. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Fig. 1 is a diagram showing an example of a hardware configuration of a response generation device 10 in an embodiment of the present invention. The response generation device 10 in Fig. 1 has a drive device 100, an auxiliary storage device 102, a memory device 103, a processor 104, an interface device 105, and the like, which are all connected to each other via a bus B.

[0013] A program for realizing the processing in the response generation device 10 is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 storing the program is set in the drive device 100, the program is installed from the recording medium 101 via the drive device 100 into the auxiliary storage device 102. However, the program does not necessarily have to be installed from the recording medium 101, but may be downloaded from another computer via a network. The auxiliary storage device 102 stores the installed program as well as necessary files, data, etc.

[0014] When an instruction to start a program is received, the memory device 103 reads out and stores the program from the auxiliary storage device 102. The processor 104 is a CPU or a GPU (Graphics Processing Unit), or a CPU and a GPU, and executes functions related to the response generation device 10 in accordance with the program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network.

[0015] Fig. 2 is a diagram showing an example of a functional configuration of the response generation device 10 in the embodiment of the present invention. In Fig. 2, the response generation device 10 has a learning unit 11 and a response generation unit 12. Each of these units is realized by a process in which one or more programs installed in the response generation device 10 are executed by a processor 104.

[0016] The learning unit 11 executes a learning process of the machine learning model m1. The machine learning model m1 is a model such as a neural network that receives as input a history of utterances in a dialogue and information (hereinafter referred to as "speaker information") indicating the characteristics of a speaker (hereinafter referred to as "responder") who responds to the history of utterances, and generates the response. In this embodiment, the "utterances" and the "responses" are text data. The history of utterances is a time-series arrangement of text data of all or a certain length of utterances made in a dialogue. The "certain length" refers to a certain number of past utterances (more precisely, a certain number of words). This length depends on the length that the machine learning model m1 (strictly speaking, a response generation model m11 described later) can receive. For example, in the case of a dialogue between two people, the history of utterances is a time-series arrangement of alternating utterances by the two speakers. The dialogue may be conducted by three or more people. The responder may be a person who has made any utterance in the history of utterances, or a person who has not made any utterance in the history of utterances.

[0017] The response generation unit 12 receives the utterance history and the speaker information of the respondent, and generates a response to the utterance history according to the speaker characteristics related to the speaker information, using the machine learning model m1. The response to the utterance history refers to a new utterance according to the context in the utterance history, and is not necessarily limited to a response to only the last utterance.

[0018] In this embodiment, since the utterance and the response are text data, the speaker's characteristics do not include voice characteristics, but include linguistic characteristics of the way of speaking and characteristics of the content of the utterance. For example, linguistic characteristics of the way of speaking include the type of filler (hesitations such as "ah", "um", "er", etc.) and the location of the filler. In addition, linguistic characteristics of the way of speaking include tone of voice (polite language, honorific language, casual speech, dialect), characteristics of the way of expressing the first person and the third person, and choice of words (which expression do you tend to choose among various expressions such as "difficult", "difficult", and "not easy" even if they have the same meaning). Characteristics of the content of the utterance refer to the characteristics of the entire response sentence, such as whether to answer "I like interesting things! I'm a hardcore Osakan!" when asked "What kind of person are you?" or to answer "I was born and raised in Asakusa, and I'm a true Edokko".

[0019] The learning unit 11 and the response generating unit 12 may be realized by using different computers.

[0020] Fig. 3 is a diagram showing an example of a configuration of a machine learning model m1 in the embodiment of the present invention. In Fig. 3, the machine learning model m1 includes a speaker embedding model m12 and a response generation model m11.

[0021] The speaker embedding model m12 is a model that takes the input 2, which is speaker information, as input, generates a speaker embedding vector v1 in the latent representation, and restores the speaker information from the speaker embedding vector v1. Therefore, the output 2 from the speaker embedding model m12 is the restored speaker information. The generated speaker embedding vector v1 is input to the response generation model m11. The speaker information is either or both of a speaker ID (a token for identifying a speaker, such as [ID-123]) and a persona sentence (a set of sentences, such as I'm from Osaka, I like funny people, and I like takoyaki). A token is the smallest unit when a sentence is divided into segments like words in Japanese. A persona sentence is a text that describes the characteristics of a speaker. In this embodiment, a token enclosed in [ ] means a token that has a special role, such as a speaker ID, a sentence end, or a blank filler.

[0022] The response generation model m11 is a model that inputs the utterance history as input 1, receives the speaker embedding vector v1 generated by the speaker embedding model m12 in a hidden layer, and generates a response as output 1. The response is a response to the input 1 and includes the features of the speaker related to the speaker embedding vector v1.

[0023] FIG. 4 is a diagram showing an example of the configuration of the speaker embedding model m12. In terms of performance, the speaker embedding model m12 is preferably a VAE (Variational Auto-Encoder) or a model with a structure similar thereto (for example, a simple auto-encoder). However, other models may be used as the speaker embedding model m12. The VAE has a function of improving the amount of information in the latent representation (embedding vector v1) generated by the model's encoder (preventing input information from being lost) by restoring the input and output to have similar contents, which is useful for improving the expressive ability of the speaker embedding vector v1 and improving the interpolation ability.

[0024] Figure 4 shows an example in which an encoder constituting a VAE generates a speaker embedding vector v1 from speaker information as input 2, and a decoder restores the speaker information from the speaker embedding vector v1 and outputs the speaker information as output 2.

[0025] Fig. 5 is a diagram for explaining a first specific example of the speaker embedding model m12. Fig. 5 shows the speaker embedding model m12 using the trained BERT as an encoder and the Transformer Decoder as a decoder.

[0026] BERT is one of the pre-trained models commonly used in natural language processing. It is known that the vectors corresponding to [CLS] tokens output by BERT strongly reflect the characteristics of the sentences input to BERT.

[0027] The Transformer is one of the neural network models used in BERT. A pre-trained Transformer Decoder is also publicly available and is often used in language generation research.

[0028] 5, BERT, which is an encoder, can input a speaker ID or a persona sentence, or both, as input 2 (speaker information). A specific format of the speaker information is, for example, as follows.

[0029] When the speaker information includes only a speaker ID, the speaker information has a format such as, for example, [CLS][ID-123][ / S][PAD][PAD][PAD].

[0030] When the speaker information contains only persona sentences, the speaker information will be in the format, for example, [CLS]I'm from Osaka[ / S][PAD][PAD][PAD].

[0031] If the speaker information includes both a speaker ID and a persona sentence, the speaker information will be in the format of [CLS][ID-123]I'm from Osaka[ / S][PAD][PAD][PAD]. In other words, you can simply connect the speaker ID and the persona sentence.

[0032] Even if the speaker information is in a format that includes both a speaker ID and a persona sentence, it is possible to express speaker information that includes only a speaker ID and speaker information that includes only a persona sentence. In this case, the one that is not included (persona sentence or speaker ID) is considered to be missing, and the missing part may be filled with a token that indicates unknown vocabulary (unknown word), such as [UNK].

[0033] For example, speaker information that includes only the speaker ID will look like this: [CLS][ID-123][UNK][ / S][PAD][PAD][PAD]. Speaker information that includes only the persona sentence will look like this: [CLS][UNK]I'm from Osaka[ / S][PAD][PAD][PAD].

[0034] The [CLS] token is required when using BERT, [ / S] is the end-of-sentence symbol, and [PAD] is padding. The padding corresponds to the part where the length of input 2 is insufficient compared to the length that can be input to the speaker embedding model m12 (the encoder of the speaker embedding model).

[0035] As an encoder, BERT processes the input as above and outputs a vector corresponding to the [CLS] token. This vector becomes the speaker embedding vector v1.

[0036] The Transformer Decoder recovers speaker information corresponding to input 2 from the speaker embedding vector v1, and outputs the recovery result as output 2. Therefore, if input 2 is a speaker ID, the Transformer Decoder recovers the speaker ID; if it is a persona sentence, the Transformer Decoder recovers the persona sentence; if it is both, the Transformer Decoder recovers both. However, the [CLS] token specific to BERT processing does not need to be recovered. Specifically, if input 2 is [CLS][ID-123][ / S][PAD][PAD][PAD], output 2 may be [ID-123][ / S][PAD][PAD][PAD].

[0037] Fig. 6 is a diagram for explaining a second specific example of the speaker embedding model m12. Fig. 6 shows a configuration example of the speaker embedding model m12 when the input of the speaker ID and the persona sentence are clearly distinguished. Input 2-1 is an input of the speaker ID only, and input 2-2 is an input of the persona sentence only.

[0038] The speaker embedding model m12 may input either input 2-1 or input 2-2 and calculate the corresponding speaker embedding vector v1, or it may input both, calculate the corresponding speaker embedding for each input 2-1 and input 2-2, and then combine them into a single speaker embedding vector v1 using techniques such as addition, averaging, max pooling, etc.

[0039] In the input 2-1, embedding (embedding layer) may be used as a method of converting the speaker ID into a speaker embedding. In the output 2-1 corresponding to the input 2-1, the speaker ID is restored from the speaker embedding vector v1. Linear combination or class classification may be used to restore the speaker ID.

[0040] The calculation of the speaker embedding vector v1 from the input 2-2 and the restoration of the persona sentence (output 2-2) from the speaker embedding vector v1 are as described in FIG.

[0041] Fig. 7 is a diagram showing a configuration example of the response generation model m11. For the response generation model m11, a model similar to an Encoder-Decoder model or a Decoder model is suitable. Fig. 7 shows the response generation model m11 as an Encoder-Decoder model. For example, the response generation model m11 may be configured using a Transformer Encoder-Decoder.

[0042] In FIG. 7, the response generation model m11 inputs the utterances made so far as input 1, receives the speaker embedding vector v1 of the speaker corresponding to output 1 from the speaker embedding model m12, and generates a response as output 1.

[0043] The speaker embedding vector v1 received from the speaker embedding model m12 may be combined or added to a vector sequence output by an encoder (e.g., a Transformer Encoder) that has received the previous utterance as input 1. In this case, the result of the combination or addition is input to a decoder (e.g., a Transformer Decoder). The decoder generates a response (output 1) based on the input vector. Alternatively, the speaker embedding vector v1 may be vector combined or added in any hidden layer of the decoder that has received the vector sequence output by the encoder.

[0044] [Study time] The learning unit 11 controls the learning of the machine learning model m1. The learning unit 11 executes the learning process using a plurality of learning data, each of which is a set of a speech history, speaker information, and a response as a correct answer (hereinafter referred to as a "correct response") in one dialogue. A correct response of certain learning data is a response to the speech history of the learning data, and is a response that includes features according to the speaker information of the learning data. The speaker information of at least a portion of the learning data includes both a speaker ID and a persona sentence. In other words, the speaker information of a portion of the learning data may be only a speaker ID or only a persona sentence.

[0045] The learning unit 11 inputs the speech history in the learning data as input 1 to the machine learning model m1, and inputs the speaker information in the learning data as input 2 to the machine learning model m1.

[0046] More specifically, the training unit 11 trains the speaker embedding model m12 (updates the parameters of the speaker embedding model m12) for each training data so that when speaker information of the training data is input as input 2, output 2 reproduces the speaker information (so that the error between input 2 and output 2 becomes zero). At this time, in order to improve the interpolation property of the speaker embedding vector v1, the training unit 11 may impose a loss function such as KL divergence on the speaker embedding vector v1 as a latent representation in the speaker embedding model m12 to which speaker information is input as input 2, thereby imposing a constraint such that the speaker embedding vector v1 as a latent representation follows a predetermined probability distribution such as a normal distribution.

[0047] The learning unit 11 also performs supervised learning (updates the parameters of the response generation model m11 and the speaker embedding model m12) for each piece of learning data such that, when the utterance history of the learning data is input as input 1 and the speaker information of the learning data is input as input 2, the output 1 reproduces the correct response of the learning data (so that the error between the output 1 calculated using a loss function and the correct response becomes 0). In other words, the error between the output 1 and the correct response affects the parameters of both the response generation model m11 and the speaker embedding model m12.

[0048] In addition, the training of the speaker embedding model m12 and the response generation model m11 may be performed simultaneously (both models may be trained simultaneously so that the error in output 1 and the error in output 2 are simultaneously set to zero in one processing), or only the speaker embedding model m12 may be trained in advance and then only the response generation model m11 may be trained, or the response generation model m11 and the speaker embedding model m12 may be trained alternately (the speaker embedding model m12 and the response generation model m11 are trained so that either the error in output 1 or the error in output 2 is set to zero in one processing).

[0049] That is, the speaker embedding model m12 in Figures 4, 5, and 6 can also be trained using only the speaker embedding model m12. In this case, the training unit 11 trains the speaker embedding model m12 by considering only the loss function of output 2 for input 2, without considering the loss function of output 1. By training only the speaker embedding model m12 in advance and then performing training taking into account the loss function of output 1, it is expected that the performance and training efficiency of the entire machine learning model m1 can be improved.

[0050] Also, training may be performed only on the response generation model m11. In that case, the training unit 11 trains the response generation model m11 by considering only the loss function of output 1 for input 1. At this time, the speaker embedding vector v1 input to the response generation model m11 may be calculated in advance, or may be a vector of all 0 or random values. By training the entire machine learning model m1 after training only the response generation model m11, it is expected that the performance and training efficiency of the entire machine learning model m1 can be improved.

[0051] The learning may be performed based on a general or well-known learning method in machine learning, and the machine learning model m1 may convert the data given as input, and the learning may be performed so that the output data is close to an ideal output according to a predetermined purpose. In the above, the ideal output is given in the form of a correct response to speaker information and a speech history.

[0052] [At the time of inference] The machine learning model m1 (speaker embedding model m12 and response generation model m11) in the following description is a trained model.

[0053] For example, when a user inputs an utterance history and speaker information, the response generator 12 generates a response to the utterance history based on information output from the machine learning model m1 (response generation model m11). The speaker information may include only a speaker ID, only a persona sentence, or both a speaker ID and a persona sentence.

[0054] The response generation unit 12 inputs the utterance history as input 1 and the speaker information as input 2 to the machine learning model m1. At this time, the speaker embedding model m12 does not need to output output 2. That is, a decoder for the speaker embedding model m12 is not essential during inference. However, it is acceptable to check whether the correct speaker has been reproduced by outputting output 2.

[0055] The machine learning model m1 (response generation model m11) outputs, as an output 1, a response corresponding to the speaker information and the speech history.

[0056] Strictly speaking, the output from the neural network as the response generation model m11 is not the text data itself indicating the response, but a probability distribution of the response, etc. Therefore, the response generation unit 12 generates text data indicating the response by applying an output (decoding) method used in general natural language processing models, such as beam search, to the output from the response.

[0057] It is possible to perform addition and average calculations using multiple speaker embedding vectors v1 generated by the speaker embedding model m12 in FIG. 4, FIG. 5, or FIG.

[0058] For example, let vector va be the speaker embedding vector v1 generated by speaker embedding model m12 to which the speaker ID of speaker A is input as speaker information of input 2. Let vector vb be the speaker embedding vector v1 generated by speaker embedding model m12 to which the persona sentence "I am kind" is input as speaker information of input 2. By adding or averaging vectors va and vb, a speaker embedding vector v1 having the characteristics of both "kind" speaker A can be generated. Furthermore, by adding or averaging speaker embedding vectors v1 based on one or more speaker IDs or persona sentences, further features can be included in speaker embedding vector v1.

[0059] Therefore, the response generation unit 12 may input multiple pieces of speaker information from a user and input each piece individually to the speaker embedding model m12. In this case, the speaker embedding model m12 generates multiple speaker embedding vectors v1. The response generation model m11 may receive a result of adding or averaging the multiple speaker embedding vectors v1 for one input of the utterance history to generate a response.

[0060] As described above, according to the present embodiment, for speaker embedding, which has conventionally been trained by maximizing the likelihood when generating a response, a speaker embedding model m12 capable of receiving both a speaker ID and a persona sentence is used, and by restoring the speaker ID and persona sentence using, for example, a VAE (structure, etc.), it is possible to obtain controllability that allows the use of persona sentences and interpolation of speaker embedding by VAE. Therefore, when learning, reproducing, and reflecting the characteristics of a specific speaker, it is possible to improve the controllability and interpolation regarding the generation of utterances with the characteristics of a specific speaker.

[0061] The following supplementary notes are further disclosed regarding the above embodiment.

[0062] (Additional note 1) Memory, at least one processor coupled to the memory; Including, The processor, A speaker embedding model is trained to receive speaker information including a token for identifying a speaker and a text showing characteristics of the speaker, so as to reproduce the speaker information; and a response generation model is trained to receive latent expressions generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue, so as to reproduce a correct response to the speaker information and the history. A response generation learning device comprising:

[0063] (Additional note 2) Memory, at least one processor coupled to the memory; Including, The processor, A speaker embedding model is input with speaker information including a token for identifying a speaker and a text showing characteristics of the speaker, and the speaker embedding model is trained to reproduce the speaker information. A response generation model is input with a latent expression generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue, and the response generation model is trained to reproduce a correct response to the speaker information and the history. speaker information is input to the speaker embedding model, and when a history of utterances in a dialogue is input to the response generation model, a response to the history of the utterances is generated based on information output by the response generation model using a latent representation of the speaker embedding model; A response generating device comprising:

[0064] (Additional note 3) a learning procedure configured to train a speaker embedding model to which speaker information including a token for identifying a speaker and a text showing characteristics of the speaker is input so as to reproduce the speaker information, and to train a response generation model to which latent expressions generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue are input so as to reproduce a correct response to the speaker information and the history; A recording medium on which a program for causing a computer to execute the above is recorded.

[0065] (Additional note 4) A speaker embedding model is input with speaker information including a token for identifying a speaker and a text showing characteristics of the speaker, and the speaker embedding model is trained to reproduce the speaker information. A response generation model is input with a latent expression generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue, and the response generation model is trained to reproduce a correct response to the speaker information and the history. a response generation step of inputting speaker information into the speaker embedding model, and generating a response to the utterance history based on information output by the response generation model using a latent representation of the speaker embedding model when a history of utterances in a dialogue is input into the response generation model; A recording medium on which a program for causing a computer to execute the above is recorded.

[0066] In this embodiment, the response generation device 10 during learning is an example of a response generation learning device.

[0067] Although the embodiment of the present invention has been described in detail above, the present invention is not limited to such specific embodiment, and various modifications and variations are possible within the scope of the gist of the present invention described in the claims. [Explanation of symbols]

[0068] 10 Response Generation Device 11 Learning Department 12 Response Generation Unit 100 Drive device 101 Recording media 102 Auxiliary storage 103 Memory device 104 processors 105 Interface device B Bus

Claims

1. a learning unit configured to train a speaker embedding model to which speaker information including a token for identifying a speaker and a text showing characteristics of the speaker is input so as to reproduce the speaker information, and to train a response generation model to which a latent expression generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue are input so as to reproduce a correct response to the speaker information and the history; A response generation learning device comprising:

2. The speaker embedding model is a Variational Auto-Encoder (VAE).

2. The response generation learning device according to claim 1,

3. The learning unit is configured to provide a constraint such that the latent representation follows a predetermined probability distribution.

3. The response generation learning device according to claim 2.

4. A speaker embedding model is input with speaker information including a token for identifying a speaker and a text showing characteristics of the speaker, and the speaker embedding model is trained to reproduce the speaker information. A response generation model is input with a latent expression generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue, and the response generation model is trained to reproduce a correct response to the speaker information and the history. a response generation unit configured to input speaker information to the speaker embedding model, and when a history of utterances in a dialogue is input to the response generation model, generate a response to the history of the utterances based on information output by the response generation model using a latent representation of the speaker embedding model; A response generating device comprising:

5. a learning procedure for training a speaker embedding model to which speaker information including a token for identifying a speaker and a text showing characteristics of the speaker is input, so that the speaker embedding model reproduces the speaker information, and training a response generation model to which latent expressions generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue are input, so that a response generation model reproduces a correct response to the speaker information and the history; A response generation learning method characterized in that the above steps are executed by a computer.

6. A speaker embedding model is input with speaker information including a token for identifying a speaker and a text showing characteristics of the speaker, and the speaker embedding model is trained to reproduce the speaker information. A response generation model is input with a latent expression generated by the speaker embedding model based on the speaker information and a history of utterances in a dialogue, and the response generation model is trained to reproduce a correct response to the speaker information and the history. a response generation step of inputting speaker information into the speaker embedding model, and generating a response to the utterance history based on information output by the response generation model using a latent representation of the speaker embedding model when a history of utterances in a dialogue is input into the response generation model; A response generating method characterized in that the above steps are executed by a computer.

7. 4. A program for causing a computer to function as the response generation learning device according to claim 1.

8. A program for causing a computer to function as the response generating device according to claim 4.

Images