Response sentence generation device, response sentence generation model learning device, their methods, and programs

Abstract

To generate a response sentence in which an individual character is reflected without inputting information about the individual character.SOLUTION: An input unit (11) inputs an input sentence and a speaker identifier which identifies a speaker. A response generation unit (12) generates a response sentence to the input sentence by inputting the input sentence and the speaker identifier to a response sentence generation model. The response sentence generation model includes a speaker model that generates a speaker embedded vector from the speaker identifier, an encoder that generates a sentence vector from the input sentence, a decoder that generates the response sentence from an attention vector indicating the content of an attention, and an attention mechanism that calculates an attention weight which is a vector of an attention tendency, from an element product of the sentence vector and the speaker embedded vector converted by a first linear conversion, and a content vector which indicates the inner status of the decoder, and generates the attention vector from the attention weight and an element product of the sentence vector and the speaker embedded vector converted by a second linear conversion which is different from the first linear conversion.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a dialogue technique for dialogue with a user, and more particularly to a technique for generating system utterances that reflect individuality. [Background technology]

[0002] With the development of dialogue systems, there is an increasing demand to impart features such as individuality and character (hereinafter collectively referred to as "personality") to dialogue systems (for example, Non-Patent Document 1). Many conventional commercial dialogue systems use a rule-based approach, and a dialogue system with individuality is constructed by preparing response rules reflecting individuality in advance. In recent dialogue systems, a method of generating responses using a neural network (hereinafter referred to as a "sentence generation method") has become common, and it is expected that a method of considering individuality in sentence generation methods will also be realized.

[0003] When considering individuality in a sentence generation method, there is a method in which individuality information that can be used in the response is input along with the input sentence. For example, when individuality is not considered, a dialogue system may generate the response sentence "I like curry rice" in response to the input sentence "What kind of food do you like?". When individuality is considered, if individuality information such as "I like fried chicken, my hobby is surfing, and I have a dog" is input along with the input sentence "What kind of food do you like?", the response sentence "I like fried chicken" can be generated. This method learns the relationship between general utterances and responses, and then generates a response sentence that reflects the individuality information input at the same time, if that information can be directly used in the response. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Ryuichiro Higashinaka, Masahiro Mizukami, Hidetoshi Kawabata, Emi Yamaguchi, Noritake Adachi, and Junji Tomita1, "Role play-based question-answering by real users for building chatbots with consistent personalities", Proceedings of the SIGDIAL 2018 Conference, pp. 264-272, Melbourne, Australia, July 2018. Summary of the Invention [Problem to be solved by the invention]

[0005] However, the unique characteristics of a specific individual (for example, when you want to give the dialogue system the characteristics of a person whose characteristics are well known, such as Oda Nobunaga or Toyotomi Hideyoshi) may be difficult to verbalize, or a response that deviates from the general relationship between utterances and responses may be required. For example, when responding to the input "Next year is the year of the monkey," if you want to respond to Toyotomi Hideyoshi's characteristics, you can expect to generate responses such as "It's my year" or "Who's the monkey!!" However, since conventional methods are only effective when information about characteristics can be reflected in the general relationship between utterances and responses, even if you input something like "My name is Toyotomi Hideyoshi. He is one of the three great heroes. He served Oda Nobunaga and unified the country. Next year is the year of the monkey," you cannot generate responses like the above.

[0006] In view of the above technical problems, an object of the present invention is to provide a dialogue technique capable of generating a response sentence that reflects a personality without inputting personality information. [Means for solving the problem]

[0007] In order to solve the above problem, a response sentence generation device of a first aspect of the present invention includes an input unit that inputs an input sentence and a speaker identifier representing a speaker, and a response sentence generation unit that generates a response sentence corresponding to the input sentence by inputting the input sentence and the speaker identifier into a response sentence generation model. The response sentence generation model includes a speaker model that generates a speaker embedding vector from the speaker identifier, an encoder that generates a sentence vector from the input sentence, a decoder that generates a response sentence from an attention vector that represents attention content, which is content that is paid attention to according to an attention tendency that indicates which part of the input sentence to focus on, and an attention mechanism that calculates an attention weight, which is a vector of attention tendency, from an element product of the sentence vector and a speaker embedding vector transformed by a first linear transformation, and a content vector that represents the internal state of the decoder, and generates an attention vector from the attention weight and an element product of the sentence vector and a speaker embedding vector transformed by a second linear transformation different from the first linear transformation.

[0008] A response sentence generation model learning device of a second aspect of the present invention includes a memory unit that stores learning data consisting of an utterance sentence, a response sentence of a specified speaker responding to the utterance sentence, and a speaker identifier representing the speaker, and a learning unit that learns a response sentence generation model that receives the utterance sentence and the speaker identifier stored in the memory unit as input and outputs a response sentence to the input utterance sentence, and the response sentence generation model includes a speaker model that generates a speaker embedding vector from the speaker identifier, an encoder that generates a sentence vector from the utterance sentence, a decoder that generates a response sentence from an attention vector that represents the content of attention, which is content that has been paid attention to according to an attention tendency that indicates which part of the utterance sentence is to be paid attention to, and an attention mechanism that calculates an attention weight, which is a vector of attention tendency, from the element product of the sentence vector and the speaker embedding vector transformed by a first linear transformation, and a content vector that represents the internal state of the decoder, and generates an attention vector from the attention weight and the element product of the sentence vector and the speaker embedding vector transformed by a second linear transformation different from the first linear transformation. Effect of the Invention

[0009] According to the present invention, a response sentence that reflects personality can be generated without inputting personality information. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram illustrating an example of the functional configuration of a response sentence generation device. [Diagram 2] FIG. 2 is a diagram illustrating a processing procedure of the response sentence generation method. [Diagram 3] FIG. 3 is a diagram illustrating an example of the functional configuration of the response sentence generation model. [Figure 4] FIG. 4 is a diagram illustrating an example of a functional configuration of a response sentence generation model learning device. [Diagram 5] FIG. 5 is a diagram illustrating a processing procedure of the response sentence generation model learning method. [Figure 6] FIG. 6 is a diagram illustrating an example of the functional configuration of a computer. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an embodiment of the present invention will be described in detail. In the drawings, components having the same functions are designated by the same reference numerals, and duplicated explanations will be omitted.

[0012] Symbols used in this text: - " should be written directly above the character immediately following it, but due to limitations in text notation, they are written immediately before the character in question. In mathematical formulas, these symbols are written in their proper position, that is, directly above the character.

[0013] [Summary of the invention] In this invention, a dialogue system using a sentence generation method is configured to assume an arbitrary speaker and generate a response sentence that reflects the personality of the speaker. In this case, the information on personality that was required to take personality into account in conventional sentence generation methods is not required. For example, from an input sentence such as "Next year is the year of the monkey," a response sentence that reflects the personality of Toyotomi Hideyoshi, "Who's the monkey!!" can be generated.

[0014] To achieve this, a framework is introduced to take into account the individual characteristics of each speaker into the attention mechanism that learns the correspondence between input and output in a neural network that learns the relationship between utterances and responses that differ for each personality, and input-output correspondences that are characteristic of each personality are learned. For example, attention tendencies that differ for each speaker, such as "this person is likely to pay attention to the word monkey" or "this is the meaning that is likely to be interpreted from the word monkey," are realized in response generation. This improves the performance of response generation that takes individuality into account (i.e., the quality of the generated response).

[0015] [Embodiment] An embodiment of the present invention comprises a response sentence generation device and method for generating a response sentence in response to an input sentence based on a user utterance in a dialogue system using a sentence generation method, and a response sentence generation model learning device and method for learning a response sentence generation model used in the response sentence generation device and method.

[0016] <Response Generation Device> As shown in Fig. 1, a response sentence generation device 1 of the embodiment receives an input sentence representing the content of a user utterance and a speaker identifier that uniquely identifies a speaker, and outputs a response sentence representing the content of a system utterance in response to the input sentence. The response sentence generation device 1 includes, for example, a model storage unit 10, an input unit 11, and a response sentence generation unit 12. The response sentence generation method of the embodiment is realized by the response sentence generation device 1 performing the processing of each step illustrated in Fig. 2.

[0017] The response sentence generation device 1 is a special device configured by loading a special program into a publicly known or dedicated computer having, for example, a central processing unit (CPU), a main storage device (RAM), and the like. The response sentence generation device 1 executes each process under the control of, for example, the central processing unit. Data input to the response sentence generation device 1 and data obtained in each process are stored in, for example, the main storage device, and the data stored in the main storage device is read out to the central processing unit as necessary and used for other processes. At least a part of the response sentence generation device 1 may be configured by hardware such as an integrated circuit. Each storage unit included in the response sentence generation device 1 can be configured by, for example, a main storage device such as a random access memory (RAM), an auxiliary storage device configured by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory, or middleware such as a relational database or a key-value store.

[0018] The model storage unit 10 stores a trained response sentence generation model. As shown in FIG. 3, the response sentence generation model 100 receives an input sentence and a speaker identifier as input, and outputs a response sentence. The response sentence generation model 100 includes, for example, a speaker model 101, an encoder 102, a decoder 103, and an attention mechanism 104. The input sentence is, for example, a spoken sentence representing the content of a question uttered by a user to a dialogue system. The speaker identifier is an identifier that uniquely identifies a person whose personality is to be reflected. The response sentence is, for example, a spoken sentence representing the content of a response from a dialogue system to a question sentence given as an input sentence.

[0019] The speaker model 101 is a trained model that receives a speaker identifier, converts the speaker identifier into a speaker embedding vector, and outputs the vector. For example, the speaker model 101 may be a model called the Speaker model described in Reference 1. [Reference 1] Jiwei Li, Michel Galley, Chris Brockett, Georgios P ​​Spithourakis, Jianfeng Gao, and Bill Dolan, "A persona-based neural conversation model," arXiv preprint, arXiv:1603.06155, 2016.

[0020] The encoder 102 receives an utterance sentence as input, converts the utterance sentence into a sentence vector, and outputs the sentence vector. The decoder 103 generates and outputs a response sentence using the attention vector output by the attention mechanism 104. The encoder 102 and the decoder 103 are the same as those used in conventional sentence generation methods. For information on conventional sentence generation methods, see Reference 2. [Reference 2] Vinyals, Oriol, and Quoc Le, "A neural conversational model," arXiv preprint, arXiv:1506.05869, 2015.

[0021] The attention mechanism 104 receives as input the speaker embedding vector output by the speaker model 101, the sentence vector output by the encoder 102, and a content vector representing the internal state of the decoder 103, and generates and outputs an attention vector. The attention mechanism 104 first uses the speaker embedding vector, the sentence vector, and the content vector to generate an attention weight, which is a vector representing which part of the input sentence to pay attention to (hereinafter referred to as "attention tendency"). Next, the attention mechanism 104 uses the attention weight, the speaker embedding vector, and the sentence vector to generate an attention vector representing the content to which attention is paid according to the attention tendency for the input sentence (hereinafter referred to as "attention content").

[0022] The difference from the attention mechanism in the conventional sentence generation method is that the speaker embedding vector is referenced in the calculation of the attention weight and the attention vector. This allows the tendency and content of attention to be changed according to the individual's personality. The tendency of attention is, for example, a feature such as "Toyotomi Hideyoshi pays strong attention to the word "monkey". The content of attention is, for example, a feature such as "Toyotomi Hideyoshi perceives the word "monkey" negatively". These features do not need to be manually added in advance, but are reflected in the attention vector by preparing training data that reflects the tendency and content of attention, specifically, a large number of sentence data linked to speakers, as training data.

[0023] Specifically, the attention mechanism 104 calculates the following formula:

[0024]

number

[0025] Here, ○ is an operator that represents the product of elements. t is a variable that indicates that the decoder is outputting the t-th word. i is a variable that indicates that the i-th word is in the input sentence consisting of N words input to the encoder. h t (dec) is a d-dimensional content vector that represents the internal state of the decoder. Note that d is the size (number of dimensions) of the computational part of the attention mechanism. H (enc) is the N×d dimensional sentence vector generated by the encoder. i (enc) ∈H (enc) is the element corresponding to the i-th word in the sentence vector. u is a d-dimensional speaker embedding vector generated by the speaker model. f(·) and g(·) are different linear transformations. f(·) and g(·) may be a first-order linear transformation, an arbitrary M-th order linear transformation, or a function such as a sigmoid function or a softsign function may be defined so that the output falls within a certain threshold such as 0 to 1 or -1 to 1, or a combination of these may be used. i is the attention weight for the i-th word in the input sentence.

[0026] That is, the attention mechanism 104 calculates the attention vector as follows. First, the attention weight a i To calculate the speaker embedding vector s u is transformed using the M-th order linear transformation f(·). The linearly transformed speaker embedding vector f(s u ) and the encoded sentence vector H (enc) Each element of h i (enc) Calculate the element product with - h i,k (enc) (The second line of the formula.) The speaker embedding vector s u Deformed by - h i,k (enc) and the decoder content vector h t (dec) Using the i-th attention weight a i (Equivalent to the fourth line of the formula). Next, to calculate the attention vector, we use the speaker embedding vector s u The linearly transformed speaker embedding vector g(s u ) and the encoded sentence vector H (enc) Each element of h i (enc) Calculate the element product with - h i,v (enc) (corresponding to the third line of the formula). The linearly transformed speaker embedding vectors are used to calculate - h i,k (enc) , - h i,v (enc) The subscripts k and v are the initials of key and value, respectively. In attention mechanisms, it is customary to call weights key and vectors that multiply weights value, so these subscripts are used. Finally, for all i, the attention weights a i and - h i,v (enc)Calculate the product of and sum them. This sum is the final output attention vector (corresponding to the first line of the formula).

[0027] The processing procedure of the response sentence generation method executed by the response sentence generation device 1 of the embodiment will be described with reference to FIG.

[0028] In step S11, an input sentence and a speaker identifier are input to the input unit 11. The input unit 11 outputs the input sentence and the speaker identifier to the response sentence generation unit 12.

[0029] In step S12, the response sentence generation unit 12 receives the input sentence and the speaker identifier from the input unit 11, and inputs the input sentence and the speaker identifier into the response sentence generation model stored in the model storage unit 10 to obtain and output a response sentence that reflects the speaker's individuality. In outputting the response sentence, a word string that becomes the response sentence is obtained by repeatedly outputting words linked to the vector obtained from the output layer of the response sentence generation model. The response sentence generation unit 12 outputs the obtained response sentence as the output of the response sentence generation device 1.

[0030] <Response sentence generation model learning device> 4, the response sentence generation model learning device 2 of the embodiment includes, for example, a learning data storage unit 20, a model learning unit 21, and a model storage unit 10. The response sentence generation model learning device 2 performs processing of each step illustrated in FIG. 5, thereby realizing the response sentence generation model learning method of the embodiment.

[0031] The response sentence generation model learning device 2 is a special device configured by loading a special program into a publicly known or dedicated computer having, for example, a central processing unit (CPU), a main storage device (RAM), etc. The response sentence generation model learning device 2 executes each process under the control of, for example, the central processing unit. Data input to the response sentence generation model learning device 2 and data obtained in each process are stored in, for example, the main storage device, and the data stored in the main storage device is read out to the central processing unit as necessary and used for other processes. At least a part of the response sentence generation model learning device 2 may be configured by hardware such as an integrated circuit. Each storage unit included in the response sentence generation model learning device 2 can be configured by, for example, a main storage device such as a random access memory (RAM), an auxiliary storage device configured by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory, or middleware such as a relational database or a key-value store.

[0032] The processing procedure of the response sentence generation model learning method executed by the response sentence generation model learning device 2 of the embodiment will be described with reference to FIG.

[0033] The learning data storage unit 20 stores learning data. The learning data includes, for example, spoken sentences, which are questions, response sentences by a specific speaker responding to the spoken sentences, and speaker identifiers representing the speakers. The learning data may be data collected from actual conversations performed in a dialogue system or the like, or may be data manually created assuming a specific person, or may be a mixture of both.

[0034] In step S20, the response sentence generation model learning device 2 reads out learning data from the learning data storage unit 20. The response sentence generation model learning device 2 inputs the read out learning data to the model learning unit .

[0035] In step S21, the model learning unit 21 uses the input learning data to learn each parameter of the neural network of the response sentence generation model 100. The learning method of the response sentence generation model is the same as the learning method of the model that generates an output using a conventional input and a speaker identifier, which is disclosed in Reference 2. That is, the softmax cross entropy for the output sentence of the model is used as a loss function, and the parameters of the encoder 102, the decoder 103, and the attention mechanism 104 are learned so as to minimize the loss. When learning the parameters of the attention mechanism 104, the parameters f and g are repeatedly updated a predetermined number of times or until a predetermined condition is satisfied. At the same time, the parameters of the speaker model 101 that converts a speaker identifier into a speaker embedding vector are learned in the same manner as the conventional speaker model. The model learning unit 21 stores the learned parameters of the response sentence generation model 100 in the model storage unit 10.

[0036] [Experimental Results] In order to measure the effect of the above embodiment, an experiment was conducted using the role-playing question-answer data disclosed in Non-Patent Document 1. Specifically, 50,000 question-answer pairs of role-playing data for three people were used as training data, and 2,000 were used as evaluation data. The number of dimensions d of the attention mechanism was set to 512, and Transformer was used as the encoder and decoder. The attention mechanism of this embodiment was implemented by replacing the self-attention and source-target attention in Transformer. A model was trained with the training data, and an answer was generated by giving the question sentence of the evaluation data. BLEU-1, BLEU-4, and PPL were used as the evaluation scale. That is, the answer sentence generated in this embodiment was compared with the correct answer sentence linked to the question of the evaluation data using BLEU-1 and BLEU-4 (the larger the value, the better). In addition, the generation probability (PPL: Perplexity) of the model for the correct answer sentence was calculated (the smaller the value, the better).

[0037] The experimental results are shown in Table 1. It can be seen that the method of this embodiment achieved the best evaluation in all evaluation indices.

[0038] [Table 1]

[0039] Although the embodiments of the present invention have been described above, the specific configuration is not limited to these embodiments, and it goes without saying that the present invention includes appropriate design changes and the like within the scope of the gist of the present invention. The various processes described in the embodiments are not only executed in chronological order according to the order described, but may also be executed in parallel or individually depending on the processing capacity of the device executing the processes or as necessary.

[0040] [Programs, recording media] When the various processing functions of each device described in the above embodiment are realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, the various processing functions of each device are realized on the computer by loading this program into the storage unit 1020 of the computer shown in Fig. 6 and operating the arithmetic processing unit 1010, the input unit 1030, the output unit 1040, etc.

[0041] The program describing the processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a non-transitory recording medium, such as a magnetic recording device or an optical disk.

[0042] The program may be distributed, for example, by selling, transferring, lending, etc. portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of a server computer and transferring the program from the server computer to other computers via a network.

[0043] A computer that executes such a program, for example, first stores a program recorded on a portable recording medium or a program transferred from a server computer in the auxiliary recording unit 1050, which is a non-transient storage device of the computer. Then, when executing a process, the computer reads the program stored in the auxiliary recording unit 1050, which is a non-transient storage device, into the storage unit 1020, which is a temporary storage device, and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from the portable recording medium and execute a process according to the program, or may execute a process according to the received program each time a program is transferred from the server computer to the computer. Alternatively, the server computer may not transfer the program to the computer, but may execute the above-mentioned process by a so-called ASP (Application Service Provider) type service that realizes a processing function only by issuing an execution instruction and obtaining a result. Note that the program in this embodiment includes information used for processing by an electronic computer and equivalent to a program (data that is not a direct instruction to a computer but has a nature that specifies the processing of a computer, etc.).

[0044] In addition, in this embodiment, the present device is configured by executing a specific program on a computer, but at least a part of the processing contents may be realized by hardware.

Claims

1. an input unit for inputting an input sentence and a speaker identifier representing a speaker; a response sentence generation unit that generates a response sentence corresponding to the input sentence by inputting the input sentence and the speaker identifier into a response sentence generation model; Including, The response sentence generation model is a speaker model that generates a speaker embedding vector from the speaker identifier; an encoder for generating a sentence vector from the input sentence; a decoder that generates a response sentence from an attention vector that indicates the content of attention, the content being attentioned according to an attention tendency that indicates which part of the input sentence should be paid attention to; an attention mechanism that calculates an attention weight, which is a vector of the attention tendency, from an element product of the sentence vector and the speaker embedding vector transformed by a first linear transformation and a content vector representing an internal state of the decoder, and generates the attention vector from the attention weight and an element product of the sentence vector and the speaker embedding vector transformed by a second linear transformation different from the first linear transformation. Response generation device.

2. a storage unit that stores learning data including an utterance sentence, a response sentence of a predetermined speaker that responds to the utterance sentence, and a speaker identifier that represents the speaker; a learning unit that learns a response sentence generation model that receives an utterance sentence and a speaker identifier stored in the storage unit as input and outputs a response sentence to the input utterance sentence; Including, The response sentence generation model is a speaker model that generates a speaker embedding vector from the speaker identifier; an encoder for generating a sentence vector from the spoken sentence; a decoder that generates a response sentence from an attention vector that indicates the content of attention, the content being attentioned according to an attention tendency that indicates which part of the utterance sentence is to be paid attention to; an attention mechanism that calculates an attention weight, which is a vector of the attention tendency, from an element product of the sentence vector and the speaker embedding vector transformed by a first linear transformation and a content vector representing an internal state of the decoder, and generates the attention vector from the attention weight and an element product of the sentence vector and the speaker embedding vector transformed by a second linear transformation different from the first linear transformation. A response sentence generation model learning device.

3. an input unit inputs an input sentence and a speaker identifier representing a speaker; a response sentence generation unit inputting the input sentence and the speaker identifier into a response sentence generation model to generate a response sentence corresponding to the input sentence; The response sentence generation model is a speaker model that generates a speaker embedding vector from the speaker identifier; an encoder for generating a sentence vector from the input sentence; a decoder that generates a response sentence from an attention vector that indicates the content of attention, the content being attentioned according to an attention tendency that indicates which part of the input sentence should be paid attention to; an attention mechanism that calculates an attention weight, which is a vector of the attention tendency, from an element product of the sentence vector and the speaker embedding vector transformed by a first linear transformation and a content vector representing an internal state of the decoder, and generates the attention vector from the attention weight and an element product of the sentence vector and the speaker embedding vector transformed by a second linear transformation different from the first linear transformation. Response generation method.

4. A storage unit stores learning data including an utterance sentence, a response sentence of a predetermined speaker responding to the utterance sentence, and a speaker identifier representing the speaker, a learning unit learning a response sentence generation model that receives an utterance sentence and a speaker identifier stored in the storage unit as input and outputs a response sentence to the input utterance sentence; The response sentence generation model is a speaker model that generates a speaker embedding vector from the speaker identifier; an encoder for generating a sentence vector from the spoken sentence; a decoder that generates a response sentence from an attention vector that indicates the content of attention, the content being attentioned according to an attention tendency that indicates which part of the utterance sentence is to be paid attention to; an attention mechanism that calculates an attention weight, which is a vector of the attention tendency, from an element product of the sentence vector and the speaker embedding vector transformed by a first linear transformation and a content vector representing an internal state of the decoder, and generates the attention vector from the attention weight and an element product of the sentence vector and the speaker embedding vector transformed by a second linear transformation different from the first linear transformation. A method for training a response sentence generation model.

5. A program for causing a computer to function as the response sentence generation device according to claim 1 or the response sentence generation model learning device according to claim 2.

Images