JP7428245B2 - Response sentence generator and program - Google Patents

Description

The present invention relates to an interaction technique for interacting with a user, and particularly to a technique for generating system utterances that reflect individuality.

With the development of dialogue systems, there is an increasing demand for providing features such as individuality and character (hereinafter collectively referred to as "individuality") to dialogue systems (for example, Non-Patent Document 1). Many conventional commercial dialogue systems use a rule system, and by preparing response rules that reflect individuality in advance, a unique dialogue system can be constructed. In recent years, in dialogue systems, methods that use neural networks to generate responses (hereinafter referred to as ``sentence generation methods'') have become commonplace, and it is hoped that a method that takes individuality into account in sentence generation methods will also be realized.

When considering personality in the sentence generation method, there is a method of inputting personality information that can be used for responses together with the input sentence. For example, suppose there is a dialogue system that generates the response sentence ``I like curry and rice'' in response to the input sentence ``What kind of food do you like?'' without considering individuality. Here, when considering personality, input information such as ``I like fried chicken when it comes to food.My hobby is surfing.I have a dog.'' along with the input sentence ``What kind of food do you like?'' By inputting this information, a response sentence such as ``I like fried chicken'' can be generated. This method learns the relationship between general utterances and responses, and then generates response sentences that reflect personality information input at the same time if that information can be directly used in the response.

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.

However, it is difficult to verbalize the unique personality of a particular individual (for example, when you want to give a dialogue system the personality of a person whose personality is well known, such as Oda Nobunaga or Toyotomi Hideyoshi), or it is difficult to express it in words. Sometimes a response that deviates from the relationship between the utterance and the response may become necessary. For example, when responding to the input ``Next year is the year of the monkey,'' a response that reflects Toyotomi Hideyoshi's personality will generate responses such as ``It's the year of the eagle!'' or ``Who's the monkey?!'' It is expected. However, conventional methods are only effective when individuality information can be reflected in the relationship between general utterances and responses. Even if you input something like, "Next year is the year of the monkey," it will not be possible to generate a response sentence like the one above.

SUMMARY OF THE INVENTION In view of the above-mentioned technical problems, an object of the present invention is to provide a dialogue technique that can generate response sentences that reflect individuality without inputting information about individuality.

In order to solve the above problems, a response sentence generation device according to a first aspect of the present invention includes an input section for inputting an input sentence and a speaker identifier representing a speaker, and an input section for inputting an input sentence and a speaker identifier representing a speaker; a response sentence generation unit that generates a response sentence by inputting it to a sentence generation model; the response sentence generation model generates a speaker model that calculates a speaker embedding vector from a speaker identifier; and a sentence vector from an uttered sentence. An encoder, a decoder that generates a response sentence using an attention vector representing the content of attention to an uttered sentence, and an attention vector that generates an attention vector using a content vector representing the internal state of the decoder, a sentence vector, and a speaker embedding vector. including a mechanism.

A response sentence generation model learning device according to a second aspect of the present invention has a learning data storage that stores learning data consisting of an uttered sentence, a response sentence in response to the uttered sentence by a predetermined speaker, and a speaker identifier representing the speaker. and a model learning unit that uses learning data to learn a response sentence generation model that receives an uttered sentence and a speaker identifier as input and outputs a response sentence in response to the uttered sentence. , a speaker model that calculates a speaker embedding vector from a speaker identifier, an encoder that generates a sentence vector from an uttered sentence, a decoder that generates a response sentence using the attention vector representing the content of attention for the uttered sentence, and a decoder. It includes an attention mechanism that generates an attention vector using a content vector representing an internal state, a sentence vector, and a speaker embedding vector.

According to this invention, a response sentence that reflects individuality can be generated without inputting individuality information.

FIG. 1 is a diagram illustrating the functional configuration of a response sentence generation device. FIG. 2 is a diagram illustrating the processing procedure of the response sentence generation method. FIG. 3 is a diagram illustrating the functional configuration of the response sentence generation model. FIG. 4 is a diagram illustrating the functional configuration of the response sentence generation model learning device. FIG. 5 is a diagram illustrating the processing procedure of the response sentence generation model learning method. FIG. 6 is a diagram illustrating the functional configuration of a computer.

Embodiments of the present invention will be described in detail below. Note that in the drawings, components having the same functions are designated by the same numbers, and redundant explanation will be omitted.

The symbol " ^- " used in a sentence should originally be written directly above the character that immediately follows it, but due to text notation restrictions, it is written immediately before the character in question. In mathematical formulas, these symbols are written in their original positions, that is, directly above the letters.

[Summary of the invention]
In the present invention, a dialogue system using a sentence generation method is configured to assume an arbitrary speaker and generate a response sentence that reflects the speaker's personality. At this time, information on individuality, which is required to take individuality into consideration in conventional sentence generation methods, is no longer necessary. For example, from the input sentence ``Next year is the year of the monkey,'' it is possible to generate a response sentence ``Who is the monkey!!'' that reflects the personality of Toyotomi Hideyoshi.

To this end, in a neural network that learns the relationship between utterances and responses that differ for each personality, we developed a framework that takes into account the characteristics of each speaker's personality for the attention mechanism that learns the correspondence between input and output. The system learns the correspondence between input and output that is characteristic of each individuality. For example, when generating response sentences, attention trends that vary depending on the speaker can be realized, such as, ``This person is likely to pay attention to the word monkey,'' or ``This person is likely to derive this kind of meaning from the word monkey.'' This improves the performance of response sentence generation that takes individuality into consideration (that is, the quality of the generated response sentence).

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

<Response sentence generation device>
As shown in FIG. 1, the response sentence generation device 1 according to the embodiment receives an input sentence representing the content of the user's utterance and a speaker identifier that uniquely identifies the speaker, and generates the content of the system utterance in response to the input sentence. Outputs the response sentence that represents. The response sentence generation device 1 includes, for example, a model storage section 10, an input section 11, and a response sentence generation section 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.

The response sentence generation device 1 is configured by loading a special program into a known or dedicated computer having, for example, a central processing unit (CPU), a main memory (RAM), etc. It is a special device. The response sentence generation device 1 executes each process under the control of, for example, a central processing unit. The data input to the response sentence generation device 1 and the data obtained through each process are stored, for example, in the main memory, and the data stored in the main memory is read out to the central processing unit as necessary. and used for other processing. At least a portion 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 includes, for example, a main storage device such as a RAM (Random Access Memory), an auxiliary storage device constituted by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory; Alternatively, it can be configured with middleware such as a relational database or key-value store.

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, 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, an uttered sentence representing the content of a question uttered by the user to the dialogue system. The speaker identifier is an identifier that uniquely identifies a person whose individuality is desired to be reflected. The response sentence is, for example, an uttered sentence expressing the content of the answer from the dialog system to the question sentence given as the input sentence.

The speaker model 101 is a trained model that receives a speaker identifier as input, converts the speaker identifier into a speaker embedding vector, and outputs the result. As the speaker model 101, for example, a model called a speaker model described in reference document 1 can be used.
[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.

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

The attention mechanism 104 receives as input the speaker embedding vector outputted by the speaker model 101, the sentence vector outputted by the encoder 102, and the 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 calculate an attention weight, which is a vector representing which part of the input sentence to pay attention to (hereinafter referred to as "attention tendency"). generate. 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 of the input sentence that has been paid attention to according to the attention tendency (hereinafter referred to as "the content of attention"). generate.

The difference from the attention mechanism in conventional sentence generation methods is that speaker embedding vectors are referenced in the calculation of attention weights and attention vectors. This changes the tendency and content of attention depending on individuality. The tendency of attention is, for example, a characteristic such as "Toyotomi Hideyoshi pays strong attention to the word monkey." The content of the warning is, for example, a characteristic such as "Toyotomi Hideyoshi views the word monkey in a negative way." These features do not need to be added manually in advance; rather, training data that reflects the tendency and content of attention, specifically, a large number of sentence data linked to speakers, can be prepared. This is reflected in the attention vector.

Specifically, the attention mechanism 104 calculates the following formula.

Here, ◯ is an operator representing the product of elements. t is a variable indicating that the decoder is outputting the tth word. i is a variable indicating that it is the i-th word in the input sentence consisting of N words input to the encoder. h _t ^(dec) is a d-dimensional content vector representing the internal state of the decoder. Note that d is the size (number of dimensions) of the calculation part of the attention mechanism. H ^(enc) is an N×d dimensional sentence vector generated by the encoder. h _i ^(enc) ∈H ^(enc) is an element corresponding to the i-th word of the sentence vector. s _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 linear transformations of first order or any M order, or can be performed using sigmoid function, softsign function, etc. You may define a function that falls within a certain threshold such as , or you may combine these functions. a _i is the attention weight for the i-th word in the input sentence.

That is, the attention mechanism 104 calculates the attention vector as follows. First, in order to calculate the attention weight a _i , the speaker embedding vector s _u is transformed using an M-order linear transformation f(·). Calculate the element product of the linearly transformed speaker embedding vector f(s _u ) and each element h _i ^{( enc) of the encoded sentence vector H (enc)} , and set it to ^- h _i,k ^(enc) ( (equivalent to the second line of the formula). Calculate the i-th attention weight a _i using ^- h _i,k ^(enc) transformed by the speaker embedding vector s _u and the decoder content vector h _t ^(dec) (corresponds to the 4th line of the formula) . Next, in order to calculate the attention vector, the speaker embedding vector s _u is transformed using an M-order linear transformation g(·). Calculate the element product of the linearly transformed speaker embedding vector g(s _u ) and each element h _i ( ^{enc) of the encoded sentence vector H (enc) ,} and set it to ^- h _i,v ^(enc) ( (equivalent to the third line of the formula). The subscripts k and v of ^- h _i,k ^(enc) and ^- h _i,v ^(enc) are calculated using each element of these encoded sentence vectors and the linearly transformed speaker embedding vector, respectively. This subscript is used because in attention mechanisms, the weight is conventionally called the key, and the vector to which the weight is applied is called the value. Finally, the product of the attention weight a _i and ^- h _i,v ^(enc) is calculated for all i, and their sum is determined. This sum is the attention vector that becomes the final output (corresponding to the first line of the formula).

With reference to FIG. 2, the processing procedure of the response sentence generation method executed by the response sentence generation device 1 of the embodiment will be described.

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.

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 personality of the speaker. When outputting a response sentence, a word string that becomes a response sentence is obtained by repeatedly outputting words associated with the vector obtained from the output layer of the response sentence generation model. The response sentence generation unit 12 outputs the obtained response sentence from the response sentence generation device 1 .

<Response sentence generation model learning device>
As shown in FIG. 4, the response sentence generation model learning device 2 of the embodiment includes, for example, a learning data storage section 20, a model learning section 21, and a model storage section 10. The response sentence generation model learning method of the embodiment is realized by the response sentence generation model learning device 2 performing the processing of each step illustrated in FIG.

The response sentence generation model learning device 2 is configured by loading a special program into a known or dedicated computer having, for example, a central processing unit (CPU), a main memory (RAM), etc. It is a special device that has been The response sentence generation model learning device 2 executes each process under the control of, for example, a central processing unit. The data input to the response sentence generation model learning device 2 and the data obtained through each process are stored, for example, in the main memory, and the data stored in the main memory is read to the central processing unit as necessary. It is output and used for other processing. The response sentence generation model learning device 2 may be configured at least in part by hardware such as an integrated circuit. Each storage unit included in the response sentence generation model learning device 2 includes, for example, a main storage device such as a RAM (Random Access Memory), and an auxiliary storage device constituted by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory. It can be configured by a device or middleware such as a relational database or a key-value store.

With reference to FIG. 5, 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.

The learning data storage unit 20 stores learning data. The learning data includes, for example, an uttered sentence that is a question, a response sentence in which a predetermined speaker responds to the uttered sentence, and a speaker identifier representing the speaker. The learning data may be a collection of conversations actually conducted in a dialog system or the like, may be created manually with a specific person in mind, or may be a mixture of these.

In step S20, the response sentence generation model learning device 2 reads learning data from the learning data storage unit 20. The response sentence generation model learning device 2 inputs the read learning data to the model learning section 21.

In step S21, the model learning unit 21 learns each parameter of the neural network of the response sentence generation model 100 using the input learning data. The learning method of the response sentence generation model is similar to the learning method of a conventional model that generates an output using an 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, decoder 103, and attention mechanism 104 are learned so as to minimize the loss. When learning the parameters of the attention mechanism 104, updating of the parameters f and g is repeated 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 the 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 parameters of the learned response sentence generation model 100 in the model storage unit 10.

[Experimental result]
In order to measure the effects of the above embodiment, an experiment was conducted using the data of the impersonation question and answer disclosed in Non-Patent Document 1. Specifically, 50,000 question-and-answer pairs from three people's impersonation data were used as learning 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 is implemented in a form that replaces the self-attention and source-target attention in the Transformer. The model was trained using the training data, and answers were generated by giving the questions from the evaluation data. The evaluation scales used were BLEU-1, BLEU-4, and PPL. That is, the answer sentences generated in this embodiment and the correct answer sentences linked to the questions in the evaluation data were compared using BLEU-1 and BLEU-4 (the larger the value, the better). We also calculated the model generation probability (PPL: Perplexity) for correct answer sentences (the smaller the value, the better).

Table 1 shows the experimental results. It can be seen that the method of this embodiment achieved the best evaluation in all evaluation indicators.

Although the embodiments of this invention have been described above, the specific configuration is not limited to these embodiments, and even if the design is changed as appropriate without departing from the spirit of this invention, Needless to say, it is included in this invention. The various processes described in the embodiments are not only executed in chronological order according to the order described, but also may be executed in parallel or individually depending on the processing capacity of the device that executes the processes or as necessary.

[Program, recording medium]
When the various processing functions of each device described in the above embodiments are realized by a computer, the processing contents of the functions that each device should have are described by a program. By loading this program into the storage unit 1020 of the computer shown in FIG. 6 and causing it to operate in the arithmetic processing unit 1010, input unit 1030, output unit 1040, etc., various processing functions in each of the above devices are realized on the computer. be done.

A program describing the contents of this process 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, an optical disk, or the like.

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

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 into the auxiliary storage unit 1050, which is its own non-temporary storage device. Store. When executing the process, this computer loads the program stored in the auxiliary storage section 1050, which is its own non-temporary storage device, into the storage section 1020, which is a temporary storage device, and executes the program according to the read program. Execute processing. In addition, as another form of execution of this program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and further, the program may be transferred to this computer from the server computer. The process may be executed in accordance with the received program each time. In addition, the above-mentioned processing is executed by a so-called ASP (Application Service Provider) type service, which does not transfer programs from the server computer to this computer, but only realizes processing functions by issuing execution instructions and obtaining results. You can also use it as Note that the program in this embodiment includes information that is used for processing by an electronic computer and that is similar to a program (data that is not a direct command to the computer but has a property that defines the processing of the computer, etc.).

Further, in this embodiment, the present apparatus is configured by executing a predetermined program on a computer, but at least a part of these processing contents may be realized by hardware.

Claims (2)

an input section for inputting an input sentence and a speaker identifier representing a speaker;
a response sentence generation unit that obtains a response 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 calculates a speaker embedding vector from a speaker identifier;
an encoder that generates a sentence vector from an uttered sentence;
a decoder that generates a response sentence using an attention vector representing the content of attention to the uttered sentence;
an attention mechanism that generates the attention vector using a content vector representing an internal state of the decoder, the sentence vector, and the speaker embedding vector;
including;
The attention mechanism is configured such that H ^(enc) is the sentence vector, N is the number of elements of the sentence vector, h _i ^(enc) is the element corresponding to the i-th word of the sentence vector, and h _t ^(dec) Let be the content vector when calculating the t-th element of the response sentence, let s _u be the speaker embedding vector, f be the first linear transformation, g be the second linear transformation, and calculate the following equation. Generate the attention vector by

Response sentence generator. A program for causing a computer to function as the response sentence generation device according to claim 1.

Images