JP7224278B2 - Apparatus, program and method for estimating response sentence to user's utterance sentence - Google Patents

Description

TECHNICAL FIELD The present invention relates to technology of a dialogue device that realizes natural dialogue with a user.

In recent years, not only smartphones and tablets but also smart speakers such as "Google Home (registered trademark)" and "Amazon Echo (registered trademark)" have been used as interactive devices with users, as well as "SOTA (registered trademark)" and "Unibo (registered trademark)". (trademark)” (hereinafter referred to as “terminal”) have been used. The terminal functions as a user interface and transmits the user's uttered voice to the interactive device as a server.
The dialogue device estimates a response sentence that will be a natural dialogue for the utterance sentence, and returns the response sentence to the terminal.
Then, the terminal replies the response sentence to the user by voice or text.
Examples of such dialogue systems include "Siri (registered trademark)" and "Talking Concier (registered trademark)."

Conventionally, there is a technique of generating a response sentence having a causal relationship with an uttered sentence and realizing a natural dialogue (for example, see Non-Patent Document 1). According to this technique, a word pair dictionary having a causal relationship can be constructed in advance, and a response sentence having a causal relationship with an utterance sentence can be preferentially selected (reranking response generation). Specifically, a word in the user's utterance sentence and a word in the response sentence are paired with a word pair dictionary. When matching is achieved (the word pair exists in the word pair dictionary), it is determined that there is a causal relationship, and this response sentence is preferentially selected.
According to Non-Patent Document 1, for example, causality pairs can be extracted and learned for the following sentences.
“Since the yen has weakened, we can expect the Japanese economy to rise from a trade perspective.”
: Causality pair {(yen depreciation) -> (economy rise)}

There is also a technique for creating a learning model using only dialogue data (utterance sentences and response sentences) having a causal relationship from a large-scale dialogue corpus (see, for example, Non-Patent Document 2).

Shohei Tanaka, Koichiro Yoshino, Katsuhito Sudoh, and Satoshi Nakamura. "Conversational Response Re-ranking Based on Event Causality and Role Factored Tensor Event Embedding" 1st Workshop NLP for Conversational AI, ACL 2019 Workshop (ConvAI). Shota Sato, Kentaro Inui. “Chat Response Learning Using Data Sampling Based on Causal Relationships” Proceedings of the 24th Annual Conference of the Language Processing Conference (March 2018)

According to the technique described in Non-Patent Literature 1, since only word pairs are matched during learning, context feature amounts other than the word pairs are not considered at all. According to the above example, the causality pair {(the yen depreciates) -> (the economy rises)} does not include any restrictive features such as "perspective of trade" or "Japan." There will be no
In addition, during operation, there is also the problem that reranking (reselection of the optimum response result) cannot be realized unless causality pairs that have been learned in advance are completely matched against actual user utterances.

According to the technique described in Non-Patent Document 2, in order to create a learning model using only dialogue data (utterance sentences and response sentences) that have a causal relationship, it relies too much on dialogue data serving as teacher data. This is not only because of the amount of training data, but also because it lacks diversity and versatility as a response sentence to a user's utterance sentence.

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide an apparatus, a program, and a method for estimating a response sentence that considers causal relationships as much as possible with respect to a user's utterance sentence without relying too much on a dialogue corpus as training data. and

According to the present invention, in a program that causes a computer installed in a dialogue device to generate a response sentence to a user's utterance sentence,
when learning,
a conjunctive particle table in which conjunctive particles that connect the preceding and following sentences in a causal relationship are registered;
learning sentence selection means for selecting learning sentences containing the conjunction particles registered in the conjunction particle table from the group of learning sentences serving as the first training data;
a classification-type first machine learning engine that learns learning sentences containing conjunctive particles as having a causal relationship;
Utterance response sentence selection means for inputting an utterance sentence and a response sentence paired in a dialogue corpus as second training data to a first machine learning engine and selecting an utterance sentence and a response sentence presumed to have a causal relationship. and,
Functioning as a second machine learning engine that inputs an utterance sentence estimated to have a causal relationship to the encoder side and inputs a response sentence to the decoder side for learning,
During operation,
It is characterized by inputting a user's uttered sentence to the second machine learning engine and causing the computer to function as interactive means for obtaining a response sentence.

According to another embodiment of the program of the present invention,
During operation,
an utterance response sentence generation means for generating an utterance response sentence by connecting a user utterance sentence input from the dialogue means and a response sentence obtained by inputting the utterance sentence to a second machine learning engine;
Presence or absence of a causal relationship is estimated by inputting the generated utterance response text into a first machine learning engine, and a causal relationship in which a response sentence of the utterance response text presumed to have a causal relationship is output to the dialogue means. It is also preferable to have a computer further function as the estimation means.

According to another embodiment of the program of the present invention,
The first machine learning engine is a categorical neural network capable of extracting the characteristics of the entire sentence that has a causal relationship,
The second machine learning engine also preferably functions as a Seq2Seq capable of extracting features between the causal utterance sentence and the response sentence.

According to another embodiment of the program of the present invention,
when learning,
Functioning as a third machine learning engine for learning by inputting the utterance sentence to the encoder side and inputting the response sentence to the decoder side for the utterance sentence and the response sentence that are paired in the dialogue corpus as the second training data,
During operation,
When the first machine learning engine estimates that there is no causal relationship, the causality estimating means inputs the utterance sentence of the utterance response sentence to the third machine learning engine and outputs the obtained response sentence to the dialogue means. It is also preferred to have the computer function to do so.

According to another embodiment of the program of the present invention,
The third machine learning engine also preferably functions as a Seq2Seq capable of extracting features between generic utterance sentences and response sentences.

According to the present invention, in a dialogue device that generates a response sentence to a user's utterance sentence,
when learning,
a conjunctive particle table in which conjunctive particles that connect the preceding and following sentences in a causal relationship are registered;
learning sentence selection means for selecting learning sentences containing the conjunction particles registered in the conjunction particle table from the group of learning sentences serving as the first training data;
a classification-type first machine learning engine that learns learning sentences containing conjunctive particles as having a causal relationship;
Utterance response sentence selection means for inputting an utterance sentence and a response sentence paired in a dialogue corpus as second training data to a first machine learning engine and selecting an utterance sentence and a response sentence presumed to have a causal relationship. and,
Functioning as a second machine learning engine that inputs an utterance sentence estimated to have a causal relationship to the encoder side and inputs a response sentence to the decoder side for learning,
During operation,
and dialogue means for inputting a user's utterance sentence to a second machine learning engine and acquiring a response sentence.

According to the present invention, in the dialogue method for a device that generates a response sentence to a user's utterance sentence,
The device
when learning,
a conjunctive particle table in which conjunctive particles that connect the preceding and following sentences in a causal relationship are registered;
A classification-type first machine that selects learning sentences containing a conjunction particle registered in a conjunction particle table from a group of learning sentences serving as first teacher data, and learns the learning sentences containing the conjunction particles as having a causal relationship. a learning engine;
A pair of utterance sentences and response sentences in the dialogue corpus, which serves as second training data, is input to the first machine learning engine, and utterance sentences and response sentences that are presumed to have a causal relationship are selected, and judged to have a causal relationship. a second machine learning engine that inputs the estimated utterance sentence to the encoder side and inputs the response sentence to the decoder side for learning;
During operation,
It is characterized by inputting a user's uttered sentence to a second machine learning engine and executing so as to obtain a response sentence.

According to the dialogue apparatus, program, and method of the present invention, it is possible to estimate a response sentence that considers causal relationships as much as possible with respect to a user's utterance sentence without relying too much on a dialogue corpus as teacher data.

1 is a functional configuration diagram of an interactive device according to the present invention; FIG. FIG. 4 is an explanatory diagram showing learning of a first machine learning engine and a second machine learning engine; 1 is a schematic functional configuration diagram of a first machine learning engine; FIG. FIG. 4 is a schematic functional configuration diagram of a second machine learning engine and a third machine learning engine; FIG. 11 is a second functional configuration diagram of the interactive device during operation in the second embodiment; FIG. 11 is an explanatory diagram showing dialogue during operation in the second embodiment; FIG. 11 is a second functional configuration diagram of the interactive device during operation in the third embodiment; FIG. 11 is an explanatory diagram showing dialogue during operation in the third embodiment;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a functional configuration diagram of a dialogue device according to the present invention.

The dialogue device 1 realizes natural dialogue with a user, and generates a response sentence to a user's utterance sentence.
According to FIG. 1, a dialogue device 1 with server functionality communicates with a terminal 2 with user interface functionality. The terminal 2, as an input/output device for the user, may acquire the user's voice with a microphone and utter it to the user with a speaker, or input a text-based utterance from the user and display a response. can be anything.
The speech recognition function may be installed in the interactive device 1 or may be installed in the terminal 2. FIG.

According to the present invention, the interactive device 1 is equipped with a plurality of machine learning engines, and is divided into <during learning> and <during operation>. Further, the interactive device 1 causes the machine learning engine to construct a learning model using a plurality of <teacher data> during learning.

<Teacher data>
According to FIG. 1, the dialogue device 1 has a conjunctive particle table 100, first teacher data 101, and second teacher data 102 as teacher data. These teacher data need not be stored by the interactive device 1 themselves, but must be input from the outside during learning.

[Conjunctive particle table 100]
The conjunctive particle table 100 registers conjunctive particles that connect sentences before and after in a causal relationship. A "conjunctive particle" is a particle that builds a causal relationship between a preamble and a postscript, and serves as a clue to the causal relationship.
For example, there are adjectives such as:
"Because of,""From,""Becauseof,""Becauseof,"
``In the background of ``, ``In response to ``, ``As a result of ``, ``In the wake of ``,
"Influence of""Causeof""If you do""If you do"
"If you don't,""In line with,""Reflecting,"

[First teacher data 101]
The first training data 101 is a large corpus of structured natural language sentences accumulated on the Internet on a large scale. This is, for example, an encyclopedia such as Wikipedia (registered trademark), which is a valid set of sentences as a natural language.
For example, suppose we have the following sentences:
"I was absent from school because my child had a fever."

[Second teacher data 102]
The second training data 102 is a group of dialogue logs consisting of utterance sentences and response sentences. This is a valid sentence group for human dialogue recognition.
For example, suppose we have the following sentences:
Utterance sentence "My child has a fever."
Response sentence "Please take a break from school"

<When learning>
The dialogue device 1 includes, as functional units during learning, a learning sentence selection unit 11, a first machine learning engine 12, an utterance response sentence selection unit 13, a second machine learning engine 14, and a third machine learning engine. and an engine 15 . These functional components are implemented by executing a program that causes a computer installed in the device to function. In addition, the flow of processing of these functional components can also be understood as a learning method of the interactive device.

FIG. 2 is an explanatory diagram showing learning of the first machine learning engine and the second machine learning engine.

[Learning text selection unit 11]
The learning sentence selection unit 11 selects learning sentences containing the “conjunctive particles” registered in the conjunctional particle table 100 from the group of learning sentences to be the first teacher data 101 . As a result, only the learning sentences obtained by deleting the connecting particles from the learning sentences containing the connecting particles and connecting them are input to the first machine learning engine 12 .
According to FIG. 2, the following training text is input to the first machine learning engine 12:
1st teacher data: "I was absent from school because my child had a fever."
Learning sentence to be input "My child had a fever and was absent from school"

[First Machine Learning Engine 12]
The first machine learning engine 12 classifies the learning sentences input from the learning sentence selection unit 11 into morphemes by morphological analysis, and then learns them as “having causal relationships”.

FIG. 3 is a schematic functional configuration diagram of the first machine learning engine.
According to FIG. 3, the first machine learning engine 12 is a categorical neural network.
At the time of learning, the first machine learning engine 12 constructs a deep learning model that comprehensively extracts the features of the expressions of the entire sentence that have a causal relationship.
During operation, the first machine learning engine 12, when inputting an utterance response sentence with unknown causality, classifies it into morphemes by morphological analysis, and then outputs the result of classification into whether there is causality or not.

[Utterance Response Sentence Selection Unit 13]
The utterance response sentence selection unit 13 inputs the utterance sentence and the response sentence that form a pair of the dialogue corpus as the second teacher data 102 to the first machine learning engine 12 . The paired utterance sentence and response sentence are input to the first machine learning engine 12 as an utterance response sentence connected as they are.

According to FIG. 2, the speech response text selection unit 13 outputs the following speech response texts to the first machine learning engine 12 .
Spoken Response Sentence "My child has a fever, please take a day off from work."
In response to this, the first machine learning engine 12 replies to the speech response text selection unit 13 with a result indicating that there is or is not a causal relationship with respect to the speech response text.
Then, the utterance response text selection unit 13 outputs only the utterance texts and response texts estimated to have a causal relationship to the second machine learning engine 14 .
Utterance sentence "My child has a fever."
Response sentence "Please take a break from school"

[Second machine learning engine 14]
The second machine learning engine 14 divides the speech sentence estimated to have a causal relationship into morphemes by morphological analysis and inputs it to the encoder side. Learn by typing.

FIG. 4 is a schematic functional configuration diagram of the second machine learning engine and the third machine learning engine.
According to FIG. 4, the second machine learning engine 14 may be specifically Seq2Seq capable of extracting features between an utterance sentence and a response sentence that have a causal relationship, seq2seq+attention, or It may be an improved model such as transform.
seq2seq is a neural network that learns replacement rules for inputting a morpheme string and outputting another morpheme string. In this way, one set of response sentences is learned for each utterance sentence. Of course, for example, LSTM (Long Short-Term Memory), which is a type of RNN (Recurrent Neural Network) capable of learning the dependency relationship of character strings, may be used.
As a result, the second machine learning engine 14 outputs a response sentence from the decoder side when the utterance sentence is input to the encoder side.

[Third Machine Learning Engine 15]
The third machine learning engine 15 divides the spoken sentences into morphemes by morphological analysis and inputs them to the encoder side for the spoken sentences and the response sentences that are pairs of the dialogue corpus, which is the second teacher data 102, and inputs them to the encoder side. A sentence is segmented into morphemes by morphological analysis and input to the decoder for learning.
In other words, unlike the second machine learning engine 14, the third machine learning engine 15 is not limited to the causal relationship, but for all of the second teacher data 102, general purpose characteristics of the relationship between the utterance sentence and the response sentence. to learn.
On the other hand, the third machine learning engine 15, like the second machine learning engine 14, may be Seq2Seq capable of extracting features between an utterance sentence and a response sentence, or may be a seq2seq+attention or transform engine. It may be an improved model such as
As a result, when the speech sentence is input to the encoder side, the third machine learning engine 15 outputs a response sentence from the decoder side.

<During operation>
[Dialogue part 16]
The dialog device 1 has a dialog section 16 as a functional section during operation.

The dialogue unit 16 receives the user's utterance sentence from the terminal 2 and outputs the utterance sentence to the second machine learning engine 14 . Then, it acquires a response sentence from the second machine learning engine 14 and transmits the response sentence to the terminal 2 . If the user interface function of the terminal 2 allows key input and display, the dialogue unit 16 can transmit and receive text-based utterance sentences and response sentences to and from the terminal 2 .

The dialogue unit 16 may have a function of recognizing a user's uttered voice and a function of converting a response sentence to a user's voice according to the user interface function of the terminal 2 . The voice recognition function converts the user's uttered voice acquired by the microphone of the terminal 2 into a text-based uttered sentence. The speech conversion function converts the response sentence obtained from the second machine learning engine 14 into a speech signal.
In this case, the dialogue unit 16 receives the user's speech data from the terminal 2 , converts the speech data into a text utterance, and inputs the utterance to the second machine learning engine 14 . It also receives a response sentence from the second machine learning engine 14 , converts the response sentence into voice data, and transmits the voice data to the terminal 2 .

FIG. 5 is a second functional configuration diagram of the interactive device during operation according to the second embodiment.

According to FIG. 2 described above, the dialogue unit 16 uses only the second machine learning engine 14 to acquire a response sentence to the utterance sentence.
On the other hand, according to FIG. 5, the dialogue device 1 uses the first machine learning engine 12 and the second machine learning engine 14 to function as functional units during operation in addition to the dialogue unit 16. It has a sentence generator 17 and a causal relationship estimator 18 . These functional components are also implemented by executing a program that causes a computer installed in the device to function. In addition, the processing flow of these functional components can also be understood as an operation method of the interactive device.

FIG. 6 is an explanatory diagram showing dialogue during operation in the second embodiment.

[Utterance Response Sentence Generation Unit 17]
The utterance response sentence generation unit 17 generates an utterance response sentence by connecting the user's utterance sentence input from the dialogue unit 16 and the response sentence obtained by inputting the utterance sentence to the second machine learning engine 14. Generate.

According to FIG. 6, it is assumed that the dialogue unit 16 receives the following utterance sentence from the terminal 2 .
Utterance sentence "My child has a fever."
The dialogue unit 16 inputs this utterance sentence to the utterance response sentence generation unit 17 .
The utterance response text generation unit 17 inputs the following input utterance text to the second machine learning engine 14 .
Utterance sentence "My child has a fever."
In response to this, the second machine learning engine 14 estimates a response sentence to the utterance sentence and returns the following response sentence.
Response sentence "Please take a break from work"
On the other hand, the utterance response text generation unit 17 connects the utterance text and the response text to generate the following utterance response text.
Spoken Response Sentence "My child has a fever, please take a day off from work."
This speech response text is input to the causality estimation unit 18 .

[Causal relationship estimation unit 18]
The causal relationship estimation unit 18 estimates whether or not there is a causal relationship by inputting the generated utterance response text into the first machine learning engine 12, and extracts the response sentence of the utterance response text that is estimated to have a causal relationship. is output to the dialogue unit 16 .

According to FIG. 6 , the causal relationship estimating unit 18 inputs the speech response text input from the speech response text generating unit 17 to the first machine learning engine 12 .
Spoken Response Sentence "My child has a fever, please take a day off from work."
On the other hand, the first machine learning engine 12 estimates whether or not there is a causal relationship, and here, the existence of a causal relationship is output to the causal relationship estimating section 18 .
At this time, since there is a causal relationship, the causal relationship estimating unit 18 outputs the response sentence of the speech response sentence to the dialogue unit 16 .
Response sentence "Please take a break from work"
By this, the dialog part 16 transmits the response sentence to the terminal 2. FIG.

FIG. 7 is a second functional block diagram of the interactive device during operation according to the third embodiment.

FIG. 7 further uses a third machine learning engine 15 compared to FIG.
The causality estimation unit 18 inputs the utterance sentences in the utterance response sentences estimated as “no causality” by the first machine learning engine 12 to the third machine learning engine 15 .
On the other hand, the third machine learning engine 15 estimates a response sentence for the utterance sentence and outputs the response sentence to the causality estimation unit 18 .
The causality estimation unit 18 then outputs the response sentence output from the third machine learning engine 15 to the dialogue unit 16 .
In this way, the causal relationship estimation unit 18 uses the third machine learning engine 15 as a general-purpose model to generate diverse response sentences even when there is no causal relationship as a response sentence to an utterance sentence. can be output.

FIG. 8 is an explanatory diagram showing dialogue during operation in the third embodiment.

According to FIG. 8, it is assumed that the following utterance sentence is input from the dialogue unit 16 to the utterance response sentence generation unit 17.
Utterance sentence "My child has a fever."
In response to this, the utterance response sentence generation unit 17 inputs the utterance sentence to the second machine learning engine 14 and obtains the following response sentence.
Response sentence "I had lunch at work"
Then, the utterance response text generation unit 17 outputs the following utterance response text connecting the utterance text and the response text to the causal relationship inferring unit 18 .
Spoken Response Sentence "My child has a fever, I had lunch at work."
On the other hand, the causality estimator 18 inputs the speech response text to the first machine learning engine 12, and it is estimated that there is no causality.
On the other hand, the causal relationship estimation unit 18 inputs the following utterance sentence in the utterance response sentence estimated to have no causal relationship to the third machine learning engine 15 .
Utterance sentence "My child has a fever."
In response to this, the causality estimation unit 18 obtains the following response sentence from the third machine learning engine 15 .
Response sentence "Do you work overtime?"
The causality estimation unit 18 then outputs the response sentence to the dialogue unit 16 .

As described in detail above, according to the interactive device, program, and method of the present invention, it is possible to estimate a response sentence that considers causal relationships as much as possible with respect to a user's uttered sentence.
In particular, according to the present invention, since the characteristics of the overall expression of the context including not only word pairs with causal relationships but also word groups around them are comprehensively learned, the response sentence to the utterance sentence is the user's sentence. feels as natural as possible for In addition, according to the present invention, even sentences that do not have a causal relationship, even if they are not in the causal dictionary, are comprehensively learned, so that a response sentence can be returned in a general purpose manner.
That is, according to the present invention, by comprehensively learning the characteristics of the overall expression of a sentence, various and versatile response sentences can be generated regardless of whether or not there is a causal relationship with the utterance sentence. can reply.

Finally, sample examples of utterance sentences and response sentences showing the difference between the non-patent document 1 and non-patent document 2 of the prior art and the present invention will be described.
<Response sentence with causal relationship to utterance sentence>
(In the case of Non-Patent Document 1)
Spoken text: "The yen has depreciated."
Refer to a dictionary of causality pairs {(the yen will depreciate) -> (the economy will rise)} Response sentence "The economy will rise."
(in the case of the present invention)
{Comprehensive learning of the features of the entire causal context}
Spoken text: "The yen has depreciated."
Response: "From the perspective of trade, we can expect the Japanese economy to rise."

<Response sentence with no causal relationship to the utterance sentence>
(In the case of Non-Patent Document 2)
Utterance sentence "I want to go to a restaurant"
{only backtracking due to lack of causality dictionary}
Response sentence "Is that so?"
(in the case of the present invention)
{Comprehensive learning of the features of the entire causal context}
Spoken text: "The yen has depreciated."
Response sentence "Do you always eat out?"

For the various embodiments of the present invention described above, various changes, modifications and omissions within the spirit and scope of the present invention can be easily made by those skilled in the art. The foregoing description is exemplary only and is not intended to be limiting. The invention is to be limited only as limited by the claims and the equivalents thereof.

1 Dialogue Device 100 Conjunctive Particle Table 101 First Teacher Data 102 Second Teacher Data 11 Learning Sentence Selector 12 First Machine Learning Engine 13 Utterance Response Sentence Selector 14 Second Machine Learning Engine 15 Third Machine Learning Engine 16 Dialog Unit 17 Utterance Response Sentence Generation Unit 18 Causal Relationship Estimation Unit 2 Terminal

Claims (7)

In a program that operates a computer installed in a dialogue device that generates a response sentence to a user's utterance sentence,
when learning,
a conjunctive particle table in which conjunctive particles that connect the preceding and following sentences in a causal relationship are registered;
learning sentence selection means for selecting learning sentences containing the conjunction particles registered in the conjunction particle table from the group of learning sentences serving as the first training data;
a classification-type first machine learning engine that learns learning sentences containing conjunctive particles as having a causal relationship;
Utterance response sentence selection means for inputting an utterance sentence and a response sentence paired in a dialogue corpus as second training data to a first machine learning engine and selecting an utterance sentence and a response sentence presumed to have a causal relationship. and,
Functioning as a second machine learning engine that inputs an utterance sentence estimated to have a causal relationship to the encoder side and inputs a response sentence to the decoder side for learning,
During operation,
A program characterized by inputting a user's utterance sentence to a second machine learning engine and causing a computer to function as dialogue means for obtaining a response sentence. During operation,
an utterance response sentence generation means for generating an utterance response sentence by connecting a user utterance sentence input from the dialogue means and a response sentence obtained by inputting the utterance sentence to a second machine learning engine;
Presence or absence of a causal relationship is estimated by inputting the generated utterance response text into a first machine learning engine, and a causal relationship in which a response sentence of the utterance response text presumed to have a causal relationship is output to the dialogue means. 2. The program according to claim 1, further causing a computer to function as estimation means. The first machine learning engine is a categorical neural network capable of extracting the characteristics of the entire sentence that has a causal relationship,
3. The program according to claim 1, wherein the second machine learning engine causes the computer to function as a Seq2Seq capable of extracting features between an utterance sentence and a response sentence that have a causal relationship. when learning,
Functioning as a third machine learning engine for learning by inputting the utterance sentence to the encoder side and inputting the response sentence to the decoder side for the utterance sentence and the response sentence that are paired in the dialogue corpus as the second training data,
During operation,
When the first machine learning engine estimates that there is no causal relationship, the causality estimating means inputs the utterance sentence of the utterance response sentence to the third machine learning engine and outputs the obtained response sentence to the dialogue means. 3. A program according to claim 2 , which causes a computer to function to do. The first machine learning engine is a categorical neural network capable of extracting the characteristics of the entire sentence that has a causal relationship,
The second machine learning engine is Seq2Seq that can extract features between utterance sentences and response sentences that have a causal relationship,
5. The program according to claim 4, wherein the third machine learning engine causes the computer to function as a Seq2Seq capable of extracting features between generic utterance sentences and response sentences. In a dialogue device that generates a response sentence to a user's utterance sentence,
when learning,
a conjunctive particle table in which conjunctive particles that connect the preceding and following sentences in a causal relationship are registered;
learning sentence selection means for selecting learning sentences containing the conjunction particles registered in the conjunction particle table from the group of learning sentences serving as the first training data;
a classification-type first machine learning engine that learns learning sentences containing conjunctive particles as having a causal relationship;
Utterance response sentence selection means for inputting an utterance sentence and a response sentence paired in a dialogue corpus as second training data to a first machine learning engine and selecting an utterance sentence and a response sentence presumed to have a causal relationship. and,
Functioning as a second machine learning engine that inputs an utterance sentence estimated to have a causal relationship to the encoder side and inputs a response sentence to the decoder side for learning,
During operation,
1. A dialogue device, comprising dialogue means for inputting a user's uttered sentence to a second machine learning engine and obtaining a response sentence. In the interaction method for a device that generates a response sentence to a user's utterance sentence,
The device
when learning,
a conjunctive particle table in which conjunctive particles that connect the preceding and following sentences in a causal relationship are registered;
A classification-type first machine that selects learning sentences containing a conjunction particle registered in a conjunction particle table from a group of learning sentences serving as first teacher data, and learns the learning sentences containing the conjunction particles as having a causal relationship. a learning engine;
A pair of utterance sentences and response sentences in the dialogue corpus, which serves as second training data, is input to the first machine learning engine, and utterance sentences and response sentences that are presumed to have a causal relationship are selected, and judged to have a causal relationship. a second machine learning engine that inputs the estimated utterance sentence to the encoder side and inputs the response sentence to the decoder side for learning;
During operation,
A device interaction method, characterized by inputting a user's utterance sentence into a second machine learning engine and executing it so as to obtain a response sentence.

Images