JP7058438B2 - Dialogue response system, model learning device and dialogue device - Google Patents

Description

The present invention relates to a dialogue response system, a model learning device and a dialogue device.

In recent years, many computers (dialogue response systems) that can interact with humans have appeared, but most of them respond to user utterances by manually creating response sentences using manually written response sentences or templates. It's just outputting. If a sentence that can be acquired from the Web (hereinafter referred to as a Web acquisition sentence) can be used for dialogue, the cost will be significantly reduced in the development of the dialogue response system, but at present, the Web acquisition sentence is effectively used in the development of the dialogue response system. not. The reason is that the Web acquisition sentence is often incomplete or unnatural as a response sentence in a dialogue and causes a sense of incongruity. For example, in the web acquisition sentence, there are many news titles that end with a noun such as "automobiles start production of engines in Thailand", and contexts such as "glass frames of these brands have a deep shine". There are sentences that do not hold without them, or sentences that contain inappropriate or redundant parts for utterances. Although the Web-acquired sentence itself is not strange, it often causes a sense of discomfort as a response to the utterance. Therefore, in order to use the Web acquisition statement in the dialogue response system, it is necessary to convert it into an appropriate form as the response statement.

Conventionally, there have been known techniques for expressing characters in a dialogue response system or changing the meaning of a sentence by converting the ending of a word. For example, as a character expression technique by flexion conversion, a technique of converting the flexion "da-masu" to "da-desu" and a technique of converting the flexion to Kansai dialect are known. As a technique to change the meaning of the sentence, for example, a technique to convert the sentence "Eat a hamburger tonight" into a suggestion-like ending "Let's eat a hamburger tonight", and a questionable technique "Do you eat a hamburger tonight?" The technique of converting to is known. However, these techniques cannot be converted into an appropriate form as a response to a user's utterance. For example, the response to the utterance "Varied tit is a wild bird known for hoarding over the winter" is "Good weather, few clouds, and you can see wild birds such as varied tit and thrush around the lake." In the case of ".", Even if the ending of the response sentence is converted, the response sentence is not appropriate for the utterance. This is because "good weather and few clouds" in the response text does not make sense, so at least this "good weather and few clouds" needs to be deleted.

Patent Document 1 describes a method for accurately converting a respected expression into a normal expression. By this method, for example, the ending "desu" or "masu" can be converted to "da".

Patent Document 2 describes a method of paraphrasing the expression according to the amount of knowledge of the interlocutor. In this way, for example, when responding to someone who is not familiar with wine, "I like Beaujolais" can be rephrased as "I like red wine".

Japanese Unexamined Patent Publication No. 2017-68879 Japanese Unexamined Patent Publication No. 2017-91368

In Patent Document 1, the respected expression can be converted into the normal expression, but the conversion is limited to the flexion. Also, it is not possible to determine when and which flexion is appropriate. In Patent Document 2, the expression is paraphrased according to the amount of knowledge of the interlocutor, but there are many cases where even if the knowledge levels match, it is not appropriate as a response sentence. For example, in Patent Document 2, unnecessary parts such as "good weather and few clouds" in the above-mentioned response sentence cannot be deleted.

Therefore, it is an object of the present invention to provide a dialogue response system, a model learning device, and a dialogue device capable of responding to an utterance in a dialogue-like manner.

The dialogue response system of the present invention includes a model learning device and a dialogue device. In the model learning device, the learning utterance sentence, the learning related sentence related to the learning utterance sentence obtained by searching the sentence in the text database, and the learning related sentence are appropriate as the response of the dialogue. A corpus storage unit that stores a learning corpus including a learning response sentence modified in various formats, and a response sentence generation model that includes related sentences in the input and outputs the response sentence using the learning corpus. It is equipped with a model learning unit for learning. The dialogue device acquires a dialogue execution related sentence related to the dialogue execution speech sentence by searching the speech sentence input unit in which the dialogue execution speech sentence is input and the sentence in the text database. Dialogue by inputting the relational sentence for dialogue execution into the response sentence generation model trained to include the acquisition part and the related sentence related to the spoken sentence as input and output the response sentence in response to the input. It is provided with a response statement generation unit that generates a response statement for execution.

The dialogue response system of the present invention is a dialogue response system including a model learning device and a dialogue device. The model learning device stores a learning corpus containing the utterances for learning, the related sentences related to the utterances for learning obtained by searching the sentences in the text database, and the suitability of the related sentences. It includes a corpus storage unit and a first learning processing unit that uses a learning corpus to learn a suitability determination model that includes utterance sentences and related sentences as inputs and outputs the degree of conformity. The dialogue device includes an utterance sentence input unit in which an utterance sentence for executing a dialogue is input, a related sentence acquisition unit that acquires a related sentence related to the utterance sentence for executing a dialogue by searching a sentence in a text database, and a related sentence acquisition unit. By inputting the spoken sentence for dialogue execution and the acquired related sentence into the conformity judgment model trained to include the spoken sentence and the related sentence in the input and output the conformity of the related sentence in response to the input. , A related sentence conformity determination unit that generates a degree of conformity is provided.

The dialogue response system of the present invention is a dialogue response system including a model learning device and a dialogue device. The model learning device includes a utterance sentence for learning, a search topic used when searching a sentence in a text database to obtain a related sentence related to the utterance sentence for learning, and a suitability of the search topic. It includes a corpus storage unit that stores a learning corpus, and a first learning processing unit that uses a learning corpus to learn a conformity determination model that includes spoken sentences and search topics in input and outputs conformity. .. The dialogue device has a utterance text input unit in which the utterance text for dialogue execution is input, and a sentence in the text database to acquire related sentences related to the utterance text for dialogue execution based on the utterance text for dialogue execution. A first topic generator that generates the search topic used when searching for, and a suitability learned to include utterances and search topics in the input and output the suitability of the search topic in response to the input. A topic suitability judgment unit that generates a degree of conformity by inputting an utterance sentence for dialogue execution and a generated search topic into the judgment model, and a plurality of suitability based on the suitability generated by the topic suitability judgment unit. By searching the text database sentences using the selection section that selects one search topic from the search topics and the selected search topic, the related sentences related to the utterance sentences for executing the dialogue are acquired. It is provided with a related sentence acquisition unit of 1.

According to the present invention, it is possible to use the data acquired from the text database as a high-quality response sentence. As a result, the range of use of the dialogue response system is greatly expanded.

It is a figure which shows the structure of the dialogue response system 1 of 1st Embodiment. It is a figure which shows the example of the utterance sentence for learning, the related sentence for learning, and the response sentence for learning of 1st Embodiment. It is a figure which shows the example of the utterance sentence for dialogue execution, the relational sentence for dialogue execution, and the response sentence for dialogue execution of 1st Embodiment. It is a figure which shows the structure of the response generation learning processing part 6 of 1st Embodiment. It is a figure which shows the structure of the response sentence generation part 9 of 1st Embodiment. It is a flowchart which shows the operation procedure of the model learning apparatus 2 of 1st Embodiment. It is a flowchart which shows the operation procedure of the dialogue apparatus 3 of 1st Embodiment. It is a figure which shows the structure of the dialogue response system 21 of 2nd Embodiment. It is a figure which shows the example of the utterance sentence for learning, the related sentence for learning, and the response sentence for learning of the 2nd Embodiment. It is a figure which shows the example of the utterance sentence for dialogue execution, the relational sentence for dialogue execution, and the response sentence for dialogue execution of the 2nd Embodiment. It is a figure which shows the structure of the response generation learning processing part 26 of 2nd Embodiment. It is a figure which shows the structure of the response sentence generation part 29 of the 2nd Embodiment. It is a flowchart which shows the operation procedure of the model learning apparatus 22 of 2nd Embodiment. It is a flowchart which shows the operation procedure of the dialogue apparatus 23 of 2nd Embodiment. It is a figure which shows the structure of the dialogue response system 31 of 3rd Embodiment. It is a figure which shows the example of the dialogue data of 3rd Embodiment, the utterance sentence for learning, the related sentence for learning, and the response sentence for learning. It is a flowchart which shows the operation procedure of the model learning apparatus 32 of 3rd Embodiment. It is a figure which shows the structure of the dialogue response system 41 of 4th Embodiment. It is a figure which shows the example of the original sentence stored in the database 1054. It is a figure which shows the example of the sentence used by the related sentence acquisition part 42 of the model learning apparatus 42. It is a figure which shows the example of the sentence used by the related sentence acquisition part 48 of the dialogue apparatus 43. It is a flowchart which shows the operation procedure of the model learning apparatus 42 of 4th Embodiment. It is a flowchart which shows the operation procedure of the dialogue apparatus 43 of 4th Embodiment. It is a figure which shows the structure of the dialogue response system 81 of 5th Embodiment. It is a figure which shows the structure of the response generation learning processing part 86 of 5th Embodiment. It is a figure which shows the structure of the dialogue response system 91 of 6th Embodiment. It is a figure which shows the structure of the response generation learning processing part 96 of 6th Embodiment. It is a figure which shows the structure of the dialogue response system 551 of the 7th Embodiment. It is a figure which shows the structure of the dialogue response system 1100 of 8th Embodiment. It is a figure for demonstrating an example of a nonconformity-related sentence. It is a figure which shows the example of the learning corpus stored in the corpus storage part 1108. It is a figure showing the neural network which realizes the conformity determination model X1 used in the conformity determination learning processing unit 1105 and the related sentence conformity determination unit 1101 of the eighth embodiment. It is a figure which shows the detail of the convolution operation in the convolutional neural network CNN1. It is a figure which shows the detail of the convolution operation in the convolutional neural network CNN2. It is a flowchart which shows the operation procedure of the model learning apparatus 1120 of 8th Embodiment. It is a flowchart which shows the operation procedure of the dialogue apparatus 1130 of 8th Embodiment. It is a figure which shows the structure of the dialogue response system 1200 of 9th Embodiment. It is a figure which shows the example of the dialogue data of 9th Embodiment, the utterance sentence for learning, the related sentence for learning, and the response sentence for learning. It is a flowchart which shows the operation procedure of the model learning apparatus 1220 of 9th Embodiment. It is a figure which shows the structure of the dialogue response system 1300 of a tenth embodiment. It is a figure which shows the example of the speech sentence for learning, the search topic, the related sentence, and the goodness of fit. (A) is a figure showing an example of a plurality of sets of learning-related sentences and learning response sentences in the learning corpus stored in the corpus storage unit 1308. (B) is a figure showing an example of a plurality of sets of a learning utterance sentence, a search topic, and a goodness of fit among the learning corpora stored in the corpus storage unit 1308. It is a figure which shows the neural network which realizes the conformity determination model X2 used in the conformity determination learning processing unit 1305 and the topic conformity determination unit 1312 of the tenth embodiment. It is a figure which shows the detail of the convolution operation in the convolutional neural network CNN3. It is a figure which shows the detail of the convolution operation in the convolutional neural network CNN4. It is a flowchart which shows the operation procedure of the model learning apparatus 1320 of 10th Embodiment. It is a flowchart which shows the operation procedure of the dialogue apparatus 1330 of a tenth embodiment. It is a figure which shows the internal structure of a model learning device or a dialogue device.

Hereinafter, embodiments will be described with reference to the drawings.
[First Embodiment]
FIG. 1 is a diagram showing the configuration of the dialogue response system 1 of the first embodiment. The dialogue response system 1 includes a model learning device 2 and a dialogue device 3.

The model learning device 2 includes a corpus generation unit 4, a corpus storage unit 5, a response generation learning processing unit 6, and a parameter storage unit 14. The corpus generation unit 4 includes an utterance sentence input unit 11, a related sentence acquisition unit 12, a learning response sentence input unit 13, and a corpus output unit 64.

The utterance sentence input unit 11 receives the input of the utterance sentence for learning created by the personnel on the development side.
The related sentence acquisition unit 12 acquires a sentence related to the utterance sentence for learning by searching the text database, and sets it as a related sentence for learning. For example, the related sentence acquisition unit 12 acquires a sentence related to a learning utterance sentence by searching a sentence posted on Web1000, which is one of the text databases, and sets it as a learning related sentence. The related sentence acquisition unit 12 acquires, for example, one sentence containing the most nouns included in the learning utterance sentence as a learning related sentence by searching the sentence posted on the Web 1000. The related sentence acquisition unit 12 acquires learning related sentences by manually or automatically searching the data on the Web 1000 by using a search engine or the like.

A sentence posted on Web1000 is a sentence posted on Web1000 at the time of acquisition of a related sentence (first case), or a sentence posted on Web1000 before the time of acquisition of a related sentence, and then a database. It is a sentence (second case) taken in. In the first case, the related sentence acquisition unit 12 directly searches the Web 1000. In the second case, the related sentence acquisition unit 12 searches the database.

The learning response sentence input unit 13 receives the input of the learning response sentence. The learning response sentence is generated by the development personnel by modifying the learning related sentence according to a predetermined rule so as to be appropriate as a response of the dialogue to the utterance. Appropriate as a dialogue response to an utterance is, for example, to maintain an important part of the dialogue response, to include a flexion that seems to be the dialogue response, to include no redundant part of the dialogue response, and to have no context. It does not include parts that do not hold, or contains particles that seem to be the response of the dialogue, and the word order that seems to be the response of the dialogue. Since the learning response sentence is created by the personnel on the development side, a high-quality learning response sentence can be obtained.

The corpus output unit 64 generates a learning corpus including a plurality of sets of learning-related sentences and learning response sentences, and writes them in the corpus storage unit 5. The corpus storage unit 5 stores a learning corpus.

The response generation learning processing unit 6 learns the response sentence generation model A in which the related sentence is input and the response sentence is output by using the learning corpus. The response sentence generation model A includes a neural network for deep learning and its parameters.

The parameter storage unit 14 stores the parameters after learning of the response sentence generation model A by the response generation learning processing unit 6. The parameter storage unit 14 stores parameters (weight matrix, etc.) after learning the neural network. The parameters after learning of the response sentence generation model A stored in the parameter storage unit 14 can be taken out to the outside via an input / output port or the like.

FIG. 2 is a diagram showing an example of a learning utterance sentence, a learning related sentence, and a learning response sentence of the first embodiment. As shown in Fig. 2, a sentence containing the noun "laundry" contained in the learning utterance sentence "I have to take in the laundry because I've been popping up" is searched from Web1000, and the learning-related sentence "In the bathtub warmer". "Best answer to dry laundry" is obtained. Then, the redundant part "quickness best answer" is deleted as the response of the dialogue included in the related sentence for learning, and the non-dialogue flexion "dry" is replaced with the dialogue-like flexion "dry". The response sentence "Let's dry the laundry with a bathtub warmer." Is created. By learning using such a learning-related sentence and a learning response sentence, a response sentence generation model A that generates an appropriate response sentence as a dialogue is generated from the related sentence.

The dialogue device 3 includes an utterance sentence input unit 7, a related sentence acquisition unit 8, a parameter storage unit 15, and a response sentence generation unit 9.

The utterance sentence input unit 7 receives input of an utterance sentence for dialogue execution from the user.
The related sentence acquisition unit 8 acquires a sentence related to the utterance sentence for dialogue execution by searching the text database, and sets it as a related sentence for dialogue execution. For example, the related sentence acquisition unit 8 acquires a sentence related to the utterance sentence for dialogue execution by searching the sentence posted on Web1000, which is one of the text databases, and sets it as the related sentence for dialogue execution. .. The related sentence acquisition unit 8 acquires, for example, one sentence including a noun included in the dialogue execution utterance sentence as the dialogue execution related sentence by searching the sentence posted on the Web 1000. The related sentence acquisition unit 8 acquires the related sentence for dialogue execution by the user's operation or automatically searching the data on the Web 1000 by using a search engine or the like.

The parameter storage unit 15 stores the parameters after learning the response sentence generation model A by the model learning device 2. For example, the parameter storage unit 15 stores parameters (weight matrix, etc.) after learning the neural network. The parameters after learning of the response sentence generation model A taken out from the model learning device 2 are input to the dialogue device 3 and stored in the parameter storage unit 15. These parameters may be distributed to the user together with the program that realizes the function of the dialogue device 3.

The response sentence generation unit 9 generates a dialogue execution response sentence by inputting the dialogue execution related sentence into the response sentence generation model A after learning. For example, the response sentence generation unit 9 can input a dialogue execution-related sentence into the neural network after learning, and use the output of the neural network as the dialogue execution response sentence.

FIG. 3 is a diagram showing an example of a dialogue execution utterance sentence, a dialogue execution related sentence, and a dialogue execution response sentence according to the first embodiment. As shown in FIG. 3, a sentence including "glasses frame" included in the dialogue execution utterance sentence "After all, the brand's spectacle frame is good." Is searched from Web1000, and the dialogue execution related sentence "such". The brand's eyeglass frames have a deep shine. " Then, the part "such" that cannot be established without the context before and after it is included in the related sentence for dialogue execution is deleted, and the ending "have" that is not like dialogue is replaced with the ending "have" that seems to be dialogue. The response statement "Branded eyeglass frames have a deep shine." Is generated.

FIG. 4 is a diagram showing the configuration of the response generation learning processing unit 6 of the first embodiment. The response generation learning processing unit 6 includes an encoder 151 and a decoder 152. The response generation learning processing unit 6 learns the Sequence to Sequence model, which is an example of the response sentence generation model A.

The encoder 151 includes a recurrent neural network developed along a time axis (t = 1 to m). Let the input vector at time t be x1 (t) and the vector of the hidden layer be h1 (t).

The hidden layer at time t receives an input vector x1 (t) representing a word of a learning-related sentence input to the input layer at time t, and a vector h1 (t-1) of the hidden layer at time (t-1). ), And the vector h1 (t) is output to the hidden layer at the time (t + 1). However, the hidden layer at time 1 receives only the input vector x1 (1). For example, the input vector x1 (1) at time 1 is a vector representing "Kyoto", and the input vector x1 (2) at time 2 is a vector representing "de". In this embodiment, a so-called word embedding vector (for example, one obtained by word2vector) or the like may be used as a vector representing a word.

When the weight matrix at time t is U1 (t) and W1 (t), the bias is b, and the activation function is f, the following equation holds.

h1 (t) = f (U1 (1) x1 (1) + b)… (1)
h1 (t) = f (U1 (t) x1 (t) + W1 (t) h1 (t-1) + b) (other than t = 1)… (2)
The decoder 152 includes a recurrent neural network developed along the time axis (t = 1 to n). The input vector at time t is x3 (t), the hidden layer vector is h3 (t), and the output vector is y3 (t).

The hidden layer at time t receives the input vector x3 (t) representing the word of the learning response sentence input to the input layer at time t, and the vector h3 (t-1) of the hidden layer at time (t-1). ), And the vector h3 (t) is output to the output layer and the hidden layer at the time (t + 1). However, the hidden layer at time 1 receives the input vector x3 (1) and the vector h1 (m) of the hidden layer at the last time m of the encoder 151. For example, the input vector x3 (1) at time 1 is a vector representing a dummy word (<s>), and the input vector x3 (2) at time 2 is a vector representing the word “Kyoto”. The output layer at time t receives the vector h3 (t) from the hidden layer at time t and outputs the output vector y3 (t). The output vector y3 (t) is a prediction of the input vector x3 (t + 1) at the time (t + 1).

When the weight matrix at time t is U3 (t), W3 (t), V3 (t), the bias is b, c, and the activation function is f, g, the following equation holds.

h3 (1) = f (U3 (1) x3 (1) + W3 (1) h1 (m) + b)… (3)
h3 (t) = f (U3 (t) x3 (t) + W3 (t) h3 (t-1) + b) (other than t = 1) ... (4)
y3 (t) = g (V3 (t) h3 (t) + c)… (5)
In the decoder 152, an error between the output sentence composed of the output vectors y3 (1) to y3 (n) and the learning response sentence (teacher data) composed of x3 (1) to x3 (n) is obtained. The error E (t) between the output vector y3 (t) at time t and the input vector x3 (t + 1) is obtained. However, at time t = n, an error E (n) between the output vector y3 (n) and the input vector x3 (1) (dummy word (<s>)) at time 1 is obtained. These errors are back-propagated to the hidden layer before the time t in the decoder 152 and to the hidden layer (time t = 1 to m) in the encoder 151. The weight matrices W3 (t), U3 (t), and V3 (t) are updated so as to minimize the sum of the errors back-propagated to the hidden layer at time t in the decoder 152. W1 (t) and U1 (t) are updated so as to minimize the sum of the errors back-propagated to the hidden layer at time t in the encoder 151. The weight matrices W3 (t), U3 (t), V3 (t), W1 (t), and U1 (t), which are the parameters of the response sentence generation model A after learning, are stored in the parameter storage unit 14.

FIG. 5 is a diagram showing the configuration of the response sentence generation unit 9 of the first embodiment. The response sentence generation unit 9 includes an encoder 153 and a decoder 154. The response sentence generation unit 9 generates a response sentence for dialogue execution by using the sequence to sequence model after learning, which is an example of the response sentence generation model A after learning.

In FIG. 5, the weight matrices W3 (t), U3 (t), V3 (t), W1 (t), and U1 (t), which are the parameters after learning the response sentence generation model A, are read from the parameter storage unit 15. And set. As these, those obtained by learning (update) by the response generation learning processing unit 6 of FIG. 4 are used.

The encoder 153 includes a recurrent neural network developed along a time axis (t = 1 to m) similar to the encoder 151. However, the input vector x1 (t) input to the input layer at time t is not a word of the learning-related sentence but a word of the dialogue execution-related sentence.

The decoder 154 includes a recurrent neural network developed along the time axis (t = 1 to n). The input vector at time t is x3 (t), the hidden layer vector is h3 (t), and the output vector is y3 (t).

The hidden layer at time t receives the output vector y3 (t-1) at time (t-1) and the vector h3 (t-1) of the hidden layer at time (t-1), and receives the vector h3 (t-1). ) Is output to the output layer and the hidden layer at the time (t + 1). However, the hidden layer at time 1 receives the vector D representing the dummy word and the vector h1 (m) of the hidden layer at the last time m of the encoder 151. The output layer at time t receives the vector h3 (t) from the hidden layer at time t and outputs the output vector y3 (t).

When the weight matrix at time t is U3 (t), W3 (t), V3 (t), the bias is b, c, and the activation function is f, g, the following equation holds.

h3 (1) = f (D + W3 (1) h1 (m) + b)… (6)
h3 (t) = f (U3 (t) y3 (t-1) + W3 (t) h3 (t-1) + b) (other than t = 1) ... (7)
y3 (t) = g (V3 (t) h3 (t) + c)… (8)
FIG. 6 is a flowchart showing the operation procedure of the model learning device 2 of the first embodiment.

In step S101, the utterance sentence input unit 11 receives the input of the utterance sentence for learning.
In step S102, the related sentence acquisition unit 12 acquires a learning-related sentence related to the learning utterance sentence from Web 1000, which is one of the text databases.

In step S103, the learning response sentence input unit 13 modifies the learning-related sentence by the operation of the personnel on the development side so that the learning-related sentence becomes appropriate as the response of the dialogue, and the learning response is generated. Receive a sentence input.

In step S104, steps S101 to S103 are repeated until all the inputs are completed, so that a plurality of sets of learning-related sentences and learning response sentences are obtained. The corpus output unit 64 generates a learning corpus including a plurality of sets of learning-related sentences and learning response sentences, and writes them in the corpus storage unit 5. The corpus storage unit 5 stores a learning corpus.

In step S105, the response generation learning processing unit 6 learns the response sentence generation model A in which the related sentence is input and the response sentence is output by using the learning corpus.

In step S106, the parameter storage unit 14 stores the parameters after learning of the response sentence generation model A by the response generation learning processing unit 6.

FIG. 7 is a flowchart showing the operation procedure of the dialogue device 3 of the first embodiment.
In step S201, the response sentence generation unit 9 reads out the parameters after learning of the response sentence generation model A by the model learning device 2 stored in the parameter storage unit 15.

In step S202, the utterance sentence input unit 7 receives input of the utterance sentence for dialogue execution from the user.

In step S203, the related sentence acquisition unit 8 acquires the dialogue execution related sentence related to the dialogue execution utterance sentence from Web1000, which is one of the text databases.

In step S204, the response sentence generation unit 9 generates a dialogue execution response sentence by inputting a dialogue execution related sentence into the post-learning response sentence generation model A obtained by using the read parameters.

As described above, according to the present embodiment, the related sentence which is a response candidate obtained based on the utterance can be converted into an appropriate response sentence. As a result, a huge number of sentences that can be acquired on the Web can be used as the response of the dialogue response system.

[Second Embodiment]
FIG. 8 is a diagram showing the configuration of the dialogue response system 21 of the second embodiment. The dialogue response system 21 includes a model learning device 22 and a dialogue device 23. The model learning device 22 of the present embodiment differs from the model learning device 2 of the first embodiment in that the learning response sentence input unit 38 and the corpus output unit 74 included in the corpus generation unit 24, and the corpus storage. A unit 25 and a response generation learning processing unit 26.

The learning response sentence input unit 38 receives the input of the learning response sentence. The learning response sentence is created by modifying the learning-related sentence according to a predetermined rule so as to be appropriate as a response of the dialogue to the current utterance by the operation of the development side personnel. Appropriate as a response to a dialogue to this utterance is appropriate as a response to a general dialogue as described in the first embodiment (including a dialogue-like ending, not including a redundant part as a dialogue, and contexts before and after. Not only does it not include parts that cannot be established without it, or it contains particles that seem to be dialogue), but it is also appropriate as a response to the utterance sentence entered this time (for example, it does not include unnatural parts).

The corpus output unit 74 generates a learning corpus including a plurality of sets of learning utterance sentences, learning related sentences, and learning response sentences, and writes them in the corpus storage unit 25. The corpus storage unit 25 stores the learning corpus.

The response generation learning processing unit 26 learns the response sentence generation model B in which the utterance sentence and the related sentence are input and the response sentence is output by using the learning corpus. The response sentence generation model B is different from the response sentence generation model A in that the utterance sentence is also input. The response sentence generation model B includes a neural network for deep learning and its parameters.

FIG. 9 is a diagram showing an example of a learning utterance sentence, a learning related sentence, and a learning response sentence of the second embodiment. As shown in Fig. 9, a sentence containing "varied tit" included in the utterance sentence for learning "Varied tit is a wild bird known for hoarding over the winter" is searched from Web1000 and used for learning. The related sentence "Good weather, few clouds, and wild birds such as varied tit and tsugumi could be seen around the lake." Then, from the learning-related sentences, the unnatural part "good weather and few clouds," is deleted as a response to this learning utterance sentence, and the flexion "done" that does not seem to be a dialogue is a suitable flexion as a response. By replacing it with "I can do it", a learning response sentence "You can also see wild birds such as yamagara and tsugumi around the lake." Is created. By learning using such an utterance sentence for learning, a related sentence for learning, and a response sentence for learning, a response sentence generation model B that generates an appropriate response sentence as a response to the utterance is generated from the utterance sentence and the related sentence. To.

The difference between the dialogue device 23 of the present embodiment and the dialogue device 3 of the first embodiment is the response sentence generation unit 29. The response sentence generation unit 29 generates a dialogue execution response sentence by inputting the dialogue execution utterance sentence and the dialogue execution related sentence into the response sentence generation model B after learning.

FIG. 10 is a diagram showing an example of a dialogue execution utterance sentence, a dialogue execution related sentence, and a dialogue execution response sentence of the second embodiment. As shown in FIG. 10, a sentence containing "pollen" contained in the dialogue execution utterance sentence "It seems that pollen has begun to fly" is searched from Web1000, and the dialogue execution related sentence "pollen makes the skin tingling." , UV rays are getting stronger, so I'm afraid to use it. " Then, from the dialogue execution related sentences, the unnatural part "because it is scary to use" is deleted as a response to the dialogue execution utterance sentence, and the ending "become" that does not seem to be the response of the dialogue is the response of the dialogue. The dialogue execution response sentence "The pollen is about to irritate the skin, and the ultraviolet rays are getting stronger." Is generated.

FIG. 11 is a diagram showing the configuration of the response generation learning processing unit 26 of the second embodiment. The response generation learning processing unit 26 includes an encoder 151 and a decoder 152, and further includes an encoder 157 and a synthesis unit 155, as in the first embodiment. The response generation learning processing unit 26 learns the Sequence to Sequence model, which is an example of the response sentence generation model B.

The encoder 157 includes a recurrent neural network developed along a time axis (t = 1 to s). An utterance sentence for learning is input to the encoder 157. Let the input vector at time t be x2 (t) and the vector of the hidden layer be h2 (t).

The hidden layer at time t receives an input vector x2 (t) representing a word of a learning utterance sentence input to the input layer at time t, and a vector h2 (t-1) of the hidden layer at time (t-1). ), And the vector h2 (t) is output to the hidden layer at the time (t + 1). However, the hidden layer at time 1 receives only the input vector x2 (1). For example, the input vector x2 (1) at time 1 is a vector representing "Kyoto", and the input vector x2 (2) at time 2 is a vector representing "ni".

When the weight matrix at time t is U2 (t) and W2 (t), the bias is b, and the activation function is f, the following equation holds.

h2 (1) = f (U2 (1) x2 (1) + b)… (9)
h2 (t) = f (U2 (t) x2 (t) + W2 (t) h2 (t-1) + b) (other than t = 1) ... (10)
The synthesizing unit 155 generates one vector hh from the vector h1 (m) of the hidden layer at the last time m of the encoder 151 and the vector h2 (s) of the hidden layer at the last time s of the encoder 157. The vector hh can be, for example, a vector obtained by adding h1 (m) and h2 (s), a vector obtained by weighting addition, and a vector including all the elements of h1 (m) and h2 (s).

The decoder 152 is the same as the first embodiment except that the hidden layer at time 1 receives the vector hh instead of the vector h1 (m) of the hidden layer at the last time m of the encoder 151.

Similar to the first embodiment, in the decoder 152, the output sentence composed of the output vectors y3 (1) to y3 (n) and the learning response sentence (teacher) composed of x3 (1) to x3 (n). The error with the data) is required. These errors are opposite to the hidden layer before time t in the decoder 152, the hidden layer in the encoder 151 (time t = 1 to m), and the hidden layer in the encoder 157 (time t = 1 to s). Propagate. The weight matrices W3 (t), U3 (t), and V3 (t) are updated so as to minimize the sum of the errors back-propagated to the hidden layer at time t in the decoder 152. W1 (t) and U1 (t) are updated so as to minimize the sum of the errors back-propagated to the hidden layer at time t in the encoder 151. W2 (t) and U2 (t) are updated so as to minimize the sum of the errors back-propagated to the hidden layer at time t in the encoder 157. The weight matrices W3 (t), U3 (t), V3 (t), W1 (t), U1 (t), W2 (t), and U2 (t), which are the parameters of the response sentence generation model B after training, are It is stored in the parameter storage unit 14.

FIG. 12 is a diagram showing the configuration of the response sentence generation unit 29 of the second embodiment. The response sentence generation unit 29 includes an encoder 153 and a decoder 154, as well as an encoder 158 and a synthesis unit 156, as in the first embodiment. The response sentence generation unit 29 generates a response sentence for dialogue execution by using the sequence to sequence model after learning, which is an example of the response sentence generation model B after learning.

In FIG. 12, the weight matrices W3 (t), U3 (t), V3 (t), W1 (t), U1 (t), W2 (t), U2 ( t) is read from the parameter storage unit 15 and set. As these, those obtained by learning (update) by the response generation learning processing unit 26 of FIG. 11 are used.

The encoder 158 includes a recurrent neural network developed along a time axis (t = 1 to s) similar to the encoder 157. However, the input vector x2 (t) input to the input layer at time t is not a word of a learning-related sentence but a word of an utterance sentence for dialogue execution.

The synthesizing unit 156 generates one vector hh from the vector h1 (m) of the hidden layer at the last time m of the encoder 153 and the vector h2 (s) of the hidden layer at the last time s of the encoder 158. The vector hh can be, for example, a vector obtained by adding h1 (m) and h2 (s), a vector obtained by weighting addition, and a vector including all the elements of h1 (m) and h2 (s).

The decoder 154 is the same as the first embodiment except that the hidden layer at time 1 receives the vector hh instead of the vector h1 (m) of the hidden layer at the last time m of the encoder 151.

FIG. 13 is a flowchart showing the operation procedure of the model learning device 22 of the second embodiment.

In step S301, the utterance sentence input unit 11 receives the input of the utterance sentence for learning.
In step S302, the related sentence acquisition unit 12 acquires a learning-related sentence related to the learning utterance sentence from Web 1000, which is one of the text databases.

In step S303, the learning response sentence input unit 13 modifies the learning-related sentence by the operation of the personnel on the development side so that the learning-related sentence becomes appropriate as the response of the dialogue, and the learning response is generated. Receive a sentence input.

In step S304, by repeating steps S301 to S303 until all the inputs are completed, a plurality of sets of learning utterance sentences, learning related sentences, and learning response sentences are obtained. The corpus output unit 74 generates a learning corpus including a plurality of sets of learning utterance sentences, learning related sentences, and learning response sentences, and writes them in the corpus storage unit 25. The corpus storage unit 25 stores the learning corpus.

In step S305, the response generation learning processing unit 26 learns the response sentence generation model B in which the utterance sentence and the related sentence are input and the response sentence is output by using the learning corpus.

In step S306, the parameter storage unit 14 stores the parameters after learning of the response sentence generation model B by the response generation learning processing unit 26.

FIG. 14 is a flowchart showing the operation procedure of the dialogue device 23 of the second embodiment.
In step S401, the response sentence generation unit 29 reads out the parameters after learning of the response sentence generation model B by the model learning device 22 stored in the parameter storage unit 15.

In step S402, the utterance sentence input unit 7 receives input of the utterance sentence for dialogue execution from the user.

In step S403, the related sentence acquisition unit 8 acquires the dialogue execution related sentence related to the dialogue execution utterance sentence from Web1000, which is one of the text databases.

In step S404, the response sentence generation unit 29 inputs the dialogue execution utterance sentence and the dialogue execution related sentence into the post-learning response sentence generation model B obtained by using the read parameters, thereby performing the dialogue execution response. Generate a statement.

As described above, according to the present embodiment, the model learning device learns the model including not only the related sentence but also the utterance sentence in the input, so that the dialogue device is more appropriate for the utterance sentence at that time. Response statement can be generated.

[Third Embodiment]
FIG. 15 is a diagram showing the configuration of the dialogue response system 31 of the third embodiment. The dialogue response system 31 includes a model learning device 32 and a dialogue device 23.

The dialogue device 23 of the present embodiment is the same as the dialogue device 23 of the second embodiment. The difference between the model learning device 32 of the present embodiment and the model learning device 22 of the second embodiment is the corpus generation unit 34.

The corpus generation unit 34 includes a dialogue data acquisition unit 39, a related sentence acquisition unit 45, a learning response sentence generation unit 63, and a corpus output unit 84.

The dialogue data acquisition unit 39 acquires dialogue data such as twitter from Web 1000, which is a kind of text database. The dialogue data consists of a user utterance sentence and a user response sentence.

The related sentence acquisition unit 45 acquires a sentence related to the user-spoken sentence included in the dialogue data from Web1000, which is one of the text databases, and uses it as a learning-related sentence. Specifically, the related sentence acquisition unit 45 acquires, for example, a sentence including at least one noun and / or at least one predicate included in the user response sentence from the data on the Web 1000 as the related sentence. That is, a sentence containing at least one noun, a sentence containing at least one predicate, or a sentence containing at least one noun and at least one predicate is acquired as a related sentence. That is, in the present embodiment, the learning-related sentence is searched not directly from the user utterance sentence but via the user response sentence which is a response to the user utterance sentence.

The learning response sentence generation unit 63 generates a learning response sentence obtained by converting a learning-related sentence into a dialogue-like sentence. Specifically, the learning response sentence generation unit 63 generates a learning response sentence by extracting a part having the same content between the user response sentence constituting the dialogue data and the learning related sentence. The learning response sentence generation unit 63 can extract a part related to a noun and / or a predicate included in the learning-related sentence from the user response sentence to obtain a learning response sentence. That is, from the user response sentence, the part related to the noun included in the learning-related sentence, the part related to the predicate included in the learning-related sentence, or the part related to the noun and the predicate contained in the learning-related sentence are extracted. Will be done. The learning response sentence generated here can be regarded as a conversion of a learning-related sentence acquired from Web 1000 into a dialogue-like sentence.

The corpus output unit 84 uses the user utterance sentence included in the dialogue data as the learning utterance sentence. The corpus output unit 84 generates a learning corpus including a plurality of sets of learning utterance sentences (see FIG. 16), learning related sentences, and learning response sentences, and writes them in the corpus storage unit 25.

FIG. 16 is a diagram showing examples of dialogue data, learning utterance sentences, learning related sentences, and learning response sentences according to the third embodiment. As shown in FIG. 16, dialogue data consisting of the user utterance "Cosmos this morning!" And the user response "Thank you. Cosmos somehow makes me feel the sign of autumn." Is acquired. A sentence containing the noun "Cosmos", the noun "Autumn sign", and the predicate "Feeling" included in the user response sentence is searched from Web1000, and the learning related sentence "Cosmos makes you feel the autumn sign" is acquired. Will be done. Then, by extracting the parts related to the noun "Cosmos", the noun "Autumn sign", and the predicate "Make me feel" included in the learning-related sentence from the user response sentence, the learning response sentence "Cosmos is". It makes me feel the sign of autumn. "Is generated. By using a part of the user response sentence of the dialogue data as the learning response sentence in this way, a learning response sentence obtained by transforming the learning-related sentence into an appropriate form as a response to the learning utterance sentence can be obtained. By learning using such an utterance sentence for learning, a related sentence for learning, and a response sentence for learning, a response sentence generation model B that generates an appropriate response sentence as a response to the utterance is generated from the utterance sentence and the related sentence. To.

FIG. 17 is a flowchart showing the operation procedure of the model learning device 32 of the third embodiment.

In step S501, the dialogue data acquisition unit 39 acquires dialogue data composed of a user utterance sentence such as twitter and a user response sentence from Web 100, which is a kind of text database. The user's utterance becomes the learning utterance.

In step S502, the related sentence acquisition unit 45 acquires a sentence related to the user-spoken sentence included in the dialogue data from Web1000, which is one of the text databases, and uses it as a learning-related sentence.

In step S503, the learning response sentence generation unit 63 generates a learning response sentence obtained by converting a learning-related sentence into a dialogue-like sentence.

By repeating steps S501 to S503 in step S504, a plurality of sets of learning utterance sentences, learning related sentences, and learning response sentences are obtained. The corpus output unit 84 generates a learning corpus including a plurality of sets of learning utterance sentences, learning related sentences, and learning response sentences, and writes them in the corpus storage unit 25. The corpus storage unit 25 stores the learning corpus.

In step S505, the response generation learning processing unit 26 learns the response sentence generation model B in which the utterance sentence and the related sentence are input and the response sentence is output by using the learning corpus.

In step S506, the parameter storage unit 14 stores the parameters after learning of the response sentence generation model B by the response generation learning processing unit 6.

As described above, according to the present embodiment, the learning corpus can be automatically generated, so that the highly accurate response sentence generation model B can be learned by using a large amount of learning data.

[Fourth Embodiment]
FIG. 18 is a diagram showing the configuration of the dialogue response system 41 of the fourth embodiment.

The dialogue response system 41 includes a model learning device 42 and a dialogue device 43. The difference between the model learning device 42 of the present embodiment and the model learning device 32 of the third embodiment is the corpus generation unit 44.

The corpus generation unit 44 includes a dialogue data acquisition unit 39, a related sentence acquisition unit 28, a learning response sentence generation unit 63, and a corpus output unit 84.

Since the dialogue data acquisition unit 39, the learning response sentence generation unit 63, and the corpus output unit 84 are the same as those in the third embodiment, the description will not be repeated.

The related sentence acquisition unit 28 includes a binary triple generation unit 51, an original sentence acquisition unit 52, and an original sentence ranking unit 53. The related sentence acquisition unit 28 selects a learning related sentence from the sentences stored in the database 1054 of the question / answering system 1051.

The database 1054 of the question / answering system 1051 stores the original sentences arranged by pattern after being acquired from Web1000.

FIG. 19 is a diagram showing an example of an original sentence stored in the database 1054. FIG. 20 is a diagram showing an example of a sentence used by the related sentence acquisition unit 28 of the model learning device 42.

As shown in FIG. 19, a binary triple is used as a pattern for organizing the original sentence. A binary triple consists of a binary pattern consisting of two variables (A, B) and a predicate, such as "A goes to B", and variables A and, such as (A, B) = (New Year, Hatsumode). Includes a pair of words corresponding to B.

The binary triple generation unit 51 generates a binary triple based on the dialogue data. Specifically, the binary triple generation unit 51 generates a binary triple based on all the nouns included in the user utterance sentence of the dialogue data and all the nouns and verbs included in the user response sentence of the dialogue data. .. For example, as shown in Fig. 20, the noun "New Year" is extracted from the user utterance "It was New Year", and the noun "Last year" and the noun from the user response sentence "Last year I went to Kyoto for the first time!" "Kyoto", the noun "Hatsumode" and the verb "go" are extracted. The binary triple generator 51 produces a binary triple "A goes to B, (A, B) = (New Year, Hatsumode)" and "A goes to B, (A, B) = (New Year, Kyoto)". Generate.

The original sentence acquisition unit 52 accesses the database 1054 of the question / answering system 1051 and acquires a plurality of original sentences that match the generated binary triple pattern. For example, as shown in FIG. 20, the binary triple "A goes to B, (A, B) = (New Year, Genbun)" or "A goes to B, (A, B) = (New Year, Kyoto)". Three original sentences (original sentence 1, original sentence 2, and original sentence 3) that match the pattern of are acquired.

The original sentence ranking unit 53 ranks a plurality of acquired original sentences according to a predetermined standard, and outputs the original sentence having the highest ranking as a learning-related sentence. Specifically, the original sentence ranking unit 53 sets the rank higher as the original sentence contains more topic words and predicates included in the user response sentence of the dialogue data. For example, the topic word can be a word other than a predetermined word. The predetermined word can be a general word that is frequently used in the dialogue, in other words, a word that is not a characteristic part of the dialogue. For example, as shown in FIG. 20, the topic words "Kyoto" and "Hatsumode" included in the user response sentence of the dialogue data and the original sentence "New Year's Day went to Kyoto for the first visit" containing the most predicate "I went". "Ta" is determined to be the first in the ranking and is output as a learning-related sentence.

The difference between the dialogue device 43 of the present embodiment and the dialogue device 33 of the third embodiment is the related sentence acquisition unit 48. The related sentence acquisition unit 48 includes a question sentence generation unit 55, an answer and original sentence acquisition unit 56, and an original sentence ranking unit 57.

FIG. 21 is a diagram showing an example of a sentence used by the related sentence acquisition unit 48 of the dialogue device 43. Here, for convenience of explanation, a case where the same example sentence as the example sentence used in the model learning device 42 is used in the dialogue device 43 will be described.

The question sentence generation unit 55 generates a question sentence by converting the dialogue execution utterance sentence into a question format according to a predetermined rule. For example, as shown in FIG. 21, the question sentence "What is the New Year going?" Is generated from the dialogue execution utterance sentence "It was New Year."

The answer and original sentence acquisition unit 56 transmits a question sentence to the question / answering system 1051. The question / answering system 1051 converts the question text into a binary triple format, and retrieves a plurality of original sentences matching the binary triple format from the database 1054. For example, the question sentence "What does New Year go?" Is converted into a binary triple "A goes to B, (A, B) = (New Year, *)" (* is an arbitrary noun). As shown in FIG. 21, three original sentences (original sentence 1, original sentence 2, original sentence 1) matching the pattern of the binary triple "A goes to B, (A, B) = (New Year, *)" from the database 1054. Sentence 3) is acquired. Further, the question / answering system 1051 generates an answer to the question sentence based on a plurality of original sentences. For example, among B included in a plurality of original sentences, the one contained most can be the answer. In FIG. 21, since all three original sentences include "Hatsumode" as B, "Hatsumode" is generated as an answer.

The question / answering system 1051 transmits an answer and a plurality of original sentences to the dialogue device 43. The answer and original sentence acquisition unit 56 acquires the answer and a plurality of original sentences sent from the question / answering system 1051. For example, as shown in FIG. 21, the answer "Hatsumode" and three original sentences are acquired.

The original sentence ranking unit 57 ranks a plurality of acquired original sentences according to a predetermined standard, and outputs the original sentence having the highest ranking as a dialogue execution related sentence. Specifically, the original sentence ranking unit 57 sets the rank higher as the original sentence contains more topic words. For example, as shown in Fig. 21, the original sentence "New Year went to Kyoto for the first time", which contains the most topic words "New Year", "Hatsumode", and "Kyoto", was selected as the first sentence in the ranking. Is done.

FIG. 22 is a flowchart showing the operation procedure of the model learning device 42 of the fourth embodiment.

In step S601, the dialogue data acquisition unit 39 acquires dialogue data composed of a user utterance sentence such as twitter and a user response sentence from Web 1000 which is a kind of text database. The user utterance sentence included in the dialogue data becomes the learning utterance sentence.

In step S602, the binary triple generation unit 51 generates a binary triple based on the dialogue data.

In step S603, the original sentence acquisition unit 52 accesses the database 1054 of the question / answering system 1051 and acquires a plurality of original sentences that match the generated binary triple pattern.

In step S604, the original sentence ranking unit 53 ranks a plurality of acquired original sentences according to a predetermined standard, and outputs the original sentence having the highest ranking as a learning-related sentence.

In step S605, the learning response sentence generation unit 63 generates a learning response sentence obtained by converting a learning-related sentence into a dialogue-like sentence.

In step S606, by repeating steps S601 to S605, a plurality of sets of learning utterance sentences, learning related sentences, and learning response sentences are obtained. The corpus output unit 84 generates a learning corpus including a plurality of sets of learning utterance sentences, learning related sentences, and learning response sentences, and writes them in the corpus storage unit 25. The corpus storage unit 25 stores the learning corpus.

In step S607, the response generation learning processing unit 26 learns the response sentence generation model B in which the utterance sentence and the related sentence are input and the response sentence is output by using the learning corpus.

In step S608, the parameter storage unit 14 stores the parameters after learning of the response sentence generation model B by the response generation learning processing unit 26.

FIG. 23 is a flowchart showing the operation procedure of the dialogue device 43 of the fourth embodiment.
In step S701, the response sentence generation unit 29 reads out the parameters after learning of the response sentence generation model B by the model learning device 42 stored in the parameter storage unit 15.

In step S702, the utterance sentence input unit 7 receives input of the utterance sentence for dialogue execution from the user.

In step S703, the question sentence generation unit 55 generates a question sentence by converting the dialogue execution utterance sentence into a question format according to a predetermined rule.

In step S704, the answer and original sentence acquisition unit 56 acquires the answer and a plurality of original sentences sent from the question / answering system 1051.

In step S705, the original sentence ranking unit 57 ranks a plurality of acquired original sentences according to a predetermined standard, and outputs the original sentence having the highest ranking as a dialogue execution related sentence.

In step S706, the response sentence generation unit 29 inputs the dialogue execution utterance sentence and the dialogue execution related sentence into the post-learning response sentence generation model B obtained by using the read parameters, thereby performing the dialogue execution response. Generate a statement.

As described above, according to the present embodiment, the quality of the response sentence to the utterance can be improved by using the existing question / response system.

[Fifth Embodiment]
FIG. 24 is a diagram showing the configuration of the dialogue response system 81 of the fifth embodiment. The dialogue response system 81 includes a model learning device 78 and a dialogue device 77.

The model learning device 78 of the present embodiment differs from the model learning device 42 of the fourth embodiment in that the original sentence ranking 49, the learning response sentence generation unit 83, the corpus output unit 85, and the corpus storage. A unit 87 and a response generation learning processing unit 86.

The original sentence ranking 49 ranks a plurality of acquired original sentences according to a predetermined standard, and outputs a plurality of original sentences having a ranking from the first place to a predetermined order as a plurality of learning-related sentences. ..

The learning response sentence generation unit 83 generates one learning response sentence from a plurality of learning related sentences. For example, the learning response sentence generation unit 83 may extract only the parts related to the nouns and verbs included in the learning-related sentence having the highest ranking from the user response sentence included in the dialogue data as the learning response sentence. good. Alternatively, the learning response sentence generation unit 83 may generate one learning response sentence by concatenating a plurality of learning related sentences in a form in which overlapping portions are not repeated.

The corpus output unit 85 generates a learning corpus including a set of a learning utterance sentence, a plurality of learning related sentences, and a learning response sentence, and writes the learning corpus into the corpus storage unit 87. The corpus storage unit 87 stores the learning corpus.

The response generation learning processing unit 86 learns a model C in which an utterance sentence and a plurality of related sentences are input and a response sentence is output by using a learning corpus. Model C differs from response sentence generation models A and B in that an utterance sentence and a plurality of related sentences are input.

The difference between the dialogue device 77 of the present embodiment and the dialogue device 43 of the fourth embodiment is the response sentence generation unit 89. The response sentence generation unit 89 generates a dialogue execution response sentence by inputting a dialogue execution utterance sentence and a plurality of dialogue execution related sentences into the model C.

FIG. 25 is a diagram showing the configuration of the response generation learning processing unit 86 of the fifth embodiment. The configuration of the response sentence generation unit 89 is also the same.

As shown in FIG. 25, the response generation learning processing unit 86 includes an encoder 151, 157 and a decoder 152, and further includes an encoder 351 and 352 and a synthesis unit 353, as in the second embodiment. ..

The first related sentence among the plurality of related sentences is input to the encoder 151. A second related sentence out of a plurality of related sentences is input to the encoder 351. A third related sentence out of a plurality of related sentences is input to the encoder 352. The synthesis unit 353 generates one vector hh from the vector of the hidden layer at the last time of each of the encoders 151, 351 and 352, 157, and supplies it to the hidden layer at time 1 of the decoder 152. Since the configurations and operations of the encoders 151, 351 and 352, 117 and the decoder 152 are the same as those described in the first and second embodiments, the description will not be repeated.

As described above, according to the present embodiment, according to the present embodiment, the model learning device learns the model including a plurality of related sentences in the input, so that the dialogue device receives the spoken sentence. A more appropriate response statement can be generated.

[Sixth Embodiment]
FIG. 26 is a diagram showing the configuration of the dialogue response system 91 of the sixth embodiment.

The dialogue response system 91 includes a model learning device 75 and a dialogue device 76.
The model learning device 75 of the present embodiment differs from the model learning device 42 of the fourth embodiment in that the question sentence generation unit 94, the learning response sentence generation unit 93, the corpus output unit 95, and the corpus. The storage unit 97 and the response generation learning processing unit 96.

The question sentence generation unit 94 generates a learning question sentence from the user's utterance sentence. The learning response sentence generation unit 93 generates a learning response sentence obtained by converting a learning-related sentence into a dialogue-like sentence. Specifically, the learning response sentence generation unit 93 includes a learning utterance sentence (user utterance sentence), a learning-related sentence output from the original sentence ranking unit 53, and a learning output from the question sentence generation unit 94. A learning response sentence is generated based on the question sentence and the topic word output from the original sentence ranking unit 53. For example, the learning response sentence generation unit 93 extracts from the user response sentence included in the dialogue data the nouns included in the learning question sentence, the nouns and verbs included in the learning related sentence, and the part related to the topic word. Then, it may be used as a response sentence for learning.

The corpus output unit 95 generates a learning corpus including a plurality of sets of learning utterance sentences, learning related sentences, learning question sentences, learning topic words, and learning response sentences, and the corpus storage unit 97. Write in. The corpus storage unit 97 stores a learning corpus.

The response generation learning processing unit 96 learns the model D in which the utterance sentence, the related sentence, the question sentence, and the topic word are input and the response sentence is output by using the learning corpus. Model D differs from other models in that it inputs utterance sentences, related sentences, question sentences, and topic words.

The difference between the dialogue device 76 of the present embodiment and the dialogue device 43 of the fourth embodiment is the response sentence generation unit 99. The response sentence generation unit 99 produces a dialogue execution utterance sentence, a dialogue execution related sentence, a dialogue execution question sentence output from the question sentence generation unit 55, and a dialogue execution topic word output from the original sentence ranking unit 57. By inputting to the model D, a response statement for executing a dialogue is generated.

FIG. 27 is a diagram showing the configuration of the response generation learning processing unit 96 of the sixth embodiment. The configuration of the response sentence generation unit 99 is also the same.

As shown in FIG. 27, the response generation learning processing unit 96 includes encoders 151 and 157, a decoder 152, and further includes encoders 451 and 452 and a synthesis unit 453, as in the second embodiment.

A question sentence is input to the encoder 451. A topic word is input to the encoder 451. The synthesis unit 453 generates one vector hh from the vector of the hidden layer at the last time of each of the encoders 151, 157, 451 and 452, and supplies it to the hidden layer at time 1 of the decoder 152. Since the configurations and operations of the encoders 151, 157, 451 and 452 and the decoder 152 are the same as those described in the first and second embodiments, the description will not be repeated.

As described above, according to the present embodiment, the model learning device learns the model including the question sentence and the topic word in the input, so that the dialogue device generates a more appropriate response sentence to the utterance sentence. can do.

[7th Embodiment]
FIG. 28 is a diagram showing the configuration of the dialogue response system 551 of the seventh embodiment. The corpus storage unit 5 stores the corpus for learning obtained by the processing in the examples so far. Further, for the related sentence acquisition unit 8, for example, the previous configuration in FIG. 1 is omitted. In this dialogue response system 551, one device performs both model learning and actual dialogue. The dialogue response system includes a corpus storage unit 5, a learning control unit 552, a response sentence generation model A, and a related sentence acquisition unit 8.

The corpus storage unit 5 is the same as the corpus storage unit 5 described in the first embodiment.
The learning control unit 552 trains the response sentence generation model A by giving learning data to the response sentence generation model A. The learning data are a learning-related sentence (input) stored in the corpus storage unit 5 and a learning response sentence (output). When the learning is completed, the network of the response sentence generation model A and the parameters are saved.

The related sentence acquisition unit 8 is the same as the related sentence acquisition unit 8 described in the first embodiment.
At the time of dialogue processing, the dialogue execution related sentence acquired by the related sentence acquisition unit 8 is input to the response sentence generation model A after learning, so that the dialogue execution response sentence based on the learning result is generated.

According to the configuration of FIG. 28, the learning can be dynamically continuously executed by using the data when the dialogue is executed. That is, the dialogue execution response sentence output from the response sentence generation model A can also be used as learning data.

Similarly, in the second to sixth embodiments, one model learning and dialogue device may perform both model learning and actual dialogue.

[Eighth Embodiment]
FIG. 29 is a diagram showing the configuration of the dialogue response system 1100 according to the eighth embodiment.

The dialogue response system 1100 includes a model learning device 1120 and a dialogue device 1130.

The model learning device 1130 of the present embodiment differs from the model learning device 2 of the first embodiment in that the related sentence conformity input unit 1103, the corpus output unit 1104, the conformity determination learning processing unit 1105, and so on. The parameter storage unit 1106.

The related sentence conformity input unit 1103 receives the input of the goodness of fit indicating the conformity of the acquired related sentence. For example, if the acquired related sentence has the same or similar topic as the learning utterance sentence, the development staff indicates the goodness of fit (for example, numerical value) indicating that the acquired related sentence is compatible. Enter "1"). If the acquired related sentence has a different topic from the learning utterance sentence, the development staff inputs the goodness of fit (for example, the numerical value "0") indicating that the acquired related sentence does not match. ..

A related sentence (hereinafter referred to as a non-conforming related sentence) in which a goodness of fit indicating non-conformity is input by the related sentence conformity input unit 1103 is sent to the corpus output unit 1104.

FIG. 30 is a diagram for explaining an example of a nonconformity-related sentence.
Sentences for learning "I'm going to Nara on a school trip""Schooltrip" and sentences including "Nara""Many junior high schools in Nara city carry out school trips to Okinawa." To. In this sentence, it is stated that the destination of the "study trip" is "Okinawa", and the topic is different from the utterance sentence for learning. In such a case, the acquired related sentence is determined to be a nonconforming related sentence.

The related sentence (hereinafter referred to as the conformity-related sentence) in which the goodness of fit indicating conformity is input by the related sentence conformity input unit 1103 is sent to the learning response sentence input unit 13 as a learning-related sentence. Similar to the first embodiment, the learning response sentence input unit 13 receives the input of the learning response sentence in which the learning-related sentence is modified into an appropriate format as the response of the dialogue, and the learning response sentence is a corpus. It is sent to the output unit 1104.

The corpus output unit 1104 includes a learning utterance sentence, a related sentence, and a plurality of sets of matching degrees of the related sentence, and further, a learning related sentence and a learning response sentence which are conformance related sentences among the related sentences. A learning corpus including a plurality of sets is generated and written to the corpus storage unit 1108.

The corpus storage unit 1108 stores the learning corpus.
FIG. 31 is a diagram showing an example of a learning corpus stored in the corpus storage unit 1108.

The conformity determination learning processing unit 1105 uses the learning utterance sentence, the related sentence, and the conformity degree stored in the corpus storage unit 1108 to include the utterance sentence and the related sentence in the input and outputs the conformity degree. The sex determination model X1 is learned. The suitability determination model X1 includes a neural network for deep learning and its parameters.

The parameter storage unit 1106 stores the parameters after learning of the conformity determination model X1 by the conformity determination learning processing unit 1105. The parameter storage unit 1106 stores parameters (weight matrix, etc.) after learning the neural network. The parameters after learning of the conformity determination model X1 stored in the parameter storage unit 1106 can be taken out to the outside via an input / output port or the like (not shown).

The dialogue device 1130 of the present embodiment differs from the dialogue device 3 of the first embodiment in the parameter storage unit 1107, the related sentence compatibility determination unit 1101, and the related sentence selection unit 1102.

The parameter storage unit 1107 stores the parameters after learning of the conformity determination model X1 by the model learning device 1120. For example, the parameter storage unit 1107 stores parameters (weight matrix, etc.) after learning of the neural network. The learned parameters of the conformity determination model X1 taken out from the model learning device 1120 are input to the dialogue device 1130 and stored in the parameter storage unit 1107. These parameters may be distributed to the user together with the program that realizes the function of the dialogue device 1130.

The goodness-of-fit determination unit 1101 generates a goodness of fit by inputting an utterance sentence for dialogue execution and an acquired related sentence into the learned conformity determination model X1. For example, when the goodness of fit as a related sentence is high, the generated goodness of fit is close to "1", and when the goodness of fit as a related sentence is low, the generated goodness of fit is close to "0". Become.

The related sentence selection unit 1102 selects a dialogue execution related sentence from a plurality of related sentences acquired by the related sentence acquisition unit 8 according to the goodness of fit, and outputs the dialogue execution related sentence to the response sentence generation unit 9. For example, the related sentence having the closest goodness of fit to "1" is selected as the related sentence for executing the dialogue.

Similar to the first embodiment, the response sentence generation unit 9 generates a dialogue execution response sentence by inputting the selected dialogue execution related sentence into the response sentence generation model A after learning.

FIG. 32 is a diagram showing a neural network that realizes the conformity determination model X1 used in the conformity determination learning processing unit 1105 and the related sentence conformity determination unit 1101 of the eighth embodiment.

The convolutional neural network CNN1 receives the learning utterance sentence which is the first input data. The convolutional neural network CNN1 includes a convolutional layer C11 and a pooling layer C12.

The first input data, the utterance sentence for learning "I'm going to Kyoto", is represented by data having a size of N × 4.

Based on the calculation results of the learning utterance sentence and the plurality of filters, in the convolution layer C11, the first convolution calculation result of M 1 × 3 size and the second convolution calculation result of M 1 × 2 size are obtained. The convolution operation result and the third convolution operation result having a size of M pieces of 1 × 1 can be obtained.

By MAX pooling of the first calculation result, the second calculation result, and the third calculation result, M first pooling results, M second pooling results, and M third pooling results in the pooling layer C12. The pooling result of is obtained. In MAX pooling, only the maximum value is output.

The convolutional neural network CNN2 receives the related sentence which is the second input data. The convolutional neural network CNN2 includes a convolutional layer C21 and a pooling layer C22.

The second input data, the related sentence "Enjoying fireworks in Kyoto", is represented by data having a size of N × 5.

Based on the operation results of the related sentence and the plurality of filters, in the convolution layer C21, the fourth convolution operation result of M 1 × 4 size and the fifth convolution operation of M 1 × 3 size are used. The result and the sixth convolution operation result of M pieces of 1 × 2 size are obtained.

By MAX pooling of the fourth calculation result, the fifth calculation result, and the sixth calculation result, M fourth pooling results, M fifth pooling results, and M sixth pooling results in the pooling layer C22. The pooling result of is obtained.

The coupling portion COE1 combines the first to sixth pooling results to generate a 6M-dimensional coupling vector.

The multilayer perceptron 500 includes an input layer 501, an intermediate layer 502, and an output layer 503.

The 6M-dimensional coupling vector data is sent to the input layer 501. The goodness of fit of the related sentence is output by the calculation of the input layer 501, the intermediate layer 502, and the output layer 503.

FIG. 33 is a diagram showing details of the convolution operation in the convolutional neural network CNN1.

In the convolutional neural network CNN1, M N × 2 size filters, M N × 3 size filters, and M N × 4 size filters are used.

By the convolution operation of the N × 4 size learning utterance sentence and the M N × 2 size filters, the first convolution operation result of M 1 × 3 sizes can be obtained.

By the convolution operation of the N × 4 size learning utterance sentence and the M N × 3 size filters, the second convolution operation result of M 1 × 2 sizes can be obtained.

By the convolution operation of the N × 4 size learning utterance sentence and the M N × 4 size filters, a third convolution operation result of M 1 × 1 sizes can be obtained.

FIG. 34 is a diagram showing details of the convolution operation in the convolutional neural network CNN2.

In the convolutional neural network CNN2, M N × 2 size filters, M N × 3 size filters, and M N × 4 size filters are used.

By the convolution operation of the related sentence of the size of N × 5 and the filter of the size of M N × 2, the fourth convolution operation result of the size of M 1 × 4 can be obtained.

By the convolution operation of the related sentence of the size of N × 5 and the filter of the size of M N × 3, the fifth convolution operation result of the size of M 1 × 3 is obtained.

By the convolution operation of the related sentence of the size of N × 5 and the filter of the size of M N × 4, the sixth convolution operation result of the size of M 1 × 3 is obtained.

At the time of learning, the goodness of fit input to the related sentence suitability input unit 1103 is used as the teacher data. The error backpropagation method updates the weights of the Multilayer Perceptron 500 and the coefficients of the filter so that the error with the teacher data is minimized. The weight of the multi-layer perceptron 500 after learning and the coefficient of the filter are stored in the parameter storage unit 1106. The weight of the multi-layer perceptron 500 after learning stored in the parameter storage unit 1106 and the coefficient of the filter are transferred to the parameter storage unit 1107.

At the time of dialogue response, the weight of the multi-layer perceptron 500 after learning and the coefficient of the filter stored in the parameter storage unit 1107 are read out and used. The goodness of fit output from the output layer 503 of the multi-layer perceptron 500 is sent to the related sentence selection unit 1102 and used for selecting the related sentence.

FIG. 35 is a flowchart showing the operation procedure of the model learning device 1120 according to the eighth embodiment.

In step S1101, the utterance sentence input unit 11 receives the input of the utterance sentence for learning.
In step S1102, the related sentence acquisition unit 12 acquires a learning-related sentence related to the learning utterance sentence from Web 1000, which is one of the text databases.

In step S1103, the related sentence conformity input unit 1103 receives the input of the goodness of fit indicating the conformity of the acquired related sentence.

In step S1104, if the value indicates that the input goodness of fit is conforming (S1104: YES), the process proceeds to step S1105, indicating that the input goodness of fit is conforming. If it is not a value (S1104; NO), the process proceeds to step S1106.

In step S1105, the learning response sentence input unit 13 is generated by modifying the learning-related sentence by the operation of the personnel on the development side so that the conformance-related sentence (learning-related sentence) becomes appropriate as the response of the dialogue. Receives the input of the learning response sentence.

In step S1106, steps S1101 to S1105 are repeated until all inputs are completed. The corpus output unit 1104 includes a learning utterance sentence, a related sentence, and a plurality of sets of matching degrees of the related sentence, and further, a learning related sentence and a learning response sentence which are conformance related sentences among the related sentences. A learning corpus including a plurality of sets is generated and written to the corpus storage unit 1108.

In step S1107, the conformity determination learning processing unit 1105 uses the learning utterance sentence, the related sentence, and the conformity degree included in the learning corpus to input the utterance sentence and the related sentence, and outputs the conformity degree as the output. The determination model X1 is learned.

In step S1108, the parameter storage unit 1106 stores the parameters after learning of the conformity determination model X1 by the conformity determination learning processing unit 1105.

In step S1109, the response generation learning processing unit 6 learns the response sentence generation model A in which the related sentence is input and the response sentence is output by using the learning related sentence and the learning response sentence included in the learning corpus. do.

In step S1110, the parameter storage unit 14 stores the parameters after learning of the response sentence generation model A by the response generation learning processing unit 6.

FIG. 36 is a flowchart showing the operation procedure of the dialogue device 1130 according to the eighth embodiment.

In step S1201, the response sentence generation unit 9 reads out the parameters after learning of the response sentence generation model A by the model learning device 1130 stored in the parameter storage unit 15.

In step S1202, the related sentence conformity determination unit 1101 reads out the parameters after learning of the compatibility determination model X1 by the model learning device 1130 stored in the parameter storage unit 1107.

In step S1203, the utterance sentence input unit 7 receives input of the utterance sentence for dialogue execution from the user.

In step S1204, the related sentence acquisition unit 8 acquires a plurality of related sentences related to the speech sentence for dialogue execution from Web 1000, which is one of the text databases.

In step S1205, the related sentence conformity determination unit 1101 inputs the dialogue execution utterance sentence and the acquired related sentence into the post-learning suitability determination model X1 obtained by using the read parameters, and thereby the goodness of fit. To generate.

In step S1206, the related sentence selection unit 1102 selects a dialogue execution related sentence from the plurality of related sentences acquired by the related sentence acquisition unit 8 according to the goodness of fit, and outputs the dialogue execution related sentence to the response sentence generation unit 9.

In step S1207, the response sentence generation unit 9 generates a dialogue execution response sentence by inputting a dialogue execution related sentence into the post-learning response sentence generation model A obtained by using the read parameters.

As described above, according to the present embodiment, when the topic of the related sentence acquired by the related sentence acquisition unit is different from the topic of the utterance sentence, such a related sentence is not used for generating the response sentence. Can be.

In addition, in this embodiment, the function of preventing the response sentence from being generated from the related sentences having different topics is added to the dialogue response system of the first embodiment, but the present invention is limited to this. It's not a thing. It can be similarly added to the dialogue response system of the second to seventh embodiments.

Further, in the present embodiment, the response sentence is generated from the conformity-related sentence, but the present invention is not limited to this. The conformity-related sentence may be directly used as the response sentence to the utterance sentence.

[9th embodiment]
FIG. 37 is a diagram showing the configuration of the dialogue response system 1200 of the ninth embodiment.

The dialogue response system 1200 includes a model learning device 1220 and a dialogue device 1230.

The model learning device 1220 of the present embodiment differs from the model learning device 32 of the third embodiment in that the related sentence acquisition unit 1201, the conformity imparting unit 1203, the corpus output unit 1104, and the corpus storage unit. The 1108, the conformity determination learning processing unit 1105, and the parameter storage unit 1106.

The related sentence acquisition unit 1201 acquires sentences related to user-spoken sentences included in the dialogue data from Web 1000, which is one of the text databases. Specifically, the related sentence acquisition unit 1201 uses, for example, a sentence including at least one noun and / or at least one predicate included in the user response sentence from the data on the Web 1000 as a matching related sentence (learning related sentence). get. The related sentence acquisition unit 1201 acquires, for example, a sentence including one noun included in the user response sentence from the data on the Web 1000 and including at least one noun not included in the user response sentence as a nonconformity related sentence.

FIG. 38 is a diagram showing examples of dialogue data, learning utterance sentences, learning related sentences, and learning response sentences according to the ninth embodiment. As shown in FIG. 38, dialogue data consisting of the user utterance "Cosmos this morning!" And the user response "Thank you. Cosmos somehow makes me feel the sign of autumn." Is acquired.

A sentence containing the noun "Cosmos", the noun "Autumn sign", and the predicate "Feeling" included in the user response sentence is searched from Web1000, and the conformance related sentence (learning related sentence) "Cosmos feels the autumn sign". Let me. "Is acquired.

In addition, a sentence containing the noun "Cosmos" included in the user response sentence and containing the noun "How to grow" not included in the user response sentence is searched from Web1000, and the nonconformity-related sentence "How to grow cosmos is difficult." To be acquired.

The goodness-of-fit addition unit 1203 assigns a goodness of fit indicating the suitability of the acquired related sentence. The goodness-of-fit addition unit 1203 assigns a goodness of fit (for example, a numerical value “1”) indicating conformity to the acquired conformity-related sentence. The goodness-of-fit addition unit 1203 assigns a goodness of fit (for example, a numerical value “0”) indicating that the acquired nonconformity-related sentence is not conforming.

A nonconformity-related sentence to which a goodness of fit indicating non-conformity is given is sent to the corpus output unit 1104.

The conformity-related sentence to which the goodness of fit indicating conformity is given is sent to the learning response sentence generation unit 63 as a learning-related sentence. Similar to the third embodiment, the learning response sentence generation unit 63 generates a learning response sentence in which the learning-related sentence is modified into an appropriate format as a dialogue response, and the learning response sentence outputs a corpus. It is sent to unit 1104.

Since the corpus output unit 1104, the corpus storage unit 1108, the conformity determination learning processing unit 1105, and the parameter storage unit 1106 are the same as those described in the eighth embodiment, the description will not be repeated.

The dialogue device 1230 of the present embodiment differs from the dialogue device 23 of the third embodiment in that the dialogue device 1230 includes a related sentence compatibility determination unit 1101 and a related sentence selection unit 1102. ..

Since the related sentence conformity determination unit 1101 and the related sentence selection unit 1102 are the same as those described in the eighth embodiment, the description will not be repeated.

FIG. 39 is a flowchart showing the operation procedure of the model learning device 1220 according to the ninth embodiment.

In step S1301, the dialogue data acquisition unit 39 acquires dialogue data composed of a user utterance sentence such as twitter and a user response sentence from Web 1000 which is a kind of text database. The user's utterance becomes the learning utterance.

In step S1302, the related sentence acquisition unit 1302 acquires a conforming related sentence (learning related sentence) and a nonconforming related sentence from the data on the Web 1000.

In step S1303, the goodness-of-fit granting unit 1203 assigns a goodness of fit indicating the conformity of the acquired related sentences (conformity-related sentences and nonconformity-related sentences).

In step S1304, if the value indicates that the given goodness of fit is conforming (S1304: YES), the process proceeds to step S1305, indicating that the given goodness of fit is conforming. If it is not a value (S1304; NO), the process proceeds to step S1306.

In step S1305, the learning response sentence generation unit 63 generates a learning response sentence obtained by converting a learning-related sentence into a dialogue-like sentence.

In step S1306, steps S1301 to S1305 are repeated until all inputs are completed. The corpus output unit 1104 includes a plurality of sets of utterance sentences for learning, related sentences, and conformity of related sentences, and further, a set of related sentences for learning and response sentences for learning, which are conforming related sentences among the related sentences. A learning corpus containing a plurality of corpora is generated and written in the corpus storage unit 1108.

In step S1307, the conformity determination learning processing unit 1105 uses the learning utterance sentence, the related sentence, and the conformity degree included in the learning corpus to input the utterance sentence and the related sentence, and outputs the conformity degree as the output. The determination model X1 is learned.

In step S1308, the parameter storage unit 1106 stores the parameters after learning of the conformity determination model X1 by the conformity determination learning processing unit 1105.

In step S1309, the response generation learning processing unit 26 uses the learning utterance sentence, the learning related sentence, and the learning response sentence included in the learning corpus to input the utterance sentence and the related sentence, and outputs the response sentence. The response sentence generation model B is learned.

In step S1310, the parameter storage unit 14 stores the parameters after learning of the response sentence generation model B by the response generation learning processing unit 26.

As described above, in the present embodiment, it is possible to automatically generate a response sentence and set the goodness of fit of the related sentence without manpower.

[10th Embodiment]
FIG. 40 is a diagram showing the configuration of the dialogue response system 1300 according to the tenth embodiment.

The dialogue response system 1300 includes a model learning device 1320 and a dialogue device 1330.

The model learning device 1320 of the present embodiment differs from the model learning device 2 of the first embodiment in that the topic generation unit 1338, the related sentence acquisition unit 1309, the topic compatibility input unit 1303, and the corpus output. A unit 1304, a corpus storage unit 1308, a conformity determination learning processing unit 1305, and a parameter storage unit 1306. The difference between the dialogue device 1330 of the present embodiment and the dialogue device 3 of the first embodiment will be described later.

FIG. 41 is a diagram showing an example of a learning utterance sentence, a search topic, a related sentence, and a goodness of fit. Hereinafter, the operations of the topic generation unit 1338, the related sentence acquisition unit 1309, and the topic suitability input unit 1303 will be described with reference to FIG. 41.

The topic generation unit 1338 generates a plurality of search topics by extracting nouns included in the learning utterance sentence and other than the predetermined excluded nouns.

As shown in FIG. 41, the topic generation unit 1338 extracts the nouns “study trip” and “Nara” from the utterance sentence for learning “I'm going to Nara on a school trip”. The topic generation unit 1338 generates the topic "study trip" as the first search topic. The topic generation unit 1338 generates the topic "Nara" as the second search topic. The topic generation unit 1338 generates the main topic "study trip" and the sub-topic "Nara" as the third search topic. The topic generation unit 1338 generates the main topic "Nara" and the sub-topic "study trip" as the fourth search topic.

The related sentence acquisition unit 1309 acquires a related sentence from Web 1000, which is one of the text databases, by using each search topic generated by the topic generation unit 1338.

As shown in FIG. 41, the related sentence acquisition unit 1309 acquires a plurality of related sentences including "study trip" by using the first search topic. For example, four related sentences of related sentences 1-1 to 1-4 are acquired.

The related sentence acquisition unit 1309 acquires a plurality of related sentences including "Nara" by using the second search topic. For example, for example, four related sentences of related sentence 2-1 to related sentence 2-4 are acquired.

The related sentence acquisition unit 1309 uses the third search topic to acquire a plurality of related sentences including the main topic "study trip" directly related to the predicate and the sub-topic "Nara" not directly related to the predicate. .. For example, four related sentences of related sentences 3-1 to 3-4 are acquired. Directly related to the predicate means that it is connected to the predicate by a case particle.

The related sentence acquisition unit 1309 uses the fourth search topic to acquire a plurality of related sentences including the main topic "Nara" directly related to the predicate and the sub-topic "study trip" not directly related to the predicate. .. For example, four related sentences of related sentences 4-1 to 4-4 are acquired.

The topic suitability input unit 1303 receives an input of a goodness of fit indicating the suitability of the search topic used for acquiring the related sentence. If the related sentence obtained by using each search topic has the same or similar topic as the learning utterance sentence, the goodness of fit indicates that the search topic is suitable (for example, the person on the development side). , Numerical value "1"), and if the topics are different, enter the goodness of fit (for example, numerical value "0") indicating that the search topic does not match.

As shown in FIG. 41, the related sentence 1-1 searched using the first search topic has the same or similar topic as the learning utterance sentence, and the related sentence 1- searched using the first search topic. When 2, 1-3, 1-4 have different topics from the utterance sentence for learning, "1", "0", "0", and "0" are input as the goodness of fit of the first search topic.

The relationship between the learning utterance sentence generated by the topic suitability input unit 1303, the search topic, and the goodness of fit is sent to the corpus output unit 1304.

A related sentence (hereinafter referred to as a non-conforming related sentence) in which a goodness of fit indicating non-conformity is input by the topic conformity input unit 1303 is not sent to the learning response sentence input unit 13.

A related sentence (hereinafter referred to as a conformity-related sentence) in which the goodness of fit indicating conformity is input by the topic conformity input unit 1303 is sent to the learning response sentence input unit 13 as a learning-related sentence. Similar to the first embodiment, the learning response sentence input unit 13 receives the input of the learning response sentence in which the learning-related sentence is modified into an appropriate format as the response of the dialogue, and the learning response sentence is a corpus. It is sent to the output unit 1304. In the example of FIG. 41, the related sentence 1-1, the related sentence 2-2, the related sentence 3-1 and the related sentence 3-2, the related sentence 3-3, the related sentence 4-2, and the related sentence 4-3 are for learning. It is sent to the response text input unit 13.

The corpus output unit 1304 generates a learning corpus including a plurality of sets of learning utterance sentences, search topics, and suitability, and a plurality of sets of learning related sentences and learning response sentences, and causes the corpus storage unit 1308 to generate a learning corpus. Write.

The corpus storage unit 1308 stores the learning corpus.
FIG. 42A is a diagram showing an example of a plurality of sets of learning-related sentences and learning response sentences in the learning corpus stored in the corpus storage unit 1308. FIG. 42B is a diagram showing an example of a plurality of sets of learning utterance sentences, search topics, and goodness of fit in the learning corpus stored in the corpus storage unit 1308.

The conformity determination learning processing unit 1305 learns the conformity determination model X2 which includes the utterance sentence and the search topic in the input and outputs the conformity degree by using the utterance sentence for learning, the search topic, and the goodness of fit. The suitability determination model X2 includes a neural network for deep learning and its parameters.

In the example of FIG. 41, for example, "going to Nara on a school trip" and "study trip" are input and "1" is output as learning 1, "going to Nara on a school trip" and "study". Learning 2 with "Travel" as input and "0" as output, Learning 3, "I'm going to Nara on a school trip" and "Study trip" as input and "0" as output, "Nara on school trip" Learning 4 is performed with "go to" and "study trip" as inputs and "0" as an output. The same is true for other search topics.

The parameter storage unit 1306 stores the parameters after learning of the conformity determination model X2 by the conformity determination learning processing unit 1305. The parameter storage unit 1306 stores parameters (weight matrix, etc.) after learning the neural network. The parameters after learning of the conformity determination model X2 stored in the parameter storage unit 1306 can be taken out to the outside via an input / output port or the like (not shown).

The dialogue device 1330 of the present embodiment differs from the dialogue device 3 of the first embodiment in that it has a parameter storage unit 1307, a topic generation unit 1368, a topic suitability determination unit 1312, and a search topic selection unit 1349. And the related sentence acquisition unit 1341.

The parameter storage unit 1307 stores the parameters after learning of the conformity determination model X2 by the model learning device 1320. For example, the parameter storage unit 1307 stores parameters (weight matrix, etc.) after learning of the neural network. The parameters after learning of the conformity determination model X2 taken out from the model learning device 1320 are input to the dialogue device 1330 and stored in the parameter storage unit 1307. These parameters may be distributed to the user together with the program that realizes the function of the dialogue device 1330.

Similar to the topic generation unit 1338, the topic generation unit 1368 extracts nouns other than the predetermined excluded nouns included in the dialogue execution utterance sentence to generate a plurality of search topics. If the dialogue execution utterance sentence is the same as the learning utterance sentence "I'm going to Nara on a school trip" in FIG. 39, the topic generation unit 1368 extracts the nouns "study trip" and "Nara". .. The topic generation unit 1368 generates the topic "study trip" as the first search topic. The topic generation unit 1368 generates the topic "Nara" as the second search topic. The topic generation unit 1368 generates the main topic "study trip" and the sub-topic "Nara" as the third search topic. The topic generation unit 1368 generates the main topic "Nara" and the sub-topic "study trip" as the fourth search topic.

The topic suitability determination unit 1312 determines the suitability of a plurality of generated search topics. That is, the topic suitability determination unit 1312 generates the goodness of fit by inputting the utterance sentence for dialogue execution and each search topic into the learned suitability determination model X2. For example, the output goodness of fit is a value close to "1" when the goodness of fit of the search topic is high, and a value close to "0" when the goodness of fit of the search topic is low.

The search topic selection unit 1349 selects the search topic having the highest goodness of fit from the plurality of search topics, and outputs the search topic to the related sentence acquisition unit 1341.

The related sentence acquisition unit 1341 acquires one related sentence from Web 1000, which is a kind of text database, by using the selected search topic. The related sentence acquisition unit 1341 includes a main topic directly related to the predicate and a subtopic not directly related to the predicate when the search topic with the highest goodness of fit is composed of a main topic and a subtopic. Get two related statements.

FIG. 43 is a diagram showing a neural network that realizes the conformity determination model X2 used in the conformity determination learning processing unit 1305 and the topic suitability determination unit 1312 of the tenth embodiment.

The convolutional neural network CNN3 receives the learning utterance sentence which is the first input data. The convolutional neural network CNN3 includes a convolutional layer C31 and a pooling layer C32.

The first input data, the utterance sentence for learning, "I'm going to Nara on a school trip," is represented by data with a size of N × 6.

Based on the calculation results of the learning utterance sentence and the plurality of filters, in the convolution layer C31, the first convolution calculation result of M 1 × 5 size and the second convolution calculation result of M 1 × 4 size are obtained. The convolution operation result and the third convolution operation result having a size of M pieces of 1 × 3 can be obtained.

By MAX pooling of the first calculation result, the second calculation result, and the third calculation result, M first pooling results, M second pooling results, and M third pooling results in the pooling layer C32. The pooling result of is obtained.

The convolutional neural network CNN4 receives a search topic which is the second input data. The convolutional neural network CNN4 includes a convolutional layer C41 and a pooling layer C42.

The search topic "Study Trip Nara", which is the second input data, is represented by data having a size of N × 2.

Depending on the calculation result of the search topic and the plurality of filters, in the convolution layer C41, the fourth convolution operation result of M 1 × 2 size and the fifth convolution operation of M 1 × 1 size The result is obtained.

By MAX pooling of the fourth calculation result and the fifth calculation result, M fourth pooling results and M fifth pooling results are obtained in the pooling layer C42.

The coupling portion COE2 combines the first to fifth pooling results to generate a 5M-dimensional coupling vector.

The multilayer perceptron 600 includes an input layer 601, an intermediate layer 602, and an output layer 603.

The data of the 5M-dimensional coupling vector is sent to the input layer 601. The goodness of fit of the search topic is output by the calculation of the input layer 601, the intermediate layer 602, and the output layer 603.

FIG. 44 is a diagram showing details of the convolution operation in the convolutional neural network CNN3.

In the convolutional neural network CNN3, M N × 2 size filters, M N × 3 size filters, and M N × 4 size filters are used.

By the convolution operation of the N × 6 size learning utterance sentence and the M N × 2 size filters, the first convolution operation result of M 1 × 5 sizes can be obtained.

By the convolution operation of the N × 6 size learning utterance sentence and the M N × 3 size filters, the second convolution operation result of the M 1 × 4 size can be obtained.

By the convolution operation of the N × 6 size learning utterance sentence and the M N × 4 size filters, a third convolution operation result of M 1 × 3 sizes can be obtained.

FIG. 45 is a diagram showing the details of the convolution operation in the convolutional neural network CNN4.

In the convolutional neural network CNN4, M N × 1 size filters and M N × 2 size filters are used.

By the convolution operation of the N × 2 size search topic and the M N × 1 size filters, a fourth convolution operation result of M 1 × 2 sizes can be obtained.

By the convolution operation of the N × 2 size search topic and the M N × 2 size filters, the fifth convolution operation result of M 1 × 1 sizes can be obtained.

At the time of learning, the goodness of fit input to the topic suitability input unit 1303 is used as teacher data. The error backpropagation method updates the weights of the Multilayer Perceptron 600 and the coefficients of the filter so that the error with the teacher data is minimized. The weight of the multi-layer perceptron 600 after learning and the coefficient of the filter are stored in the parameter storage unit 1306. The weight of the multi-layer perceptron 600 after learning stored in the parameter storage unit 1306 and the coefficient of the filter are transferred to the parameter storage unit 1307.

At the time of dialogue response, the weight of the multi-layer perceptron 600 after learning and the coefficient of the filter stored in the parameter storage unit 1307 are read out and used. The goodness of fit output from the output layer 603 of the multi-layer perceptron 600 is sent to the search topic selection unit 1349 and used for selecting the search topic.

FIG. 46 is a flowchart showing the operation procedure of the model learning device 1320 according to the tenth embodiment.

In step S1401, the utterance sentence input unit 11 receives the input of the utterance sentence for learning.
In step S1402, the topic generation unit 1338 generates a plurality of search topics by extracting nouns included in the learning utterance sentence and other than the predetermined excluded nouns.

In step S1403, the related sentence acquisition unit 1309 acquires a related sentence from Web 1000, which is one of the text databases, using each search topic generated by the topic generation unit 1338.

In step S1404, the topic suitability input unit 1303 receives the input of the goodness of fit indicating the suitability of the search topic used for acquiring the related sentence.

In step S1405, if the value indicates that the input goodness of fit is conforming (S1405: YES), the process proceeds to step S1406, indicating that the input goodness of fit is conforming. If it is not a value (S1405; NO), the process proceeds to step S1407.

In step S1406, the learning response sentence input unit 13 is generated by modifying the learning-related sentence by the operation of the personnel on the development side so that the conformance-related sentence (learning-related sentence) becomes appropriate as the response of the dialogue. Receives the input of the learning response sentence.

In step S1407, steps S1401 to S1406 are repeated until all inputs are completed. The corpus output unit 1304 generates a learning corpus including a plurality of sets of learning utterance sentences, search topics, and suitability, and a plurality of sets of learning related sentences and learning response sentences, and causes the corpus storage unit 1308 to generate a learning corpus. Write.

In step S1408, the conformity determination learning processing unit 1305 includes the utterance sentence and the search topic in the input by using the learning utterance sentence, the search topic, and the goodness of fit stored in the learning corpus, and outputs the goodness of fit. Learn the goodness-of-fit determination model X2.

In step S1409, the parameter storage unit 1306 stores the parameters after learning of the conformity determination model X2 by the conformity determination learning processing unit 1305.

In step S1410, the response generation learning processing unit 6 uses the learning-related sentence and the learning response sentence stored in the learning corpus to input the related sentence and output the response sentence as the response sentence generation model. Learn A.

In step S1411, the parameter storage unit 14 stores the parameters after learning of the response sentence generation model A by the response generation learning processing unit 6.

FIG. 47 is a flowchart showing the operation procedure of the dialogue device 1330 according to the tenth embodiment.

In step S1501, the response sentence generation unit 9 reads out the parameters after learning of the response sentence generation model A by the model learning device 1330 stored in the parameter storage unit 15.

In step S1502, the topic suitability determination unit 1312 reads out the parameters after learning of the suitability determination model X2 by the model learning device 1330 stored in the parameter storage unit 1307.

In step S1503, the utterance sentence input unit 7 receives input of the utterance sentence for dialogue execution from the user.

In step S1504, the topic generation unit 1368 extracts nouns other than the predetermined excluded nouns included in the dialogue execution utterance sentence, and generates a plurality of types of search topics.

In step S1505, the topic suitability determination unit 1312 generates a goodness of fit by inputting an utterance sentence for dialogue execution and each search topic into the post-learning suitability determination model X2 obtained by using the read parameters. do.

In step S1506, the search topic selection unit 1349 selects one search topic from the plurality of search topics according to the goodness of fit, and outputs it to the related sentence acquisition unit 1341.

In step S1507, the related sentence acquisition unit 1341 acquires one related sentence related to the speech sentence for dialogue execution from Web 1000, which is one of the text databases, by searching using the selected search topic.

In step S1508, the response sentence generation unit 9 generates a dialogue execution response sentence by inputting a dialogue execution related sentence into the post-learning response sentence generation model A obtained by using the read parameters.

As described above, according to the present embodiment, when the search topic used for acquisition by the related sentence acquisition unit is inappropriate, it is possible to prevent the related sentence from being acquired using such a search topic. can.

In the present embodiment, a function for preventing related sentences from being acquired from an inappropriate search topic is added to the dialogue response system of the first embodiment, but the present invention is limited to this. is not. It can be similarly added to the dialogue response system of the second to eighth embodiments.

Further, in the present embodiment, the response sentence is generated from the related sentence acquired by using the search topic having a high degree of conformity, but the present invention is not limited to this. As the response sentence to the utterance sentence, the related sentence acquired by using the search topic with high goodness of fit may be directly used.

[Eleventh Embodiment]
The components of the model learning device and the interactive device according to the above-described embodiment can be realized by computer hardware and a computer program executed on the computer hardware.

FIG. 48 is a diagram showing an internal configuration of a model learning device or a dialogue device.
With reference to FIG. 48, the model learning device or dialogue device includes a bus 3300, a CPU (Central Processing Unit) 3301, a ROM (Read Only Memory) 3302, a RAM (Random Access Memory) 3303, and an HDD (Hard Disk). Drive) 3304, DVD drive 3306 to which DVD (Digital Versatile Disk) 3307 is connected, network I / F3309, keyboard 3310, mouse 3311, memory port 3312 to which removable memory 3313 is connected, and display 3308. To prepare for.

The keyboard 3310 and the mouse 3311 accept input from development personnel or users. The model learning program or dialogue program for realizing the functions described in the above-described embodiment on the computer is transferred from the DVD 3307, the removable memory 3313, and the network I / F 3309 to the HDD 3304. The ROM 3302 stores, for example, a boot program or the like. The RAM 3303 stores a program being executed, work data thereof, and the like. The network I / F3309 is connected to the Internet 3400 and can acquire data on the Web 1000.

[Modification example]
The present invention is not limited to the above embodiment. The present invention also includes the following modifications.

(1) In the above embodiment, the related sentence search unit searches for sentences posted on the Web, but the present invention is not limited to this. The related sentence search unit of the model learning device may search for sentences that are created for a purpose other than the purpose used by the model learning device and are included in the data that can be used by the public. The related sentence search unit of the dialogue device may search for sentences that are created for a purpose other than the purpose used in the dialogue and are included in the data that can be used by the public. As a result, sentences included in newspaper data recorded and distributed on DVDs or CDs can be used as related sentences instead of the Web.

(2) Not only the weight matrix but also parameters such as bias b, c, activation functions f, and g may be stored in the parameter storage units 14 and 15.

(3) As an example of a model that outputs a response statement, a model composed of a recurrent neural network is used, but the model is not limited to this, and a model using a neural network other than the recurrent neural network may be used. good. As an example of a model for determining related sentences and search topics, a model composed of a convolutional neural network and a multi-layer perceptron was used, but the model is not limited thereto.

(4) It is also possible to perform preprocessing such as attaching a tag related to emotion to the original sentence and execute learning together with the tag information. As a result, it is possible to construct a dialogue response system that outputs based on emotions.

(5) It is also possible to learn a model in which a question sentence is input and an answer sentence expressing an advice and an answer to the question sentence is output, and a dialogue response system for outputting an advice and an answer to the question sentence can be constructed.

(6) In the first to sixth embodiments, the model learning device supplies the parameters after learning the model (neural network) to the dialogue device, but the present invention is not limited to this. The model learning device may supply data representing the model itself, that is, a configuration of a neural network constituting an encoder and a decoder, and parameters after training of the neural network to the dialogue device. The dialogue device generates a dialogue execution response statement by constructing a model based on the data representing the supplied model itself. The same applies to the eighth to tenth embodiments.

It should be considered that the embodiments disclosed this time are exemplary in all respects and not restrictive. The scope of the present invention is shown by the scope of claims rather than the above description, and is intended to include all modifications within the meaning and scope equivalent to the scope of claims.

1,21,31,41,81,91,551,1100,1200,1300 Dialogue response system, 2,22,32,42,75,78,1120,1220,1320 Model learning device, 3,23,43, 76,77,1130,1230,1330 Dialogue device, 4,24,34,44,82,92,1140,1240,1340 Corpus generator, 5,25,87,97,1108,1308 Corpus storage, 6, 26,86 Response generation learning processing unit, 7,11 utterance sentence input unit, 8,12,28,45,48,1201,1341 Related sentence acquisition unit, 9,29,89 Response sentence generation unit, 13,38 For learning Response sentence input unit, 14, 15, 1106, 1107, 1306, 1307 parameter storage unit, 39 dialogue data acquisition unit, 33, 63, 83, 93 learning response sentence generation unit, 49, 53, 57 original sentence ranking unit, 51 Binary triple generator, 52 Original sentence acquisition unit, 55 Question sentence generation unit, 56 Answer and original sentence acquisition unit, 64,74,84,85,95,1104,1304 Corpus output unit, 151,153,157,158 , 351,352,451,452 Encoder, 152,154 Decoder, 552 Learning Control Unit, 1000 Web, 1051 Question / Answer System, 1054 Database, 1101 Related Sentence Conformity Judgment Unit, 1102 Related Sentence Selection Unit, 1103 Related Sentence Conformity Gender input unit, 1105, 1305 Conformity judgment learning processing unit, 1202 Conformity granting unit, 1303 topic conformity input unit, 1338, 1368 topic generation unit, 1312 topic conformity judgment unit, 1349 search topic selection unit, 3300 bus, 3301 CPU, 3302 ROM, 3303 RAM, 3304 HDD, 3306 DVD drive, 3307 DVD, 3308 display, 3309 network I / F, 3310 keyboard, 3311 mouse, 3312 memory port, 3313 removable memory, 3400 Internet, 500,600 multi-layer perceptron, 501,601 input layer, 502,602 intermediate layer, 503,603 output layer, CNN1, CNN2, CNN3, CNN4 convolutional neural network, C11, C21 convolutional layer, C12, C22 pooling layer, COE1, COE2 joint.

Claims (10)

The learning utterance sentence, the learning related sentence related to the learning utterance sentence obtained by searching the sentence in the text database, and the learning related sentence are in an appropriate format as a response of the dialogue. A corpus storage unit that stores a learning corpus including a modified learning response sentence,
Using the learning corpus, a model learning device including a learning processing unit that includes a related sentence as an input and learns a response sentence generation model that outputs a response sentence, and a model learning device.
The utterance text input section where the utterance text for dialogue execution is input, and
A related sentence acquisition unit that acquires a related sentence for dialogue execution related to the spoken sentence for executing the dialogue by searching a sentence in the text database, and a related sentence acquisition unit.
Dialogue by inputting the related sentence for executing the dialogue into the response sentence generation model trained to include the related sentence related to the utterance sentence as an input and output the response sentence in response to the input. A dialogue response system consisting of a dialogue device equipped with a response statement generator that generates a response statement for execution. It is a model learning device that learns the response sentence generation model used in the dialogue device that automatically generates the response sentence for the utterance sentence.
The learning utterance sentence, the learning related sentence related to the learning utterance sentence obtained by searching the sentence in the text database, and the learning related sentence are in an appropriate format as a response of the dialogue. A corpus storage unit that stores a learning corpus including a modified learning response sentence,
A model learning device including a learning processing unit that uses the learning corpus to learn a response sentence generation model that includes a related sentence as an input and outputs the response sentence as an output. The corpus storage unit stores a learning corpus including the utterance sentence for learning, one or more related sentences for learning, and the response sentence for learning.
The model learning device according to claim 2, wherein the learning processing unit learns a response sentence generation model that includes the utterance sentence and the related sentence as an input and outputs the response sentence by using the learning corpus. It is a dialogue device that automatically generates a response sentence to an utterance sentence.
The utterance text input section where the utterance text for dialogue execution is input, and
A related sentence acquisition unit that acquires a related sentence for dialogue execution related to the spoken sentence for executing the dialogue by searching a sentence in the text database, and a related sentence acquisition unit.
Dialogue by inputting the related sentence for executing the dialogue into the response sentence generation model trained to include the related sentence related to the utterance sentence as an input and output the response sentence in response to the input. A dialogue device equipped with a response statement generator that generates a response statement for execution. The response sentence generation model includes the utterance sentence and the related sentence as the input.
The response sentence generation unit generates a response sentence for executing the dialogue by inputting the utterance sentence for executing the dialogue and one or a plurality of related sentences for executing the dialogue into the response sentence generation model. Item 4. The dialogue device according to item 4. A sentence including at least one nomenclature and / or at least one predicate included in the user response sentence of dialogue data consisting of a user spoken sentence and a user response sentence is acquired as a related sentence for learning, and the user response sentence is described as described above. The model learning device according to claim 2, further comprising a corpus generation unit that extracts parts related to nomenclatures and / or predicates included in related sentences for learning and uses them as response sentences for learning. A dialogue response system consisting of a model learning device and a dialogue device.
The model learning device is
A corpus storage unit that stores a learning corpus including a learning utterance sentence, a related sentence related to the learning utterance sentence acquired by searching a sentence in a text database, and a degree of conformity of the related sentence. When,
Using the learning corpus, it is provided with a first learning processing unit that learns a conformity determination model that includes utterance sentences and related sentences in the input and outputs the goodness of fit.
The dialogue device is
The utterance text input section where the utterance text for dialogue execution is input, and
A related sentence acquisition unit that acquires a related sentence for dialogue execution related to the spoken sentence for executing the dialogue by searching a sentence in the text database, and a related sentence acquisition unit.
The utterance sentence for executing the dialogue and the acquired said sentence were added to the conformity determination model trained to include the utterance sentence and the related sentence in the input and output the degree of conformity of the related sentence in response to the input. It is equipped with a related sentence conformity determination unit that generates a degree of conformity by inputting a related sentence for executing a dialogue .
The corpus storage unit further stores a learning corpus including a learning response sentence in which the learning related sentence having a high goodness of fit of the related sentence is modified into an appropriate format as a response of the dialogue .
The model learning device further
A second learning processing unit for learning a response sentence generation model that includes the related sentence in the input and outputs the response sentence by using the related sentence for learning and the response sentence for learning is provided.
The dialogue device further
Dialogue by inputting the related sentence for executing the dialogue into the response sentence generation model trained to include the related sentence related to the utterance sentence as an input and output the response sentence in response to the input. An interactive response system that includes a response statement generator that generates response statements for execution . A sentence containing at least one noun and / or at least one predicate included in the user response sentence of the dialogue data consisting of the user utterance sentence and the user response sentence is acquired as the related sentence having a high degree of conformity.
The seventh aspect of claim 7, further comprising a corpus generation unit that acquires a sentence including one noun included in the user response sentence and at least one noun not included in the user response sentence as the related sentence having low goodness of fit. Dialogue response system. A dialogue response system consisting of a model learning device and a dialogue device.
The model learning device is
A learning corpus that includes a learning utterance sentence, a search topic used when searching a sentence in a text database to obtain a related sentence related to the learning utterance sentence, and a degree of conformity of the search topic. A corpus memory that memorizes
Using the learning corpus, a first learning processing unit for learning a conformity determination model that includes utterance sentences and search topics as inputs and outputs the goodness of fit is provided.
The dialogue device is
The utterance text input section where the utterance text for dialogue execution is input, and
A first topic generation unit that generates a search topic used when searching a sentence in a text database to acquire a related sentence related to the utterance sentence for executing the dialogue based on the utterance sentence for executing the dialogue. When,
The utterance sentence for executing the dialogue and the generated search are added to the conformity determination model trained to include the utterance sentence and the search topic in the input and output the goodness of fit of the search topic in response to the input. A topic suitability judgment unit that generates a goodness of fit by inputting a topic,
A selection unit that selects one search topic from a plurality of search topics based on the goodness of fit generated by the topic suitability determination unit, and a selection unit.
It comprises a first related sentence acquisition unit that acquires a dialogue execution related sentence related to the dialogue execution utterance sentence by searching a sentence in the text database using the selected search topic .
The corpus storage unit modifies the learning-related sentences related to the learning utterance sentences obtained by searching the sentences in the text database and the learning-related sentences into an appropriate format as a response to the dialogue. Memorize the learning corpus, including the learning response sentences that have been made,
The model learning device further
A second learning processing unit for learning a response sentence generation model that includes the related sentence in the input and outputs the response sentence by using the related sentence for learning and the response sentence for learning is provided.
The dialogue device further
Dialogue by inputting the related sentence for executing the dialogue into the response sentence generation model trained to include the related sentence related to the utterance sentence as an input and output the response sentence in response to the input. An interactive response system that includes a response statement generator that generates response statements for execution . The model learning device is
A second topic generation unit that generates the search topic based on the utterance sentence for learning,
It is provided with a second related sentence acquisition unit that acquires a related sentence related to the utterance sentence for learning by searching a sentence in a text database using the generated search topic.
The first topic generation unit and the second topic generation unit generate a search topic consisting of a main topic and a subtopic, and generate a search topic.
When the search topic consists of the main topic and the subtopic, the first related sentence acquisition unit and the second related sentence acquisition unit include the main topic directly related to the predicate, and the predicate includes the predicate. The dialogue response system according to claim 9 , wherein a related sentence including a subtopic that is not directly related is acquired.

Images