Program, device and method for resetting dialogue history and generating response sentence
The program and method address dialogue breakdowns and repetitive conversations by resetting latent variables and using advanced learning techniques to generate responsive sentences on new topics, ensuring coherent and engaging interactions.
Patent Information
- Application Number
- JP2021190599
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-11-24
- Publication Date
- 2024-12-09
- Estimated Expiration
- 2041-11-24
AI Technical Summary
Existing dialogue systems struggle to maintain coherent conversation when users introduce unexpected topics or continue indifferent conversations due to training based on past dialogue history, leading to breakdowns or stagnation.
A program and method utilizing a recurrent neural network that resets latent variables based on past dialogue history when dialogue breaks or becomes stagnant, incorporating features like hierarchical vocabulary, positive/negative learning, and facial expression analysis to generate responsive sentences on new topics.
Enables continuous and engaging dialogue by addressing unexpected topic changes and avoiding repetitive agreements, enhancing conversation richness and coherence.
Smart Images

Figure 0007596253000001 
Figure 0007596253000002 
Figure 0007596253000003
Abstract
Description
[Technical field]
[0001] The present invention relates to a technique for a dialogue device that realizes natural dialogue with a user. [Background technology]
[0002] The dialogue device can generate a response sentence that is a natural dialogue such as a chat in response to a sentence spoken by a user. The sentence spoken by the user may be text-based or may be a sentence in which a spoken voice is converted into text. The response sentence of the system may also be text-based or may be a sentence in which text is spoken by speech synthesis. Examples of such dialogue devices include "Siri (registered trademark)" and "Shabette Concierge (registered trademark)."
[0003] Conventionally, in order to realize natural dialogue such as casual conversation, there is a technology of a Seq2Seq model composed of an encoder / decoder using LSTM (Long Short-Term Memory) (see, for example, Non-Patent Document 1). According to this technology, the encoder inputs a word token sequence of an utterance sentence and saves a latent variable (vector representation) by LSTM. The decoder generates a word token sequence of a response sentence using the latent variable of the hidden layer created by the encoder. However, since there are relatively many backchannels such as "That's nice," "That's right," and "That's right" in the learning data of general dialogue, the decoder tends to generate safe backchannels probabilistically.
[0004] In addition, there is a one-to-many learning technology that uses past dialogue history (pairs of utterances and responses) as an encoder-decoder model (see, for example, Non-Patent Document 2). This technology can determine the timing of topic transitions and decide the next topic. In other words, it enables in-depth dialogue even within the same topic. ,Next This allows the system to automatically learn the transition probabilities of hidden conditions, resulting in richness in the diversity, information, and character of utterances. [Prior art documents]
Non-licensed literature
[0005]
Non-licensed literature 1
Non-licensed Document 2
Non-licensed Document 4
[0006] However, according to the techniques described in Non-Patent Documents 1 and 2, the dialogue may break down, for example, as described below. <Case Study 1> Sentence: I'm hungry Response: Do you want something to eat? Sentence: I want to eat gyoza. Response: Gyoza is delicious. Sentence: I suddenly got a stomach ache <- (suddenly changes topic) Response: ★ Shall I help you deliver the gyoza? <- (Dialogue breaks down) In this way, if a user suddenly talks about a different topic in the middle of a conversation, the conversation will break down due to being too influenced by the previous topic. This is because the encoder-decoder is trained based on the history of past conversations and cannot respond to sudden changes in topic.
[0007] There are also cases where disinterested dialogue continues, for example, as follows: <Case Study 2> Sentence: I'm hungry Response: Do you want something to eat? Sentence: I want to eat gyoza. Response: Gyoza is delicious. Sentence: Right Response: That's true <- (nod) Sentence: Okay. Response: ★Like <- (answer) In this way, the continuous use of backchannels that indicate agreement leads to a disinterested dialogue that ultimately makes it impossible to change the topic of conversation. This is also because the encoder-decoder is trained only on past backchannels.
[0008] Therefore, an object of the present invention is to provide a program, device, and method that can generate a response sentence to a user's spoken sentence without breaking down the dialogue and without continuing the interjections. [Means for solving the problem]
[0009] According to the present invention, there is provided a program for causing a computer to function so as to generate a response sentence in response to an utterance sentence, the program comprising: In a training stage, utterances and responses in teacher data are input in a time series to construct a learning model based on latent variables, and in an estimation stage, a recurrent neural network outputs a response each time an utterance serving as estimation data is input in a time series; a dialogue learning engine that constructs a learning model by associating utterances and valid responses as teacher data in a training stage, and inputs utterances as estimation data and responses output from a recurrent neural network in an estimation stage and outputs a validity score; As an estimation stage, If the score output from the dialogue learning engine is below a certain threshold, it is considered a dialogue breakdown. A dialogue relationship determination means for determining a dialogue history resetting means for resetting a latent variable based on a past dialogue history for the recurrent neural network when the dialogue relationship determining means determines that the answer is true; and make it work. The recurrent neural network outputs a response sentence estimated only from the utterances after the latent variables based on the past dialogue history are reset. The present invention is characterized in that the computer is made to function in such a manner.
[0012] According to another embodiment of the program of the present invention, A backchannel message registering means for registering backchannel messages in advance is provided; In the estimation step, the dialogue relation determination means determines that the response sentence output from the recurrent neural network is true as a backchannel succession when the backchannel sentence registered in the backchannel sentence registration means continues in time series a predetermined number of times or more. It is also preferable to function the computer in such a manner.
[0013] According to another embodiment of the program of the present invention, A vocabulary database in which vocabulary is conceptually organized hierarchically; a current topic word selection means for identifying a previous topic word based on a past dialogue history and selecting another current topic word that belongs to the same category as the previous topic word; and make it work. The recurrent neural network resets latent variables based on the past dialogue history, and then outputs a response based on the current topic. It is also preferable to make the computer further function as described above.
[0014] According to another embodiment of the program of the present invention, In the training stage, a learning model is constructed by associating utterances with positive or negative flags as teacher data, and in the estimation stage, an utterance positive / negative learning engine is used to input utterances as estimation data and output positive or negative flags. and make it work. The dialogue relation determination means determines that the topic change is true when the flag output from the speech sentence positive / negative learning engine is determined to be negative in the estimation stage. It is also preferable to function the computer in such a manner.
[0015] According to another embodiment of the program of the present invention, In the training stage, a learning model is constructed by associating a video of the facial expression of the user speaking with a positive or negative flag as teacher data, and in the estimation stage, a facial image positive / negative learning engine is used that inputs a video of the facial expression of the user speaking and outputs a positive or negative flag. and make it work. The dialogue relation determination means determines the topic change as true when the flag output from the face image positive / negative learning engine is determined as negative in the estimation stage. It is also preferable to function the computer in such a manner.
[0016] According to another embodiment of the program of the present invention, A vocabulary database in which vocabulary is conceptually organized hierarchically; a current topic word selection means for identifying a previous topic word based on a past dialogue history and selecting a current topic word that is an antonym or a different category of the previous topic word; and make it work. The recurrent neural network resets latent variables based on the past dialogue history, and then outputs a response based on the current topic. It is also preferable to make the computer further function as described above.
[0017] According to another embodiment of the program of the present invention, A Recurrent Neural Network (RNN) is a Curriculum Learning engine, a Long Short Term Memory (LSTM), or a Gated Recurrent Unit (GRU). It is also preferable to function the computer in such a manner.
[0018] According to the present invention, there is provided a dialogue device for generating a response sentence in response to an utterance sentence, In a training stage, utterances and responses in teacher data are input in a time series to construct a learning model based on latent variables, and in an estimation stage, a recurrent neural network outputs a response each time an utterance serving as estimation data is input in a time series; a dialogue learning engine that constructs a learning model by associating utterances and valid responses as teacher data in a training stage, and inputs utterances as estimation data and responses output from a recurrent neural network in an estimation stage and outputs a validity score; As an estimation stage, If the score output from the dialogue learning engine is below a certain threshold, it is considered a dialogue breakdown. A dialogue relationship determination means for determining a dialogue history resetting means for resetting a latent variable based on a past dialogue history for the recurrent neural network when the dialogue relationship determining means determines that the answer is true; having The recurrent neural network outputs a response sentence estimated only from the utterances after the latent variables based on the past dialogue history are reset. It is characterized by:
[0019] According to the present invention, there is provided a dialogue method for an apparatus for generating a response sentence in response to an utterance sentence, comprising the steps of: The device is In a training stage, utterances and responses in teacher data are input in a time series to construct a learning model based on latent variables, and in an estimation stage, a recurrent neural network outputs a response each time an utterance serving as estimation data is input in a time series; In the training stage, a learning model is constructed by associating utterances and correct response sentences as teacher data, and in the estimation stage, a dialogue learning engine is inputted with utterances as estimation data and response sentences outputted from a recurrent neural network, and outputs a correct score. having As an estimation stage, If the score output from the dialogue learning engine is below a certain threshold, it is considered a dialogue breakdown. a first step of determining First Step When the answer is determined to be true by 、 The recurrent neural network is the third step that outputs the response sentence estimated only from the utterance after the latent variables based on the past dialogue history are reset. Run It is characterized by: Effect of the Invention
[0020] According to the program, device, and method of the present invention, a response sentence can be generated in response to a user's utterance, which does not cause a breakdown in dialogue and does not involve continued backchanneling. [Brief description of the drawings]
[0021] [Figure 1] FIG. 2 is a functional configuration diagram of the dialogue device according to the present invention. [Diagram 2] FIG. 13 is a functional block diagram of a dialogue device further having a dialogue learning engine that outputs a score. [Diagram 3] FIG. 13 is a functional configuration diagram of a dialogue device further including a backchannel message registration unit. [Figure 4] FIG. 13 is an explanatory diagram showing a search of a vocabulary database. [Diagram 5] FIG. 13 is a functional configuration diagram of a dialogue device further having a spoken sentence positive / negative learning engine. [Figure 6] FIG. 13 is a functional configuration diagram of a dialogue device further having a face image positive / negative learning engine. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0022] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0023] FIG. 1 is a functional block diagram of a dialogue device according to the present invention.
[0024] 1, the dialogue device 1 generates a system response sentence in response to a sentence spoken by a user. The dialogue device 1 has a teacher database 101, a recurrent neural network 11, a dialogue relation determination unit 12, and a dialogue history reset unit 13. These functional components are realized by executing a program that causes a computer mounted on the device to function. The process flow of these functional components can also be understood as a dialogue method.
[0025] [Teacher Database 101] The teacher database 101 is a past dialogue corpus in which valid responses to utterances are recorded in a time-series correspondence. The teacher database 101 is used in the training stage of the recurrent neural network 11.
[0026] The recurrent neural network 11 inputs the utterance sentence and the response sentence as a language feature vector generated for each morpheme based on morphological analysis. Specifically, the language feature vector is a word converted into a distributed representation (word embedding representation) of a high-dimensional vector by a neural network. For example, Word2vec (registered trademark) can be used.
[0027] The teacher database 101 may be a database in which utterances and responses are recorded in advance as linguistic feature vectors. Furthermore, the teacher database 101 may be a text-based record of utterances and responses, which may be converted into language feature vectors when input to the recurrent neural network 11.
[0028] [Recurrent Neural Networks 11] A recurrent neural network (RNN) 11 functions as a training and estimation stage as follows. (Training phase) The spoken sentences and response sentences in the training data are input in chronological order, and a learning model based on latent variables is constructed. (Estimation stage) Each time an utterance sentence that will become estimation data is input in time series, a response sentence is output.
[0029] Specifically, the recurrent neural network 11 may be a Curriculum Learning engine, a Long Short Term Memory (LSTM), or a Gated Recurrent Unit (GRU).
[0030] Curriculum learning is a deep learning technique that first learns relatively easy examples and then gradually increases the difficulty of the examples, thereby improving prediction accuracy compared to learning random examples. Curriculum learning is particularly effective when learning is difficult using stochastic gradient descent in LSTM, for example. LSTM is capable of learning long-term dependencies, and combines the input gate and forget gate into one gate called an "update gate." It has an internal state and can transition to the next state based on the input value and state at each point in time. Similarly, the GRU also helps to retain memory of features from measurements taken many steps ago by introducing a forget-update gate, which effectively creates a shortcut path between each time step, making it faster. Of course, the recurrent neural network 11 is not limited to this, and may be a Seq2Seq-based network capable of extracting general features between utterances and responses, or may be a network based on seq2seq+attention or transform. er It may be an improved model such as
[0031] According to Fig. 1, the recurrent neural network 11 is configured by arranging N encoders and decoders and training them by collectively using the past N pairs of utterances and responses. The encoders are trained to input utterances and the decoders are trained to output responses. The utterance and response sentences at the current time N are first trained in the encoder-decoder at the N time point. Then, when the utterance and response sentences at the future time point N+1 are input, the utterance and response sentences at the current time point N are shifted as the utterance and response sentences at the past time point N-1. At this time, they are input to the encoder-decoder at the past time point N-1. Also, the latent variable z constructed by the encoder-decoder at the past time point N-1 is input to the encoder at the current time point N. That is, the encoder at the current time point N is trained to input the utterance sentence at the current time point N and the latent variable z at the past time point N-1, and to output the response sentence at the current time point N from the decoder at the current time point N. Similarly, the encoder for the past time point N-1 is trained to input the utterance sentence for the past time point N-1 and the latent variable z for the past time point N-2, and to output the response sentence for the past time point N-1 from the decoder for the past time point N-1.
[0032] [Dialogue relationship determination unit 12] In the estimation stage, the dialogue relation determination unit 12 determines whether or not the dialogue relation between the "utterance sentence" of the estimation data and the "response sentence" output from the recurrent neural network 11 satisfies a predetermined condition. Here, the predetermined condition may be determined to be true when any one of the following four conditions is met. <Condition 1: When the dialogue breaks down due to the relationship between the spoken sentence and the response sentence> <Condition 2: When a certain number of consecutive interjections are used between an utterance and a response> <Condition 3: When the utterance is judged to be negative> <Condition 4: When the user's face image is judged to be negative>
[0033] [Dialogue history reset part 13] When the dialogue relation determination unit 12 determines that the result is true, the dialogue history reset unit 13 resets the latent variables of the recurrent neural network 11 based on the past dialogue history. After the latent variables based on the past dialogue history are reset, the recurrent neural network 11 outputs a response sentence estimated only from the current utterance sentence. This allows the response sentence to be generated only from new topic topics without being influenced by past topic topics.
[0034] <Condition 1: When the dialogue breaks down due to the relationship between the spoken sentence and the response sentence> FIG. 2 is a functional block diagram of a dialogue device further having a dialogue learning engine that outputs a score. 2, compared with FIG. 1, an interactive learning engine 14 is further included.
[0035] [Dialogue Learning Engine 14] The dialogue and learning engine 14 functions as follows for the training and estimation stages. (Training stage) A learning model is constructed by matching spoken sentences and correct response sentences as teaching data. (Estimation stage) Input the spoken sentences and response sentences that serve as estimation data, and output a validity score. The dialogue and learning engine 14 may be a machine learning engine such as, for example, Bidirectional Encoder Representations from Transformers (BERT).
[0036] In this case, the dialogue relation determination unit 12 determines that a dialogue breakdown is true when the score output from the dialogue learning engine is equal to or smaller than a predetermined threshold in the estimation stage. That is, the dialogue breakdown is classified according to the predetermined threshold. Then, the dialogue history reset unit 13 resets the latent variables of the recurrent neural network 11 based on the past dialogue history.
[0037] As a result, the above-mentioned case 1 is improved as follows. <Case Study 1> Sentence: I'm hungry Response: Do you want something to eat? Sentence: I want to eat gyoza. Response: Gyoza is delicious. Sentence: I suddenly got a stomach ache <-(Here, we generate the response "May I help you deliver the gyoza?" When the user selects the utterance "My stomach suddenly started hurting," the score drops. <- (Judge that the dialogue has broken down and reset the latent variables based on the past dialogue history) <- (Generate a response to only the spoken sentence "My stomach suddenly hurts") Response: Are you feeling okay? Are you going to buy some stomach medicine?
[0038] <Condition 2: When a certain number of consecutive interjections are used between an utterance and a response> FIG. 3 is a functional configuration diagram of a dialogue device further having a backchannel message registration unit. 3, compared to FIG. 2, a backchannel message registration unit 102, a vocabulary database 103, and a current topic word selection unit 15 are further included.
[0039] [Response registration section 102] The backchannel statement registration unit 102 registers backchannel statements in advance. Specifically, it registers colloquialisms that indicate agreement, such as "That's right," "That's right," "It's good," and "I like it."
[0040] In response to this, in the estimation stage, the dialogue relationship determination unit 12 determines that the response sentence output from the recurrent neural network 11 is true as a backchannel if the backchannel sentences registered in the backchannel sentence registration unit 102 are consecutive in chronological order a predetermined number of times or more. Then, the dialogue history reset unit 13 resets the latent variables of the recurrent neural network 11 based on the past dialogue history.
[0041] [Vocabulary Database 103] The vocabulary database 103 conceptually organizes vocabulary in a hierarchical structure.
[0042] FIG. 4 is an explanatory diagram showing a search of a vocabulary database. According to FIG. 4, a vocabulary database is a collection of a large number of natural language terms, hierarchically structured as topical topics. For example, the subordinate concepts of "artificial objects" correspond to "food," "medicine," and "clothing." In addition, "grilled meat," "boiled fish," and "dumplings" correspond to the subordinate concepts of "artificial objects" -> "food" -> "cooking" -> "other / cooking." Furthermore, for example, "Chinese medicine," "cold medicine," and "stomachache medicine" correspond to the subordinate concepts of "artificial object" -> "drug" -> "third type."
[0043] [Current Topic Word Selection Part 15] The current topic word selection unit 15 identifies a "previous topic word" based on the past dialogue history, and selects another "current topic word" that is in the same category as the previous topic word, using the vocabulary database 103. Even if the previous topic word is an unknown vocabulary that is not registered in the vocabulary database 103, it can be linked to the closest synonymous vocabulary as a language feature vector (e.g., Word2Vec). Then, the recurrent neural network 11 outputs a response sentence based on the current topic word after the latent variables based on the past dialogue history are reset.
[0044] The current topic word selection unit 15 may select the current topic word in various ways as follows. (1) Select a current topic word from the previous topic word in the same or different category. (2) Select the current topic word that is a higher concept (or lower concept) from the previous topic word. (3) Select an antonym of the previous topic word as the current topic word.
[0045] According to FIG. 4, it is assumed that the word "gyoza" existing in the vocabulary database 103 is identified as a previous topic word. As a result, the above-mentioned case 2 is improved as follows. <Case Study 2> Sentence: I'm hungry Response: Do you want something to eat? Sentence: I want to eat gyoza. Response: Gyoza is delicious. Sentence: Right Response: That's true <- (1st interjection) Sentence: Okay. <- (Here, if you generate a response sentence "Like", the backchannel sentences will be repeated.) <- (determines that a backchannel is being spoken and resets latent variables based on past dialogue history) <- (select the current topic word "yakiniku" from the previous topic word "gyoza" in the same category) <-(Generate a response to the current topic word "yakiniku") Response: By the way, would you like to know the latest information on Yakiniku? Here, the user is determined to be disinterested if they respond with two consecutive responses, but this can be set as appropriate.
[0046] <Condition 3: When the utterance is judged to be negative> FIG. 5 is a functional configuration diagram of a dialogue device further having an utterance sentence positive / negative learning engine. 5, compared to FIG. 1, a spoken sentence positive / negative learning engine 16 is further included.
[0047] [Speech Positive / Negative Learning Engine 16] The speech sentence positive / negative learning engine 16 functions as follows in a training phase and an estimation phase. (Training stage) A learning model is constructed by associating spoken sentences with positive or negative flags as teaching data. (Estimation stage) A spoken sentence that will be used as estimation data is input, and a positive or negative flag is output. Specifically, the spoken sentence positive / negative learning engine 16 may be a classifier such as a Support Vector Machine or the Cloud Natural Language API provided by Google (registered trademark) (see, for example, Non-Patent Document 3).
[0048] In the estimation stage, when the flag output from the positive / negative learning engine is determined to be negative, the dialogue relation determination unit 12 determines that the topic has changed to true. Then, the dialogue history reset unit 13 resets the latent variables of the recurrent neural network 11 based on the past dialogue history.
[0049] This results in the following improvements, for example: <Case Study 2> Sentence: I'm hungry Response: Do you want something to eat? Sentence: I want to eat gyoza. Response: Gyoza is delicious. Sentence: I can't make gyoza right away. <-(This is a negative change of topic, based on past conversation history. Reset the latent variables) <- (select the current topic word "yakiniku" from the previous topic word "gyoza" in the same category) <-(Generate a response to the current topic word "yakiniku") Response: By the way, would you like to know the latest information on Yakiniku?
[0050] <Condition 4: When the user's face image is judged to be negative> FIG. 6 is a functional configuration diagram of a dialogue device further having a face image positive / negative learning engine. 6, in comparison with FIG. 1, a face image positive / negative learning engine 17 is further included.
[0051] [Facial image positive / negative learning engine 17] The face image positive / negative learning engine 17 functions as follows in a training phase and an estimation phase. (Training stage) As teacher data, a learning model is constructed by associating footage of a user's facial expression while speaking with a positive or negative flag. (Estimation stage) An image of the facial expression of the user speaking is input, and a positive or negative flag is output. Existing technology that determines three emotional polarities (positive / negative / neutral) from a user's facial image can be used (see, for example, Non-Patent Document 4). A video is composed of multiple frames (still images), and the emotional polarity is determined from the user's facial expression shown in each frame, and the most frequently occurring emotional polarity is determined.
[0052] Specifically, time-series video features are extracted from the user's facial video. Video features include, for example, 68 facial feature points, AU (Action Unit) features, gaze, facial expression, etc. For example, AU features define multiple types of movements for the three areas of the face: lips, eyes, and eyebrows. This is the smallest unit of facial movement used in the facial expression description method known as FACS (Facial Action Coding System), and is composed of a combination of these AUs. The face image positive / negative learning engine 17 trains by associating these image features with positive or negative flags.
[0053] In the estimation stage, when the flag output from the positive / negative learning engine is determined to be negative, the dialogue relation determination unit 12 determines that the topic has changed to true. Then, the dialogue history reset unit 13 resets the latent variables of the recurrent neural network 11 based on the past dialogue history.
[0054] At this time, it is also preferable that the current topic word selection unit 15 uses the vocabulary database 103 to select a current topic word that is an antonym of the previous topic word or is in a different category. This results in the following improvements, for example: <Case Study 2> Sentence: I'm hungry Response: Do you want something to eat? Sentence: Shall we eat some dumplings from the fridge? Response: Gyoza is delicious. (Facial image: judged as negative, reset latent variables based on past dialogue history) <-(Select the current topic word "clothing" from the previous topic word "gyoza" which is in a different category) <-(Generate a response to the current topic word "clothing") Response: By the way, do you want to know more about the hottest jackets this winter?
[0055] As described above in detail, the program, device, and method of the present invention can generate a response sentence to a user's utterance that does not break down the dialogue and does not require continuous backchanneling. In other words, it realizes a casual conversation that combines a wide range of topics with in-depth dialogue.
[0056] Furthermore, this will enable, for example, "the system to communicate sufficiently with the user," making it possible to contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), which is to "build resilient infrastructure, promote sustainable industrialization and foster innovation."
[0057] Regarding the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical idea and perspective of the present invention can be easily made by those skilled in the art. The above description is merely an example and is not intended to be restrictive in any way. The present invention is limited only by the scope of the claims and their equivalents. [Explanation of symbols]
[0058] 1. Interactive device 101 Teacher Database 102 Response Registration Section 103 Vocabulary Database 104 Teacher Database 11 Recurrent Neural Networks 12 Dialogue Relationship Determination Unit 13 Dialogue history reset section 14. Dialogue Learning Engine 15 Current Topic Selection Section 16 Speech positive / negative learning engine 17 Facial image positive / negative learning engine
Claims
1. A program for causing a computer to function to generate a response sentence in response to an utterance sentence, In a training stage, utterances and responses in teacher data are input in a time series to construct a learning model based on latent variables, and in an estimation stage, a recurrent neural network outputs a response each time an utterance serving as estimation data is input in a time series; a dialogue learning engine that constructs a learning model by associating utterances and valid responses as teacher data in a training stage, and inputs utterances as estimation data and responses output from a recurrent neural network in an estimation stage and outputs a validity score; a dialogue relation determination means for determining, as an estimation step, that a dialogue breakdown is true when a score output from the dialogue learning engine is equal to or smaller than a predetermined threshold; a dialogue history resetting means for resetting a latent variable based on a past dialogue history for the recurrent neural network when the dialogue relationship determining means determines that the answer is true; and make it work. The recurrent neural network outputs a response sentence estimated only from the utterances after the latent variables based on the past dialogue history are reset. A program that causes a computer to function in such a manner.
2. A backchannel message registering means for registering backchannel messages in advance is provided; In the estimation step, the dialogue relation determination means determines that the response sentence output from the recurrent neural network is true as a backchannel succession when the backchannel sentence registered in the backchannel sentence registration means continues in time series a predetermined number of times or more.
2. The program according to claim 1, which causes a computer to function as described above.
3. A vocabulary database in which vocabulary is conceptually organized hierarchically; a current topic word selection means for identifying a previous topic word based on a past dialogue history and selecting another current topic word that belongs to the same category as the previous topic word; and make it work. The recurrent neural network resets latent variables based on the past dialogue history, and then outputs a response based on the current topic.
3. The program according to claim 2, further causing the computer to function as follows.
4. In the training stage, a learning model is constructed by associating utterances with positive or negative flags as teacher data, and in the estimation stage, an utterance positive / negative learning engine is provided that inputs utterances as estimation data and outputs positive or negative flags. and make it work. The dialogue relation determination means determines that the topic change is true when the flag output from the speech sentence positive / negative learning engine is determined to be negative in the estimation stage.
2. The program according to claim 1, which causes a computer to function as described above.
5. In the training stage, a learning model is constructed by associating a video of the facial expression of the user speaking with a positive or negative flag as teacher data, and in the estimation stage, a facial image positive / negative learning engine is used that inputs a video of the facial expression of the user speaking and outputs a positive or negative flag. and make it work. The dialogue relation determination means determines the topic change as true when the flag output from the face image positive / negative learning engine is determined as negative in the estimation stage.
2. The program according to claim 1, which causes a computer to function as described above.
6. A vocabulary database in which vocabulary is conceptually organized hierarchically; a current topic word selection means for identifying a previous topic word based on a past dialogue history and selecting a current topic word that is an antonym or a different category of the previous topic word; and make it work. The recurrent neural network resets latent variables based on the past dialogue history, and then outputs a response based on the current topic.
3. The program according to claim 2, further causing the computer to function as follows.
7. A Recurrent Neural Network (RNN) is a Curriculum Learning Engine, a Long Short Term Memory (LSTM), or a Gated Recurrent Unit (GRU).
7. The program according to claim 1, which causes a computer to function as follows.
8. A dialogue device for generating a response sentence in response to an utterance sentence, In a training stage, utterances and responses in teacher data are input in a time series to construct a learning model based on latent variables, and in an estimation stage, a recurrent neural network outputs a response each time an utterance serving as estimation data is input in a time series; a dialogue learning engine that constructs a learning model by associating utterances and valid responses as teacher data in a training stage, and inputs utterances as estimation data and responses output from a recurrent neural network in an estimation stage and outputs a validity score; a dialogue relation determination means for determining, as an estimation step, that a dialogue breakdown is true when a score output from the dialogue learning engine is equal to or smaller than a predetermined threshold; a dialogue history resetting means for resetting a latent variable based on a past dialogue history for the recurrent neural network when the dialogue relationship determining means determines that the answer is true; having The recurrent neural network outputs a response sentence estimated only from the utterances after the latent variables based on the past dialogue history are reset.
13. An interactive device comprising:
9. A dialogue method for a device that generates a response sentence in response to an utterance sentence, comprising: The device is In a training stage, utterances and responses in teacher data are input in a time series to construct a learning model based on latent variables, and in an estimation stage, a recurrent neural network outputs a response each time an utterance serving as estimation data is input in a time series; In the training stage, a learning model is constructed by associating utterances and correct response sentences as teacher data, and in the estimation stage, a dialogue learning engine is inputted with utterances as estimation data and response sentences outputted from a recurrent neural network, and outputs a correct score. having A first step of estimating the dialogue breakdown as true when the score output from the dialogue learning engine is equal to or smaller than a predetermined threshold; A second step of resetting a latent variable based on a past dialogue history for the recurrent neural network when the first step determines that the answer is true; The recurrent neural network is the third step that outputs a response sentence estimated only from the utterances after the latent variables based on the past dialogue history are reset. The present invention is characterized by carrying out the following steps.
Citation Information
Patent Citations
Dialogue robot and robot control program
JP2018173456A
Compression of recurrent neural network models
JP2019512760A
Dialog execution program, device and method for controlling response utterance generation by response word determination
JP2020126186A
Voice guidance system and voice guidance method
JP2020149264A
