JP2023074852A - Conversation database - Google Patents

Abstract

To allow a non-task type chatbot terminal to have a dialog with a user without failure.SOLUTION: A dialog method between a user H and a non-task type chatbot terminal 1 for having a dialog with a user, includes the steps of, by the terminal 1, (1) acquiring an utterance of the user, (2) estimating an utterance intention of the user to attach an utterance intention label to the utterance of the user, (3) when a kind of an intention of a response in the non-task type chatbot is indicated by a response label, referring to a conversation database storing a plurality of response patters formed in an arrangement of the response labels for each utterance intention label of the user and storing a selection probability of each response pattern, to select the response pattern corresponding to the utterance intention label attached in (2) on the basis of the selection probability of the response pattern, (4) selecting a response phrase corresponding to each response label forming the response pattern, and (5) creating a response sentence on the basis of the response phrase.SELECTED DRAWING: Figure 3

Description

The present invention relates to conversation databases and dialogue methods.

The promotion of communication through daily conversation plays an important role in building and maintaining interpersonal relationships. Particularly in the elderly, maintaining opportunities for social interaction with others through conversation is thought to affect various health promotions such as improving QOL and maintaining memory.

However, in recent years, due to the effects of the declining birthrate and aging population, as well as the trend toward nuclear families, opportunities for elderly people to communicate with their families are decreasing. In addition, with the shrinking of local community activities, it is becoming less common for the elderly to have horizontal connections such as "neighborhood associations." Therefore, even if the elderly have a desire to "talk", there is a problem that they cannot have the opportunity to communicate as much as they thought. This applies not only to elderly people living alone, but also to nursing care facilities and housing for the elderly where elderly people live.

In response to the problem that the elderly have few opportunities to communicate, attempts are being made to substitute systems such as robots and dialogue agents equipped with artificial intelligence as conversation partners for the elderly.

Systems equipped with artificial intelligence can be broadly classified into two types. One is a system such as a smart speaker that performs utterances according to tasks such as "Tell me about today's weather" and "Tell me about XX". This system is called a "task chatbot." The other is a system that allows conversation as an activity to spend time with the other party without a clear problem such as daily conversation or casual chat. The system is called a "non-task chatbot."

Non-task chatbots have a much higher degree of freedom in choosing words than task-type chatbots, so there is a high possibility of dialogue failure where the content of the conversation between the user and the non-task-type chatbot terminal does not mesh. . Therefore, in recent years, a dialogue system has been proposed that uses deep learning technology to learn a huge amount of text data, thereby making it difficult for the dialogue to fail. For example, especially in 2020, Facebook announced BlenderBot based on Transformer (Non-Patent Document 1). However, since BlenderBot is a dialogue model based on deep learning, it sometimes requires a huge amount of text data, and the learning data depends on data collected through the Internet, etc., rather than actual conversations between humans. Resulting in. Therefore, it is difficult to realize "human-like conversations" that humans actually have by using a dialogue model based on deep learning alone.

On the other hand, a dialogue system has been proposed that uses a model generated based on an actual human dialogue corpus (conversation data between humans) to interact with humans in a human conversation style (Patent Document 1). This technology predicts words that match the question (words in collocation relationships) from the mapping relationship between the keywords in the question and the keywords in the answers generated based on the human dialogue corpus, and predicts the word matching with the grammatical structure. Generate an answer sentence by adjusting the order. However, in this method, the dialogue system cannot read the intention of human speech, and there is a possibility that the contents of the speech do not correspond to the intention of human speech. In addition, since sentences are generated by adjusting the order of words from the grammatical structure after extracting keywords, there is a possibility that sentences that do not hold in context may be generated.

Japanese Patent No. 6726800

Roller et al. Recipes for building an open-domain chatbot, arXiv:2004.13637(2020)

In contrast to the above-described prior art, an object of the present invention is to eliminate the breakdown of dialogue when a non-task type chatbot engages in a dialogue such as chatting with a user.

In order to eliminate the dialogue breakdown with the user in the non-task type chatbot terminal, the present inventors have found that instead of stochastically outputting the reply phrase of the chatbot to the user's utterance, it is necessary to select a reply phrase. Prior to this, it was found that it is effective to stochastically consider the pattern of the type of reply intention of the chatbot terminal, and the present invention was completed.

That is, the present invention is a conversation database used for generating a conversation between a user and a non-task type chatbot terminal,
An utterance intention label indicating the type of utterance intention of the user, an utterance intention storage unit storing example sentences corresponding to each utterance intention label, and a reply label indicating the type of reply intention by the non-task type chatbot terminal. In this case, a plurality of response patterns formed by one response label or a sequence of multiple response labels are stored for each user's utterance intention label, and the probability that each response pattern is selected in response to the user's utterance. A response pattern storage unit that stores
provides a conversation database with

The present invention also provides a non-task chatbot terminal interaction method for interacting with a user, comprising:
A non-task type chatbot terminal
(1) obtaining a user's utterance;
(2) a step of referring to the utterance intention storage unit of the conversation database according to claim 1, estimating the user's utterance intention, and attaching an utterance intention label to the user's utterance;
(3) By referring to the response pattern storage section of the conversation database described in claim 1, a response pattern corresponding to the utterance intention label attached in (2) is determined based on the probability of the response pattern stored in the response pattern storage section. the step of selecting by
(4) selecting a reply phrase corresponding to each reply label forming the reply pattern selected in step (3);
(5) To provide an interactive method for creating a reply sentence based on the reply phrase selected in step (4).

According to the database of the present invention, when a non-task type chatbot terminal responds, it is possible to stochastically select a response pattern formed of response labels representing the type of intention of the response, corresponding to the type of intention of the user's utterance. Therefore, it is possible to create a reply phrase considering the intention of the reply. Therefore, it is less likely that the chatbot terminal will not match the user's utterance intentions with the chatbot terminal's reply intentions, and the chatbot terminal can continue the non-task-type dialogue with the user.

FIG. 1 is a schematic diagram of a dialogue between a user and a non-task type chatbot terminal. FIG. 2 is a block diagram of the conversation database of the embodiment. FIG. 3 is a flow chart of an interaction method between a user and a non-task type chatbot terminal of the embodiment.

The present invention will be described in detail below with reference to the drawings.
(Conversation database configuration)
For example, as shown in FIG. 1, when an elderly user H and a non-task chatbot terminal 1 interact with the conversation database of the present invention, the non-task chatbot terminal 1 can generate a conversation. database that can be used. The conversation database of the present invention can be used not only for a non-task type chatbot terminal to have a dialogue with a user, but also for a task-type dialogue system, for example, a non-routine dialogue system in which a user interacts with a questionnaire survey device. It can also be used in interpersonal conversations with various automatic devices, such as when you make a response.

FIG. 2 is a block diagram of conversation database 10 of one embodiment of the present invention. This database 10 comprises a utterance intention storage section 11, a reply pattern storage section 12, and further comprises a reply phrase storage section 13 as necessary.

The utterance intention storage unit 11 stores an utterance intention label indicating the type of utterance intention of the user, and an example sentence corresponding to each utterance intention label. The user's utterance intention label simply expresses the user's utterance intention in short words. The user's utterance intention label may be encoded, for example, as shown in the same table. The types of user's utterance intention labels are not limited to the 10 types listed in Table 1, but by roughly classifying utterance intentions in daily conversations of the elderly into these 10 types, it is possible to make remarks that understand the other party's intentions. is.

Table 2 is an example sentence of a user's utterance corresponding to the user's utterance intention label shown in Table 1. A user's utterance intention label and its example sentence are associated with each other and stored in a database, so that the non-task type chatbot terminal 1 acquires the user's utterance and refers to the utterance intention storage unit 11 to refer to the user's utterance. Intention can be estimated.

On the other hand, the reply label is a label indicating the type of reply intention when the non-task chatbot terminal 1 replies to the user's statement. Table 3 is a list of types of reply labels stored in the reply pattern storage unit 12 of this embodiment. Although 12 types of response labels are listed in Table 3, the types of response labels in the present invention are not limited to these.

The response pattern storage unit 12 stores, in the responses from the non-task chatbot terminal 1, the rate at which a response pattern formed by one response label or a sequence of multiple response labels is selected for each intention of the user's utterance ( hereinafter also referred to as selection probability). That is, the response pattern storage unit 12 stores a plurality of response patterns for each user's utterance intention label, and also stores the probability that each response pattern is selected by the user's utterance intention label. For example, Table 4 shows the response patterns in the non-task chatbot terminal 1 when the user's utterance intention label is information notification, and the selection probabilities of the response patterns.

Although Table 4 lists response patterns with up to three response labels, the number of response labels forming a response pattern in the present invention is not limited to this. Although the number may be four or more, it is sufficient if the number of reply labels forming the reply pattern is three or less so that the non-task type chatbot terminal 1 can give a correct reply. That is, the number of response labels forming a response pattern can be four or more. We found that by not changing the theme of the dialogue, there are many cases where multiple reply labels are incorporated that express reactions that are not related to the breakdown of the conversation. Therefore, as a method for creating a reply message of the non-task chatbot terminal 1, it is preferable to set the number of response labels forming a reply pattern to 3 or less in order to increase the processing speed of creating a reply message. On the other hand, the number of reply labels that form the reply pattern should be plural, in order to keep the conversation between the user H and the non-task type chatbot terminal 1 from becoming a task type and to continue a daily conversation. is preferred. In the dialogue method of the present invention, the fact that the response pattern selected based on the selection probability is ultimately formed of one response label does not hinder daily conversation.

After the response pattern is selected according to the user's utterance intention label in the process of the non-task chatbot terminal 1 creating a response sentence, the response phrase storage unit 13 stores one or more responses forming the response pattern. The reply phrases are accumulated and stored so that the reply phrase corresponding to the label can be selected.

For example, Table 5 is an example of reply phrases stored in the reply phrase storage unit 13 corresponding to the reply label "reaction (Re)". In Table 5, [named expression] is a notation for quoting when the user used a specific expression such as "Kao Museum" in the previous statement.

As will be described later, the response phrase corresponding to the response pattern can be selected by, for example, a document generation method using Transformer-based deep learning technology, or by calculating the TF-IDF value and calculating the cosine similarity. Since it can also be formed by a method or the like, the reply phrase storage unit 13 is provided as necessary in the database of the present invention. However, for response labels such as "Reaction", "Consent/Understanding", "Agree", "Favorable response", "Follow", etc. It is preferable to store corresponding response phrases in the response phrase storage unit 13 . As a result, one randomly selected from the reply phrases stored in the reply phrase storage unit 13 can be used as the reply phrase corresponding to the reply label, making it possible to reply easily.

In addition, for more flexible response labels such as “my story and thoughts,” “information notification,” and “answer,” we created a data set of similar user utterances and responses from human conversation data. It is preferable to store them in the phrase storage unit 13 .

(How to create a conversation database)
The conversation database 10 is preferably created by acquiring actual conversation data between people. As a result, the human conversation structure, that is, the arrangement of the speaker's utterance intention and the corresponding listener's response intention occurring in an actual conversation between humans, more specifically, the user's utterance intention label and its utterance intention label It is possible to implement a sequence of response patterns formed by one or more response labels in a non-task type chatbot terminal corresponding to .

Acquisition of actual conversation data between humans assumes a dialogue between a non-task type chatbot terminal and a user, and a "listener" who plays the role of a non-task type chatbot terminal and a "speaker" who plays the role of the user. It can be done by having a daily conversation to say what you want to talk about for a certain period of time without setting a particular task and collecting conversation data. The number of "speakers" is preferably as large as possible in order to increase the variation of the content of the conversation. are “listeners”, and the “listeners” talk with 22 “speakers” about everyday topics such as hobbies and recent activities. We were able to obtain enough conversation data for the bot terminal to continue a conversation with an elderly user without failure. Conversation data for creating the conversation database of the present invention does not require a huge amount of text data that is required when constructing a dialogue system using deep learning technology.

Note that the age of the “speaker” when collecting the conversation data may be determined according to the age of the user who is expected to interact with the non-task chatbot terminal 1 . For example, by using children as "speakers" when collecting conversation data, it is possible to obtain conversation data that enables a non-task chatbot terminal to continue daily conversations with children without disruption.

The collected conversation data between humans is transcribed as text data in order to be stored in the conversation database 10 . It is not necessary to write back responses such as "yes", "yes", "yeah".

In order to store the user's utterance intention label and its example sentences in the utterance intention storage unit 11, the utterance intention label is determined for the utterance content of the speaker in the collected conversation data, and the utterance is set as the example sentence of the utterance intention label. 2, a set of user's utterance intention labels and example sentences is accumulated in the utterance intention storage unit 11. As shown in FIG. In this case, for example, when one statement intention is expressed in multiple sentences, such as "It's been a week. It's been a long time." When there is a nuance to the question as a whole, such as "I went to ,", a plurality of sentences may be combined as an example sentence of one utterance intention label.

On the other hand, in the response pattern storage unit 12, a plurality of response patterns of the non-task type chatbot terminal for each user's utterance intention label and the probability that each response pattern is selected as a response to the user's utterance are accumulated. A response label is determined for the listener's response in the collected conversation data, and the response is used as an example of the response label. If the reply consists of multiple sentences, the reply label and its sequence are determined for each sentence of the reply. In this way, the reply pattern storage unit 12 stores the reply patterns formed by the determined reply labels or the example sentences corresponding to the reply patterns.

In determining the response label, the response may be, for example, a "reaction" such as "Is that so?" , these are common in terms of reactions, but it is preferable to define separate response labels such as "reaction", "satisfaction/understanding", and "agree", respectively. As a result, it becomes possible to create various reply phrases, and as a result, it becomes possible to take detailed measures for the user.

In the listener's response to the actual conversation data, the rate at which the response label or the response pattern formed by the arrangement thereof is selected for each utterance intention label of the speaker is calculated as the selection probability, and the selection probability is also stored in the response pattern storage unit 12. do. In this way, the response pattern storage unit 12 accumulates the response pattern selection probability for each user's utterance intention, as shown in Table 4, for example.

If the theme of the conversation is more in-depth than everyday conversation, such as "arguing about a certain event," the user's utterance intention labels of 10 types shown in Table 1 are not enough, and "affirming the other party's utterance" is not enough. ” and “denial of the other party's utterance” can be added to the utterance intention labels shown in Table 1. On the other hand, depending on the theme of the conversation, there is a possibility that "self-deprecating remarks" is not included in the remark intention labels shown in Table 1. Therefore, the user's utterance intention label can be appropriately changed according to the collected conversation data. Similarly, the type of reply label can be changed as appropriate according to the collected conversation data.

In addition, when the speaker's response label is "Answer" such as "My hobby is reading", the listeners are more likely to ask a "question" next, and there is a tendency for the listeners to dig deeper into the speaker's answers. . On the other hand, when the speaker's utterance is an "answer including the nuance of no" such as "I don't really like soccer", the listener tends to select a "response" such as "That's right" once. A more human-like conversation structure can be implemented by subdividing the utterance intentions of the speaker and the listener. On the other hand, for example, if it is necessary to reduce the speaker's response label for the purpose of speeding up the response speed of the chatbot terminal by speeding up the calculation processing of the system, the closely related "answer" and " Two utterance intentions of "an answer including the nuance of no" may be integrated into one utterance intention of "answer".

(Method of Creating Response Phrase Storage Section)
Based on the human-to-human conversation data described above, reply phrases are collected for each listener's reply label.

If the reply label is "information notice", a data set of the speaker's utterances and the listener's reply phrases is accumulated, as shown in Table 6, for example. Similarly, data sets are created for reply labels such as "my thoughts" and "answer", and stored in the reply phrase storage unit.

When the listener's response label is "question", the speaker's utterance is morphologically analyzed, and by referring to the information, the 5W1H information lacking in the speaker's content is addressed with a response phrase. Or create a data phrase stock of questions about 5W1H information based on human conversation data. For example, as shown in Table 7, a reply phrase of "question" is created for the missing 5W1H information.

Conversation data between humans, even if the response labels are "Provide topic", "Greeting", "Farewell greeting", "Reaction", "Consent/Understanding", "Agree", "Favorable reaction", and "Follow" It is possible to respond by preparing phrase stock from. For example, Table 8 is an example of phrase stock when "topic offer" is selected as a reply label.

(Dialogue method)
The dialogue method of the present invention is performed using the conversation database described above. FIG. 3 is a flowchart of the interaction method of one embodiment of the present invention.

In this interactive method, as step (1), the non-task chatbot terminal 1 acquires the user's speech. For example, if the non-task chatbot terminal 1 has a voice input function, as shown in FIG. is recorded, the recorded data is converted into text by a speech recognition application, and acquired as text data. It is also possible to record an utterance using the voice input function of a smartphone or tablet terminal, convert the recorded data into text using a general voice recognition application, and acquire the text. In addition, the user's utterance may be obtained as text information using an interactive tool using the Internet such as LINE (registered trademark). may be obtained.

FIG. 3 shows that "Have you ever been to the Kao Museum?" is acquired as the user's statement.

Next, as step (2), the user's statement intention is estimated from the user's statement content acquired in (1) by referring to the statement intention storage unit of the conversation database of the present invention, and a statement intention label is attached to the user's statement. attached. Specifically, a data set of the user's utterance intention and the corresponding sentence stored in the utterance intention storage unit 11 is used to create a utterance intention estimation model, and the user's utterance intention is acquired from the model. The utterance intention estimation model can be created using a Transformer-based machine learning method such as BERT (Bidirectional Encoder Representations from Transformers). It can also be created using a cloud-based API service that extracts information from machine learning, such as Microsoft's LUIS (Language Understanding).

Microsoft's LUIS was used in this embodiment. In LUIS, all data sets of example sentences corresponding to user utterance intentions stored in the conversation database are input and learned in advance. This makes it possible to estimate and output the user's utterance intention by inputting an input sentence not included in the data set. For example, when the sentence "Have you ever been to the Kao Museum?" Using this utterance intention label "information notification", a reply sentence is created in a later step.

Next, as a step (2'), a named entity is extracted. The named entity is extracted by morphologically analyzing the user's utterance obtained in step (1). Here, the morphological analysis is an analysis method for dividing an input sentence into linguistically meaningful minimum units based on information such as parts of speech of words registered in advance based on a dictionary or the like. For example, when the user's input sentence is "Have you ever been to the Kao Museum?" It can be divided into morphemes such as /wa/ari/masu/ne/. From this information, the word "Kao Museum", which is a proper noun, is extracted. The unique expressions obtained here are used as necessary in the "step of obtaining a reply phrase for each reply label" described later. Unlike the dialogue method described in Patent Document 1, it is not necessary to treat named entities as keywords.

Next, as step (3), the reply pattern storage unit 12 of the conversation database of the present invention is referenced to obtain a reply label, and the reply pattern corresponding to the utterance intention label attached in (2) is stored in the reply pattern storage unit 12. based on the probabilities of the response patterns stored in . Specifically, the response pattern selection probability corresponding to the user's utterance intention label stored in the response pattern storage unit 12 of the conversation database 10 is referred to, and the response pattern is determined according to the selection probability. More specifically, when the user's input sentence is "Have you been to the Kao Museum?", the user's utterance intention label is output as "information notification" in step (2) above. Therefore, the response pattern is determined from the probability of selection of the response pattern to the “information notification” stored in the response pattern storage unit 12 .

To select according to the selection probability means to make a random selection each time so that the selection probability is reached when repeated many times. Specifically, the first row of Table 4 has a probability of 3.4%, and in this case, the program is programmed so that 34 selections are made when 1,000 repetitions are performed.

More specifically, in this embodiment, random numbers are obtained using a random number generation program that randomly generates numbers from 0 to 999 with equal probability. Each acquired number is associated with each row in advance according to the probability of Table 4. That is, 0 to 33 correspond to the first line, 34 to 30 correspond to the second line, and the next 40 correspond to the third line. For example, if a random number "80" is obtained by the random number generation program, the third row is selected. topic provision (Wa)”.

By doing so, it is possible to realize selection according to the selection probability, which converges to the target selection probability if it is repeated a sufficiently large number of times.

Various methods are known for the selection according to the probability. In addition to the methods described here, for example, hardware is also used to generate random numbers. Selection using so-called pseudo-random numbers is also effective in realizing natural conversation aimed at by the present invention.

Next, as step (4), a reply phrase corresponding to each reply label constituting the reply pattern selected in step (3) is selected.

As an example, in step (3), it is assumed that a response pattern made up of three response labels of "reaction (Re)", "information notification (In)", and "question (Q)" is selected (see Table 4). shaded area). In this case, first, a reply phrase corresponding to the reply label "reaction (Re)" is created.

When the conversation database 10 has a reply label and a reply phrase storage unit 13 that stores reply phrases corresponding to the reply labels, the reply The phrase storage unit 13 is referred to, and one is randomly selected from the response phrases (Table 5) to the “reaction (Re)” stored in the response phrase storage unit 13 . The response phrase corresponding to the response label "reaction (Re)" can be created by calculating the TF-IDF value, calculating the cosine similarity, or using the Transformer-based deep A method using learning technology may be used.

At this time, if a reply phrase containing [named word] is selected, the named word obtained from the user's input sentence is substituted. If the user's input sentence does not contain a named entity, the response phrase containing the named entity should not be selected. For example, if the user's input sentence is "Have you ever been to the Kao Museum?", the reply phrase "Are you at the Kao Museum?" . In addition, response labels other than "reaction", such as "satisfaction/understanding", "agree", "favorable response", and "follow", For the reply label such as "Greeting", similarly to "Reaction", it is possible to create a reply phrase by selecting one from the stock of reply phrases stored in the reply phrase storage unit 13 at random.

When "information notice" is selected as the reply label, morphological analysis is performed on all sentences with the reply label "information notice" stored in the reply phrase storage unit 13 of the conversation database 10, and for each morpheme Calculate the TF-IDF value. TF (Term Frequency) is an index representing how important a word is in a target sentence. It is calculated as the ratio of the number of occurrences of the word to the total number of words in the entire sentence. For example, words such as "to do" and "to be" appear frequently, so the TF indicates a high value. On the other hand, IDF (Inverse Document Frequency) is an index representing how characteristic a word is. For example, words such as "Kao Museum" that do not appear in the entire text show high prices in the IDF. TF-IDF is the value of the product of these, and can be said to be an index indicating the feature amount of each sentence.

Then, the cosine similarity is calculated based on the TF-IDF value of the user's input sentence and the TF-IDF value of the user's utterance in the dataset, and the response phrase that most closely corresponds to the user's utterance is selected. If the cosine similarity is the closest, it means that the utterance in the database closest to the user's input sentence can be referred to and the corresponding reply phrase can be selected.

For example, in response to the user's input sentence "Have you been to the Kao Museum?", if you want to refer to the reply phrase that responds to the response label "Information notification", you can enter "Have you been to the Kao Museum?" The user's utterance showing the highest value of cosine similarity for the TF-IDF vector of "I have been to the Kao Museum in the past" in Table 6. Therefore, "I went to Kao's factory tour." corresponding to the statement is selected as a reply phrase.

It is possible to prepare contextually consistent phrases by outputting reply phrases corresponding to utterances in the data set using the same method for reply labels such as "my talk and thoughts" and "answers."

In addition, the present technology does not refer to the data recorded in the response phrase storage unit 13 to select the corresponding sentence, but creates the response phrase by generating the sentence using the Transformer-based deep learning technology. is also possible. In the case of the method of referring to data, if there is no utterance similar to the user's input sentence in the data set, it is possible to refer to an irrelevant utterance, but by introducing Transformer-based machine learning , the user's utterances are used as input information, and response phrases corresponding to each response label can be learned to generate new sentences, making it possible to handle a wider range of input information.

As a method of creating a reply phrase when "question (Q)" is selected as the reply label, first, the sentence input by the user is morphologically analyzed. For each morpheme, it is checked whether the user has stated 5W1H information, one piece of 5W1H information that the user has not stated is specified, and this information is selected from the "question" phrase stock stored in the reply phrase storage section of the conversation database. Refer to phrases that ask about information. For example, if the user's input sentence is "Have you ever been to the Kao Museum?" ” can also be read. On the other hand, since information other than location (Where) is lacking, it can be seen that it is possible to ask about other 4W1H information. These pieces of information are stored in the reply phrase storage unit 13, and the question phrase is referred to from there. For example, when the reason (Why) is selected from the 4W1H information, the phrase "What was the reason?" prepared from conversation data between humans obtained in advance is selected.

Next, as step (5), a reply sentence is created based on the reply phrase for each reply label selected in step (4). Normally, a reply sentence can be created by connecting the reply phrases for each reply label selected in step (4) in the order of the reply labels forming the reply pattern. For example, in response to the user's input sentence, "Have you ever been to the Kao Museum?" is the response phrase "Is it Kao Museum?" If the response phrase is "What was the reason for this?", by connecting these three response phrases, the response sentence will be "Is it the Kao museum? I went to the Kao factory tour. Did something happen?" is output.

Thus, according to the dialogue method of the present invention, the dialogue between the non-task type chatbot terminal 1 and the user H can be continued without failure.

On the other hand, in the commercially available dialogue method that does not consider the reply intention label, for example, in the next dialogue between the chatbot terminal and the user, the utterance 2 of the chatbot terminal fails.

Chatbot terminal (utterance 1) "Does the guest have a favorite song? If so, please tell me."
Guest (utterance 1) "I like Mr. A."
Chatbot terminal (utterance 2) "It seems that you are interested. Female singers' songs are easy to evoke sympathy from women."

According to the dialogue method of the present invention, the guest's message "I like Mr. A." is given the utterance intention label "answer", and the reply patterns are the reply labels "reaction" and "my thoughts". Suppose a reply pattern formed by is selected. In that case, since the phrase corresponding to "reaction" is selected as the first response label, the phrase "You seem to like it" is not selected. According to the dialogue method of the present invention, a phrase such as "Mr. A, isn't it? Recently, female singers have been very active, isn't it?"

As with the input text, the response text may be returned to the user as a system voice by using speech synthesis technology, or may be displayed in a message application such as LINE (registered trademark). By using a personal computer, it may be displayed visually on a display.

1 non-task type chatbot terminal 10 conversation database 11 utterance intention storage unit 12 reply pattern storage unit 13 reply phrase storage unit H user

Claims (5)

A conversation database used for generating a conversation between a user and a non-task type chatbot terminal,
An utterance intention label indicating the type of utterance intention of the user, an utterance intention storage unit storing example sentences corresponding to each utterance intention label, and a reply label indicating the type of reply intention by the non-task type chatbot terminal. In this case, a plurality of response patterns formed by one response label or a sequence of multiple response labels are stored for each user's utterance intention label, and the probability that each response pattern is selected in response to the user's utterance. A response pattern storage unit that stores
Conversation database with . A reply phrase storage section storing reply labels and reply phrases corresponding to the reply labels is provided, and the reply phrase storage section stores reply phrases into which a named entity is inserted and reply phrases into which a named entity is not inserted. A conversation database according to claim 1. 3. The conversation database according to claim 1 or 2, wherein a reply pattern to one utterance intention label is formed by three or less reply labels. A method for interacting with a non-task chatbot terminal for interacting with a user, comprising:
A non-task type chatbot terminal
(1) obtaining a user's utterance;
(2) referencing the utterance intention storage section of the conversation database according to claim 1, estimating the user's utterance intention, and attaching an utterance intention label to the user's utterance;
(3) By referring to the response pattern storage section of the conversation database described in claim 1, a response pattern corresponding to the utterance intention label attached in (2) is determined based on the probability of the response pattern stored in the response pattern storage section. the step of selecting by
(4) selecting a reply phrase corresponding to each reply label forming the reply pattern selected in step (3);
(5) A dialogue method that performs a step of creating a reply sentence based on the reply phrase selected in step (4). 4. In the step of selecting the reply phrase of (4), the non-task type chatbot terminal selects the reply phrase by referring to a database storing reply labels and reply phrases corresponding to the reply labels. How to interact as described.

Images