JP6769405B2 - Dialogue system and dialogue method - Google Patents

Description

The present invention relates to a dialogue system, and more particularly to a question answering system that generates an answer to a question from a user.

In a question answering system, it is desired to generate an appropriate answer to a question written by a user in a natural sentence.

In Patent Document 1, a discriminator that has been machine-learned to discriminate between a result of analyzing a natural sentence input by a user and a function required in each situation is used. Then, by comparing these, the functions required by the user from the system are extracted.

In Patent Document 2, the degree of ambiguity of a word or phrase in a user-input question sentence is obtained, and a question sentence for specifying the meaning of the word or phrase with a high degree of ambiguity is generated. This is repeated to generate the appropriate answer.

Japanese Unexamined Patent Publication No. 2016-151928 Japanese Unexamined Patent Publication No. 2004-110524

However, in Patent Document 1, when the user's input is ambiguous, an appropriate answer cannot be generated. Further, in Patent Document 2, since a narrowing-down dialogue occurs even when a self-evident matter is omitted, a redundant dialogue between the user and the system occurs.

An object of the present invention is to identify a question and generate an appropriate answer even when the question in natural sentences from the user is ambiguous in the dialogue system.

The first aspect of the present invention is
How to get the text of the user's question,
A storage means for associating and holding a plurality of predetermined questions, one or more attribute items representing the attributes of the plurality of questions, and an attribute value representing the contents of the attribute items for each of the plurality of questions.
An extraction means for extracting the attribute item or the attribute value or a keyword related to these synonyms from the text.
Using the keyword and the storage means, a first intent understanding means for finding a candidate for a question that matches the text among the plurality of questions.
A second means of understanding the intention to obtain a score indicating the degree of agreement between each of the plurality of questions and the text for each of the plurality of questions.
A question specifying means for identifying a question from the processing result of the first intention understanding means or the second intention understanding means,
With
The question identification means
If the question is uniquely determined from the processing result of the second intention understanding means, the user's question is specified from the processing result of the second intention understanding means.
If the question is not uniquely determined from the processing result of the second intention understanding means, the user's question is specified from the processing results of the first intention understanding means and the second intention understanding means.
It is a dialogue system.

According to such a configuration, the question identifying means asks a question from a question candidate obtained by the first intention understanding means or a score obtained by the second intention understanding means even if the question sentence of the user is ambiguous. Can be identified. Therefore, an appropriate answer can be returned.

In the present invention, the storage means holds a plurality of predetermined questions. Further, the storage means holds "attribute items" representing the attributes (features) of the plurality of questions as items for identifying the plurality of questions. Further, the storage means holds an "attribute value" which is a value (content) actually taken by each attribute.

For example, in the case of a "question answering system for car troubles", it is preferable that the "attribute item" is an item that enables any question regarding car troubles to be specified. For example, the "attribute item" can be [<part>, <symptom>, <question format>] or the like. Further, in the above example, if a predetermined question "What should I do if the engine makes an abnormal noise?" Is provided, the "attribute value" corresponding to this question is ["engine", " It can be "abnormal noise", "how"], etc.

In the present invention, it is preferable that the question identifying means determines that a question is uniquely determined when there is one question in which the normalized value of the score is equal to or greater than a predetermined threshold value.

Further, in the present invention, in the question specifying means, the question using the second intention understanding means is specified by asking a question for which the normalized value of the score is equal to or more than a predetermined threshold value for all the plurality of questions. It is preferably done by identifying it as a user question. Further, in the question identifying means, the identification of the question using the first intention understanding means and the second intention understanding means is a value obtained by normalizing the score for the question candidates obtained by the first intention understanding means. It is preferable to ask a question in which is equal to or greater than a predetermined threshold value by identifying it as a user's question.

In this way, even if the question cannot be uniquely specified only by the processing result of the second intention understanding means, the question identifying means specifies the question by using the processing result of the first intention understanding means. be able to.

In the present invention, it is possible to further provide a question sentence generation means for generating a question sentence to the user for the attribute item for which the keyword is not extracted by the extraction means when the question is not uniquely determined by the question identification means. preferable.

As a result, it is possible to ask about attribute items and the like for which keywords are not extracted from the question first input by the user, so that even if the user's question is ambiguous, it can be dealt with. As a specific example, it is assumed that a question such as "something is wrong with the engine" is input by the user in the above-mentioned "question answering system for car trouble". In this case, since the keyword "engine" is extracted by the extraction means, it can be determined that the attribute item <part> is a question related to "engine". Here, since the attribute item <symptom> is unknown, it is possible to generate a question sentence asking "what kind of symptom is it?" Using the attribute item <symptom> for which the keyword has not been extracted.

In the present invention, the question sentence generating means generates a question sentence by using the attribute value of the attribute item in which a plurality of types of the attribute values exist among the attribute items whose keywords have not been extracted by the extraction means. ,be able to.

As a result, the question sentence generation means can generate a question sentence about the content of a specific attribute value, so that redundant dialogue with the user can be avoided. Specifically, as a result of narrowing down the attribute item <parts> to the "engine" in the above question, for example, the question sentence generation means has "A" for the attribute value of the attribute item <symptom> for which the keyword has not been extracted. Is it? Or is it B? ”Can be generated. Further, the question sentence generation means uses the attribute value of the attribute item having a plurality of types of attribute values. For example, when the attribute value of the attribute item <symptom> is only one type of "xxx", "xxx" is used. It is possible to avoid the generation of redundant question sentences such as "Is it?".

In the present invention, the question sentence generating means targets only the questions whose normalized value of the score is equal to or more than a predetermined value among the plurality of questions, and the attribute item having the maximum average amount of information (entropy). It is possible to generate a question sentence using the attribute value of.

In this way, the question sentence generation means can narrow down the questions that are expected to be closer to the user's question, and can determine which attribute item to ask based on the average amount of information. It is possible to minimize the number of dialogues required to identify.

Further, in the present invention, the question sentence generating means can also use the attribute value as a question sentence for a plurality of attribute items among the attribute items.

Further, in the present invention, the processing of the question sentence generating means, the first intention understanding means, the second intention understanding means, and the question specifying means can be repeated until the question is uniquely determined by the question specifying means. ..

Further, in the present invention, it is preferable that the second intention understanding means obtains the score by using a classifier generated by machine learning.

The present invention can also be regarded as a dialogue system including at least a part of the above means. The present invention can also be regarded as an interactive method for executing at least a part of the above processing. The present invention can also be regarded as a computer program for causing a computer to execute this method, or as a computer-readable storage medium in which the computer program is stored non-temporarily. Each of the above means and treatments can be combined with each other as much as possible to constitute the present invention.

For example, one aspect of the present invention is
Executed by a computer having a storage means for associating a plurality of predetermined questions, one or more attribute items representing the attributes of the plurality of questions, and an attribute value representing the contents of the attribute items for each of the plurality of questions. It ’s a dialogue method,
Get step to get the text of the user's question, and
An extraction step of extracting a keyword related to the attribute item or the attribute value or a synonym thereof from the text.
Using the keyword and the storage means, a first intent understanding step of finding a candidate for a question that matches the text among the plurality of questions.
A second intent understanding step of obtaining a score representing the degree of agreement between each of the plurality of questions and the text for each of the plurality of questions.
A question identification step that identifies a question from the processing result of the first intention understanding step or the second intention understanding step,
Including
In the question identification step
If the question is uniquely determined from the processing result of the second intention understanding step, the user's question is specified from the processing result of the second intention understanding step.
If the question is not uniquely determined from the processing result of the second intention understanding step, the user's question is specified from the processing results of the first intention understanding step and the second intention understanding step.
It is a dialogue method.

According to the present invention, in the dialogue system, even when a question in a natural sentence from a user is ambiguous, it is possible to identify the question and generate an appropriate answer.

FIG. 1 is a diagram showing a configuration of a voice dialogue system according to an embodiment. FIG. 2 is a diagram showing an example of the configuration of the storage unit according to the embodiment. FIG. 3 is a diagram showing an example of processing of the machine learning type classification unit according to the embodiment. FIG. 4 is a diagram showing an example of processing of the question specifying unit according to the embodiment. FIG. 5 is a flowchart showing an example of the flow of voice dialogue processing according to the embodiment. FIG. 6 is a flowchart showing an example of the flow of the question sentence generation process according to the embodiment. FIG. 7 is a diagram showing an example of a method of narrowing down the questions in the question sentence generation process according to the embodiment. FIG. 8 is a diagram showing an example of a method of selecting an attribute item in the question sentence generation process according to the embodiment.

Hereinafter, embodiments of the present invention will be described in detail exemplarily with reference to the drawings. The following description is an exemplary description of the present invention, and the present invention is not limited to the following embodiments.

<System configuration>
FIG. 1 is a diagram showing a configuration of a voice dialogue system according to the present embodiment. The voice dialogue system according to the present embodiment includes a voice dialogue device 100 and a smartphone 200.

The voice dialogue device 100 according to the present embodiment is an information processing device (computer) including an arithmetic device, a storage device, an input / output device, and the like. When the arithmetic unit executes the program stored in the storage device, the acquisition unit 110, the storage unit 120, the extraction unit 130, the rule-based classification unit 140, the machine learning type classification unit 150, and the question identification unit of the voice dialogue device 100 are executed. Functions such as unit 160, question sentence generation unit 170, and answer generation unit 180 are provided.

Further, the smartphone 200 according to the present embodiment includes an information processing device (computer) including a microphone, a speaker, and the like. The smartphone 200 acquires the user's utterance content (question), converts the utterance content into text, and then outputs the utterance content to the voice dialogue device 100. Further, the smartphone 200 acquires a question sentence and an answer sentence from the voice dialogue device 100 and outputs the question sentence and the answer sentence to the user by voice or the like.

In the present embodiment, the acquisition unit 110 acquires the text of the user's question text via the smartphone 200.

The storage unit 120 stores a plurality of predetermined questions, one or more attribute items representing the attributes of the plurality of questions, and an attribute value which is a value (content) actually taken by the attribute item for each of the plurality of questions. Hold.

FIG. 2 shows an example of the storage unit 120 in the “question answering system for car trouble” according to the present embodiment. In the example of FIG. 2, a plurality of questions (ID) 121, attribute item 122 (parts), attribute item 123 (symptom), and attribute item 124 (question format) are stored in the storage unit 120. Further, in the example of FIG. 2, the storage unit 120 corresponds to the attribute values "engine" and "tire" corresponding to <parts>, the attribute values "xxx" and "yyy" corresponding to <symptom>, and <question format>. The attribute values "how" and "what" to be used are stored.

Although the example of FIG. 2 shows the case where there are three attribute items, the number of attribute items may be any number. Further, the content of the attribute item and the attribute value of the storage unit 120 is not particularly limited. Further, in FIG. 2, a table is used as the configuration of the storage unit 120, but the data structure is not particularly limited. The storage unit 120 may be configured so that a predetermined question can be specified.

In addition, there may be a plurality of questions in which the contents of all the attribute values of the storage unit 120 match. For example, even if there are two or more questions whose attribute values of the attribute items [<parts>, <symptoms>, <question format>] in FIG. 2 are ["engine", "abnormal noise", "how"]. Good. Even in such a configuration, as will be described later, the storage unit 120 can be used to obtain question candidates. Further, the storage unit 120 may be created manually, or may be created by using machine learning or the like. Further, the contents of the storage unit 120 may be fixed, or may be added, updated, deleted, or the like.

The extraction unit 130 extracts an attribute item or attribute value of the storage unit 120 or a keyword related to these synonyms from the input text. Here, when the user asks the following ambiguous question, the extraction unit 130 extracts the keyword "engine" related to the attribute value in the question text by referring to the storage unit 120 shown in FIG.
"Samething strange about the engine"

The extraction method is not particularly limited. For example, the extraction unit 130 may extract all words and the like included in the text by using morphological analysis and the like.

The rule-based classification unit 140 (first intention understanding means) seeks a question candidate that matches the user's question from the extraction results of the storage unit 120 and the extraction unit 130. First, the rule-based classification unit 140 refers to the keywords extracted by the extraction unit 130 and the attribute values of the storage unit 120. Then, the rule-based classification unit 140 outputs a question having an attribute value related to the keyword as a candidate for the user's question. Here, the attribute item from which the keyword is extracted is referred to as the attribute item group A, and the attribute item from which the keyword is not extracted is referred to as the attribute item group B.

For example, when the user asks an ambiguous question as described above, first, the rule-based classification unit 140 refers to the keyword "engine" of the question text extracted by the extraction unit 130 and the attribute value of the storage unit 120. Then, the rule-based classification unit 140 asks for a question having an attribute value related to the keyword "engine" as a candidate. When the storage unit 120 has the configuration shown in FIG. 2, questions 1 to 5 in which "engine" is set in the attribute value of <part> are output as question candidates. In this case, the attribute item <part> is included in the attribute item group A, and the attribute items <symptom> and <question format> are included in the attribute item group B.

The machine learning type classification unit 150 (second intention understanding means) obtains a degree of agreement (score) between each of the predetermined questions and the text output from the acquisition unit 110. In the present embodiment, the machine learning type classification unit 150 calculates the score using a classifier generated by using machine learning. As for the method of creating a classifier using machine learning, an existing method is used. An example of the score obtained by the classifier is shown in FIG. 3 (A).

The question identification unit 160 identifies a user's question by using the processing result of the rule-based classification unit 140 or the machine learning type classification unit 150. In the present embodiment, when a question is uniquely determined from the processing result of the machine learning type classification unit 150, the question identification unit 160 identifies the question as a user's question. When the user's question is not uniquely determined from the processing result of the machine learning type classification unit 150, the question identification unit 160 makes a determination using the rule-based classification unit 140 and the machine learning type classification unit 150. The details of the question specifying unit 160 will be described in the processing explanation using the flowchart of FIG.

The question sentence generation unit 170 generates a question sentence to the user when the question is not uniquely determined in the question identification unit 160. Then, the generated question sentence is output to the user via the smartphone 200. In the present embodiment, the question sentence generation unit 170 generates a question sentence by using the attribute value of the attribute item group B for the question candidates obtained from the calculation result of the rule-based classification unit 140. The details of the question sentence generation unit 170 will be described in the processing description using the flowcharts of FIGS. 5 and 6.

The answer generation unit 180 generates an answer to the question specified by the question identification unit 160. Then, the generated answer is output to the user via the smartphone 200. In the present embodiment, the answer generation unit 180 generates an answer sentence with reference to a second storage means (not shown) in which the question and the answer are associated with each other on a one-to-one basis.

<Processing content>
Similar to the above, the processing in the voice dialogue system according to the present embodiment will be described with reference to FIG. 5 with reference to an example in which a question such as "something is wrong with the engine" is input from the user.

In step S102, the acquisition unit 110 acquires the text of the user's question via the smartphone 200.

In step S104, the machine learning type classification unit 150 calculates the degree of coincidence (score) between each of the plurality of predetermined questions and the acquired text.

In step S106, the question identification unit 160 determines whether or not the question is uniquely determined by using the processing result of step S104. Specifically, the question identification unit 160 normalizes the scores calculated for all of the plurality of questions in step S104 so that the total value is 1, and the normalized value (confidence) is a predetermined threshold value. Determine whether the above questions are uniquely determined. Here, the predetermined threshold value is a value of certainty that can be regarded as a question intended by the user if it is higher than that. FIG. 3 (B) is an example of the certainty obtained by normalizing the scores for all the questions of FIG. 3 (A). When the question is uniquely determined, the question specifying unit 160 identifies the question as a user's question and proceeds to step S118. If the question is not uniquely determined, the process proceeds to step S110.

In step S108, when it is determined that the question is not uniquely determined as a result of the processing in step S106 (S106-NO), the extraction unit 130 extracts the attribute item or attribute value of the storage unit 120 from the text acquired in step S102. Or extract keywords related to these synonyms.

In step S110, the rule-based classification unit 140 calculates the candidate of the user's question by referring to the extracted keyword and the storage unit 120.

In step S112, the question identification unit 160 determines whether or not the question is uniquely determined from the processing result of step S110. Specifically, the score is normalized again only for the questions calculated as candidates. Then, the question specifying unit 160 determines whether or not a question whose obtained value (certainty) is equal to or higher than a predetermined threshold value is uniquely determined. In this embodiment, the same threshold value is used in step S106 and step S112, but different values may be used. When the question is uniquely determined, the question identification unit 160 identifies the question as a user's question and proceeds to step S118. If the question is not uniquely determined, the process proceeds to step S114. FIG. 4 (A) shows an example in which the scores are normalized for all the predetermined questions. FIG. 4B shows an example in which the score is normalized only for the question candidates calculated by the rule-based classification unit 140.

In step S114, when it is determined in step S112 that the question is not uniquely determined (S112-NO), the question sentence generation unit 170 generates a question sentence to the user. In the present embodiment, the question sentence generation unit 170 generates a question sentence using the attribute values of the attribute items included in the attribute item group B. Then, the generated question sentence is output to the user via the smartphone 200. The details of step S114 will be described in the processing description using the flowchart of FIG. 6 described later.

If a response from the user is detected after the system asks the user in step S116, the rule-based classification unit 140 calculates the question candidates again and proceeds to step S112. Then, the processes of steps S112 to 116 are repeated until the question is uniquely determined.

In step S118, when the question is uniquely determined in the above step (S106-YES, S112-YES), the answer generation unit 180 gives an answer suitable for the question. The answer is output to the user via the smartphone 200.

(Question sentence generation process)
Details of the process in step S114 will be described with reference to FIGS. 6 and 7.

<< Step S1141 >>
In step S1141, the question sentence generation unit 170 narrows down the questions. Specifically, the question sentence generation unit 170 selects a question whose certainty is equal to or higher than a predetermined value. Here, the predetermined value is a value of certainty that if it is less than that, it can be regarded as not a question intended by the user. In the present embodiment, the predetermined value uses a value less than the above-mentioned predetermined threshold value, but may be equal to or more than the predetermined threshold value. FIG. 7 is an example in which the question sentence generation unit 170 narrows down the questions with a predetermined value of 0.10. In the example of FIG. 7, the question sentence generation unit 170 narrows down the five questions calculated in step S110 to four questions 1 to 4.

<< Step S1142 >>
In step S1142, the question sentence generation unit 170 selects an attribute item to be used for the question sentence. First, the question sentence generation unit 170 pays attention to the attribute item group B in which the keyword is not extracted. Then, the question sentence generation unit 170 selects, among the attribute items included in the attribute item group B, an attribute item having a plurality of types of attribute values corresponding to the questions narrowed down by the process of step S1141. In the example of FIG. 7, the question sentence generation unit 170 first pays attention to the attribute items <symptom> and <question format> included in the attribute item group B. Then, an attribute item having a plurality of types of attribute values corresponding to questions 1 to 4 narrowed down for each attribute item is selected. Specifically, there are four types of attribute values corresponding to the attribute item <symptom>: "xxx", "yyy", "zzz" and "www", and the attribute value corresponding to the attribute item <question format> is "how". There is only one type of. Therefore, in the example of FIG. 7, the question sentence generation unit 170 selects the attribute item <symptom> in which a plurality of attribute values exist.

<< Step S1143 >>
In step S1143, it is determined whether or not there is one attribute item selected from the processing result of step S1142. When there is only one selected attribute item, the question sentence generation unit 170 sets the attribute item as the attribute item used in the question sentence, and proceeds to step S1145. If there are two or more selected attribute items, the process proceeds to step S1144.

<< Step S1144 >>
In step S1144, when the attribute item is not determined to be one in step S1143 (S1143-NO), the question sentence generation unit 170 determines the attribute item to be used for the question sentence from the plurality of attribute items selected in step S1142. ..

Here, in step S1145, which will be described later, the question sentence generation unit 170 generates a question sentence using two types of attribute values for the attribute item determined in step S1144. For example, for the attribute item <symptom>, "A or B?" (If there are three or more attribute values, "A or B? Or another symptom?") Generate a question sentence.

Therefore, as a method of determining the attribute item in step S1144, the question sentence generation unit 170 can determine the attribute item in consideration of, for example, the average amount of information (entropy) when asking the above question. Specifically, the question sentence generation unit 170 calculates the average amount of information using the certainty of each attribute value for each attribute item, and generates a question sentence using the attribute item having the maximum obtained value. Hereinafter, a method of determining an attribute item in consideration of the average amount of information will be described with reference to FIG.

Ｅは属性項目ごとの平均情報量、Ｐ_１からＰ_３は所定の属性値で集約した確信度の合計値である。 FIG. 8 is an example in which a plurality of attribute items are selected in step S1142. In such a case, the question sentence generation unit 170 obtains the average amount of information using the following equation 1 from the certainty that the questions narrowed down in step S1141 are normalized again.

E is the average amount of information for each attribute item, P ₃ from P ₁ is the total value of the confidence that aggregates a predetermined attribute value.

Here, the question sentence generation unit 170 selects two types of attribute values. For example, in the case of the attribute item <symptom> of the storage unit 320 shown in FIG. 8, the question sentence generation unit 170 selects “xxx” for the first attribute value and “yyy” for the second attribute value. In the present embodiment, when there are three or more types of attribute values, the top two types of attribute values are used in descending order of the score, as in the process of step S1145 described later, but the method of selecting the attribute values is particularly high. Not limited. For example, the question sentence generation unit 170 may randomly select two types of attribute values.

Then, let P _{1 be} the total value of the certainty levels aggregated by the _first attribute value, and P _{2 be} the total value of the certainty levels aggregated by the second attribute value. If there are three or more types of attribute values, the total value of the certainty of all the attribute values after the _third is P ₃ (P ₃ = 1-P _1- P ₂ ).

In the example of the storage unit 320 shown in FIG. 8, the average amount of information of the attribute item <symptom> is 1.57 (P ₁ = 0.31, P ₂ = 0.28, P ₃ = 0.41), and the attribute item < The average amount of information in the question format> is 0.99 (P ₁ = 0.45, P ₂ = 0.55, P ₃ = 0.00). Therefore, the question sentence generation unit 170 determines the attribute item used for question sentence generation as the attribute item <symptom> having the maximum average amount of information.

The method of selecting the attribute value used for calculating the average amount of information is not particularly limited. For example, the question sentence generation unit 170 may randomly select the attribute value. Further, the types of attribute values used for calculating the average information amount are not limited to two types. For example, the question sentence generation unit 170 may calculate the average information amount using one type or three or more types of attribute values. Good.

In addition to the above-mentioned method of considering the average amount of information, as the method of determining the attribute item in step S1144, for example, the question sentence generation unit 170 may randomly select the attribute item. Further, the priority may be set in advance for the attribute items, and the question sentence generation unit 170 may select the attribute items having high priority according to the priority. Further, the question sentence generation unit 170 may select, for example, the attribute item having the most types according to the type of the attribute value.

<< Step S1145 >>
In step S1145, as described above, a question sentence is generated using two types of attribute values of the attribute items determined in step S1143 (S1143-YES) or step S1144. For example, when the attribute values "xxx" and "yyy" of the attribute item <symptom> are used, the question sentence generation unit 170 generates the following question sentence.
"Is it xxx or yyy? Or is it some other symptom?"
Here, in the present embodiment, when the question sentence generation unit 170 has three or more types of attribute values, the top two types of attribute values are used in descending order of the score, but the method of selecting the attribute values is not particularly limited. .. For example, the question sentence generation unit 170 may randomly select two types of attribute values.

In this embodiment, an example in which two types of attribute values are used in the question sentence has been described, but the number is not particularly limited. For example, the question sentence generation unit 170 may generate a question sentence using all the attribute values of the corresponding attribute items.

<Advantageous effect of this embodiment>
According to the present embodiment, even when the user's question is ambiguous and it is necessary to answer the question, the number of dialogues required to identify the question can be minimized by asking the question in consideration of the amount of information obtained. ..

<Modification example>
The data acquired by the acquisition unit 110 is not limited to text. For example, the acquisition unit 110 may acquire the voice of the user's utterance or the like. Further, the acquisition unit 110 may acquire data by communication or the like.

Further, in the present embodiment, an example in which the voice dialogue device system uses two classification units (intention understanding units) has been described, but the configuration of each is not particularly limited. The first classification section may be configured to use a plurality of questions and a table or the like that holds their characteristics, etc., and the second classification section may have scores for each of the plurality of questions and the user's question ( Any configuration can be used as long as the degree of coincidence) can be obtained. For example, the voice dialogue device 100 may obtain the score by using a rule-based classifier different from that used by the rule-based classification unit 140 in calculating the score.

Further, the voice dialogue device 100 does not have to have a second storage means (not shown) used in the answer generation unit 180. For example, the answer may be provided in the storage unit 120. Further, the voice dialogue device 100 does not have to have the answer generation unit 180, and in this case, the question identification unit 160 may specify the question and generate the answer.

Further, in the present embodiment, in step S114, the question sentence generation unit 170 generates a question sentence to the user by using (the attribute value of) any one of the attribute items, but the attribute item used for the question sentence. Is not limited to one. For example, the question sentence generation unit may ask about <symptom> and at the same time about <question format>. By generating question sentences for a plurality of attribute items in this way, the number of dialogues required for the question can be further reduced.

Further, in the present embodiment, in step S114, the question sentence generation unit 170 has described an example of generating a question sentence using any of the attribute values included in the attribute item group B, but the method of generating the question sentence is this. Not limited to. Any question text that can identify the user's question will do. For example, in the above example, the question sentence generation unit 170 uses the "engine" which is a trivial content (attribute item group A) from the user's question to ask a question sentence such as "what kind of trouble is the engine?" It may be generated. Further, a question sentence that does not include a word or the like provided in the storage unit 120 may be generated. For example, in the above example, using a keyword other than the keyword "engine" in the question, "something is wrong" specifically. You may generate a question sentence such as "What kind of trouble is it?"

The input to the smartphone 200 is not limited to voice, and may be, for example, text input by the keyboard. Further, the output from the smartphone 200 to the user may be text or the like. Further, the smartphone 200 according to the present embodiment may play a role of converting voice into text, and may be replaced with, for example, a PC or the like. Further, the question answering system does not have to use the smartphone 200. For example, the voice dialogue device 100 may be provided with a microphone, a speaker, a display, or the like to directly acquire the user's question, or may directly ask and answer the user's question.

Further, in the present embodiment, an example using the conviction after normalization has been described, but the present invention is not limited to this. For example, the question may be specified using the score itself obtained by the machine learning type classification unit 150. In this case, the above-mentioned predetermined threshold value and the like may be provided for the score.

Further, in the present embodiment, the explanation has been given by taking the "question answering system for car trouble" as an example, but the subject of the question is not limited to this.

<Others>
The configurations of the above-described embodiments and modifications can be appropriately combined and used within a range that does not deviate from the technical idea of the present invention. Further, the present invention may be realized by appropriately modifying it without departing from the technical idea.

100: Voice dialogue device 110: Acquisition unit 120, 220, 320: Storage unit 130: Extraction unit 140: Rule-based classification unit 150: Machine learning type classification unit 160: Question identification unit 170: Question generation unit 180: Answer generation Department 200: Smartphone

Claims (11)

How to get the text of the user's question,
A storage means for associating and holding a plurality of predetermined questions, one or more attribute items representing the attributes of the plurality of questions, and an attribute value representing the contents of the attribute items for each of the plurality of questions.
An extraction means for extracting the attribute item or the attribute value or a keyword related to these synonyms from the text.
Using the keyword and the storage means, a first intent understanding means for finding a candidate for a question that matches the text among the plurality of questions.
A second means of understanding the intention to obtain a score indicating the degree of agreement between each of the plurality of questions and the text for each of the plurality of questions.
A question identifying means for identifying a question from the processing result of the first intention understanding means or the second intention understanding means,
With
The question identification means
If the question is uniquely determined from the processing result of the second intention understanding means, the user's question is specified from the processing result of the second intention understanding means.
If the question is not uniquely determined from the processing result of the second intention understanding means, the user's question is specified from the processing results of the first intention understanding means and the second intention understanding means.
Dialogue system. The question identifying means determines that a question is uniquely determined when there is one question in which the normalized value of the score is equal to or greater than a predetermined threshold value.
The dialogue system according to claim 1. In the question specifying means, the question using the second intention understanding means is specified.
A question in which the normalized value of the score is equal to or higher than a predetermined threshold for all of the plurality of questions is asked by identifying it as a user's question.
In the question identifying means, the identification of the question using the first intention understanding means and the second intention understanding means is performed.
A question for which the value obtained by normalizing the score is equal to or greater than a predetermined threshold value is specified as a user's question for the question candidates obtained by the first intention understanding means.
The dialogue system according to claim 1 or 2. A question sentence generation means for generating a question sentence to a user for the attribute item for which a keyword has not been extracted by the extraction means when the question is not uniquely determined by the question identification means.
The dialogue system according to any one of claims 1 to 3, further comprising. The question sentence generation means generates a question sentence by using the attribute value of the attribute item in which a plurality of types of the attribute values exist among the attribute items for which keywords have not been extracted by the extraction means.
The dialogue system according to claim 4. The question sentence generating means targets only the questions whose normalized value of the score is equal to or greater than a predetermined value among the plurality of questions, and the attribute value of the attribute item having the maximum average amount of information (entropy). Generate a question sentence using
The dialogue system according to claim 5. The question sentence generation means uses the attribute value as a question sentence for a plurality of attribute items among the attribute items.
The dialogue system according to claim 6. The processing of the question sentence generating means, the first intention understanding means, the second intention understanding means, and the question specifying means is repeated until the question is uniquely determined by the question specifying means.
The dialogue system according to any one of claims 4 to 7. The second intention understanding means obtains a score using a classifier generated by machine learning.
The dialogue system according to any one of claims 1 to 8. Executed by a computer having a storage means for associating a plurality of predetermined questions, one or more attribute items representing the attributes of the plurality of questions, and an attribute value representing the contents of the attribute items for each of the plurality of questions. It ’s a dialogue method,
Get step to get the text of the user's question, and
An extraction step of extracting a keyword related to the attribute item or the attribute value or a synonym thereof from the text.
Using the keyword and the storage means, a first intent understanding step of finding a candidate for a question that matches the text among the plurality of questions.
A second intent understanding step of obtaining a score representing the degree of agreement between each of the plurality of questions and the text for each of the plurality of questions.
A question identification step that identifies a question from the processing result of the first intention understanding step or the second intention understanding step,
Including
In the question identification step
If the question is uniquely determined from the processing result of the second intention understanding step, the user's question is specified from the processing result of the second intention understanding step.
If the question is not uniquely determined from the processing result of the second intention understanding step, the user's question is specified from the processing results of the first intention understanding step and the second intention understanding step.
How to interact. A program for causing a computer to perform each step of the method according to claim 10.

Images