JP6604542B2 - Dialogue method, dialogue program and dialogue system - Google Patents

Description

  The present disclosure relates to a dialog method in a dialog system that responds to a user's speech, a dialog program that responds to a user's speech, and a dialog system that responds to a user's speech.

  2. Description of the Related Art In recent years, a technique for efficiently exchanging a system and a user by creating a model that represents user preferences has been proposed.

  For example, the spoken dialogue system shown in Patent Document 1 uses a paraphrase used when a keyword to be recognized is included in a response sentence, a response type indicating the type of the response sentence, and a paraphrase word. A condition for selecting a response type is recorded, a paraphrase and a response sentence template for the recognized keyword are determined based on the condition for selecting a paraphrase and a response type, and the determined response sentence template A response sentence is generated by inserting a paraphrase word into.

  In addition, the conventional speech understanding system shown in Patent Document 2 receives an electronic program guide (EPG) information, processes the EPG information, receives a dictation request and a knowledge extraction unit that forms a program database. A speech recognizer that translates text into a series of text information consisting of multiple words, a natural language processor that accepts a series of text information and processes words to interpret the semantic content of dictation requests, and a sufficient number of keyword slots And a conversation controller that analyzes the task frame to determine whether or not is entered and queries the user for additional information to enter into an empty slot.

JP 2008-39928 A JP 2000-250575 A

  However, in the conventional speech understanding system, the value to be input to the slot of the task frame is directly asked to the user, the answer obtained from the user is confirmed again, and the value of the slot is determined. For this reason, the dialogue time between the system and the user becomes long, and the processing time of the system becomes long.

  The present disclosure has been made in order to solve the above-described problem, and can provide a dialogue method, a dialogue program, and a dialogue program capable of reducing the dialogue time between the dialogue system and the user and reducing the processing time of the dialogue system. The purpose is to provide a dialogue system.

  An interaction method according to an aspect of the present disclosure is an interaction method used in an interaction system that responds to a user's utterance, and is necessary for executing a task that generates a response sentence for the user's utterance A plurality of nodes are stored in association with each other, utterance information indicating the utterance content of the user is acquired, a first node corresponding to the utterance information is identified from the plurality of nodes, and the identified first One second node is selected based on the weight value associated with each of the plurality of second nodes from among the plurality of second nodes associated with the selected node, A response sentence corresponding to the first second node is generated.

  According to this configuration, a plurality of nodes necessary for executing a task for generating a response sentence in response to a user's utterance are stored in association with each other. Utterance information indicating the user's utterance content is acquired. A first node corresponding to the utterance information is identified from among the plurality of nodes. One second node is selected from a plurality of second nodes associated with the identified first node based on a weight value associated with each of the plurality of second nodes. The Then, a response sentence corresponding to the selected second node is generated.

  Therefore, it is not necessary to generate a question sentence for causing the user to select one second node from among the plurality of second nodes, and based on the weight value associated with each of the plurality of second nodes. Since the response sentence corresponding to the selected second node is generated, the dialogue time between the dialogue system and the user can be shortened, and the processing time of the dialogue system can be shortened.

  In the above interactive method, the weight value may represent a probability that the plurality of second nodes have been selected by the user in the past.

  According to this configuration, the weight value represents the probability that a plurality of second nodes have been selected by the user in the past, and thus the weight value can be easily calculated.

  In the above interactive method, the second node having the probability larger than a predetermined value may be selected from the plurality of second nodes.

  According to this configuration, since the second node having a probability greater than a predetermined value is selected from among the plurality of second nodes, it is possible to easily select one second node.

  In the above interactive method, if there is no second node having the probability greater than a predetermined value among the plurality of second nodes, the user is selected from the plurality of second nodes. You may generate the response sentence for making it.

  According to this configuration, when there is no second node having a probability greater than a predetermined value among the plurality of second nodes, a response statement for causing the user to select one of the plurality of second nodes is provided. Since it is generated, even if it is not possible to select one second node, the user can select it.

  Further, in the above interactive method, information indicating the user's answer to the response sentence is acquired, and the user's answer is an answer for selecting one second node among the plurality of second nodes. Depending on whether or not, the weight value may be updated.

  According to this configuration, information indicating the user's answer to the response sentence is acquired. The weight value is updated according to whether or not the user's answer is an answer for selecting one second node among the plurality of second nodes.

  Therefore, since the weight value is updated every time the user uses the system, it is possible to select one second node corresponding to the usage status of the user.

  Further, in the above interactive method, the second node of one of the plurality of second nodes associated with the first node of the first of the plurality of first nodes, and the first of the plurality of first nodes. The weight value is associated with a combination with each of a plurality of second nodes associated with other first nodes among the nodes, and the first second node is identified. And when the first second node is specified, the first second node and each of the plurality of second nodes associated with the other first node One second node may be selected from the plurality of second nodes associated with the other first node based on a weight value associated with the combination of the first and second nodes.

  According to this configuration, one second node of the plurality of second nodes associated with the first node of the plurality of first nodes, and the plurality of first nodes A weight value is associated with each combination with a plurality of second nodes associated with the other first nodes. It is determined whether one second node has been identified. When one second node is specified, a weight associated with a combination of one second node and each of a plurality of second nodes associated with the other first node Based on the value, one second node is selected from among a plurality of second nodes associated with other first nodes.

  Therefore, one second node of the plurality of second nodes associated with one first node of the plurality of first nodes and the other of the plurality of first nodes One second node can be selected according to a combination with each of a plurality of second nodes associated with the first node.

  A dialogue program according to another aspect of the present disclosure is a dialogue program that responds to a user's utterance, and includes a plurality of nodes necessary for executing a task of generating a response sentence with respect to the user's utterance. A storage unit that stores each of them in association with each other; an acquisition unit that acquires utterance information indicating the utterance content of the user; a specifying unit that specifies a first node corresponding to the utterance information from the plurality of nodes; Based on a weight value associated with each of the plurality of second nodes, a second one of the plurality of second nodes associated with the first node identified by the identifying unit The computer is caused to function as a selection unit that selects a node and a generation unit that generates a response sentence corresponding to the first second node selected by the selection unit.

  According to this configuration, a plurality of nodes necessary for executing a task for generating a response sentence in response to a user's utterance are stored in association with each other. Utterance information indicating the user's utterance content is acquired. A first node corresponding to the utterance information is identified from among the plurality of nodes. One second node is selected from a plurality of second nodes associated with the identified first node based on a weight value associated with each of the plurality of second nodes. The Then, a response sentence corresponding to the selected second node is generated.

  Therefore, it is not necessary to generate a question sentence for causing the user to select one second node from among the plurality of second nodes, and based on the weight value associated with each of the plurality of second nodes. Since the response sentence corresponding to the selected second node is generated, the dialogue time between the dialogue system and the user can be shortened, and the processing time of the dialogue system can be shortened.

  A dialog system according to another aspect of the present disclosure is a dialog system that responds to a user's utterance, and includes a plurality of nodes necessary for executing a task of generating a response sentence with respect to the user's utterance. A storage unit that stores each of them in association with each other; an acquisition unit that acquires utterance information indicating the utterance content of the user; a specifying unit that specifies a first node corresponding to the utterance information from the plurality of nodes; Based on a weight value associated with each of the plurality of second nodes, a second one of the plurality of second nodes associated with the first node identified by the identifying unit And a generation unit that generates a response sentence corresponding to the first second node selected by the selection unit.

  According to this configuration, a plurality of nodes necessary for executing a task for generating a response sentence in response to a user's utterance are stored in association with each other. Utterance information indicating the user's utterance content is acquired. A first node corresponding to the utterance information is identified from among the plurality of nodes. One second node is selected from a plurality of second nodes associated with the identified first node based on a weight value associated with each of the plurality of second nodes. The Then, a response sentence corresponding to the selected second node is generated.

  Therefore, it is not necessary to generate a question sentence for causing the user to select one second node from among the plurality of second nodes, and based on the weight value associated with each of the plurality of second nodes. Since the response sentence corresponding to the selected second node is generated, the dialogue time between the dialogue system and the user can be shortened, and the processing time of the dialogue system can be shortened.

  According to the present disclosure, it is possible to shorten the dialogue time between the dialogue system and the user and reduce the processing time of the dialogue system.

It is a figure for demonstrating the outline | summary of the speech dialogue system in this Embodiment. It is a figure which shows the structure of the speech dialogue system in this Embodiment. It is a figure which shows an example of a judgment condition table. It is a flowchart for demonstrating the voice dialogue process of the voice dialogue system in this Embodiment. It is a flowchart for demonstrating the weight value update process of the speech dialogue system in this Embodiment. It is a figure for demonstrating the difference between the voice dialogue process of the voice dialogue system in this Embodiment, and the voice dialogue process of the conventional voice dialogue system. It is a figure which shows an example of the semantic network of the voice dialogue system in the modification of this Embodiment. It is a flowchart for demonstrating the voice dialogue process of the voice dialogue system in the modification of this Embodiment. It is a figure showing an example of the dialogue sentence by the dialogue method used in the speech dialogue system of a prior art example. It is a figure showing an example of the dialogue sentence by the dialogue method used in the voice dialogue system concerning this indication.

  Embodiments of the present invention will be described below with reference to the accompanying drawings. The following embodiments are examples embodying the present invention, and do not limit the technical scope of the present invention.

  FIG. 1 is a diagram for explaining the outline of the voice interaction system according to the present embodiment.

  The example shown in FIG. 1 shows an example of a semantic network used when selling drinks. The semantic network shown in FIG. 1 includes a plurality of nodes necessary for executing a task for generating a response sentence in response to a user's utterance. The plurality of nodes are associated with each other. Relation information indicating the relation between the two nodes is given to the two nodes associated with each other. The relationship information is information indicating that one node is in a lower-level concept with respect to the other node, and one node is an indispensable item for executing a task related to the concept included in the other node. Information indicating that one node is an item arbitrarily set for a task related to a concept included in the other node, and information indicating that one node is the value of the other node including.

  For example, the node 11 indicating “offee” and the node 12 indicating “drink” are subordinate to the node 12 indicating “drink” (the is-a relationship). ) Are associated with each other. Note that the node 11 indicating “coffee” is also called a domain.

  In addition, the node 14 indicating “size” and the node 12 indicating “link” include relationship information indicating that the node 14 indicating “size” is an essential item for the node 12 indicating “link”. It is associated. The node 15 indicating “quantity” and the node 12 indicating “link” are items in which the node 15 indicating “quantity” is arbitrarily set with respect to the node 12 indicating “link”, and is notified by the user. Relation information indicating that the node is determined in each case is associated.

  The node 17 indicating “sugar” and the node 11 indicating “offee” are indispensable items for the node 17 indicating “sugar” to execute a task on the node 11 indicating “offee”. Is associated with the relationship information. The node 18 indicating “temperature” and the node 11 indicating “offset” indicate that the node 18 indicating “temperature” is an indispensable item for executing a task on the node 11 indicating “offset”. The relationship information shown is associated. The node 17 indicating “sugar” and the node 18 indicating “temperature” are called essential slots or simply slots.

  Relation information indicating that the node 19 indicating “hot” is the value of the node 18 indicating “temperature” is associated with the node 19 indicating “hot” and the node 18 indicating “temperature”. The node 20 indicating “cold” and the node 18 indicating “temperature” are associated with the relationship information indicating that the node 20 indicating “cold” is the value of the node 18 indicating “temperature”. The node 19 indicating “hot” and the node 20 indicating “cold” are nodes selected for the node 18 indicating “temperature”. The node 19 indicating “hot” and the node 20 indicating “cold” are called slot values.

  When a user purchases a drink and chooses coffee, it must be determined whether sugar is required or whether it is hot or cold, and if these are not determined, the drink is served The task of generating a response sentence cannot be executed in response to the user's utterance. That is, temperature is an essential node (slot) to accomplish the task, and the system needs to determine the value of the slot (in this case, cold or hot).

  In the conventional system, the system asks the user whether the coffee temperature is hot or cold, and the user's answer is determined by voice recognition. When the user selects hot, the system asks again whether it is hot or not, determines the user's answer by voice recognition, and determines the temperature of the coffee.

  On the other hand, in the system of the present disclosure, regarding the temperature of the coffee, whether the user has selected hot or cold in the past without asking the user whether the system should be hot or cold. Depending on the weight value, a weight value is assigned to each, and whether to make it hot or cold is inquired according to the weight value. For example, if cold has been selected with a probability of 60% in the past and hot has been selected with a probability of 40%, the system will ask the user whether or not to make it cold, and the user's answer will be voice recognition. Judge and determine the temperature of the coffee. In this case, it is not necessary to reconfirm the temperature as compared with the conventional system, the interaction time between the system and the user can be shortened, and the processing time of the system can be shortened.

  If a question with specific and correct contents can be generated, it becomes easy to acquire information necessary for executing a task from the user. For example, when a user orders coffee, the system limits the user's answer to positive or negative expressions such as “yes” or “no” by asking “Are you sure you want hot coffee?” can do.

  As another case, for example, consider a case where the user places an order for a set menu that specifies “cheese burger set”. At this time, when the system asks the user about the type of drink in the set menu, the user asks an alternative question such as “Which is cork or orange juice?” Based on the probability. This makes it easy to guide the user's answer to an answer that includes content that can be accepted by the system, such as “It is cork” or “It is orange juice.” That is, when the system asks an alternative question, the user is encouraged not to make an unexpected answer. As a result, the possibility that the user uses expressions that can be accepted by the system is higher than when the conventional technique is applied, and information can be more reliably acquired from the user.

  FIG. 2 is a diagram showing a configuration of the voice interaction system according to the present embodiment. The speech dialogue system includes a speech recognition unit 101, a natural language processor 102, a memory 103, a conversation management unit 104, and a speech synthesis unit 105.

  The memory 103 includes a semantic network storage unit 111, a weight value management table storage unit 112, and a determination condition table storage unit 113.

  The semantic network storage unit 111 stores in advance a semantic network connecting a plurality of nodes. The semantic network storage unit 111 stores a plurality of nodes necessary for executing a task for generating a response sentence in response to a user's utterance.

  The weight value management table storage unit 112 stores the value of the slot included in the semantic network and the weight value in association with each other.

  The determination condition table storage unit 113 stores the number of selectable slot values, the condition for selecting the slot value, the slot value obtained when the condition is satisfied, and the template representing the response sentence in association with each other. Yes.

  The voice recognition unit 101 converts input voice acquired by a microphone (not shown) into text information. The voice recognition unit 101 recognizes a user's utterance and converts it into text information.

  The natural language processor 102 acquires speech information (text information) indicating the content of the user's speech. The natural language processor 102 specifies a first node (slot) corresponding to the utterance information from the plurality of nodes. The natural language processor 102 analyzes a series of text information output by the speech recognition unit 101 to understand the meaning content and the intention of the user's utterance. The natural language processor 102 understands the utterance content using language understanding knowledge stored in a language understanding database (not shown), for example. The natural language processor 102 extracts meaningful words from the text information. The natural language processor 102 searches the semantic network stored in the semantic network storage unit 111, and if the extracted word exists in the semantic network, the natural language processor 102 sets the slot and slot related to the task specified by the extracted word. A plurality of associated slot values are extracted from the semantic network.

  The natural language processor 102 includes a syntax analysis unit 131 and a memory access unit 132. The syntax analysis unit 131 performs a process of extracting words from the utterance contents of the user converted into text. The memory access unit 132 searches the semantic network stored in the memory 103 for the words extracted by the syntax analysis unit 131, extracts slots and the like, and extracts the extracted slots from the conversation management unit 104 (conversation generation unit 121). ).

  The memory access unit 132 may be included in the conversation management unit 104. The natural language processor 102 extracts words from the utterance contents of the user converted into text and outputs them to the memory access unit of the conversation management unit 104. The memory access unit of the conversation management unit 104 extracts slots and the like from the semantic network. May be.

  The conversation management unit 104 includes a conversation generation unit 121 and a weight value update unit 122. The conversation generation unit 121 corresponds to each of the plurality of second nodes from the plurality of second nodes (slot values) associated with the first node (slot) specified by the natural language processor 102. Based on the assigned weight value, one second node is selected. The weight value represents the probability that a plurality of second nodes have been selected by the user in the past. The conversation generation unit 121 selects a second node having a probability that is greater than a predetermined value among the plurality of second nodes. The conversation generator 121 generates a response sentence corresponding to the selected second node (slot value). When there is no second node having a probability greater than a predetermined value among the plurality of second nodes, the conversation generation unit 121 generates a response sentence for causing the user to select one of the plurality of second nodes. Generate.

  The conversation generation unit 121 acquires a weight value associated with each of the plurality of slot values extracted by the natural language processor 102 from the weight value management table, and based on the acquired weight value, one slot value To decide. The conversation generation unit 121 generates a response sentence corresponding to the determined slot value. At this time, the conversation generation unit 121 refers to the determination condition table stored in the determination condition table storage unit 113 and determines whether or not the determination condition corresponding to the number of slot values that can be selected for a certain slot is satisfied. to decide. When the determination condition is satisfied, the conversation generation unit 121 inserts the slot value into a response sentence template prepared in advance, and generates a response sentence.

  FIG. 3 is a diagram illustrating an example of the determination condition table.

  As shown in FIG. 3, the determination condition table associates the number of selectable slot values, the condition for selecting the slot value, the slot value obtained when the condition is satisfied, and the template representing the response sentence. ing.

For example, if there are _two selectable slot values “v ₁ ” and “v ₂ ”, and v ₁ is larger than 50% and v ₂ is smaller than 50%, the conversation generating unit 121 sets v ₁ as a slot. Select as value. In addition, the conversation generation unit 121 has two selectable slot values “v ₁ ” and “v ₂ ”, and when v ₁ is smaller than 50% and v ₂ is larger than 50%, v ₂ is slotted. Select as value. Then, when _{the v 1} has been selected as a slot value, conversation generating unit 121, _"v or is _one you _{want? (Do you want v 1?} )" To generate a response sentence. On the other hand, if _{the v 2} has been selected as a slot value, conversation generating unit 121, _"v Do you _{2 want? (Do you want v 2?} )" To generate a response sentence.

Further, the conversation generating unit 121 has two selectable slot values “v ₁ ” and “v ₂ ”, and when v ₁ and v ₂ are 50%, v ₁ and v ₂ are set as slot values. select. Then, when _{the v 1} and _{v 2} is selected as the slot value, conversation generating unit 121, _"v Do you want to either of ₁ or _{v 2? (How about v 1} or v 2?)" Generates a response sentence To do.

Further, the conversation generating unit 121 has a plurality of selectable slot values “v ₁ ”, “v ₂ ”,..., “V _x ”, and any one of the slot values v _i is larger than 50%. , V _i are selected as slot values. Then, when _{the v i} has been selected as a slot value, conversation generation unit 121 generates a response sentence _{"v i} do you _{want? (Do you want v i?} )".

Further, the conversation generating unit 121 has a plurality of selectable slot values “v ₁ ”, “v ₂ ”,..., “V _x ”, and any one of the slot values v _i is larger than 40%. If any one of the slot values v _j different from the slot value v _i is larger than 40%, v _i and v _j are selected as the slot values. Then, when _{the v i} and _{v j} has been selected as a slot value, conversation generating unit 121, generates a response sentence _{"v i} or _v Which do you _{want? (How about v i or v} j?) Of the _j" To do.

Further, the conversation generating unit 121 has a plurality of selectable slot values “v ₁ ”, “v ₂ ”,..., “V _x ”, and when any slot value v _i is smaller than 40%, Do not select a slot value. If no slot value is selected, the conversation generation unit 121 generates a response sentence “What XX (slot name) do you want?” ”.

  Further, the natural language processor 102 acquires text information indicating a user's answer to the response sentence. The natural language processor 102 determines whether the text information is a positive answer such as “Yes” or a negative answer such as “No”. The natural language processor 102 outputs floor information indicating whether the user's answer is positive or negative to the weight value update unit 122. The weight value updating unit 122 updates the weight value according to whether or not the user's answer is a reply that selects one second node among the plurality of second nodes. That is, when the answer information is a positive answer, the weight value update unit 122 recalculates and updates the probabilities associated with a plurality of selectable slot values. On the other hand, when the answer information is a negative answer, the conversation generation unit 121 generates a response sentence for causing the user to select one of a plurality of slot values.

  The voice synthesis unit 105 converts the response sentence generated by the conversation management unit 104 into voice. The voice converted by the voice synthesizer 105 is output from a speaker (not shown).

  In the voice dialogue system shown in FIG. 2, one device may include the voice recognition unit 101, the natural language processor 102, the memory 103, the conversation management unit 104, and the voice synthesis unit 105. Further, the speech recognition unit 101, the natural language processor 102, the memory 103, the conversation management unit 104, and the speech synthesis unit 105 may be distributed over a plurality of devices. For example, a terminal device may include a speech recognition unit 101 and a speech synthesis unit 105, and a server connected to the terminal device through a network may include a natural language processor 102, a memory 103, and a conversation management unit 104. Good.

  Next, the voice dialogue process of the voice dialogue system in the present embodiment will be described.

  FIG. 4 is a flowchart for explaining the voice dialogue processing of the voice dialogue system in the present embodiment.

  First, in step S1, the natural language processor 102 acquires a word from the utterance information indicating the user's utterance content.

  Next, in step S2, the natural language processor 102 searches the semantic network stored in the semantic network storage unit 111 and specifies the extracted word based on the relationship information associated with each node. The nodes (slot and slot value) related to the task to be executed are extracted from the semantic network.

  Next, in step S3, the conversation generation unit 121 determines a slot in which a value needs to be input in order to execute the task.

  Next, in step S <b> 4, the conversation generation unit 121 acquires the weight values of a plurality of slot values associated with the determined slot from the weight value management table.

  Next, in step S <b> 5, the conversation generation unit 121 refers to the determination condition table stored in the determination condition table storage unit 113 and determines whether there is a weight value that satisfies the determination condition. If it is determined that there is a weight value that satisfies the determination condition (YES in step S5), in step S6, the conversation generating unit 121 determines the slot value to be a slot value corresponding to the weight value that satisfies the determination condition. To do.

  Next, in step S7, the conversation generator 121 generates a confirmation response sentence using the determined slot value. The confirmation response text is a response text for confirming to the user whether the determined slot value is acceptable.

  On the other hand, when it is determined that there is no weight value satisfying the determination condition (NO in step S5), the conversation generation unit 121 generates a request response sentence in step S8. The request response text is a response text requesting the user to select a desired slot value from a plurality of selectable slot values.

  Subsequently, the weight value update processing of the voice interaction system in the present embodiment will be described.

  FIG. 5 is a flowchart for explaining the weight value update processing of the voice interaction system according to the present embodiment.

  First, in step S <b> 11, the weight value update unit 122 confirms the slot value included in the response sentence generated by the conversation generation unit 121.

  Next, in step S <b> 12, the weight value update unit 122 acquires response information indicating whether or not the user's response to the response sentence is affirmative from the natural language processor 102.

  Next, in step S13, the weight value update unit 122 determines whether or not the answer information is a positive answer. Here, when it is determined that the answer information is a positive answer (NO in step S13), in step S14, the weight value updating unit 122 acquires a new slot value. At this time, when there are two selectable slot values, the weight value updating unit 122 acquires a slot value that has not been presented to the user as a new slot value. When there are three or more selectable slot values, the weight value update unit 122 acquires the slot value selected by the user as a new slot value.

  On the other hand, when it is determined that the answer information is a positive answer (YES in step S13), in step S15, the weight value update unit 122 recalculates the weight value.

Here, a method of calculating the weight value will be described. First, the initial value of the weight value is stored in the weight value management table before the weight value is calculated by the weight value update unit 122. X slot values v ₁ , v ₂ ,..., V _x can be selected for a certain slot, the number of users is n, and the number of users who have selected each slot value is N ₁ , Assuming N ₂ ,..., N _x , the weight value (probability) of each slot value is expressed by N ₁ / n, N ₂ / n,..., N _x / n. In this case, the number N ₁ of the user who selected the number n and the slot value of the _user, N _2, · · ·, the N _x, an arbitrary number is assigned. For example, the number of people N ₁ , N ₂ ,..., N _x may be set based on past statistical data. Further, the initial value of the weight value may be set to the same value. For example, when two slot values can be selected, the initial value of the weight value (probability) of each slot value is 50%. It may be set.

Further, when the weight value is recalculated by the weight value update unit 122, the weight value update unit 122 adds 1 to the number of users n and 1 to the number N _x of the selected slot values, Recalculate the weight values of all selectable slot values. For example, when the slot value v ₂ is selected, the weight values (probabilities) of the slot values v ₁ , v ₂ ,..., V _x are N ₁ / (n + 1), (N ₂ +1) / (n + 1). ),..., N _x / (n + 1).

  Next, in step S16, the weight value update unit 122 stores the recalculated weight value in the weight value management table storage unit 112, and updates the weight value in the weight value management table.

  FIG. 6 is a diagram for explaining the difference between the voice dialogue processing of the voice dialogue system in the present embodiment and the voice dialogue processing of the conventional voice dialogue system. FIG. 6 shows an example of a voice interaction process when a user orders a hamburger set at a hamburger store.

  First, the user utters “Please give a hamburger set” to the voice interaction system.

  In a conventional speech dialogue system, a word “hamburger set” is extracted from a user's utterance, and a task frame corresponding to the extracted word is specified. In this case, a task frame that provides a hamburger set is specified. Next, the conventional voice interaction system specifies a slot necessary for executing the specified task frame, and asks the user which slot value to use among a plurality of slot values corresponding to the specified slot. In the example shown in FIG. 6, the slot is a drink, and the slot value is coke, tea, orange juice, or the like. The conventional voice dialogue system creates a response sentence 405 "What do you want to drink?" In response to this, the user answers “Please give me coke”. Furthermore, in order to confirm the content uttered by the user, the conventional voice dialogue system creates a response sentence “Is the drink a coke?” And outputs it as a voice. Then, the user answers “Yes”. When an affirmative answer is obtained from the user, the conventional spoken dialogue system sets the slot value of the task frame and executes the task frame. At this time, if a value is input to all of the essential slots in the task frame, the task corresponding to the task frame is executed. If no value is entered in all of the required slots, a question or the like is made to prompt the user to enter a value.

  On the other hand, in the voice interaction system according to the present embodiment, the word “hamburger set” is extracted from the user's utterance, and the nodes (domain, slot, and slot value) related to the task specified by the extracted word are extracted from the semantic network. Extract. In the example shown in FIG. 6, the domain 400 is “hamburger set”, the slot 401 is “drink”, the slot values 402, 403, and 404 are “coke”, “tea”, “orange juice”, and the like. It is.

  Next, the voice interaction system in the present embodiment determines a slot in which a value needs to be input in order to execute the task. Here, the determined slot is a drink. Next, the voice interaction system according to the present embodiment acquires weight values of a plurality of slot values associated with the determined slot from the weight value management table. In the example shown in FIG. 6, the weight value of coke as a slot value is 60%, the weight value of tea as a slot value is 20%, and the weight value of orange juice as a slot value is 5%.

  Next, the voice interaction system according to the present embodiment determines whether there is a weight value that satisfies the determination condition. In this case, since the weight value of coke is 60%, the spoken dialogue system in the present embodiment determines that there is a weight value that satisfies the determination condition. Next, the voice interaction system in the present embodiment determines the slot value to be “coke”. Then, the voice interaction system in the present embodiment creates a response sentence 406 “Do you want to coke?” And outputs it as a voice. In response to this, the user answers “Yes”. When an affirmative answer is obtained from the user, the voice interaction system in the present embodiment executes a task of generating a response sentence for the user's utterance when providing the hamburger set.

  As described above, in the conventional system, the system asks the user what to drink, and the user's answer is determined by voice recognition. When the user selects coke, the conventional system asks again whether coke is acceptable, determines the user's answer by voice recognition, and determines a drink.

  On the other hand, in the system of the present disclosure, the system assigns a weight value to each according to which drink the user has selected in the past without asking the user what to drink. Ask whether to coke according to the weight value. For example, when coke has been selected with a probability of 60% in the past, the system of the present disclosure confirms to the user whether the drink can be coke. And the system of this indication judges a user's answer by voice recognition, and if a reply which affirms from a user is obtained, it will determine a drink as coke.

  In this case, the system according to the present disclosure does not need to confirm the drink again as compared with the conventional system, can reduce the interaction time between the system and the user, and can reduce the processing time of the system.

  Next, a modified example of the voice interaction system in the present embodiment will be described.

  FIG. 7 is a diagram illustrating an example of a semantic network of the voice interaction system according to the modification of the present embodiment. The semantic network shown in FIG. 7 shows an example of a semantic network used when searching for restaurants.

  In FIG. 7, a node 21 indicating “restaurant” is linked to a node 22 indicating “area” and a node 23 indicating “type”. The node 22 indicating “region” and the node 23 indicating “type” are indispensable slots. The node 22 indicating “region” is linked to a node 24 indicating “north” and a node 25 indicating “south”. A node 24 indicating “north” and a node 25 indicating “south” are slot values of the node (slot) 22 indicating “region”. The node 23 indicating “kind” is linked to a node 26 indicating “Indian cuisine”, a node 27 indicating “Chinese cuisine”, and a node 28 indicating “American cuisine”. A node 26 indicating “Indian cuisine”, a node 27 indicating “Chinese cuisine”, and a node 28 indicating “American cuisine” are slot values of a node (slot) 23 indicating “type”.

  Further, in the modification shown in FIG. 7, the slot values of different slots are linked, and the node (slot value) 24 indicating “north” is the node (slot value) 26 indicating “Indian food” and “Chinese food”. Is linked to a node (slot value) 27 indicating "A" and a node (slot value) 28 indicating "American cuisine". For example, a weight value of 30% is given to the connection between the node 24 indicating “north” and the node 26 indicating “Indian cuisine”. For example, a weight value of 60% is assigned to the connection between the node 24 indicating “north” and the node 27 indicating “Chinese food”. Furthermore, for example, a weight value of 10% is assigned to the connection between the node 24 indicating “north” and the node (slot value) 28 indicating “American cuisine”. That is, in the past, the probability that Indian food was selected after the northern region was selected was 30%, and the probability that Chinese food was selected after the northern region was selected was 60%. The probability that American cuisine is selected after the region is selected is 10%.

  In the system of the present disclosure, when the user speaks “I'm looking for a restaurant at the north part of town.”, The user speaks to a Chinese restaurant. The user's answer is judged by voice recognition, and the restaurant is determined.

  The weight value management table storage unit 112 includes one second node of the plurality of second nodes associated with the first node of the plurality of first nodes, and the plurality of first nodes. A weight value is stored in association with each combination with a plurality of second nodes associated with other first nodes of one node. Here, the first node is a slot, and the second node is a slot value.

  The natural language processor 102 determines whether one second node has been identified. When one second node is specified, the conversation generation unit 121 determines the combination of one second node and each of a plurality of second nodes associated with other first nodes. Based on the weight values associated with each other, one second node is selected from the plurality of second nodes associated with the other first nodes.

  Next, the voice dialogue process of the voice dialogue system in the modification of the present embodiment will be described.

  FIG. 8 is a flowchart for explaining the voice dialogue processing of the voice dialogue system in the modification of the present embodiment.

  Since the process of step S21 and step S22 is the same as the process of step S1 and step S2 shown in FIG. 4, description is abbreviate | omitted.

  Next, in step S23, the natural language processor 102 determines whether or not any slot in the semantic network has been specified. For example, in the example shown in FIG. 7, a slot indicating “region” is specified from the content of the user's utterance. If it is determined that an arbitrary slot has been specified (YES in step S23), the natural language processor 102 selects the specified slot in step S24.

  Next, in step S25, the natural language processor 102 selects the next slot linked to the identified slot. For example, in the example shown in FIG. 7, a slot indicating “type” linked to a slot indicating “region” is selected.

  Next, in step S26, the conversation generation unit 121 acquires, from the weight value management table, the weight values associated with the slot value of the identified slot and the plurality of slot values of the selected next slot. To do. The weight value management table includes one slot value among a plurality of slot values associated with one slot among a plurality of slots and a plurality of slots associated with other slots among the plurality of slots. The weight values are stored in association with the combinations with the respective slot values.

  Next, in step S27, the conversation generation unit 121 refers to the determination condition table stored in the determination condition table storage unit 113 and determines whether there is a weight value that satisfies the determination condition. If it is determined that there is a weight value that satisfies the determination condition (YES in step S27), in step S28, the conversation generating unit 121 determines the slot value to be a slot value corresponding to the weight value that satisfies the determination condition. To do. For example, in the example shown in FIG. 7, the slot value indicating “Chinese cuisine” is determined.

  Next, in step S29, the conversation generator 121 generates a confirmation response sentence using the determined slot value. The confirmation response text is a response text for confirming to the user whether the determined slot value is acceptable. For example, in the example shown in FIG. 7, a confirmation response sentence “How about a Chinese restaurant?” Is generated.

  On the other hand, if it is determined that an arbitrary slot has not been specified (NO in step S23), or if it is determined that there is no weight value satisfying the determination condition (NO in step S27), conversation generation is performed in step S30. The unit 121 generates a request response sentence. The request response text is a response text requesting the user to select a desired slot value from a plurality of selectable slot values. For example, in the example shown in FIG. 7, when the user speaks “Looking for a restaurant”, the voice interaction system needs to determine slots indicating “region” and “type”. Therefore, the conversation generation unit 121 generates a request / response sentence for selecting the slot value of either the “region” or “type” slot, for example. ? "Reply-to-request" or "Indian, Chinese or American? Request response sentence is generated.

  If it is determined in step 23 that an arbitrary slot has not been specified, the voice interaction process may be terminated.

  Further, the weight value in the present embodiment represents the probability that each of the selectable slot values has been selected by the user in the past, but the present disclosure is not particularly limited thereto, and each of the selectable slot values is A value may be assigned to. For example, when the slot value is selected by the user, the weight value update unit 122 may add 1 to the weight value of the selected slot value.

  The voice interactive system may set an arbitrary weight value for the slot value. For example, by setting the weight value of the slot value of a product that the dealer wants to sell particularly higher than the weight value of the slot value of other products, it is possible to recommend the product that the seller wants to sell to the user.

  In addition, the voice interactive system may set an arbitrary weight value depending on time (season). For a slot whose frequency of reference varies greatly depending on the time, the weight value that has been updated by the update process up to that time may be set to an arbitrary value in consideration of the influence of the time when that time comes. Also, weight value information corresponding to each period may be prepared in advance, and the weight values of all slots may be changed to arbitrary values when the period comes. At this time, the changed value is not fixed until the time has passed, and the weight value update processing may be applied after setting.

  In this embodiment, text information converted from a user's voice is used. However, the present disclosure is not particularly limited thereto, and text information directly input by an input device such as a keyboard or a touch panel is used. Also good.

  In addition, the voice interaction system according to the present embodiment may include a speaker specifying unit that specifies a user who has spoken. In this case, the weight value management table stores a slot value and a weight value in association with each specified user. Thereby, the response sentence according to each user can be produced | generated, and the dialog time between a system and a user can be shortened more.

  In the voice interaction system of the present embodiment, the period or number of times for updating the weight value may be set. In this case, the spoken dialogue system may stop updating the weight value when a predetermined period has elapsed since the start of updating the weight value. Further, the voice interaction system may stop updating the weight value when the number of weight value updates reaches a predetermined number. As the number of weight value updates increases, the weight value may converge to a certain value. Therefore, the processing load of the voice interactive system can be reduced by setting the period or number of times for updating the weight value.

  In addition, since there are products that are sold only for a predetermined season or a predetermined period, the voice interaction system according to the present embodiment sets a slot value that can be selected only for a predetermined season or a predetermined period. Alternatively, the weight value may be updated only for a predetermined period.

  In order to quantitatively confirm the effect of the present invention, a simulation experiment was conducted. FIG. 9A and FIG. 9B are diagrams showing examples of dialogs executed under two different conditions. FIG. 9A and FIG. 9B both show an example of the dialogue performed between the store clerk and the customer in the same situation in the hamburger shop.

  In the dialogue example shown in FIG. 9A, a question sentence on the store clerk side is generated according to a question method (condition) used by a conventional voice dialogue system, and a customer side response is made. In the dialogue example shown in FIG. 9B, a question sentence on the store clerk side is generated according to a question method (condition) used by the voice dialogue system according to the present disclosure, and a customer side response is made. Hereinafter, explanation will be made while comparing the way (conditions) of these two questions.

  9A and 9B, “Would you like side salad or French fries?” (Conversation sentence 1001) corresponds to “Would you like French fries?” (Conversation sentence 2001). At this time, when the content indicated by the conversation sentence 2001 is output as a question from the spoken dialogue system, it is determined that the probability that “french fries” is ordered as “meal side” is equal to or greater than a threshold value. The customer's utterances are “French fries” of the conversation sentence 1002 and “Yes” of the conversation sentence 2002, respectively, and both the question and the answer are shorter in the example shown in FIG. 9B, which is more efficient. Similarly, in the conversation sentence 2005 corresponding to “Would you like large, small or medium?” In the conversation sentence 1005, based on the statistics that the order probability of “medium” is high, the conversation sentence 1005 is changed to “Would you like medium”. In the conventional example, the response of the customer is “medium” of the conversation sentence 1006, whereas in the example of the voice dialogue system in the present disclosure, The conversation sentence 2006 is “Yes”, and the response content is shortened.

  On the other hand, in the example of the voice dialogue system according to the present disclosure, “Would you like coke?” In the conversation sentence 2007 is a proposal type in contrast to the “What kind of meal drink would you like?” In the conversation sentence 1007 in the conventional example. This is a short question compared to the conventional example. However, the customer's response to this is “Hi-orange lavaburst” of the conversation sentence 1008 in the conventional example, whereas “No. Hi-orange lavaburst” of the conversation sentence 2008 is the example of the spoken dialogue system of the present disclosure. "In the example of the speech dialogue system of the present disclosure, the conversation sentence is partially longer. When there are no multiple options that are expected to have a particularly high probability, there is no significant difference between the question and the answer, such as the conversation sentence 1003 and the conversation sentence 2003, and the conversation sentence 1004 and the conversation sentence 2004. The contents of the orders accepted by the conversation examples shown in FIGS. 9A and 9B are the order acceptance contents 1010 and the order contents 2010, respectively, which are exactly the same results.

  When the total number of characters of the two dialogues performed in this way is counted, it becomes 330 characters (1009) and 273 characters (2009), respectively, and depending on the question method (condition) used by the voice dialogue system of the present disclosure, It can be seen that the number of characters in the conversation is smaller when the question sentence is generated, that is, the conversation is shorter. Here, although both were compared using the number of characters, the time required for the conversation can be estimated by multiplying the number of characters by a predetermined coefficient. As mentioned above, if the customer's answer to the store's proposal-type question is negative, the time required for the conversation is partially increased, but the customer's response to the store-side proposal-type question is high. Since a yes / no question is asked only when it can be expected to be affirmative in probability, the time required for the conversation is longer when a proposal type question is output than with a conventional question. It can be said that long cases occur rarely (low probability).

  In the same manner as this example, a simulation was performed on the order of 100 cases. As a result, the total number of characters was 28746 characters by the conventional method and 26168 characters by the method according to the present disclosure. In the method according to the present disclosure, the total dialogue length is reduced by 11.4% compared to the conventional example, and it was confirmed that the dialogue method in the voice dialogue system according to the present disclosure has a large effect.

  The dialogue method, the dialogue program, and the dialogue system according to the present disclosure can reduce the dialogue time between the dialogue system and the user, can reduce the processing time of the dialogue system, and respond to the user's utterance. The present invention is useful as a dialogue method in a dialogue system, a dialogue program that responds to a user's utterance, and a dialogue system that responds to a user's utterance.

DESCRIPTION OF SYMBOLS 101 Speech recognition part 102 Natural language processor 103 Memory 104 Conversation management part 105 Speech synthesizer 111 Meaning network storage part 112 Weight value management table storage part 113 Judgment condition table storage part 121 Conversation generation part 122 Weight value update part 131 Syntax analysis part 132 Memory access part

Claims (6)

A dialogue method used in a dialogue system that responds to a user's utterance,
A plurality of nodes necessary for executing a task for generating a response sentence in response to the user's utterance are stored in association with each other,
Obtaining utterance information indicating the utterance content of the user;
Identifying a first node corresponding to the utterance information from the plurality of nodes;
Based on a weight value associated with each of the plurality of second nodes, a second node is selected from the plurality of second nodes associated with the identified first node. Selected,
It generates a response sentence according to the selected second node of said 1,
The weight value represents a probability that the plurality of second nodes have been selected by the user in the past,
A second node having the probability larger than a predetermined value is selected from the plurality of second nodes;
How to interact. A dialogue method used in a dialogue system that responds to a user's utterance,
A plurality of nodes necessary for executing a task for generating a response sentence in response to the user's utterance are stored in association with each other,
Obtaining utterance information indicating the utterance content of the user;
Identifying a first node corresponding to the utterance information from the plurality of nodes;
Based on a weight value associated with each of the plurality of second nodes, a second node is selected from the plurality of second nodes associated with the identified first node. Selected,
Generating a response sentence corresponding to the selected first second node;
The weight value represents a probability that the plurality of second nodes have been selected by the user in the past,
If there is no second node having the probability greater than a predetermined value among the plurality of second nodes, a response sentence for causing the user to select one of the plurality of second nodes is generated. ,
To-talk method. A dialogue method used in a dialogue system that responds to a user's utterance,
A plurality of nodes necessary for executing a task for generating a response sentence in response to the user's utterance are stored in association with each other,
Obtaining utterance information indicating the utterance content of the user;
Identifying a first node corresponding to the utterance information from the plurality of nodes;
Based on a weight value associated with each of the plurality of second nodes, a second node is selected from the plurality of second nodes associated with the identified first node. Selected,
Generating a response sentence corresponding to the selected first second node;
Obtaining information indicating the user's answer to the response sentence;
Updating the weight value according to whether or not the user's answer is an answer for selecting one second node of the plurality of second nodes;
To-talk method. A dialogue method used in a dialogue system that responds to a user's utterance,
A plurality of nodes necessary for executing a task for generating a response sentence in response to the user's utterance are stored in association with each other,
Obtaining utterance information indicating the utterance content of the user;
Identifying a first node corresponding to the utterance information from the plurality of nodes;
Based on a weight value associated with each of the plurality of second nodes, a second node is selected from the plurality of second nodes associated with the identified first node. Selected,
Generating a response sentence corresponding to the selected first second node;
A second node of a plurality of second nodes associated with a first node of the plurality of first nodes and a second one of the plurality of first nodes. The weight value is associated with a combination with each of a plurality of second nodes associated with one node;
Determining whether the first second node has been identified;
When the first second node is specified, the first second node is associated with a combination of each of the plurality of second nodes associated with the other first node. Selecting one second node from among the plurality of second nodes associated with the other first node based on the determined weight value;
To-talk method. An interactive program that responds to user utterances,
A storage unit for storing a plurality of nodes necessary for executing a task of generating a response sentence in response to the user's utterance;
An acquisition unit for acquiring utterance information indicating the utterance content of the user;
A specifying unit for specifying a first node corresponding to the utterance information from the plurality of nodes;
Based on the weight value associated with each of the plurality of second nodes from among the plurality of second nodes associated with the first node identified by the identifying unit, A selection unit for selecting two nodes;
Causing the computer to function as a generation unit that generates a response sentence corresponding to the first second node selected by the selection unit ;
The weight value represents a probability that the plurality of second nodes have been selected by the user in the past,
The selection unit selects a second node having the probability larger than a predetermined value among the plurality of second nodes.
Interactive program. An interactive system that responds to user utterances,
A storage unit for storing a plurality of nodes necessary for executing a task of generating a response sentence in response to the user's utterance;
An acquisition unit for acquiring utterance information indicating the utterance content of the user;
A specifying unit for specifying a first node corresponding to the utterance information from the plurality of nodes;
Based on the weight value associated with each of the plurality of second nodes from among the plurality of second nodes associated with the first node identified by the identifying unit, A selection unit for selecting two nodes;
A generation unit that generates a response sentence corresponding to the first second node selected by the selection unit;
Equipped with a,
The weight value represents a probability that the plurality of second nodes have been selected by the user in the past,
The selection unit selects a second node having the probability larger than a predetermined value among the plurality of second nodes.
Dialog system.

Images