Patent Application
20190392005
The present application claims priority from Japanese application JP 2018-119325, filed on Jun. 22, 2018, the contents of which is hereby incorporated by reference into this application.
The present invention relates to a speech dialogue system, a model creating device, and a model creating method.
As a related text dialogue system (hereinafter, related system), there is a system which outputs a plurality of question sentences to a user and displays information based on a plurality of answer sentences input by the user. For example, when the related system is used to provide a service of displaying a riding time, the related system prompts a user to input a place of departure and a destination and displays a riding time based on information on the input departure place and destination.
For example, an example of techniques relating to the related system includes a technique described in JP-A-2015-225402. JP-A-2015-225402 describes an information retrieval device that includes: a storage unit which stores a plurality of response contents including an assumed answer and an asking-back question to lead to the assumed response; a reception unit which receives a user question; a retrieval unit which retrieves the plurality of response contents on the basis of the user question received by the reception unit and acquires either the assumed answer or the asking-back question corresponding to the user question; and an output unit which outputs a response content acquired by the retrieval part.
In the technique described in JP-A-2015-225402, it is necessary to previously determine the order of user questions. Therefore, as a speech dialogue system that appropriately selects and outputs answer sentences and question sentences in response to the user questions, attempts have been made to construct a speech dialogue system that includes a slot value extraction unit and a plurality of slot value extraction models. However, it is necessary to manually create a large number of assumed input character strings used to create the slot value extraction models, which results in a problem of complicated operation.
An object of the invention is to automatically create a plurality of slot value extraction models.
In order to solve the problems, the invention provides a speech dialogue system that converts an input speech to be input into information of an input character string, creates an output character string containing information of an answer sentence or a question sentence based on the converted information of the input character string, converts information of the created output character string into a synthetic speech, and outputs the converted synthetic speech as an output speech. The speech dialogue system includes: a value list in which a plurality of values indicating candidates of a character string assumed in advance, which are information constituting a character string, and a plurality of value identifiers that identify each of the plurality of values are stored in association; an answer sentence list in which each of a plurality of slots indicating an identifier that identifies the information constituting the character string and each of the plurality of value identifiers are stored in association, and each of the plurality of slots and each of the plurality of value identifiers are stored in association with one or more answer sentences; a peripheral character string list in which each of the plurality of slots and each of a plurality of peripheral character strings arranged adjacent to each of the plurality of slots are stored in association; a storage unit that stores a plurality of assumed input character strings assumed in advance and a plurality of slot value extraction models including one or more of the slots and the values associated with each of the plurality of assumed input character strings; a slot value extraction unit that compares a similarity between the input character string and each of the assumed input character strings in the plurality of slot value extraction models, estimates a position of a slot in the input character string based on a slot associated with an assumed input character string having a high degree of similarity, and extracts a value corresponding to the estimated position of the slot from the input character string; a learning data creating unit that creates first learning data based on the value list, the answer sentence list, and the peripheral character string list; and a model creating unit that creates a first slot value extraction model based on the first learning data and stores the created first slot value extraction model in the storage unit as a model belonging to the plurality of slot value extraction models.
According to the invention, a plurality of slot value extraction models can be automatically created. As a result, work cost required for creating the slot value extraction models can be reduced.
An embodiment of the invention will be described in detail below with reference to drawings.
(Configuration of Speech Dialogue System 2000)
The speech processing system 3000 includes a speech input unit 10 that includes a microphone or the like and from which a speech is input, a speech recognition unit 20 that removes a sound (noise) other than a speech from a speech 100 input from the speech input unit 10 and converts the speech from which the noise has been removed into character string information (input character string 200), a speech synthesis unit 60 that creates a synthetic speech 400 according to an output character string 300 output from the text dialogue system 1000, and a speech output unit 70 that includes a speaker and the like and outputs a predetermined synthetic speech from the synthetic speech 400 created by the speech synthesis unit 60.
The text dialogue system 1000 includes a text dialogue support device 1200 and a model creating device 1100. The text dialogue support device 1200 is connected to the speech processing system 3000 and transmits the corresponding output character string 300 to the speech processing system 3000 by performing predetermined information processing based on the input character string 200 received from the speech processing system 3000.
The text dialogue support device 1200 includes a slot value extraction unit 30, a value identifier estimation unit 40, an answer narrow-down unit 50, a plurality of slot value extraction models 500, a value list 510, an answer sentence list 520, and a question sentence list 530. The slot value extraction unit 30 refers to the plurality of slot value extraction models 500, estimates an identifier (hereinafter referred to as slot) related to information included in the input character string 200, and extracts a character string (hereinafter referred to as value) related to the slot from the input character string 200. The value identifier estimation unit 40 compares the degree of similarity between the value and a plurality of assumed values registered in advance in the value list 510. Within the value list 510, when there is an assumed value having a high degree of similarity to the value, the value identifier estimation unit 40 determines the identifier of the assumed value (hereinafter, referred to as value identifier) as the value identifier of the value.
The answer narrow-down unit 50 determines whether value identifiers of slots necessary for information display have been prepared. For example, when value identifiers of slots necessary for displaying a riding time have been prepared, the answer narrow-down unit 50 outputs an answer sentence (a character string describing the riding time) associated with the value identifiers. On the other hand, when the value identifiers of the slots are not prepared, the answer narrow-down 50 outputs a question sentence (for example, where is the place of departure?) prompting the user to input information related to the missing slot (for example, <departure place>).
The model creating device 1100 is an information processing device used by an administrator or the like of the speech dialogue system 2000 and the text dialogue system 1000, and creates the slot value extraction model 500 to which the slot value extraction unit 30 refers. The model creating device 1100 includes a learning data creating unit 80, a model creating unit 90, a peripheral character string list 540, and a plurality of learning data 550. The learning data creating unit 80 transmits and receives information to and from the text dialogue support device 1200, acquires information recorded in the value list 510 and the answer sentence list 520, and creates a plurality of the learning data 550 necessary for creating the slot value extraction model 500 based on the information recorded in the value list 510, the answer sentence list 520, and the peripheral character string list 540. The model creating unit 90 creates the slot value extraction model 500 from the learning data 550 by performing conversion processing on the learning data 550, for example, performing processing by machine learning, and transmits the created slot value extraction model 500 to the text dialogue support device 1200.
The plurality of slot value extraction models 500, the value list 510, the answer sentence list 520, the question sentence list 530, the peripheral character string list 540, and the plurality of pieces of learning data 550 are stored in a storage unit configured by the main storage device 12 or the auxiliary storage device 13. In addition, the slot value extraction unit 30, the value identifier estimation unit 40, the answer narrow-down unit 50, the learning data creating unit 80, and the model creating unit 90 can achieve functions thereof through, for example, executing various processing programs (a slot value extraction program, a value identifier estimation program, an answer narrow-down program, a learning data creating program, and a model creating program) stored in the main storage device 12 or the auxiliary storage device 13 by the CPU.
(Process Flow of Speech Dialogue System 2000)
Next, the process flow of the speech dialogue system 2000 according to the first embodiment of the invention will be described.
Next,
As described above, by a series of process flow, the speech 100 of the dialogue partner input to the speech input unit 10 can be converted into the information of the input character string 200, and the information of the converted input character string 200 can be transmitted to the text dialogue system 1000. The information of the output character string 300 output from the text dialogue system 1000 can be converted into the synthetic speech 400, and the converted synthetic speech 400 can be played from the speech output unit 70 to the dialogue partner.
(Process Flow of Text Dialogue System 1000)
Next, the process flow of the text dialogue system 1000 will be described.
For example, when information of “I would like to go to Tokyo Station” is input as the input character string 200, the slot value extraction unit 30 compares the degree of similarity between the input character string 200 and the assumed input character string 502 of the slot value extraction model 500 of
Next, when the information of slot and value is received from the slot value extraction unit 30, the value identifier estimation unit 40 refers to the value list 510, and compares the degree of the similarity between the received value and the assumed value 512. When the degree of similarity is high, the value identifier 511 corresponding to the assumed value 512 is estimated, and information of the estimation result (value identifier) and information of the value are transferred to the answer narrow-down unit 50 (S31). For example, the value identifier estimation unit 40 estimates “<Tokyo station>” as the value identifier 511 when the received value is “Tokyo Station”.
Next, when the information (“<Tokyo station>”) of the estimation result (value identifier) and the information (“Tokyo Station”) of the value are received from the value identifier estimation unit 40, the answer narrow-down unit 50 refers to the answer sentence list 520, and determines whether the value identifiers of the slots necessary for information display have been prepared (S32, S33). For example, when the value identifiers of the slots necessary for displaying the riding time (for example: the value identifier of the slot <destination> is <Tokyo Station>, the value identifier of the slot <place of departure> is <Katsuta Station>) have been prepared, the answer narrow-down unit 50 outputs, for example, the information of “The riding time is approximately 2 hours.” as the answer sentence 523 associated with the value identifiers (“<Tokyo Station>”, “<Katsuta Station>”) (S34), and the processing in this routine ends.
On the other hand, when there is only a value identifier “<Tokyo Station>” indicating <destination> and the value identifiers of the slots necessary for display of riding time have not been prepared, the answer narrow-down unit 50 refers to the question sentence list 530 and outputs, for example, the information of “where is the place of departure” as the question sentence 532 prompting the user to input information related to the missing slot (for example, <place of departure>) (S35). Next, the answer narrow-down unit 50 records the information of the acquired value identifier in a memory (storage unit) (S36), and the processing in this routine ends.
As described above, by a series of process flow of the text dialogue system 1000, a plurality of question sentences are output to the user, appropriate information display are possible based on a plurality of answer sentences input by the user.
(Process Flow of Model Creating Device 1100)
Next, the process flow of the model creating device 1100 according to the first embodiment of the invention will be described.
(Method for Creating Learning Data 550)
In order to create an assumed input character string, the learning data creating unit 80 acquires a plurality of value identifiers associated with one answer sentence in the answer sentence 523 from the answer sentence list 520 (S40). Next, the learning data creation unit 80 selects N (N=1 to Nmax (predefined maximum value)) value identifier(s) from the acquired multiple value identifiers to create combinations (S41), and creates permutations for each created combination (S42). For example, when there are two value identifiers associated with the answer sentence 523, for example, M21=[<Katsuta Station>, <Tokyo Station>], M22=[<Tokyo Station>, <Katsuta Station>]) are created as permutations using two value identifiers such as “<Katsuta Station>” and “<Tokyo Station>”; and for example, M11=[<Katsuta Station>], M12=[<Tokyo Station>] are created as permutations using one value identifier.
Next, the learning data creation unit 80 determines whether permutations of the value identifiers have been created for all answer sentences (S43). In step S43, when a negative determination result is obtained, the process flow of the learning data creating unit 80 proceeds to step S40, and the processes of steps S40 to S43 are repeated. On the other hand, when a positive determination result is obtained in step S43, the learning data creating unit 80 selects one permutation from the permutations created in step S42 (S44), and selects one value identifier of the selected permutation (S45).
Next, the learning data creating unit 80 refers to the value list 510 based on the value identifier selected from the permutation, and acquires, from the assumed value 512 in the value list 510, a value such as “Katsuta Station” as the value associated with the value identifier (for example, <Katsuta Station>) of the permutation such as M21=[<Katsuta Station>, <Tokyo Station>] (S46).
At this time, the learning data creating unit 80 refers to the answer sentence list 520 based on the value identifier selected from the permutation, and acquires, from the slot and value identifier 522 in the answer sentence list 520, a slot such as “<place of departure>” as the slot associated with a value identifier (for example, <Katsuta Station>) of the permutation such as M21=[<Katsuta Station>, <Tokyo Station>]) (S47). Further, the learning data creating unit 80 refers to the peripheral character string list 540 based on the obtained slot “<place of departure>”, and acquires, from the slot peripheral character string 542 in the peripheral character string list 540, a peripheral character string such as “from @” as the peripheral character string associated with the acquired slot “place of departure” (S48).
Next, based on the value (“Katsuta Station”) acquired in step S46, the slot (<place of departure>) acquired in step S47, and the peripheral character string (“from @”) acquired in step S48, the learning data creating unit 80 creates a character string such as C1=“from Katsuta Station” in which the value such as “Katsuta Station” is inserted into a value insertion position of the peripheral character string such as “@” (S49).
Next, the learning data creating unit 80 determines whether character strings have been created for all the value identifiers in the permutation (S50). When a negative determination result is obtained in step S50, the process flow of the learning data creating unit 80 proceeds to step S45, and the processes of steps S45 to S50 are repeated.
At this time, the learning data creating unit 80 acquires, from the assumed value 512 in the value list 510, a value such as “Tokyo station” as the value associated with another value identifier of the permutation M21 such as <Tokyo Station>. Further, the learning data creating unit 80 acquires, from the slot and value identifier 522 in the answer sentence list 520, a slot such as “<destination>” as the slot associated with the other value identifier such as <Tokyo Station>. Still further, the learning data creating unit 80 refers to the peripheral character string list 540 based on the acquired slot “<destination>”, and acquires, from the slot peripheral character string 542 in the peripheral character string list 540, a peripheral character string such as “I want to go to @” as the peripheral character string associated with the acquired slot “<destination>”. At this time, the learning data creating unit 80 creates a character string such as C2=“I want to go to Tokyo Station” in which the value (for example, “Tokyo Station”) is inserted into the value insertion position of the peripheral character string.
On the other hand, when a positive determination result is obtained in step S50, the learning data creating unit 80 combines the character strings created from the value identifiers to create information of the assumed input character string (S51). For example, the learning data creating unit 80 combines the character strings created from the value identifiers included in the permutation to create an assumed input character string, for example, C1+C2=“I want to go from Katsuta Station to Tokyo station”.
Next, the learning data creating unit 80 determines whether the assumed input character strings have been created for all the permutations (S52). When a negative determination result is obtained in step S52, the process flow of the learning data creating unit 80 proceeds to step S45, and the processes of steps S44 to S52 are repeated. On the other hand, when a positive determination result is obtained in step S52, the learning data creating unit 80 creates, as learning data (first learning data) 550, data associated with the slots and values used for creating a plurality of assumed input character strings and associated with the assumed input character strings (S53), and the processing in this routine ends.
At this time, for each combination of the permutations of the value identifiers, the learning data creating unit 80 respectively acquires the values associated with the value identifiers of elements belonging to the permutations of the value identifier from the value list 510 as the values of elements, acquires the slots associated with the value identifiers of elements from the answer sentence list 520 as the slots of elements, and acquires the peripheral character strings associated with the slots of elements from the peripheral character string list 540 as the peripheral character strings of elements. Then, the learning data creating unit 80 creates the character strings of elements by combining the acquired values of elements and the acquired peripheral character strings of elements, creates a plurality of assumed input character strings by combining the character strings of elements, and creates the first learning data 550 associated with the assumed input character strings and the slots and values of elements based on the plurality of created assumed input character strings and the slots and values of elements used for creating the plurality of assumed input character strings.
(Model Creating Method)
The model creating unit 90 creates a slot value extraction model (first slot value extraction model) 500 according to the learning data (first learning data) 550. In the slot value extraction model 500, the assumed input character string and the slot and value defined in advance are registered. For example, the learning data 550 and the slot value extraction model 500 may be the same. Further, the slot value extraction model 500 may be created by machine learning (for example, the method of conditional random fields) using the assumed input character string of the learning data 550 and the slot and value as inputs.
According to the present embodiment, a plurality of slot value extraction models can be automatically created. As a result, the work cost required for creating the slot value extraction models can be reduced.
According to the second embodiment, highly accurate slot value extraction can be achieved by switching between a plurality of slot value extraction models (first and second slot value extraction models) in the speech dialogue system 2000 described in the first embodiment. Further, the work cost required for creating the plurality of slot value extraction models is reduced.
In the first embodiment, when the value identifiers of the slots necessary for information display have not been prepared, the answer narrow-down unit 50 refers to the question sentence list 530 and outputs a question sentence (for example, where is the place of departure?) that prompts the user to input information related to the missing slot (for example, <place of departure>). In contrast, in order to extract a slot value with high accuracy from an input character string of a dialogue partner, the slot value extraction unit 30 according to the second embodiment uses a slot value extraction model (second slot value extraction model) in which only an assumed input character string related to an acquired slot is not included. Since only the assumed input character string related to the acquired slot is not included in the slot value extraction model, there is no possibility that the slot value extraction unit erroneously extracts the acquired slot. Therefore, accuracy of slot value extraction according to the second embodiment is higher than that in the first embodiment.
Further, in order to reduce work cost necessary for creating a plurality of slot value extraction models, the learning data creating unit 80 according to the second embodiment creates the second learning data by only removing the assumed input character string related to the specific slot from the learning data (first learning data) 550 created in the first embodiment. Then, the model creating unit 90 creates the second slot value extraction model from the second learning data.
Specifically, in the case of the learning data 550 created in the first embodiment, the learning data creating unit 80 creates combinations, for example, two types, in which N (N=1 to M−1) slots are selected from all the slots (M=2) (S60). Next, the learning data creating unit 80 selects one combination from the combinations (two types) created in step S60, and for the selected combination, the learning data (the second learning data) 550(2A, 2B) is created, in which only the assumed input sentence (assumed input character string) related to the slot not included in the combination is removed from the learning data 550 (S61), as is shown in
According to the present embodiment, highly accurate slot value extraction can be achieved by switching the plurality of slot value extraction models from the first slot value extraction model to the second slot value extraction model in the speech dialogue system 2000 described in the first embodiment. In addition, the work cost required for creating the plurality of slot value extraction models can be reduced.
In order to extract a slot value with high accuracy from an input character string of a dialogue partner, the slot value extraction unit 30 according to the third embodiment switches a slot value extraction model to be used from a first slot value extraction model to a third slot value extraction model based on a dialogue log. An example of the dialogue log is shown in
The ID 561 is an identifier for uniquely identifying the dialogue log. The question sentence 562 is information for managing a question sentence for a user. In the question sentence 562, for example, information of “Where is the destination?” is registered. The slot 563 is information for managing the probability (ratio) of the slot included in the question sentence 562. For example, as indicated by “1” in the ID 561, “-” (no question output) is shown as the question sentence 562, and when the probability of including the information of “<place of departure>” is “20%”, the information of “20%” is registered in <place of departure> 564. As indicated by “2” in the ID 561, “Where is the destination?” is shown as the question sentence 562, and when the probability of including the information of “<place of departure>” is “0%”, the information of “0%” is registered in <place of departure> 564. As indicated by “3” in the ID 561, “Where is the place of departure?” is shown as the question sentence 562, and when the probability of including the information of “<place of departure>” is “80%”, the information of “80%” is registered in <place of departure> 564. As indicated by “4” in the ID 561, “When is the departure time?” is shown as the question sentence 562, and when the probability of including the information of “<place of departure>” is “0%”, the information of “0%” is registered in <place of departure> 564.
The dialogue log shows probabilities of respective slots of being included in the input character string of the dialogue partner. For example, when there is no question sentence output of the text dialogue system 1000 (“1” in the ID 561), the probability that only the character string related to <place of departure> 564 in the slot 563 is included in the input character string 200 of the dialogue partner is “20%” which is equal to or higher than a threshold (for example, 10%), and the probability that only the character string related to <destination> 565 in the slot 563 is included in the input character string 200 is “80%” which is equal to or higher than the threshold. Therefore, in order to improve accuracy of slot value extraction, in the slot value extraction of the input character string 200 when there is no output of the question sentence, the slot value extraction unit 30 uses the slot value extraction model 550 (see
Similarly, in the slot value extraction of the input character string 200 for the question sentence “Where is the destination?”, the slot value extraction unit 30 uses the slot value extraction model 550 (see
In addition, in the slot value extraction of the input character string 200 for the question sentence “Where is the place of departure?”, the slot value extraction unit 30 uses the slot value extraction model 550 (see
In addition, in the slot value extraction of the input character string 200 for the question sentence “When is the departure time?”, the slot value extraction unit 30 uses the slot value extraction model 550 (see
Therefore, based on the dialogue log 560, it is necessary to manage the slot value extraction model 550 in which the assumed input character string related to the specific slot is registered with a management table.
At this time, the learning data creating unit 80 creates the learning data related to the specific slot based on the dialogue log 560 in order to reduce the work cost necessary for creating the plurality of slot value extraction models 500 (see
According to the present embodiment, highly accurate slot value extraction can be achieved by switching the plurality of slot value extraction models from the first slot value extraction model to the third slot value extraction model in the speech dialogue system 2000 described in the first embodiment. In addition, the work cost required for creating the plurality of slot value extraction models can be reduced.
While the invention made by the inventor has been described in detail based on the embodiments, the invention is not limited thereto, and various modifications can be made without departing from the scope of the invention. For example, the value list 510 and the answer sentence list 520 may be arranged in the model creating device 1100.
The invention can be widely applied to a dialogue system in which voice and text are input such as a dialogue robot equipped with a speech dialogue system and a chat bot equipped with a text dialogue system.
The configurations, functions, and the like may be achieved entirely or partially by hardware, for example, by designing them in an integrated circuit. In addition, the configurations, functions, and the like may be achieved by software by interpreting and executing a program for achieving each function by a processor. Information such as programs, tables and files for realizing the functions may be recorded and stored in a storage device such as a memory, a hard disk, or a solid state drive (SSD), or in a recording medium such as an integrated circuit (IC) card, a secure digital (SD) memory card, or a digital versatile disc (DVD).
| Number | Date | Country | Kind |
|---|---|---|---|
| 2018-119325 | Jun 2018 | JP | national |
