JP6097776B2 - Word selection device, method, and program - Google Patents

Description

  The present invention relates to a word selection device, method, and program, and more particularly, to a word selection device, method, and program for specifying input utterance content.

  Conventionally, the reliability of a recognition candidate sentence has been mainly used to narrow down the user utterance sentence from the recognition candidate sentence of the user utterance. The reliability is an index indicating how likely the recognition candidate sentence is from the viewpoint of an acoustic model or a language model, and a calculation method has been proposed in the prior art (Patent Document 1). On the other hand, there is a limit to narrowing down the user utterance sentence only by the reliability, and a method for specifying the content of the user utterance by combining information other than information obtained from the speech recognition engine has been proposed.

  Conventionally, there is a method of scoring words of recognition candidate sentences using past utterance history and word category information and determining user utterance sentences based on the scores (Non-patent Document 1).

  There is also a method of proposing a strategy to identify the user's utterance by using the hierarchical structure of the address and the address information of the facility and conducting a map search with car navigation etc. while proceeding with the dialogue while asking confirmation (Non-patent document 2).

JP 2013-072922 A

Takaki Fujiwara, Toshihiko Ito, Kenji Araki, Mitsuhiko Kai, Tatsuhiro Konishi, Yukihiro Ito, “Spoken Language Understanding Method Using Recognition Reliability and Dialog History”, IEICE Transactions, Vol.J89-D No.7 , Pp.1493-1503, 2006 Kitaoka Norihide, Yano Hirotoshi, Nakagawa Seiichi, “Natural and Efficient Spoken Dialogue Strategy for Repairing Misrecognition”, Information Processing Society of Japan Research Report. SLP, Spoken Language Information Processing, pp.37-42, 2006

  However, in the technique of Non-Patent Document 1, a large amount of utterance history rich in word variations is required, and in the case of a service including a plurality of domains such as recipe search and map search, a dialogue strategy is set to a domain. There is a problem that it is necessary to think every time.

  Further, in the technique of Non-Patent Document 2, there is a problem that it is difficult to narrow down by this method when there is a search condition that does not have a hierarchical structure in the search of the domain targeted by the service.

  Further, in the conventional technology, external data such as past dialogue history and information on the hierarchical structure of words is required. In addition, in order to deal with various domains, it is necessary to determine which content word corresponds to each domain and to determine a strategy for answering questions. Furthermore, in the past, when a question was returned from a system to a user, the dialogue was mainly a confirmation such as “Are you sure?”. However, there is a problem that it is difficult to understand how the user should respond when the presented candidate is different from the user's intention to speak.

  The present invention has been made in order to solve the above-mentioned problems, and an object thereof is to provide a word selection device, method, and program for appropriately selecting a question-returning word for specifying input utterance content. To do.

  In order to achieve the above object, the word selection device according to the first aspect of the present invention specifies for each of a specific character string for each of a plurality of recognition candidate sentences that are voice recognition results for the input user's voice data. At least one extracted from the recognition candidate sentence by the dialogue function based on at least one dialogue function for extracting the specific character string that is predetermined for the dialogue scenario that is a rule for performing the operation of A content analysis unit that generates a content array storing the specific character string, and the user based on the at least one specific character string stored in the content array generated by the content analysis unit A word candidate extraction unit that selects words to be replied to.

  A word selection method according to a second aspect of the present invention is a word selection method in a word selection device including a content analysis unit and a word candidate extraction unit, wherein the content analysis unit is configured to generate a voice for input voice data of a user. For each of a plurality of recognition candidate sentences that are recognition results, at least one for extracting the specific character string, which is predetermined for a dialogue scenario that is a rule for performing a specific action on each of the specific character strings. Based on two dialogue functions, a content array storing at least one specific character string extracted from the recognition candidate sentence by the dialogue function is generated, and the word candidate extraction unit is generated by the content analysis unit A word to be asked to the user is selected based on the at least one specific character string stored in the content array.

  According to the first and second inventions, the content analysis unit performs a specific operation for each of the specific character strings for each of a plurality of recognition candidate sentences that are voice recognition results for the input user's voice data. At least one specific character string extracted from the recognition candidate sentence by the interactive function based on at least one interactive function for extracting the specific character string, which is predetermined for the interactive scenario that is a rule for performing The stored content array is generated, and the word candidate extraction unit selects a word to be asked to the user based on at least one specific character string stored in the generated content array.

  In this way, for each of a plurality of recognition candidate sentences that are voice recognition results for the input user's voice data, a content array is generated based on at least one dialogue function predetermined for the dialogue scenario, Appropriate selection of the word to be queried to identify the input utterance content by selecting the word to be queried to the user based on at least one specific character string stored in the generated content array Can do.

  Moreover, in the first invention, the word candidate extraction unit includes a content function generated by the content analysis unit and a dialog function corresponding to the dialog function obtained by extracting the specific character string stored in the content array. If a content table that represents a combination of names as one row is created and only one type of interactive function name exists in the created content table, there is a row that matches the content array in the created content table When the matching rows are integrated, and there is a row in which the content array is in an inclusive relationship in the content table, among the rows in the inclusive relationship, a specific stored in the content array When the row with the smaller number of character strings is deleted, and there is only one column in which the specific character string is different among the columns corresponding to the elements of the content array in the content table, The identification included in the difference column Each string may be selected as the ask back word.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said word selection apparatus.

  As described above, according to the word selection device, method, and program of the present invention, a dialogue scenario is determined in advance for each of a plurality of recognition candidate sentences that are voice recognition results for the input user's voice data. Generating a content array based on at least one interactive function and selecting a word to be queried to the user based on at least one specific character string stored in the generated content array It is possible to appropriately select a question-returning word for specifying the uttered content.

It is a figure which shows the example of a dialogue function. It is a block diagram which shows the functional structure of the word selection apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the example of a content arrangement | sequence. It is a figure which shows the example of a content arrangement | sequence. It is a figure which shows the example of a content arrangement | sequence. It is a figure which shows the example of a content table. It is a figure which shows the example of a content table. It is a figure which shows the example of a content table. It is a flowchart figure which shows the word selection processing routine in the word selection apparatus concerning the 1st Embodiment of this invention. It is a flowchart figure which shows the determination processing routine of the question return method based on the reliability in the word selection apparatus concerning the 1st Embodiment of this invention. It is a flowchart figure which shows the determination processing routine of the question return method based on the dialogue function in the word selection apparatus concerning the 1st Embodiment of this invention. It is a flowchart figure which shows the decision processing routine of the question return method in the word selection apparatus concerning the 1st Embodiment of this invention. It is a flowchart figure which shows the execution processing routine of the question return in the word selection apparatus concerning the 1st Embodiment of this invention. It is a block diagram which shows the functional structure of the word selection apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart figure which shows the word selection process routine in the word selection apparatus concerning the 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, a service that achieves the purpose of the user through a voice dialogue between the user and the system is assumed. Hereinafter, the user utterance refers to a speech utterance by a service user, and a sentence that is correctly converted into text is referred to as a “user utterance sentence”. As a result of recognizing user utterances in the speech recognition engine, a plurality of user utterance sentence candidates may be obtained, which are referred to as “recognition candidate sentences”. In addition, when the system synthesizes and speaks, it is called “system utterance”. The text that is the basis of composition is called a “system utterance”.

<Principle of the word selection device according to the present embodiment>
First, the principle of the word selection device according to this embodiment will be described. The word selection device according to the present embodiment analyzes a plurality of candidates for speech recognition results of speech input from a user, and extracts necessity / unnecessity of question answering and selection of selected crotch presented by question answering. Note that the word selection device according to the present embodiment uses a dialog strategy described in advance in order to perform a normal dialog.

  Next, terms used in the description of the word selection device according to the present embodiment will be described.

  “Question return” means that the system determines the user utterance sentence through a dialogue that presents the selected crotch in the system utterance and asks the user to select it in order to correctly set the user utterance sentence. For example, assume that there are two recognition candidate sentences as shown in (1) and (2) below.

(1) “I want to know Thai curry recipes”
(2) “I want to know the recipe of vegetable curry”

  The act of selecting a user utterance sentence from the above two recognition candidate sentences (1) and (2) leads to specifying the user's utterance intention. If an erroneous recognition candidate sentence is selected, the search result presented by the system is different from the user's intention, leading to a decrease in user satisfaction. Therefore, let us consider that the correct recognition candidate sentence is selected by the user through a question-and-answer dialogue.

  There are two types of question answering strategies: full text answering and word answering. The full text query return is a method in which a plurality of recognition candidate sentences are enumerated as they are. In this case, a system utterance sentence “Do you want to know the recipe for Thai curry or do you want to know the recipe for vegetable curry?” Is generated, synthesized, and uttered. The user speaks again while being aware of the difference, and the system determines the user utterance sentence in response. The full-text query return strategy is effective in situations where the user's utterances are entered so as not to make a mistake in each letter, and is expected to be used in a mail creation service by voice input.

  On the other hand, word query return is a user utterance sentence only by selecting a part of the user utterance sentence by extracting only the parts with different contents from the plurality of recognition candidate sentences and enumerating them. Is to decide. For example, in the case of the above (1) and (2), a system utterance sentence “Is it Thai curry or vegetable curry?” Is generated, and this is synthesized and uttered. The user can determine the user utterance by speaking the word presented as an option. The word query return strategy is effective for database searches and web searches for various domains, and is expected to be used for searching information using a search application such as a smartphone or signage, and for entering a destination in a car navigation system. In the word selection device according to the present embodiment, this word query is focused.

  In the word query return strategy, it is necessary to limit the word presented as the selected crotch to only words related to the user's utterance intention. For example, it is assumed that there are three recognition candidate sentences as shown in (3) to (5) below.

(3) “I want to know curry recipes”
(4) “I want to know the recipe of curry”
(5) “I want to hear curry recipes”

  Each of the recognition candidate sentences in (3) to (5) above is partially different in word, but the user's intention to search for a curry recipe is the same. In addition, regardless of the user utterance sentence, the system behavior of searching the recipe database with the keyword of the dish name of curry is the same.

  Therefore, it is not necessary to determine a user's utterance sentence without a word or phrase without fail for a service that is suitable for a word query return strategy, and it is a problem if the part related to the user's intention and system behavior can be recognized correctly. Absent. Since the behavior of a normal system is determined by the user's intention, if the user's intention can be correctly recognized, it is considered that there is no problem even if other parts are recognized incorrectly.

  However, when the candidate word sentences are returned using different words by comparing the recognition candidate sentences, in the case of the recognition candidate sentences (3) to (5) above,

"Is it""or" is "?"
"Do you want to know" or "Do you want to hear"?

  This is a question that is unnecessary for the user. In this case, the user may feel unsure what is being asked, and may feel uncomfortable with unnecessary questions. For this reason, when taking a word query return strategy, it is necessary to extract a portion related to the user's intention from each recognition candidate sentence and compare it to determine whether or not there is a query return.

  A “dialogue scenario” defines a strategy for dialogue. In the dialogue scenario, the service provider or the operator describes the behavior of the system according to the user utterance. This is prepared in advance according to the service, and the dialogue between the user and the system is controlled according to the dialogue scenario. Specifically, the dialogue scenario is a rule for performing a specific operation based on a character string, a word, morpheme information, or a time as a specific key. Note that the dialogue scenario used in the present embodiment prescribes a rule for performing a specific operation based on a specific character string or word.

  A “dialogue function” is used as a part for describing a dialogue scenario. A function that extracts information necessary for dialogue from a user's utterance sentence, or a function that selects and generates an appropriate next system response sentence or system action from the extracted result. Here, it is called a function, but it is a function that is called according to the dialogue scenario. If it is a function that takes a user utterance sentence or a recognition candidate sentence as input and performs processing to advance the dialogue, the implementation method is It doesn't matter. In addition, a character string, a word, morpheme information, time, or the like serving as a specific key used in a dialogue scenario can be acquired by the dialogue function. Note that the interactive function used in the present embodiment is a function for extracting a specific character string or word.

  In the present embodiment, the text analysis result is input together with the user utterance sentence or the recognition candidate sentence. However, the present invention is not limited to this. Further, not only the user utterance sentence or the recognition candidate sentence but also arbitrary information necessary for proceeding with the dialogue may be given as input. For example, sensor information, image information, past usage history, time information, other services, information obtained from other databases, and the like can be mentioned.

  In this embodiment, an example in which the output of the interactive function is an array of words is shown. This is to easily perform the operation of comparing a plurality of outputs, and the output format is not limited to this. An example of the interactive function is shown in FIG. The interactive function in the example of FIG. 1 is a part of the interactive function used when describing an interactive scenario in a service that allows recipe search and restaurant search.

  Using the word extracted by this interactive function as a search condition, it is combined with a function that generates a search expression for searching the recipe database for item number 1 and item number 2, and a restaurant database for item number 3. It is assumed that it is actually used in a dialogue scenario.

  Here, the process up to extraction of a word serving as a search condition is defined as one interactive function. However, the process up to obtaining a search result based on the extracted result may be defined as one interactive function. The dialogue function may be of any length or definition unit as long as some processing result is output for the input user utterance sentence or recognition candidate sentence. In addition, it is not necessary to create a dialog function specially in order to realize the function of returning a query, and it is assumed that a function that has a calling relationship is already reused in a dialog scenario.

  In the word selection device according to the present embodiment, by comparing the processing result with an interactive function that allows each recognition candidate sentence to obtain a processing result, whether or not to return a question to the user and options to be used for the question return are selected. decide.

<Configuration of Word Selection Device According to First Embodiment>
Next, the configuration of the word selection device according to the first embodiment will be described. As shown in FIG. 2, the word selection device 100 according to the first embodiment includes a CPU, a RAM, and a ROM that stores a program for executing a word selection processing routine described later and various data. Can be configured. This word selection device is functionally configured to include an input unit 10, a calculation unit 20, and an output unit 90 as shown in FIG.

  The input unit 10 accepts user voice data input from a microphone.

  The calculation unit 20 includes a dialogue function storage unit 22, a speech recognition unit 24, a reliability determination unit 26, a content analysis unit 28, a word candidate extraction unit 30, a question answering processing unit 32, and a speech synthesis unit 34. , Including.

The dialogue function storage unit 22 includes a dialogue function name consisting of dialogue function names n corresponding to each of the dialogue functions f _n determined in advance for the dialogue system targeted by the word selection device 100 according to the first embodiment. A set is stored. More specifically, a dialogue function name set of dialogue functions that have a calling relationship in the dialogue scenario that is the target in the first embodiment and includes recognition candidate sentences (to be described later) as input information is stored. Further, each of the dialogue functions f _n corresponding to the dialogue function name n is stored in the dialogue function storage unit 22. In the first embodiment, the interactive function f _n corresponding to the interactive function name can be called based on the interactive function name n.

  The voice recognition unit 24 performs voice recognition on the user's voice data received by the input unit 10, and the reliability of the voice data is assigned to the top N recognition candidate sentences (N-best). A set of recognition candidates to which the reliability is added is acquired as a recognition candidate group, and is output to the reliability determination unit 26. Here, the reliability indicates that the absolute value of the reliability value is meaningful, and the higher the value, the more reliable the recognition candidate. The reliability calculation method is calculated using, for example, the method disclosed in Patent Document 1. It should be noted that the analysis result in the speech recognition unit 24 is required to include the text data of the speech data, and the recognition candidate group has the highest reliability and is the user utterance sentence. It is necessary to include not only information related to one recognition candidate sentence that is highly likely, but also information on a plurality of recognition candidate sentences (N-best) from which the recognition candidate sentence is selected. The value of N is assumed to be predetermined. In addition, the data holding method for each recognition candidate is not limited.

  Further, the voice recognition unit 24 performs voice recognition on the user's voice data received by the input unit 10 in response to the output of the voice data of the question answering processing unit 32, and acquires a reaction user utterance sentence. , Output to the inquiry processing unit 32. Here, when a plurality of recognition candidates are obtained for the reaction user utterance sentence, a response user utterance sentence obtained by narrowing the recognition candidates from the word included in the reliability or the recognition candidate sentence is used.

  The reliability determination unit 26 narrows down the recognition candidate sentences included in the recognition candidate group input from the speech recognition unit 24 based on the reliability of the recognition candidate sentence to obtain a recognition candidate sentence group. In addition, the reliability determination unit 26 determines the method for returning the acquired recognition candidate sentence group based on the number of recognition candidate sentences (number of elements) included in the recognition candidate sentence group. Or set to “hold” and output to the content analysis unit 28 together with the recognition candidate sentence group. Here, “recurrent utterance” means that a system utterance for prompting the user to utter again is reproduced as in “Please ask again”, and the process returns to the process of the first voice recognition unit 24. Further, “determined” refers to a state in which a user uttered sentence can be determined without asking a question. “Pending” refers to a state in which a method for answering a question cannot be determined at this stage, but is determined in a process described later.

  Specifically, first, the reliability determination unit 26 calculates the reliability threshold stored in the memory (not shown) and the reliability of the recognition candidate for each of the recognition candidates included in the recognition candidate group. If the comparison result is equal to or greater than the threshold, the recognition candidate sentence included in the recognition candidate is added to the recognition candidate sentence group. Then, the reliability determination unit 26 sets a question answering method based on the number of recognition candidate sentences (number of elements) included in the acquired recognition candidate sentence group. When the number of recognition candidate sentences is 0, the question answering method is set to “recurrent speech”. When the number of recognition candidate sentences is 1, the question answering method is set to “confirmed”, and the recognition candidate sentences of the recognition candidate sentence group are set as user utterance sentences. When the number of recognition candidate sentences is larger than 1 (2 or more), the question answering method is set to “hold”.

  When the question answering method input from the reliability determination unit 26 is “pending”, the content analysis unit 28 determines the recognition candidate sentence group based on each dialogue function stored in the dialogue function storage unit 22. Sets the method for returning the recognition candidate sentence group.

  Specifically, first, the content analysis unit 28 performs text analysis for each of the recognition candidate sentences included in the recognition candidate sentence group. Here, the text analysis may be analysis related to a recognition candidate sentence and words included in it, such as language processing such as morphological analysis, dependency analysis, specific expression extraction, predicate normalization, and categorization of field words. .

Next, for each of the dialogue function names n stored in the dialogue function storage unit 22, the content analysis unit 28 changes the recognition candidate sentence t and the dialogue function f _n called based on the dialogue function name n, Input the text analysis result of the recognition candidate sentence. When an output result is obtained from the interactive function f _n , the output array is generated as a content array оut (t, n). 3, 4, and 5 show examples of the content array оut (t, n). FIG. 3 to FIG. 5 are examples for each recognition candidate sentence group when three different patterns of user utterances are input. The example of FIG. 3 shows an example in the case where there are three recognition candidate sentences, and shows the result when input to the interactive function shown in FIG. 4 and 5 show an example in which there are two recognition candidate sentences, and the same results as in FIG. 3 are shown.

  Next, the content analysis unit 28 sets a query return method based on the number of acquired content arrays оut (t, n). Here, if the number of the content array оut (t, n) is 0, the question answering method is set to “recurrent speech”. When the number of content arrays оut (t, n) is 1, the question answering method is set to “confirmed”, and the recognition candidate sentence corresponding to the element of the content array is set as the user utterance sentence. When the number of content arrays оut (t, n) is larger than 1 (2 or more), the inquiry method is set to “hold”, and each of the content arrays оut (t, n) The dialogue function name corresponding to each element of the content array оut (t, n) and each recognition candidate sentence are output to the word candidate extraction unit 30.

  When the question return method input from the content analysis unit 28 is “pending”, the word candidate extraction unit 30 and each of the input content array оut (t, n) and the content array оut (t, n) Based on the dialogue function name corresponding to each element and each of the recognition candidate sentences, a content table is created, and based on the content table, “repeated utterance”, “word question return”, or “confirmed” Any one of the answering methods is determined. Here, “word query return” is to determine a user utterance sentence by playing back a system utterance in which a part of user utterances is enumerated and asking the user to select it.

  Specifically, first, the word candidate extraction unit 30 sets each input content array оut (t, n) as one line, and recognizes the recognition candidate sentence t, the interactive function name n, and the content array оut (t, n). Create a table of contents that associates. The content table has the same number of rows as the number of content arrays оut (t, n). In each row, columns with no elements are blank. Examples of content tables are shown in FIG. 6, FIG. 7, and FIG. 6 corresponds to the content array оut (t, n) of FIG. 3, FIG. 7 corresponds to the content array оut (t, n) of FIG. 4, and FIG. 8 shows the content array оut (t, t of FIG. , n). The item numbers in the first column of FIGS. 6 to 8 are described for convenience of explanation, and the magnitude of the values does not make sense.

Next, the word candidate extraction unit 30 determines whether or not there are a plurality of types of dialogue function names n in the created content table. Here, when the word candidate extraction unit 30 determines that there are a plurality of types of dialogue function names n as in the example illustrated in FIG. 8, the question answering method is set to “recurrent speech”. On the other hand, when the word candidate extraction unit 30 determines that there are not a plurality of types of dialogue function names n as in the examples shown in FIGS. 6 and 7, the content array оut (t, n) If there is a line that completely matches the element of, delete one of the lines. That is, the rows in which the contents of all the elements in the content array оut (t, n) overlap are integrated. For example, in FIG. 6, the contents of the content array оut (t ₁ , Recipe_food_menu) of the row of item number 1 and all the elements of the content array оut (t ₃ , Recipe_food_menu) of the row of item number 3 are complete. Match. For this reason, either the item number 1 row or the item number 3 row is deleted. Either method can be deleted, and a method such as selecting the one with higher reliability corresponding to the recognition candidate sentence or selecting one existing on the content table can be considered. In the first embodiment, the case where the row of item number 3 is deleted will be described below.

Next, if there is a row in the content table that is in an inclusive relationship with the content of the content array оut (t, n) in the content table, the word candidate extraction unit 30 is a subset row (the one with the smaller number of elements). ) Is deleted. For example, referring to FIG. A specific word included in an element because the contents of the array of item number 4 content array оut (t ₄ , Recipe_food_menu) and the contents of the item array 5 content array оut (t ₅ , Recipe_food_menu) are inclusive relation Delete the line of item number 5 with a small number of.

Next, the word candidate extraction unit 30 determines whether or not there is a difference in value for each column corresponding to the element of the content array оut (t, n) in the content table. The word candidate extraction unit 30 counts the columns having different values among the columns corresponding to the elements of the content array iout (t, n), and sets them as c. For example, in the case of the example of FIG. 6, the element of the content array оut (t ₁ , Recipe_food_menu) of the remaining row of item number 1 is compared with the element of the content array оut (t ₂ , Recipe_food_menu) of the row of item number 2. Then, since only column 1 corresponding to one element of the content array оut (t, n) has a different value, c = 1. Here, when c = 0, the word candidate extraction unit 30 sets the question return method to “determined”, and sets the recognition candidate sentences of the elements remaining in the content table as user utterance sentences. In addition, when c = 1, the word candidate extraction unit 30 sets the question answering method to “word question answering”, generates a word candidate array composed of values in the different columns, and obtains the word candidate array and the content table. The data is output to the question return processing unit 32. In the example of FIG. 6, the word candidate sequence is (“vegetable curry”, “thai curry”). Further, the word candidate extraction unit 30 sets the question answering method to “recurrence” when c> 1 (c ≧ 2).

  The question answering unit 32 asks the user to answer questions as necessary based on the set question answering method type input from the reliability determination unit 26, the content analysis unit 28, or the word candidate extraction unit 30. A system utterance to be performed is performed, a user utterance sentence is determined, and output from the output unit 90. Note that the question answering method is always entered. When the question answering method is “Confirmed”, the user utterance sentence, and when the question answering method is “word answering”, the word candidate array and the contents table are asked. Input to the return processing unit 32.

  Specifically, when the input type of the set question answering method is “word answer”, the question answering unit 32 selects the input using the inputted word candidate array. Generate a system utterance sentence that prompts For example, if each word included in the word candidate sequence is connected by a punctuation mark and the sentence “Which is?” Is connected at the end, in the case of the example in FIG. 6, “Which is vegetable curry or Thai curry? Is created. The system utterance sentence may be any sentence as long as it prompts selection from the word group of the word candidate array. Also, when displaying on the screen instead of presenting it to the user by voice, the description is changed to match the interface.

  Next, the question answering processing unit 32 outputs the created system utterance sentence to the speech synthesis unit 34, and the speech data generated in the speech synthesis unit 34 is output from the output unit 90 and reproduced.

  Next, the inquiry processing unit 32 determines whether the reaction user utterance input from the speech recognition unit 24 includes only one element of the word candidate sequence. If the response processing unit 32 determines that the obtained reaction user utterance does not contain only one element of the word candidate sequence, the same voice data is output from the output unit 90 again. Repeat the playback process. Here, an upper limit value of the number of repetitions may be set in advance, and when the upper limit value is exceeded, the question answering method may be “recurrent speech”. On the other hand, when the inquiry processing unit 32 determines that the obtained reaction user utterance includes only one element of the word candidate sequence, the word candidate determined to be included Is obtained from the contents table and set as a user utterance sentence. For example, in the case of FIG. 6, when the user utters “vegetable”, the recognition candidate sentence “I want to know the vegetable curry recipe” in the row of item number 1 of the contents table is set as the user utterance sentence. The user utterance sentence is output from the output unit 90.

  In addition, the question answering processing unit 32 outputs a user utterance sentence from the output unit 90 when the inputted question answering method is “confirmed”.

  Further, when the input set question answering method is “recurrence”, the question answering processing unit 32 generates a system utterance sentence that prompts the recurrence utterance and outputs it to the speech synthesizer 34. Next, the voice data corresponding to the system utterance sentence prompting the recurrent utterance acquired from the voice synthesizer 34 is output from the output part 90 and reproduced, and null is set in the user utterance sentence.

  The speech synthesizer 34 has a function of generating and outputting synthesized speech data based on the given text and its features. In the first embodiment, text-format system utterances input from the question answering processing unit 32 are input, synthesized speech data is generated, and output to the question answering processing unit 32.

  The input may include parameters relating to speech attributes such as voice quality, intonation, accents, etc., or a single text sentence in which these parameters are embedded in the system utterance sentence. For example, the SSML format can be used. The format of the synthesized voice data is not limited. The speech synthesizer 34 may have any configuration as long as it receives information including system utterances and outputs synthesized speech data. In the first embodiment, the technique disclosed in Patent Document 3 (Japanese Patent Laid-Open No. 2012-237925) is used.

<Operation of the word selection device according to the first embodiment>
Next, the operation of the word selection device 100 according to the first embodiment will be described. When the voice data of the user input from the microphone is received in the input unit 10, the word selection processing routine shown in FIGS. 9 to 13 is executed by the word selection device 100.

  First, in step S <b> 100 of FIG. 9, the voice recognition unit 24 acquires a dialogue function name set stored in the dialogue function storage unit 22.

  Next, in step S102, the voice recognition unit 24 performs voice recognition on the voice data received by the input unit 10, and acquires N recognition candidate sentences.

  Next, in step S104, the speech recognition unit 24 calculates the reliability for each of the N recognition candidate sentences acquired in step S102.

  Next, in step S106, the speech recognition unit 24 adds the reliability of the recognition candidate sentence acquired in step S104 to each of the N recognition candidate sentences acquired in step S102, and acquires a recognition candidate group. .

  Next, in step S108, the reliability determination unit 26 acquires a recognition candidate sentence group from the recognition candidate group acquired in step S106 based on the reliability of each of the recognition candidate sentences included in the recognition candidate group, A question answering method is set based on the number of recognition candidate sentences included in the acquired recognition candidate sentence group.

  Next, in step S110, the content analysis unit 28 sets a question answering method based on the conversation function name set obtained in step S100, the recognition candidate sentence group and the question answering method obtained in step S106. .

  Next, in step S112, the word candidate extraction unit 30 determines a question answering method based on the question answering method acquired in step S110.

  Next, in step S114, the question return processing unit 32 executes the question return processing based on the question return method acquired in step S112, and ends the word selection processing.

  The question answering method determination process based on the reliability in step S108 will be described in detail with reference to FIG.

  In step S200 of FIG. 10, the reliability determination unit 26 acquires a reliability threshold stored in a memory (not shown).

  Next, in step S202, the reliability determination unit 26 determines recognition candidates to be processed from the recognition candidate group acquired in step S106.

  Next, in step S204, the reliability determination unit 26 determines whether or not the reliability of the recognition candidate to be processed is equal to or greater than the threshold acquired in step S200. When the reliability determination unit 26 determines that the reliability of the recognition candidate to be processed is greater than or equal to the threshold, the determination process for the question return method based on the reliability proceeds to step S206. On the other hand, when the reliability determination unit 26 determines that the reliability of the recognition candidate to be processed is less than the threshold value, the determination process of the question answering method based on the reliability proceeds to step S208.

  Next, in step S206, the reliability determination unit 26 adds the recognition candidate sentence of the recognition candidate to be processed to the recognition candidate sentence group.

  Next, in step S208, the reliability determination unit 26 determines whether or not the processing from step S202 to step S204 or step S206 has been completed for all of the recognition candidates included in the recognition candidate group acquired in step S106. To do. If the reliability determination unit 26 determines that the processing from step S202 to step S204 or step S206 has been completed for all of the recognition candidates included in the recognition candidate group acquired in step S106, a question based on the reliability is obtained. The return method determination processing moves to step S210. On the other hand, when the reliability determination unit 26 determines that the processing from step S202 to step S204 or step S206 has not been completed for all the recognition candidates included in the recognition candidate group acquired in step S106, the reliability is determined. The determination process of the question return method based on the degree proceeds to step S202, changes the recognition candidate to be processed, and repeats the processes from step S204 to step S208.

  Next, in step S210, the reliability determination unit 26 determines whether or not the number of recognition candidate sentences included in the recognition candidate sentence group acquired in step S206 is zero. When the reliability determination unit 26 determines that the number of recognition candidate sentences included in the recognition candidate sentence group is 0, the determination process of the question return method based on the reliability proceeds to step S212. On the other hand, when the reliability determination unit 26 determines that the number of recognition candidate sentences included in the recognition candidate sentence group is not 0, the determination process of the question return method based on the reliability proceeds to step S214.

  Next, in step S212, the reliability determination unit 26 sets the question answering method to “recurrent speech”, and ends the question answering method determination process based on the reliability.

  In step S214, the reliability determination unit 26 determines whether or not the number of recognition candidate sentences included in the recognition candidate sentence group acquired in step S206 is one. When the reliability determination unit 26 determines that the number of recognition candidate sentences included in the recognition candidate sentence group is 1, the determination process of the question return method based on the reliability proceeds to step S216. On the other hand, when the reliability determination unit 26 determines that the number of recognition candidate sentences included in the recognition candidate sentence group is not 1, the determination process of the question return method based on the reliability proceeds to step S220.

  Next, in step S216, the reliability determination unit 26 sets the inquiry method to “determined”.

  Next, in step S218, the reliability determination unit 26 sets the recognition candidate sentence included in the recognition candidate sentence group acquired in step S206 as a user uttered sentence, and ends the determination process of the question return method based on the reliability. To do.

  In step S220, the reliability determination unit 26 sets the question return method to “hold”, and ends the question return method determination process based on the reliability.

  The question answering method determination process based on the interactive function in step S110 will be described in detail with reference to FIG.

  In step S300 of FIG. 11, the content analysis unit 28 determines whether or not the question return method acquired in step S108 is “pending”. When the content analysis unit 28 determines that the question answering method is “pending”, the question answering method determination process based on the dialogue function proceeds to step S302. On the other hand, when the content analysis unit 28 determines that the question answering method is not “pending”, the question answering method determination process based on the interactive function is terminated.

  Next, in step S302, the content analysis unit 28 determines a recognition candidate sentence to be processed among the recognition candidate sentences included in the recognition candidate sentence group acquired in step S206.

  Next, in step S304, the content analysis unit 28 performs text analysis on the recognition candidate sentence to be processed.

  Next, in step S306, the content analysis unit 28 determines a dialogue function name to be processed from the dialogue function name set acquired in step S100.

Next, in step S307, the content analysis unit 28 adds the recognition candidate sentence t to be processed and the text analysis result acquired in step S304 to the interaction function f _n called based on the interaction function name n to be processed. Enter.

  Next, in step S308, the content analysis unit 28 determines whether an output result is obtained in step S307. When the content analysis unit 28 determines that an output result has been obtained, the determination process of the question answering method based on the interactive function proceeds to step S310. On the other hand, when the content analysis unit 28 determines that the output result has not been obtained, the determination process of the question return method based on the dialogue function proceeds to step S312.

  In step S310, the content analysis unit 28 generates the output array acquired in step S307 as a content array itut (t, n).

  In step S312, the content analysis unit 28 determines whether or not the processing from step S306 to step S308 or step S310 has been completed for all the dialog function names included in the dialog function name set acquired in step S100. When the content analysis unit 28 determines that the processes from step S306 to step S308 or step S310 have been completed for all the dialog function names included in the dialog function name set, the question return determination process based on the dialog function Proceeds to step S314. On the other hand, when the content analysis unit 28 determines that the processing from step S306 to step S308 or step S310 has not been completed for all the dialogue function names included in the dialogue function name set, the question based on the dialogue function is used. In the return determination process, the process proceeds to step S306, the dialogue function name to be processed is changed, and the process from step S307 to step S312 is repeated.

  Next, in step S314, the content analysis unit 28 determines whether or not the processing from step S302 to step S312 has been completed for all the recognition candidate sentences included in the recognition candidate sentence group acquired in step S206. When the content analysis unit 28 determines that the processing from step S302 to step S312 has been completed for all the recognition candidate sentences, the determination process of the question return method based on the interactive function proceeds to step S316. On the other hand, when the content analysis unit 28 determines that the processing from step S302 to step S312 has not been completed for all the recognition candidate sentences, the determination processing of the question return method based on the interactive function proceeds to step S302. The process proceeds to change the recognition candidate sentence to be processed, and repeats the processing from step S304 to step S314.

  Next, in step S316, the content analysis unit 28 determines whether or not the number of content arrays iout (t, n) acquired in step S310 is zero. When the content analysis unit 28 determines that the number of acquired content arrays iout (t, n) is 0, the determination process of the question return method based on the interactive function proceeds to step S318. On the other hand, when the content analysis unit 28 determines that the number of acquired content arrays оut (t, n) is not 0, the determination process of the question return method based on the interactive function proceeds to step S320.

  Next, in step S318, the content analysis unit 28 sets the question answering method to “recurrent speech”, and ends the question answering method determination process based on the interactive function.

  In step S320, the content analysis unit 28 determines whether or not the number of content arrays iout (t, n) acquired in step S310 is one. When the content analysis unit 28 determines that the number of acquired content arrays iout (t, n) is 1, the determination process of the question return method based on the interactive function proceeds to step S322. On the other hand, when the content analysis unit 28 determines that the number of acquired content arrays iout (t, n) is not 1, the determination process of the question return method based on the interactive function proceeds to step S326.

  Next, in step S322, the content analysis unit 28 sets the inquiry method to “determined”.

  Next, in step S324, the content analysis unit 28 sets a recognition candidate sentence corresponding to the only content array оut (t, n) acquired in step S310 as a user utterance sentence, and a question answering method based on an interactive function. This determination process ends.

  In step S326, the content analysis unit 28 sets the question answering method to “hold” and ends the question answering method determination process based on the interactive function.

  The question return method determination process in step S112 will be described in detail with reference to FIG.

  In step S400 of FIG. 12, the word candidate extraction unit 30 determines whether or not the question answering method acquired in step S108 or step S110 is “pending”. When the word candidate extraction unit 30 determines that the question answering method is “pending”, the question answering method determination process proceeds to step S402. On the other hand, when the word candidate extraction unit 30 determines that the question answering method is not “hold”, the question answering method determination process ends.

  Next, in step S402, the word candidate extraction unit 30 creates a content table based on the content array acquired in step S310.

  Next, in step S404, the word candidate extraction unit 30 determines whether there are a plurality of types of dialogue function names included in the content table acquired in step S402. If the word candidate extraction unit 30 determines that there are a plurality of types of dialogue function names included in the content table, the question answering method determination process proceeds to step S406. On the other hand, if the word candidate extraction unit 30 determines that there is only one interactive function name included in the content table, the question return method determination process proceeds to step S408.

  Next, in step S406, the word candidate extraction unit 30 sets the question answering method to “recurrent speech”, and ends the question answering method determination process.

  In step S408, the word candidate extraction unit 30 determines whether or not there is a row in the content table acquired in step S402 that completely matches all the elements of the content array оut (t, n). If the word candidate extraction unit 30 determines that there is a row in the content table in which all the elements of the content array оut (t, n) are completely identical, the inquiry return method determination process proceeds to step S410. On the other hand, if the word candidate extraction unit 30 determines that there is no line in the content table in which all the elements of the content array оut (t, n) completely match, the question return method determination process proceeds to step S412. To do.

  Next, in step S410, the word candidate extraction unit 30 performs integration so as to leave only one completely matching line acquired in step S408, and deletes each of the other completely matching lines.

  Next, in step S412, the word candidate extraction unit 30 has an inclusion relationship in the set of content arrays оut (t, n) between the row pairs of the content table acquired in step S402 or step S410. It is determined whether or not. If the word candidate extraction unit 30 determines that there is an inclusion relationship in the set of the content array оut (t, n) between the row pairs of the content table, the question return method determination processing proceeds to step S414. Transition. On the other hand, if the word candidate extraction unit 30 determines that there is no inclusion relationship in the set of content arrays оut (t, n) between the row pairs of the content table, The process proceeds to step S416.

  Next, in step S414, the word candidate extraction unit 30 deletes a line having a small number of words included in the element from the line pairs having the inclusion relation acquired in step S412.

  Next, in step S416, the word candidate extraction unit 30 has a difference in value for each column corresponding to the element of the content array оut (t, n) in the content table acquired in step S402, step S410, or step S414. It is determined whether or not there is. When the word candidate extraction unit 30 determines that there is a difference in value for each column corresponding to the element of the content array оut (t, n) in the content table, the question answering method determination process proceeds to step S422. Transition. On the other hand, when the word candidate extraction unit 30 determines that there is no difference in value for each column corresponding to the element of the content array оut (t, n) in the content table, The process proceeds to S418.

  Next, in step S418, the word candidate extraction unit 30 sets the question answering method to “confirmed”.

  Next, in step S420, the word candidate extraction unit 30 sets the recognition candidate sentence of the element remaining in the contents table acquired in step S402, step S410, or step S414 as the user utterance sentence, The decision process is terminated.

  In step S422, the word candidate extraction unit 30 counts the number of columns c having the difference acquired in step S416.

  Next, in step S424, the word candidate extraction unit 30 determines whether or not the number of columns c acquired in step S422 is 1. If the word candidate extraction unit 30 determines that the number of columns c is 1, the determination process of the question answering method proceeds to step S426. On the other hand, if the word candidate extraction unit 30 determines that the number of columns c is not 1, the question answering method determination process proceeds to step S406.

  Next, in step S426, the word candidate extraction unit 30 sets the question answering method to “word answering”.

  Next, in step S428, the word candidate extraction unit 30 generates a word candidate array composed of the column values with differences acquired in step S416, and ends the question return method determination process.

  The inquiry execution process in step S114 will be described in detail with reference to FIG.

  In step S500 of FIG. 13, the question return processing unit 32 determines whether or not the question return method acquired in step S108, step S110, or step S112 is “word question return”. When the question answering processing unit 32 determines that the acquired question answering method is “word answering”, the question answering execution process proceeds to step S502. On the other hand, if the question answering processing unit 32 determines that the acquired question answering method is not “word answering”, the question answering execution process proceeds to step S518.

  Next, in step S <b> 502, the question return processing unit 32 generates a system utterance sentence that prompts selection based on the word candidate sequence acquired in step S <b> 428, and outputs it to the speech synthesis unit 34.

  Next, in step S <b> 504, the speech synthesizer 34 generates speech data corresponding to the system utterance sentence acquired in step S <b> 502 and outputs the speech data to the question answering processor 32.

  Next, in step S506, the inquiry processing unit 32 outputs the audio data acquired in step S504 from the output unit 90 and reproduces it.

  Next, in step S508, the voice recognition unit 24 determines whether or not the user's utterance has been accepted. If the speech recognition unit 24 determines that the user's utterance has been received, the query execution process proceeds to step S510. On the other hand, when the voice recognition unit 24 determines that the user's utterance has not been accepted, the inquiry execution process repeats step S508.

  Next, in step S510, the voice recognition unit 24 performs voice recognition on the voice data of the user acquired in step S508, and acquires a reaction user utterance.

  Next, in step S512, the inquiry processing unit 32 includes only one element (word candidate) of the word candidate array acquired in step S428 in the reaction user utterance sentence acquired in step S508. Determine whether or not. When the question-reply processing unit 32 includes only one element of the word candidate sequence in the reaction user utterance sentence, the question-reply execution process proceeds to step S514. On the other hand, if the response user utterance sentence 32 does not include any element of the word candidate array or includes two or more elements of the word candidate array, The return execution process proceeds to step S506, and the process is repeated.

  Next, in step S514, the inquiry processing unit 32 corresponds to the word candidate included in the reaction user utterance sentence acquired in step S512 from the content table acquired in step S402, step S410, or step S414. A recognition candidate sentence to be acquired is acquired, and the recognition candidate sentence is set as a user utterance sentence.

  Next, in step S516, the question return processing unit 32 outputs the user utterance sentence acquired in step S514, step S420, step S324, or step S218 from the output unit 90, and ends the question return execution process.

  In step S518, the inquiry return processing unit 32 determines whether or not the inquiry return method acquired in step S108, step S110, or step S112 is “confirmed”. If the question answering processing unit 32 determines that the acquired question answering method is “confirmed”, the question answering execution process proceeds to step S516. On the other hand, when the question answering processing unit 32 determines that the acquired question answering method is not “confirmed”, the question answering execution process proceeds to step S519.

  Next, in step S 519, the question return processing unit 32 generates a system utterance sentence that prompts a recurrent utterance and outputs it to the speech synthesizer 34.

  Next, in step S520, the speech synthesizer 34 generates speech data corresponding to the system utterance sentence acquired in step S519, and outputs the speech data to the question return processing unit 32.

  Next, in step S522, the inquiry processing unit 32 outputs the audio data acquired in step S520 from the output unit 90 and reproduces it.

  Next, in step S524, the question return processing unit 32 sets null to the user utterance sentence.

  Next, in step S526, the voice recognition unit 24 determines whether or not the user's utterance has been accepted. When the voice recognition unit 24 determines that the user's utterance has been received, the inquiry execution process proceeds to step S102 in FIG. On the other hand, when the voice recognition unit 24 determines that the user's utterance has not been accepted, the question return execution process repeats step S526.

  As described above, according to the word selection device according to the first embodiment, the dialogue scenario is predetermined for each of a plurality of recognition candidate sentences that are the speech recognition results for the input user's speech data. , By generating a content array based on at least one interactive function, and by selecting a word to query the user based on at least one specific string stored in the generated content array It is possible to appropriately select a question-returning word for specifying the utterance content.

  In addition, the contents of user utterances can be narrowed down without using external information.

  In addition, instead of thinking about a strategy specialized in answering questions, by diverting a normal dialogue strategy, it is possible to reduce the cost of separately creating a dialogue strategy only for answering questions.

  In addition, by determining the content using the dialogue strategy, it is possible to determine whether or not a query is necessary based not only on the surface layer of the recognition candidate sentence but also on the content, thereby reducing unnecessary questions.

  Moreover, by diverting the dialogue strategy, it is possible to identify which part of the content necessary for proceeding with the dialogue is ambiguous, so that it is possible to present the selected crotch to the user and answer the question.

  In addition, by having the above-described content analysis unit and word candidate extraction unit, unnecessary question answering can be performed by judging the content of a plurality of recognition candidate sentences using the processing result of the dialogue function that realizes a normal dialogue strategy. A user's utterance sentence can be determined without performing. In addition, it is possible to extract a selected crotch of a word to be presented to the user when a question needs to be returned, and to secure a user utterance sentence by a selection act.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, in the first embodiment, the case of using three types of “word query”, “determined”, and “recurrent speech” has been described as the method of query return, but the present invention is not limited to this. The return method is not limited to any of the three types of “word inquiry”, “determined”, and “recurrent speech”, but may be used in combination with another type and a function for discriminating it.

  Further, in the first embodiment, the case has been described in which the content analysis unit 28 uses all the dialogue functions stored in the dialogue function storage unit 22. However, the present invention is not limited to this. For example, when the priority is set for each of the dialogue functions stored in the dialogue function storage unit 22, the dialogue functions are used in the priority order for the target bilingual candidate sentence, and output first. It is also possible to hold only the result of the interactive function obtained as described above and proceed to the processing of the next candidate translation sentence. In this way, a result reflecting the priority can be obtained. Also, if there is a conversation that is processed preferentially in the conversation scenario, the conversation function name is arranged before the name of the conversation function corresponding to the process, and the conversation function with the lower priority is placed behind. Content analysis that reflects In addition, when priority is not set for the interactive function, it is possible to add a process for holding the results of all the interactive functions for which output is obtained and comparing the results in another process. For example, the content array that is the output result of the interactive function may be held as a result only for the content array that includes the largest number of elements.

  Further, in the first embodiment, the case has been described where the reliability uses a reliability that is meaningful for the absolute value of the value, but is not limited thereto. For example, although the absolute value of the value has no meaning, a reliability having different properties such as a difference or a ratio having a significance in comparing the degree of reliability between recognition candidate sentences may be used. In this case, in the above-described reliability determination unit 26, it is necessary to partially change the processing of the reliability determination unit 26 that requires a comparison method change, such as comparing the reliability of each factor candidate. Note that the processing method is not limited as long as a reliable recognition candidate can be selected by some method.

  In the first embodiment, the case where the recognition candidate is a combination of the recognition candidate sentence and the reliability of the recognition candidate sentence has been described. However, the present invention is not limited to this. For example, a recognition candidate may accompany a recognition candidate sentence and information such as text reading kana information and part-of-speech information.

  In the first embodiment, the speech recognition unit may have any configuration as long as it has a function of inputting speech data and outputting a recognition candidate group. For example, you may use methods, such as patent document 1 or patent document 2 (Unexamined-Japanese-Patent No. 2012-032538).

  In the first embodiment, the case where the fixed value stored in the memory is used as the reliability threshold value has been described. However, the present invention is not limited to this. For example, it may be specified each time it is called or may be changed according to the surrounding environment where the service is used, the user, and the like.

  In the first embodiment, the case where the content table is realized in the form of a table has been described. However, the present invention is not limited to this. For example, any method that can compare the output results of interactive functions in the respective recognition candidate sentences may be used.

  Further, in the first embodiment, an example of a service that assumes an answering question by voice has been described, but the present invention is not limited to this. For example, when displaying the contents of an inquiry on a signage or a smartphone, a system response sentence or an interface for prompting selection may be created and realized as a function of displaying on a screen.

  In the first embodiment, it is assumed that the system responds by voice, but the present invention is not limited to this. For example, the system side may interact with the user using an interface other than voice, such as displaying text on the screen.

  In the first embodiment, the case where each element of the content array output based on the interactive function is a word has been described. However, the present invention is not limited to this. For example, each element of the content array output based on the interactive function may be a character string.

  Next, a word selection device according to the second embodiment will be described.

  The second embodiment is different from the first embodiment in that the reliability is not used. In addition, about the structure and effect | action similar to the word selection apparatus concerning 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

<Configuration of Word Selection Device According to Second Embodiment>
Next, the configuration of the word selection device according to the second embodiment will be described. As shown in FIG. 14, a word selection device 200 according to the second embodiment includes a CPU, a RAM, and a ROM that stores a program for executing a word selection processing routine described later and various data. Can be configured. This word selection device is functionally configured to include an input unit 10, a calculation unit 220, and an output unit 90 as shown in FIG.

  The calculation unit 220 includes a dialogue function storage unit 22, a voice recognition unit 224, a content analysis unit 28, a word candidate extraction unit 30, a question answering processing unit 32, and a voice synthesis unit 34. Yes.

  The speech recognition unit 224 performs speech recognition on the user's speech data received by the input unit 10, acquires N recognition candidate sentences (N-best) as recognition candidate sentence groups for the speech data, and analyzes the contents. To the unit 28. In addition, the voice recognition unit 224 sets the question answering method to “hold” and outputs it to the content analysis unit 28. Note that other processes in the speech recognition unit 224 are the same as those performed in the speech recognition unit 24 in the word selection device according to the first embodiment, and thus description thereof is omitted.

<Operation of the word selection device according to the second embodiment>
Next, the operation of the word selection device 200 according to the second embodiment will be described. When the user's voice data input from the microphone is received by the input unit 10, the word selection processing routine shown in FIG.

  First, in step S600 of FIG. 15, the speech recognition unit 224 sets the N recognition candidate sentences acquired in step S102 as a recognition candidate sentence group.

  Next, in step S602, the voice recognition unit 224 sets the question answering method to “hold”.

  As described above, according to the word selection device according to the second embodiment, the dialogue scenario is determined in advance for each of a plurality of recognition candidate sentences that are voice recognition results for the input user's voice data. , By generating a content array based on at least one interactive function, and by selecting a word to query the user based on at least one specific string stored in the generated content array It is possible to appropriately select a question-returning word for specifying the utterance content.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do.

,
DESCRIPTION OF SYMBOLS 10 Input part 20 Operation part 22 Dialog function memory | storage part 24 Speech recognition part 26 Reliability determination part 28 Content analysis part 30 Word candidate extraction part 32 Processing part 34 Speech synthesis part 90 Output part 100 Word selection apparatus 200 Word selection apparatus 220 Operation part 224 Voice recognition unit

Claims (5)

The identification is predetermined for a dialogue scenario that is a rule for performing a specific operation for each specific character string for each of a plurality of recognition candidate sentences that are voice recognition results for the input voice data of the user. based on the plurality of interactive functions to extract the type of string, for each of the plurality of interactive functions, from the recognition candidate sentence of at least one of said specific type have been extracted by the interaction function string A content analysis unit for generating a content array storing
A word to be queried to the user is selected based on the at least one particular type of character string stored in the content array generated for each of the plurality of interactive functions by the content analysis unit. A word candidate extraction unit;
Including a word selection device. The identification is predetermined for a dialogue scenario that is a rule for performing a specific operation for each specific character string for each of a plurality of recognition candidate sentences that are voice recognition results for the input voice data of the user. A content analysis unit that generates a content array storing at least one specific character string extracted from the recognition candidate sentence by the interactive function, based on at least one interactive function for extracting the character string;
A word candidate extraction unit that selects a word to be replied to the user based on the at least one specific character string stored in the content array generated by the content analysis unit;
Only including,
The word candidate extraction unit includes a combination of the content sequence generated by the content analysis unit and the dialogue function name corresponding to the dialogue function from which the specific character string stored in the content sequence is extracted as one line. Create a table of contents to represent
When there is only one type of interactive function name in the created content table, when there is a row in which the content array matches in the created content table, the matching rows are integrated,
When there is a row in which the content array is in an inclusion relationship in the content table, the row having the smaller number of specific character strings stored in the content array is deleted from the rows in the inclusion relationship. ,
Of the columns corresponding to the elements of the content array in the content table, when there is only one column where the specific character string is different, the specific character string included in the difference column is A word selection device that selects each word as a question-returning word . A word selection method in a word selection device including a content analysis unit and a word candidate extraction unit,
The content analysis unit determines in advance a dialogue scenario that is a rule for performing a specific operation on each specific character string for each of a plurality of recognition candidate sentences that are voice recognition results for the input user's voice data. It is, on the basis of the plurality of interactive functions for extracting character strings of a particular type, for each of the plurality of interactive functions, at least one of said extracted by the conversation function from the recognition candidate sentence Create a content array that stores a particular type of string,
The word candidate extraction unit is configured to inform the user based on the at least one specific type of character string stored in the content array generated for each of the plurality of interactive functions by the content analysis unit. Select a word to ask
Word selection method. A word selection method in a word selection device including a content analysis unit and a word candidate extraction unit,
The content analysis unit determines in advance a dialogue scenario that is a rule for performing a specific operation on each specific character string for each of a plurality of recognition candidate sentences that are voice recognition results for the input user's voice data. Generating a content array storing at least one specific character string extracted from the recognition candidate sentence by the interactive function based on at least one interactive function for extracting the specific character string. And
The word candidate extraction unit selects a word to be queried to the user based on the at least one specific character string stored in the content array generated by the content analysis unit;
Word selection method ,
The word candidate extraction unit selects a combination of the content array generated by the content analysis unit and the dialog function name corresponding to the dialog function from which the specific character string stored in the content array is extracted. Create a table of contents that represents as one row,
When there is only one type of interactive function name in the created content table, when there is a row in which the content array matches in the created content table, the matching rows are integrated,
When there is a row in which the content array is in an inclusion relationship in the content table, the row having the smaller number of specific character strings stored in the content array is deleted from the rows in the inclusion relationship. ,
Of the columns corresponding to the elements of the content array in the content table, when there is only one column where the specific character string is different, the specific character string included in the difference column is A word selection method for selecting each as a word to be asked back.   The program for functioning a computer as each part of the word selection apparatus of Claim 1 or Claim 2.