Learning device, learning method, and learning program

Abstract

The present invention provides a learning device (100) comprising: an accepting unit (131) for acquiring language data depicted in a form enabling elements constituting a text to be substituted, and accepting designation of a method for expanding the language data and creating text; and a learning unit (132) for learning a model relating to a natural language, using the text created by expanding the language data on the basis of the designation accepted by the accepting unit.

Description

[Technical Field]

[0001] This disclosure relates to a learning device, learning method, and learning program for performing machine learning related to language processing. [Background technology]

[0002] With the widespread adoption of smartphones and smart speakers, dialogue systems that respond appropriately to user speech and actions are becoming increasingly common. These dialogue systems utilize machine learning models generated by analyzing speech and text data.

[0003] For example, a known method utilizes texts written in EBNF (Extended Backus Naur Form), an extension of the predefined grammatical rule BNF (Backus Naur Form), for model generation and phonological feature extraction in dialogue systems. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2003-255989 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] Conventional technology allows for the efficient collection and use of text data, such as sentences, for training dialogue systems.

[0006] However, when new learning is performed to use a natural language model learned by a method such as RNN (Recurrent Neural Network) or Transformer for a specific application such as a discriminator, the data described in BNF or the like cannot be used as learning data as it is. That is, since the data described in BNF or the like is described according to specific rules, it is difficult to use it as learning data for machine learning. On the other hand, when the data described in BNF or the like is expanded to create individual sentences, the number of sentences becomes enormous, and a problem occurs in that the learning takes too much time.

[0007] Therefore, the present disclosure proposes a learning device, a learning method, and a learning program that can efficiently learn using data described by a predetermined grammar rule.

Means for Solving the Problem

[0008] In order to solve the above problems, a learning device according to one aspect of the present disclosure includes a reception unit that receives language data described in a replaceable form of elements constituting a sentence and receives a specification of a method for expanding the language data to create a sentence, and a learning unit that learns a model related to a natural language using the sentence created by expanding the language data based on the specification received by the reception unit.

Brief Description of the Drawings

[0009] [Figure 1] It is a diagram showing an overview of information processing according to the first embodiment. [Figure 2] It is a diagram (1) for explaining expansion processing according to the first embodiment. [Figure 3] It is a diagram showing a configuration example of a learning device according to the first embodiment. [Figure 4] It is a diagram showing an example of a language data storage unit according to the first embodiment. [Figure 5] It is a diagram showing an example of a replacement rule storage unit according to the first embodiment. [Figure 6]Figure (2) illustrates the unfolding process according to the first embodiment. [Figure 7] This is a flowchart (1) showing the processing flow according to the first embodiment. [Figure 8] This is a flowchart (2) showing the processing flow according to the first embodiment. [Figure 9] This figure shows an overview of the information processing according to the second embodiment. [Figure 10] Figure (1) illustrates the unfolding process according to the second embodiment. [Figure 11] Figure (2) illustrates the unfolding process according to the second embodiment. [Figure 12] This is a flowchart showing the processing flow according to the second embodiment. [Figure 13] This figure shows an example of information processing related to a modified version of the second embodiment. [Figure 14] This is a hardware configuration diagram showing an example of a computer that implements the functions of a learning device. [Modes for carrying out the invention]

[0010] The embodiments will be described in detail below with reference to the drawings. In the following embodiments, the same parts will be denoted by the same reference numerals to avoid redundant explanations.

[0011] This disclosure will be explained in the order of the items shown below. 1. First Embodiment 1-1. Overview of Information Processing According to the First Embodiment 1-2. Configuration of the learning device according to the first embodiment 1-3. Procedure for processing according to the first embodiment 1-4. Modified examples according to the first embodiment 2. Second Embodiment 2-1. Overview of Information Processing According to the Second Embodiment 2-2. Configuration of the learning device according to the second embodiment 2-3. Procedure for processing according to the second embodiment 2-4. Modified Examples According to the Second Embodiment 3. Other Embodiments 4. Effects of the learning device related to this disclosure 5. Hardware Configuration

[0012] (1. First Embodiment) (1-1. Overview of information processing according to the first embodiment) An example of information processing according to the first embodiment of this disclosure will be explained using Figure 1. Figure 1 is a diagram showing an overview of the information processing according to the first embodiment. The information processing according to the first embodiment is performed by the learning system 1 shown in Figure 1.

[0013] As shown in Figure 1, the learning system 1 includes a learning device 100, a programmer terminal 10, and a user terminal 75. The learning device 100 is an example of an information processing device that performs the learning process related to this disclosure, and is, for example, a server. The programmer terminal 10 is an information processing device used by a programmer 20 who performs settings related to machine learning, and is, for example, a personal computer. The user terminal 75 is an information processing device used by a user who uses the speech recognition device 50 provided by the learning device 100, and is, for example, a smartphone or a tablet terminal.

[0014] Note that each device in Figure 1 conceptually represents a function in the learning system 1, and can take various forms depending on the embodiment. For example, the setting terminal 10 may be one or more terminal devices. Also, the setting terminal 10 and the user terminal 75 may be the same terminal.

[0015] In the example shown in Figure 1, the learning device 100 has a speech classifier 50, which is a natural language model pre-trained using methods such as RNN or Transformer. The programmer 20 configures the speech classifier 50 to update it using new training data in order to use it for a specific purpose. For example, the programmer 20 trains the speech classifier 50 to be a classifier that can identify whether or not an utterance made by user 70 indicates some kind of intention. Specifically, the programmer 20 configures the machine learning model so that when user 70 makes an utterance related to "locking," the speech classifier 50 outputs a result determining whether the utterance made by user 70, as input to the speech classifier 50, includes the intention to "check the lock" or does not include that intention.

[0016] When performing such learning, the learning device 100 requires a sufficient number of utterance examples to improve the accuracy of utterance recognition. Furthermore, when performing new learning using a model pre-trained with methods such as RNN or Transformer, the data used for that learning must be in the same format as that used for pre-training. Here, as a method for describing utterance content, i.e., sentences, there is a method such as BNF (Backus Naur Form) that describes the elements constituting the sentence in a replaceable format. With such a method, the elements constituting the sentence can be expanded into multiple parts, so a large number of sentences can be created from a small amount of information. For this reason, if utterance examples described with BNF or the like can be used for learning, the programmer 20 can proceed with the learning process efficiently. However, if the format is described according to specific rules such as BNF, new learning cannot be performed using a model pre-trained with ordinary sentences. On the other hand, if data described with BNF or the like is expanded to create individual sentences, the number of sentences becomes enormous, resulting in the problem of learning taking too long. In other words, there is a challenge in the learning process of natural language models to efficiently learn using data described with predetermined grammatical rules.

[0017] Therefore, the learning device 100 enables efficient model training using data described with predetermined grammatical rules such as BNF by executing the learning process related to this disclosure. Below, the outline of the learning process related to this disclosure will be explained in flow order using Figure 1, which shows the learning process related to this disclosure in a schematic block diagram.

[0018] As shown in Figure 1, the programmer 20 configures the learning process on the learning device 100 via the programmer terminal 10 (step S1). At this time, the programmer 20 uses the utterance examples 30 that the learning device 100 has previously stored for learning. In the example in Figure 1, the utterance examples 30 are data described using predetermined grammatical rules such as BNF.

[0019] The learning device 100 acquires an example utterance 30 according to the instructions of the setter 20 (step S2). The learning device 100 also acquires a BNF rule 40, which is a grammatical rule applied to the example utterance 30 (step S3). Then, the learning device 100 expands on the example utterance 30 according to the BNF rule 40 and generates a sentence that can be used for learning (step S4).

[0020] Here, we will explain the expansion of data described in BNF using Figure 2. Figure 2 is a diagram (1) for explaining the expansion process according to the first embodiment. For example, as shown in data example 80, suppose an example of utterance example 30 is the data "{HOUSE}'s {ROOM}, lock check" or "{HOUSE}'s {ROOM} lock confirmation". "{HOUSE}" is one of the elements that make up a sentence and has multiple options. The individual features of the options are called connections. Such expandable elements are called rules. That is, the data "{HOUSE}'s {ROOM}, lock check" has two rules: "{HOUSE}" and "{ROOM}".

[0021] "{HOUSE}" is an element that represents the concept of house, and features such as "my house, home, my place, my sister's house, Grandma's house, ..." constitute the options. Individual terms (features) that represent the concept of house, such as "my house," "home," and "Grandma's house," are connected. This shows that the element "{HOUSE}" can be expanded to the number of options, or in other words, the number of features defined as connections, according to the rules associated with the element.

[0022] Similarly, "{ROOM}" consists of options such as "kotatsu room, shared room, guest room, closet, bathroom, ..." and so on. "{ROOM}" can also be expanded as many times as there are features defined as connections.

[0023] For example, the data "{HOUSE}'s {ROOM}, lock check" can be expanded into sentences as many times as there are combinations of choices for each individual element, such as "Check the lock on the kotatsu room in my house" or "Check the lock on the closet in my house." In other words, the number of expansions for data described in BNF can be calculated based on the number of times the rule appears in the example sentences and the number of choices for each individual element.

[0024] To explain the rule again according to the general BNF notation, "A := ii|jj|kk" indicates that "there are 3 choices derived from rule A (ii, jj, kk)", meaning that the number of expansions of rule A is 3. For example, in example sentence X, "Tomorrow is A", rule A appears once, so the total number of expansions of example sentence X is 3. Also, for example, in example sentence Y: "Tomorrow is A, A, and A", rule A appears 3 times, so the total number of expansions of example sentence Y is 3*3*3 (=3). 3 ) becomes 27.

[0025] Furthermore, when another rule appears in the concatenation derived by a certain rule, the number of expansions of the concatenation is multiplied by the total number of expansions of that other rule. For example, assume there is a rule "B := m A m | ii". In this case, the number of expansions of rule B is obtained by adding the number of expansions of rule A, which is 3, to the concatenation of independent options "ii", resulting in a number of expansions of 4. For example, in the case of "Example sentence Z: 'B is A'", since the number of expansions of rule A is 3 and the number of expansions of rule B is 4, the total number of expansions of example sentence Z is 3 * 4, which is 12.

[0026] That is, the total number of expansions "D" all " when expanding the language data described in BNF a " can be expressed as a polynomial using the number of expansions "D" b " of rule A, the number of expansions "D" all " of rule B, etc. For example, assume that example sentence X, example sentence Y, and example sentence Z are included in the language data. Then the total number of expansions is (the number of expansions of example sentence X) + (the number of expansions of example sentence Y) + (the number of expansions of example sentence Z), and the polynomial "D" all = D a + D a 3 + D b * D a ". Note that "D" b " can be expressed as "D" a + 1".

[0027] As described above, since addition and multiplication occur based on each rule for the data described in BNF, an extremely large number of example sentences are generated when all are expanded. Generally, a large number of example sentences are required when generating a learning model. However, when newly learning in order to use an already generated model for some purpose, if the learning data is extremely large, the learning time will be too long and there is a risk of lacking practicality.

[0028] Therefore, in step S1, the learning device 100 receives a specification regarding expansion from the setter 20. As an example, the learning device 100 receives a specification of the number of learning data (number of samples) expected by the setter 20 ("D" expectedThe learning device 100 accepts the specification of the number of expansions for the language data.

[0029] The learning device 100 then calculates the number of expansions to be equal to the total number of expansions specified by the setter 20 by imposing certain constraints on the polynomial used in the expansion. Specifically, the learning device 100 sets an expansion coefficient "R" to limit the number of choices derived from each rule by a certain multiplier. a " and "R b We will introduce "[...]."

[0030] A polynomial whose total expansion is adjusted using expansion coefficients is given by equation (1) below.

[0031]

number

[0032] In equation (1), "D a " and "D b Since " is a constant, the learning device 100 will use the number of expansions "D" specified by the setter 20. expected The expansion coefficient "R" is used to make the value of "" as close as possible to the specified number. a " and "R b The only thing to decide is ". Note that in equation (1), when example sentence Z is expanded, rule A is also included in the expansion of rule B, so the expansion coefficient for rule A also affects the number of expansions of rule B. For this reason, when an expansion coefficient is introduced, "D b " is as shown in equation (1), "D a *R a This is represented as "+1".

[0033] When the learning device 100 receives a specification of the total number of expansions from the setter 20, the expansion coefficient "R" in equation (1) a " and "R b The learning device 100 determines "D" and performs the expansion to achieve the specified total number of expansions. expected If a specific number such as "2000" is specified for ", then "R" will be set so that the solution to equation (1) is 2000. a " and "Rb The specific value of " is determined. At this time, the learning device 100 determines "R" so that all rules are spread evenly according to the number of choices. a =R b It may also follow constraints such as "R a " and "R b If the problem cannot be solved analytically, an approximate value is calculated using methods such as gradient descent.

[0034] Returning to Figure 1, the process from step S4 onwards will be explained. The learning device 100 uses equation (1) to expand the data with each rule ultimately limited by the expansion coefficient, and uses the generated sentences to train the speech classifier 50 (step S5).

[0035] For example, the learning device 100 trains the speech classifier 50 to become a classifier that can identify whether an input utterance corresponds to the intention of "checking that the door is locked" (indicated by "○" in Graph 60) or does not correspond to that intention (indicated by "×" in Graph 60), as shown in Graph 60.

[0036] Subsequently, the learning device 100 receives the utterance from user 70 via the user terminal 75 (step S6). The learning device 100 inputs the utterance from user 70 into the learned utterance discriminator 50 and returns to user 70 the result of determining whether the utterance indicates a positive or negative emotion (step S7).

[0037] As described above, the learning device 100 acquires language data in a format in which the elements constituting a sentence can be replaced, such as a context-free grammar (CFG) like BNF, and receives a specification from the user 20 for a method of creating a sentence by expanding the language data (such as specifying the total number of expansions). Then, the learning device 100 uses the sentence created by expanding the language data based on the received specification to train the speech classifier 50, which is a model related to natural language.

[0038] This allows the learning device 100 to use data described by BNF or the like as training data in new training of a natural language model that has already been trained using other methods such as Transformer. Furthermore, by accepting specifications such as the total number of expanded sentences from the user 20, the learning device 100 can arbitrarily adjust the balance between training time and accuracy. Specifically, instead of training a vast amount of text with all the data expanded, the learning device 100 can use a number of sentences specified by the user 20 for training, so that the speech classifier 50 can be trained to the training time and accuracy desired by the user 20. In other words, the learning device 100 can efficiently train using data described by predetermined grammatical rules.

[0039] (1-2. Configuration of the learning device according to the first embodiment) Next, the configuration of the learning device 100 will be described. Figure 3 is a diagram showing an example of the configuration of the learning device 100 according to the first embodiment.

[0040] As shown in Figure 3, the learning device 100 includes a communication unit 110, a storage unit 120, and a control unit 130. The learning device 100 may also include an input unit (e.g., a keyboard or mouse) for receiving various operations from an administrator or other person managing the learning device 100, and a display unit (e.g., a liquid crystal display) for displaying various information.

[0041] The communication unit 110 is implemented, for example, by a NIC (Network Interface Card). The communication unit 110 is connected to the network N (for example, the Internet) by wire or wireless connection and transmits and receives information with the configurator terminal 10 and the user terminal 75 via the network N.

[0042] The memory unit 120 is implemented by, for example, semiconductor memory elements such as RAM (Random Access Memory) and flash memory, or by storage devices such as hard disks and optical discs. The memory unit 120 includes a language data storage unit 121, a substitution rule storage unit 122, and a learning model parameter storage unit 123. Each memory unit will be described in turn below.

[0043] As shown in Figure 4, the language data storage unit 121 stores language data used in the learning process. The language data corresponds to, for example, the utterance example 30 shown in Figure 1. Figure 4 is a diagram showing an example of the language data storage unit 121 according to the first embodiment. In the example shown in Figure 4, the language data storage unit 121 has items such as "sentence ID," "corresponding intent," and "sentence body." Note that in Figures 4 and 5, the information stored in the storage unit 120 is sometimes conceptually shown as "A01" and "G01," but in reality, each piece of information described later is stored in the storage unit 120.

[0044] The "Sentence ID" is identification information used to identify individual language data. The "Corresponding Intent" indicates which intent the sentence should be learned as an example sentence. The "Sentence Body" shows information about the natural language sentence itself, which may include replaceable elements (rule IDs).

[0045] Next, the substitution rule storage unit 122 will be described. Figure 5 shows an example of the substitution rule storage unit 122 according to the first embodiment. As shown in Figure 5, the substitution rule storage unit 122 stores the substitution content of the replaceable elements of the sentences that serve as training data. In the example shown in Figure 5, the substitution rule storage unit 122 has items such as "rule ID" and "derivative content". For example, the substitution rule storage unit 122 corresponds to the BNF rule 40 shown in Figure 1. Although the BNF rule is shown as an example in this embodiment, the substitution rule storage unit 122 may also store information about rules of other grammatical rules.

[0046] The "Rule ID" is identification information used to identify replaceable elements. The "Derivation Content" contains information about which options and connections the element can be expanded into.

[0047] Next, the learning model parameter storage unit 123 will be described. The learning model parameter storage unit 123 stores weight information of the learning results according to the model used in the classifier. The learning device 100 updates these weights according to the model calculations during learning. Specific examples of classifier models include classification using a Transformer alone, a combination of a Transformer and logistic regression, and a combination of weights from an RNN result and logistic regression.

[0048] Returning to Figure 3, the explanation continues. The control unit 130 is implemented by, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), etc., which executes a program stored inside the learning device 100 (for example, the learning program according to this disclosure) using RAM (Random Access Memory) or the like as a working area. The control unit 130 is also a controller and may be implemented by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

[0049] The reception unit 131 receives a specification for a method of creating a document by acquiring language data in which the elements constituting the document are described in a replaceable format and expanding the language data.

[0050] Language data that describes the elements constituting a text in a replaceable format is, for example, a speech example described in BNF. That is, the reception unit 131 accepts a specification of a method for creating a text by expanding the language data described in accordance with BNF. For example, the reception unit 131 obtains language data from the language data storage unit 121. Alternatively, the reception unit 131 may obtain language data tailored to a specific learning purpose from the user terminal 10.

[0051] The reception unit 131 accepts a specification of the number of sentences to be created by expanding language data, as an example of specifying a method for creating sentences by expanding language data. In other words, the reception unit 131 accepts the total number of data to be expanded that are described in BNF. Specifically, the total number of expansions is given by "D" as shown in equation (1). expected This is the value that will be substituted into ".

[0052] Furthermore, the reception unit 131 may accept specifications for expansion weights for each rule corresponding to an element included in the language data.

[0053] In the explanation of Figure 1, the learning device 100 is configured such that all rules in the example sentences are evenly distributed according to the number of choices, as shown in "R a =R b The expansion coefficient "R" is under conditions such as " a " and "R b An example of how to determine this was shown. However, depending on the example sentence, there may be a mix of elements that the programmer 20 wants to expand many times and elements that they want to expand few times. In such cases, the programmer 20 sets a high expansion coefficient for one element and a low expansion coefficient for other elements. In other words, the programmer 20 sets the weighting for the expansion of one element to be heavier and the weighting for the expansion of other elements to be lighter. Specifically, if the programmer 20 considers that "{HOUSE}" is not important for learning but "{ROOM}" is important in the data "{HOUSE} lock confirmation", then the programmer 20 sets the expansion coefficient for "{HOUSE}" to be small and the expansion coefficient for "{ROOM}" to be large.

[0054] The learning unit 132 learns a natural language model using sentences created by expanding language data based on the specifications received by the reception unit 131. The natural language model is, for example, an utterance classifier that identifies the intent of an utterance. In this disclosure, an utterance classifier is shown as an example of a natural language model, but the natural language model is not limited to an utterance classifier and may be a generator that produces a response to an utterance, etc.

[0055] The learning unit 132, for example, when it receives a specification for the number of sentences to be created by expanding the language data, trains the model using the sentences created by expanding the language data based on the specified number of sentences. Specifically, the learning unit 132 uses the "D" shown in equation (1). expected When a value to be substituted into " is specified, the expansion coefficient "R" of equation (1) will be used to approximate that value. a " and "R b The learning unit 132 then determines the expansion coefficient and expands the data using the expansion number obtained by multiplying it by the determined expansion coefficient, and obtains sentences to be used as training data. Subsequently, the learning unit 132 uses the obtained sentences to train (update) the speech classifier 50.

[0056] Furthermore, when the learning unit 132 receives a specification of weights for expansion from the setter 20, it trains the model using sentences created by expanding language data based on those weights. Specifically, the learning unit 132 uses the expansion coefficient "R" determined according to the weights. a " and "R b The value is substituted into equation (1) to expand the language data. Then, the learning unit 132 uses the expanded text to train (update) the speech classifier 50. As a result, the learning unit 132 can obtain training data that aligns with the intentions desired by the programmer 20, and thus create a model that better suits the programmer 20's objectives.

[0057] Furthermore, the learning unit 132 may also expand the language data by assigning weights to each element for expansion according to the number of characters in the features that constitute the elements to be replaced included in the language data. The number of characters in an element is, for example, the average number of characters in the choices derived by the rule corresponding to the element. In the example in Figure 2, the average number of characters in the element "{HOUSE}" is determined by calculating the average number of characters in the choices such as "atashinchi (5 characters)" and "ie (2 characters)". For example, if the average number of characters in the element "{HOUSE}" is "4" and the average number of characters in the element "{ROOM}" is "6", the learning unit 132 will determine the expansion coefficients for each to be in a ratio of "4:6". The reason for calculating in this way is that it is assumed that words with longer character counts are more likely to have a greater impact on learning. In other words, the learning unit 132 can perform more accurate learning by adjusting the expansion weights according to the number of characters, thus enabling more efficient learning.

[0058] Alternatively, the learning unit 132 may classify each element contained in the language data based on the similarity of the features that constitute the element, and then extract features in order from the classified groups to create a text that expands the language data. In other words, the learning unit 132 prioritizes the expansion of each rule in a way that increases the variance as much as possible.

[0059] This point will be explained using Figure 6. Figure 6 is Figure (2) for illustrating the expansion process according to the first embodiment. Graph 150 in Figure 6 schematically shows the state in which the rule derivation options are classified (clustered) according to the expansion coefficient.

[0060] An example of a clustering method is shown below. First, the learning unit 132 calculates the features of the sequence of rule derivation options (i.e., words and phrases). For example, the learning unit 132 calculates a "Bag of chars" as a feature of the expanded result for the rule. Here, "Bag of chars" refers to a vector in which the frequency of occurrence of each character included in the features is one element. Note that if features for each word or phrase can be obtained, the learning unit 132 may calculate features using methods other than "Bag of chars," such as "CBOW" or "Skip-gram." Next, the learning unit 132 determines the actual number of rule expansions by multiplying the number of rule derivation options by the expansion coefficient, and performs k-means clustering with the actual number of expansions as k.

[0061] In the example in Figure 6, the learning unit 132 determines the number of clusters, or k, in k-means clustering to be "3". In this case, the learning unit 132 obtains clusters as shown in Graph 150. Specifically, as shown in Figure 6, the learning unit 132 obtains a first cluster with center point 152, a second cluster with center point 158, and a third cluster with center point 162. Note that while random numbers are usually used as the initial values ​​for k-means, in this case the results change each time features are calculated. Therefore, methods such as Var-Part, which sequentially divides the data into clusters using high-dimensional values ​​of variance to create the initial clusters, may be applied.

[0062] Then, when the learning unit 132 expands the rules, it selects unused connections one by one from the generated clusters and expands them as strings corresponding to the rules. Specifically, the learning unit 132 selects feature 154 from the first cluster, which is the connection closest to the center point 152 within the cluster, and expands it. Next, the learning unit 132 selects feature 160 from the second cluster, which is the connection closest to the center point 158 ​​within the cluster, and expands it. Next, the learning unit 132 selects feature 164 from the third cluster, which is the connection closest to the center point 162 within the cluster, and expands it. After that, the learning unit 132 selects feature 156 from the first cluster, which is the connection next closest to the center point 152 after feature 154, and expands it. In other words, the learning unit 132 sorts the expansion results within each cluster in order of proximity to the center point and adopts the expansion results in order of proximity to the center point. Furthermore, once the learning unit 132 selects features within the same cluster, it does not use the selected features again until all features within that cluster have been used in the expansion result. This allows the learning unit 132 to suppress the decrease in the variance of the features in the training data that occurs when similar features are concentrated in the expansion of the rules.

[0063] For example, in the element "{HOUSE}" shown in Figure 2, "uchicchi" and "uchi" are extremely similar, and even if they are expanded consecutively, there is no difference in the feature quantities of the generated example sentences, which may make them unsuitable as training data. Therefore, as shown in Figure 6, the learning unit 132 clusters the features of each rule according to their feature quantities (i.e., similarity) and expands them by selecting features sequentially from each cluster. As a result, the learning unit 132 can efficiently acquire example sentences composed of features with dispersed feature quantities, enabling highly accurate learning with relatively little training data.

[0064] The determination unit 133 uses the speech classifier learned by the learning unit 132 to determine the intent of the utterance. For example, the determination unit 133 inputs the user's utterance intent obtained via the user terminal 75 into the speech classifier and determines the intent of the utterance based on the output result. The determination unit 133 may determine various results, not limited to the intent of the utterance, depending on the speech classifier (model) learned by the learning unit 132. For example, the determination unit 133 may perform automatic speech recognition (ASR) processing or other natural language understanding (NLU) processing. For example, the determination unit 133 may identify important words or phrases in the utterance. For example, the determination unit 133 may decompose the acquired speech into morphemes via ASR or NLU, and determine what intent or attributes each morpheme possesses.

[0065] The transmitting unit 134 transmits various types of information. For example, the transmitting unit 134 transmits the result determined by the determination unit 133 to the user terminal 75. The transmitting unit 134 may also transmit the results of the deployment based on the specifications received from the setter 20, or learning results, etc., to the setter terminal 10.

[0066] (1-3. Procedure for processing according to the first embodiment) Next, the procedure for processing according to the first embodiment will be explained using Figures 7 and 8. First, the procedure for the learning process according to the first embodiment will be explained using Figure 7. Figure 7 is a flowchart (1) showing the flow of processing according to the first embodiment.

[0067] As shown in Figure 7, the learning device 100 acquires speech examples to be used for learning (step S101). Next, the learning device 100 determines whether or not it has received a specification for development from the setter 20 (step S102). If no specification has been received (step S102; No), the learning device 100 waits until a specification is received.

[0068] On the other hand, if the designation is accepted (step S102; Yes), the learning device 100 develops speech examples in accordance with the designation (step S103). For example, the learning device 100 develops speech examples based on the total number of examples specified by the setter 20.

[0069] Then, the learning device 100 learns a classifier that was previously generated using a method such as Transformer, using the expanded utterance examples (step S104).

[0070] Next, the procedure for the determination process according to the first embodiment will be explained using Figure 8. Figure 8 is a flowchart (2) showing the flow of the process according to the first embodiment.

[0071] The learning device 100 determines whether or not it has received a request from the user 70 for identification related to speech (step S201). If it has not received a request (step S201; No), the learning device 100 waits until it receives a request.

[0072] On the other hand, if the request is accepted (step S201; Yes), the learning device 100 inputs the text corresponding to the utterance obtained from the user 70 into the speech classifier 50 (step S202).

[0073] Next, the learning device 100 acquires an identification result based on the output from the speech classifier 50 and transmits the acquired result to the user 70 (step S203). If the text acquired from the user 70 is unidentifiable, the learning device 100 may send an error message to the user 70 indicating that the text is unidentifiable.

[0074] (1-4. Modified examples according to the first embodiment) The learning process according to the first embodiment described above may be modified in various ways. Modifications of the first embodiment are described below.

[0075] In the first embodiment, when the learning device 100 receives a specification of the total number of expansions from the user 20, an example is shown in which the expansion coefficient is set to adjust the number of expansions. However, the learning device 100 may expand the language data without setting the expansion coefficient.

[0076] For example, the learning device 100 may unpack all the language data and randomly extract a number of sentences from the unpacked sentences that is specified by the setter 20. With this method, the learning device 100 does not need to determine the unpacking coefficient, and therefore the processing speed can be improved.

[0077] (2. Second Embodiment) (2-1. Overview of information processing according to the second embodiment) Next, a second embodiment will be described. In the second embodiment, the learning device 100 provides a user interface to the setter 20 and accepts setting specifications through the user interface.

[0078] Figure 9 shows an overview of the processing according to the second embodiment. In the example shown in Figure 9, the learning device 100 provides a user interface 90 to the programmer 20 and displays the user interface 90 on the programmer's terminal 10.

[0079] The user 20 specifies the time required for the speech classifier 50 to learn (referred to as the "setting reflection time") or the accuracy of the speech classifier 50 in the user interface 90 (step S11).

[0080] For example, the user 20 can specify the setting reflection time or accuracy by moving a slider displayed on the user interface 90. For example, the user 20 can specify a longer setting reflection time or higher accuracy. Alternatively, the user 20 can specify a shorter setting reflection time, a longer reflection time, or lower accuracy.

[0081] After acquiring the utterance example 30 and BNF rules (steps S12 and S13), the learning device 100 determines the total number of expansions of the utterance example 30 according to the setting reflection time or target accuracy specified by the setter 20 (step S14). Specifically, if the setting reflection time is specified to be long or the accuracy to be high, the learning device 100 expands the utterance example 30 so that the total number of expansions is relatively large. Alternatively, if the setting reflection time is specified to be short or the accuracy to be low, the learning device 100 expands the utterance example 30 so that the total number of expansions is relatively small. The process of determining the number of expansions for each element from the total number of expansions is the same as in the first embodiment. That is, if the setting reflection time is specified to be long by the setter 20, the learning device 100 adjusts the number of expansions by determining that the value of the expansion coefficient is relatively large. Alternatively, if the setting reflection time is specified to be short by the setter 20, the learning device 100 adjusts the total number of expansions by determining that the value of the expansion coefficient is relatively small. The learning device 100 then trains the speech classifier 50 based on the sentences generated by expanding the utterance example 30. To determine the corresponding total expansion count from the target accuracy, relationship information between the total expansion count and accuracy, which has been previously evaluated using existing pairs of representative language data and evaluation data for other purposes, is used.

[0082] As described above, the learning device 100 according to the second embodiment receives deployment specifications from the programmer 20 via the user interface 90. This allows the programmer 20 to intuitively specify deployments without having to input a specific number for the total number of deployments, thus enabling learning to be performed with a simple procedure. The learning device 100 may also display, along with a slider, the estimated learning completion time when learning is performed based on the programmer 20's specifications in the user interface 90.

[0083] (2-2. Configuration of the learning device according to the second embodiment) The functional configuration of the learning device 100 according to the second embodiment is the same as that of the first embodiment, so it is not shown in the illustration. Hereinafter, only the processing configuration that differs from the first embodiment will be described.

[0084] In the control unit 130 according to the second embodiment, the reception unit 131 receives a specification of the learning time when learning is performed using sentences created by expanding language data. For example, the reception unit 131 receives a specification of the learning time from the user 20 via the user interface 90. The learning unit 132 then learns the speech classifier 50 using sentences created by expanding language data based on the specified learning time.

[0085] Furthermore, the reception unit 131 accepts the specification of the accuracy to be estimated in the speech classifier 50, which is learned by the learning unit 132. The learning unit 132 learns the speech classifier 50 using sentences created by expanding language data based on the specified accuracy.

[0086] Furthermore, after receiving a request for judgment from the user 20 to the speech classifier 50 trained by the learning unit 132 based on the specified accuracy, the receiving unit 131 may receive a specification from the user 20 in response to the judgment, indicating that the accuracy estimated in the model further trained by the learning unit 132 should be improved. In this case, the learning unit 132 further trains the speech classifier 50 using sentences created by expanding language data, based on the specification to improve accuracy.

[0087] In other words, the reception unit 131 first receives the accuracy specification from the user 20, then presents the speech classifier 50 learned with that specification to the user 20 on the user interface 90, and receives speech input from the user 20. Then, the reception unit 131 receives a response from the user 20 who has entered the speech, on the user interface 90, indicating that they would like to further improve the accuracy.

[0088] For example, the reception unit 131 displays a button such as "Continue learning" on the user interface 90 and accepts a response from the user 20 by having them press it. In this case, the reception unit 131 decides to use more sentences as training data, assuming that the user 20 desires further accuracy improvement. The learning unit 132 then performs more expansions compared to the number used in the previous expansion, and trains the speech classifier 50. For example, the learning unit 132 may pre-set the number of sentences to be expanded in stages, such as 100, 200, or 400, and sequentially increase the number of sentences used for training according to the response from the user 20. In this case, the learning unit 132 may also allow the user 20 to try identifying the utterance with the speech classifier 50 using the number of sentences expanded in the previous expansion, even during sequential learning. Furthermore, since the learning unit 132 takes time to train the speech classifier 50 in proportion to the number of utterance examples (sentences), the user interface 90 may display a time indicator such as "XX minutes remaining until the next accuracy improvement." This point will be explained using Figures 10 and 11.

[0089] Figure 10 is a diagram (1) illustrating the expansion process according to the second embodiment. In the example shown in Figure 10, the screen 91 of the user terminal 10 displays the message "Applying settings. HH:MM until accuracy is improved for the next 1000 samples," the sentence "I wonder if it will be sunny tomorrow," entered by the user 20 to test the speech classifier 50, and the judgment button 93.

[0090] In other words, the example screen 91 shown in Figure 10 indicates that the learning device 100 is performing learning using an additional 1000 sentences based on the instructions from the user 20. At this time, the user 20 can see the remaining learning time (HH displays hours, MM displays minutes) in the "HH:MM" section of the display 92. Furthermore, if the user 20 wants to test the accuracy of the speech classifier 50 before learning 1000 samples, they can determine the current accuracy of the speech classifier 50 by entering any sentence and pressing the judgment button 93.

[0091] Another example will be explained using Figure 11. Figure 11 is a diagram (2) illustrating the deployment process according to the second embodiment. In the example shown in Figure 11, the screen 94 of the user terminal 10 displays a message such as "Settings being applied. Target accuracy up to 90% HH:MM". That is, in the example of screen 94 shown in Figure 11, the learning device 100 is performing learning to improve the judgment accuracy of the speech classifier 50 to 90% based on the setting of the user 20. At this time, the user 20 can see the remaining learning time in "HH:MM" on screen 94, as in the example of Figure 10. Also, as in the example of Figure 10, if the user 20 wants to try the speech classifier 50 before the accuracy is improved, they can enter any sentence and press the judgment button to see the current accuracy of the speech classifier 50.

[0092] Thus, the learning device 100 according to the second embodiment receives information from the user 20 via the user interface 90 regarding the degree of accuracy of the speech classifier 50 and the number of sentences to be used for learning (i.e., the number of expansions). In other words, the learning device 100 allows the user 20 to specify in detail the accuracy of the speech classifier 50 and the time required for learning, thereby improving the usability of the learning process.

[0093] (2-3. Procedure for processing according to the second embodiment) Next, the procedure for the process according to the second embodiment will be explained using Figure 12. Figure 12 is a flowchart showing the flow of the process according to the second embodiment.

[0094] As shown in Figure 12, the learning device 100 acquires speech examples to be used for learning (step S301). Subsequently, the learning device 100 determines whether or not it has received a specification from the user 20 regarding the development, including accuracy and learning time (step S302). If no specification has been received (step S302; No), the learning device 100 waits until a specification is received.

[0095] On the other hand, if the designation is accepted (Step S302; Yes), the learning device 100 develops speech examples according to the designation (Step S303). Then, the learning device 100 trains the speech classifier using the developed speech examples (Step S304).

[0096] Next, the learning device 100 presents the learning results to the user 20 (step S305). For example, the learning device 100 presents the learning results to the user 20 by displaying the results on the user interface 90 or by having the speech classifier 50, which reflects the learning results, perform a judgment via screens 91 and 94.

[0097] Subsequently, the learning device 100 determines whether or not it has received a change in specifications from the setter 20 (step S306). For example, the learning device 100 determines whether or not it has received a new specification from the setter 20 to improve the accuracy of the speech classifier 50 or to increase the number of speech examples.

[0098] If the change in specifications is accepted (step S306; Yes), the learning device 100 returns to step S303 and develops speech examples according to the specifications. On the other hand, if the change in specifications is not accepted (step S306; No), the learning device 100 considers that the request of the setter 20 has been fulfilled and terminates the learning process.

[0099] (2-4. Modified example according to the second embodiment) In the second embodiment, an example was shown in which the learning device 100 receives specifications regarding accuracy and expansion number from the user 20. Furthermore, an example was shown in which the learning device 100 presents the user 20 with the results of learning using the specified expansion number. Here, the learning device 100 may also accept specifications from the user 20 after presenting the expansion results. This point will be explained using Figure 13.

[0100] Figure 13 shows an example of information processing according to a modified version of the second embodiment. The example shown in Figure 13 shows the learning device 100 displaying an example of a sentence created by expanding on an example of an utterance on screen 95.

[0101] In this example, the learning device 100 acquires language data described in BNF, then expands a predetermined amount of language data to create example sentences. The learning device 100 then displays the created example sentences in a manner that is recognizable to the user 20. For example, the learning device 100 creates an amount of example sentences that can be displayed on any user interface (such as a display window) on the screen of the user terminal 10, and displays the created example sentences.

[0102] In other words, the learning device 100 displays the results of expanding the language data on a user interface used by the programmer 20 who will be using the speech classifier 50, and then accepts a specification from the programmer 20. The learning device 100 then uses the sentences containing the expansion results displayed on the user interface to train the speech classifier 50. This allows the programmer 20 to know in advance how many example sentences will be used for training and what kind of example sentences will be used, thus enabling efficient training.

[0103] Furthermore, when the learning device 100 expands the text to fit the screen of the user terminal 10, if the display area changes due to screen size, orientation, or scaling, the number of items expanded may be dynamically changed to match the display area.

[0104] (3. Other Embodiments) The processes described in each embodiment above may be carried out in various other forms besides those described above.

[0105] In the embodiments described above, the user terminal 75 is shown as an example of a smartphone or the like. However, the user terminal 75 is not limited to a smartphone; it may also be a device that has the function of recognizing the user's speech and making predetermined responses, such as a smart speaker. Furthermore, the user terminal 75 may be a wearable device such as a watch or glasses. In addition, the user terminal 75 may be implemented by various smart devices that have information processing functions. For example, the user terminal 75 may be a smart home appliance such as a television, air conditioner, or refrigerator, a smart vehicle such as an automobile, a drone, or a household robot.

[0106] Furthermore, while the above embodiments show examples of creating natural language models by training them using methods such as RNNs and Transformers, the training methods are not limited to these, and various known methods may be used.

[0107] Furthermore, among the processes described in each of the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, specific names, and information including various data and parameters shown in the above document and drawings can be changed at will unless otherwise specified. For example, the various information shown in each figure is not limited to the information shown.

[0108] Furthermore, the components of each illustrated device are functionally conceptual and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions. For example, the reception unit 131 and the determination unit 133 may be integrated.

[0109] Furthermore, the embodiments and modifications described above can be combined as appropriate, provided that the processing content is not inconsistent.

[0110] Furthermore, the effects described herein are merely illustrative and not limiting; other effects may also occur.

[0111] (4. Effects of the learning device related to this disclosure) As described above, the learning device according to this disclosure (learning device 100 in the embodiment) comprises a reception unit (reception unit 131 in the embodiment) and a learning unit (learning unit 132 in the embodiment). The reception unit acquires language data (e.g., utterance examples in the embodiment) in which the elements constituting a text are replaced, and accepts a specification for a method of creating a text by expanding the language data. The learning unit learns a natural language model (e.g., utterance classifier 50 in the embodiment) using the text created by expanding the language data based on the specification received by the reception unit.

[0112] Thus, the learning device relating to this disclosure can use data described by other grammatical rules as training data in the new training of a natural language model created by any method. In other words, the learning device can use language data that is not normally available for training, and therefore can efficiently train natural language models.

[0113] The reception unit also accepts the number of sentences to be created by expanding the language data. The learning unit then uses the sentences created by expanding the language data based on the specified number of sentences to train the model.

[0114] This allows the learning device to use a specified number of sentences for learning, rather than having to learn from a vast amount of unpacked data, thus reducing the time required for learning.

[0115] Furthermore, the reception unit accepts specifications for the weighting of each element to be replaced in the language data. The learning unit trains the model using sentences created by expanding the language data based on the specified weightings for expansion.

[0116] As a result, the learning device uses data that has been adjusted, such as increasing the number of expansions for elements that the user wants to emphasize, to create a model that meets the user's requirements.

[0117] Furthermore, the learning unit assigns weights to each element for expansion based on the number of characters in the features that constitute the element to be replaced in the language data, and then expands the language data.

[0118] This allows the learning device to use data that focuses on elements with a large number of characters that are considered important to the text, thereby increasing the likelihood of creating a highly accurate model.

[0119] Furthermore, the learning unit classifies each element contained in the language data based on the similarity of the features that constitute that element, and then extracts features sequentially from the classified groups to create a text that expands upon the language data.

[0120] This allows the learning device to use data with distributed similarity among the unfolded sentences for training, thereby increasing the likelihood of creating a highly accurate model.

[0121] The reception department also accepts requests for the training time when using sentences created by expanding language data. The training department then trains the model using sentences created by expanding language data based on the specified training time.

[0122] This allows the learning device to complete learning processes, which typically require long processing times, within the time frame requested by the user, thereby improving the user experience.

[0123] Furthermore, the reception unit accepts the specification of the accuracy to be estimated in the model trained by the learning unit. The learning unit trains the model using sentences created by expanding language data based on the specified accuracy.

[0124] This allows the learning device to create a model with the accuracy required by the user.

[0125] Furthermore, the reception unit receives a request from the user for a judgment on the model trained by the learning unit based on the specified accuracy, and then receives a specification from the user in response to that judgment, which is to improve the accuracy estimated in the model further trained by the learning unit. Based on the specification to improve accuracy, the learning unit further trains the model using sentences created by expanding the language data.

[0126] This allows the learning device to terminate or continue the learning process in response to feedback from the user, thereby creating a model that meets the user's requirements.

[0127] Furthermore, the reception unit displays the results of expanding the language data on the user interface used by the user who will be using the model, and then accepts the user's specifications. The learning unit uses the text, including the expansion results displayed on the user interface by the reception unit, to train the model.

[0128] This allows the learning device to train using example sentences that the user has recognized in advance, enabling it to create a model that meets the user's needs.

[0129] Furthermore, the reception desk accepts requests for a method to create text by expanding language data written in accordance with the BNF (Backus-Naur form).

[0130] This allows the learning device to use language data described according to the rules of BNF for training, enabling it to efficiently acquire a relatively large number of speech examples and proceed with learning smoothly.

[0131] (5. Hardware Configuration) The information devices such as the learning device 100 according to each embodiment described above are realized by a computer 1000 having a configuration such as that shown in Figure 14. The learning device 100 according to this disclosure will be described below as an example. Figure 14 is a hardware configuration diagram showing an example of a computer 1000 that realizes the functions of the learning device 100. The computer 1000 has a CPU 1100, RAM 1200, ROM (Read Only Memory) 1300, HDD (Hard Disk Drive) 1400, a communication interface 1500, and an input / output interface 1600. The parts of the computer 1000 are connected by a bus 1050.

[0132] The CPU 1100 operates based on programs stored in the ROM 1300 or HDD 1400, and controls various parts. For example, the CPU 1100 loads the programs stored in the ROM 1300 or HDD 1400 into the RAM 1200 and executes processing corresponding to various programs.

[0133] ROM1300 stores boot programs such as the BIOS (Basic Input Output System) executed by CPU1100 when computer 1000 starts up, as well as programs that depend on the computer 1000's hardware.

[0134] HDD1400 is a computer-readable recording medium that non-temporarily stores programs executed by CPU1100 and data used by such programs. Specifically, HDD1400 is a recording medium that stores a learning program related to this disclosure, which is an example of program data 1450.

[0135] The communication interface 1500 is an interface for the computer 1000 to connect to an external network 1550 (e.g., the Internet). For example, the CPU 1100 can receive data from other devices or transmit data it has generated to other devices via the communication interface 1500.

[0136] The input / output interface 1600 is an interface for connecting the input / output device 1650 and the computer 1000. For example, the CPU 1100 receives data from input devices such as a keyboard or mouse via the input / output interface 1600. The CPU 1100 also transmits data to output devices such as a display, speaker, or printer via the input / output interface 1600. The input / output interface 1600 may also function as a media interface for reading programs recorded on a predetermined recording medium (media). Examples of media include optical recording media such as DVDs (Digital Versatile Discs) and PDs (Phase Change Rewritable Disks), magneto-optical recording media such as MOs (Magneto-Optical Disks), tape media, magnetic recording media, or semiconductor memory.

[0137] For example, when the computer 1000 functions as a learning device 100 according to the first embodiment, the CPU 1100 of the computer 1000 realizes functions such as the control unit 130 by executing a learning program loaded on the RAM 1200. The HDD 1400 stores the learning program according to this disclosure and data in the storage unit 120. The CPU 1100 reads and executes the program data 1450 from the HDD 1400, but as another example, these programs may be obtained from other devices via an external network 1550.

[0138] Furthermore, this technology can also be configured as follows. (1) A receiving unit that receives a specification for a method of creating a document by acquiring language data in which the elements constituting the document are described in a replaceable format, and expanding the said language data, A learning unit learns a natural language model using text created by expanding the language data based on the designation received by the reception unit, A learning device equipped with the following features. (2) The aforementioned reception unit is The system accepts a specification for the number of sentences to be created by expanding the aforementioned language data. The aforementioned learning unit, The model is trained using the sentences created by expanding the language data based on the specified number of sentences. The learning device described in (1) above. (3) The aforementioned reception unit is For each element to be replaced included in the aforementioned language data, the weighting for expansion is specified. The aforementioned learning unit, The model is trained using sentences created by expanding the language data based on the specified weightings for the expansion. The learning device described in (1) or (2) above. (4) The aforementioned learning unit, Based on the number of characters in the features that constitute the elements to be replaced in the aforementioned language data, a weight is assigned to each element for expansion, and the language data is expanded accordingly. A learning device as described in any of (1) to (3) above. (5) The aforementioned learning unit, For each element included in the aforementioned language data, a classification is performed based on the similarity of the features constituting that element, and by sequentially extracting features from the classified groups, a text is created by expanding the language data. A learning device as described in any of (1) to (4) above. (6) The aforementioned reception unit is The system accepts the specification of the learning time when performing learning using the text created by expanding the aforementioned language data. The aforementioned learning unit, The model is trained using sentences created by expanding the language data based on the specified learning time. A learning device as described in any of (1) to (5) above. (7) The aforementioned reception unit is The learning unit accepts a specification for the accuracy to be estimated in the model learned by the learning unit, The aforementioned learning unit, The model is trained using sentences created by expanding the language data based on the specified accuracy. A learning device as described in any of (1) to (6) above. (8) The aforementioned reception unit is After receiving a request from the user for a judgment on the model trained by the learning unit based on the specified accuracy, the learning unit receives a request from the user to improve the accuracy estimated in the model further trained by the learning unit as a response to the judgment. The aforementioned learning unit, Based on the designation to improve accuracy, the model is further trained using the sentences created by expanding the language data. The learning device described in (7) above. (9) The aforementioned reception unit is After displaying the results of expanding the aforementioned language data on the user interface used by the user utilizing the model, the system receives the aforementioned specification from the user. The aforementioned learning unit, The model is trained using the text containing the results of the expansion displayed on the user interface by the reception unit. A learning device as described in any of (1) to (8) above. (10) The aforementioned reception unit is The system accepts specifications for a method of creating text by expanding the language data, which is written in accordance with BNF (Backus-Naur form). A learning device as described in any of (1) to (9) above. (11) Computers We accept requests for a method to obtain language data in which the elements constituting a text are described in a replaceable format, and then expand that language data to create a text. The system uses the text created by expanding the language data based on the received specifications to train a natural language model. Learning methods. (12) Computers, A receiving unit that receives a specification for a method of creating a document by acquiring language data in which the elements constituting the document are described in a replaceable format, and expanding the said language data, A learning unit learns a natural language model using text created by expanding the language data based on the designation received by the reception unit, A learning program designed to function as such. [Explanation of Symbols]

[0139] 1. Learning System 10. User terminal 20 Setters 50 Speech Identifiers 70 users 75 User terminals 100 Learning Devices 110 Communications Department 120 Storage section 121 Language data storage unit 122 Replacement Rule Storage Unit 123 Learning Model Parameter Storage Unit 130 Control Unit 131 Reception Department 132 Learning Department 133 Judgment section 134 Transmitter

Claims

1. A receiving unit that receives a specification for a method of creating multiple sentences by expanding the language data, which is language data in which the elements constituting a sentence are described in a form that allows for substitution of the elements, and which includes features that are arbitrary language expressions that can be substituted for the elements, and stores the language data in a memory unit, and by substituting the elements with features, A learning unit newly trains a natural language model using the text created by expanding the language data by replacing one or more of the elements with the features, based on the designation received by the reception unit which is information instructing how to expand the language data into multiple sentences, as new training data for using the natural language model for a specific purpose. Equipped with, The aforementioned reception unit is The above specification accepts the total number of expansions of the above text. The aforementioned learning unit, For each element included in the aforementioned language data, a classification is performed based on the similarity of features that replace that element, and from the multiple classified groups, one feature is extracted from each group to create a text that expands the language data to satisfy the aforementioned specifications. Learning device.

2. A receiving unit that receives a specification for a method of creating multiple sentences by expanding the language data, which is language data in which the elements constituting a sentence are described in a form that allows for substitution of the elements, and which includes features that are arbitrary language expressions that can be substituted for the elements, and stores the language data in a memory unit, and by substituting the elements with features, A learning unit newly trains a natural language model using the text created by expanding the language data by replacing one or more of the elements with the features, based on the designation received by the reception unit which is information instructing how to expand the language data into multiple sentences, as new training data for using the natural language model for a specific purpose. Equipped with, The aforementioned reception unit is The system accepts the specification of the learning time when performing learning using the text created by expanding the aforementioned language data. The aforementioned learning unit, Based on the specified learning time, the number of expansions of the language data is adjusted, and the model is trained using the sentences created according to the adjusted number of expansions. Learning device.

3. A receiving unit that receives a specification for a method of creating multiple sentences by expanding the language data, which is language data in which the elements constituting a sentence are described in a form that allows for substitution of the elements, and which includes features that are arbitrary language expressions that can be substituted for the elements, and stores the language data in a memory unit, and by substituting the elements with features, A learning unit newly trains a natural language model using the text created by expanding the language data by replacing one or more of the elements with the features, based on the designation received by the reception unit which is information instructing how to expand the language data into multiple sentences, as new training data for using the natural language model for a specific purpose. Equipped with, The aforementioned reception unit is The above specification accepts the total number of expansions of the above text, and for each element to be replaced included in the language data, it accepts a weighting, which is a numerical value for adjusting the number of expansions of each element, and the weighting is according to the number of characters of the feature that replaces the element to be replaced included in the language data. The aforementioned learning unit, The model is trained using sentences created by expanding the language data based on the specified weights. Learning device.

4. A receiving unit that receives a specification for a method of creating multiple sentences by expanding the language data, which is language data in which the elements constituting a sentence are described in a form that allows for substitution of the elements, and which includes features that are arbitrary language expressions that can be substituted for the elements, and stores the language data in a memory unit, and by substituting the elements with features, A learning unit newly trains a natural language model using the text created by expanding the language data by replacing one or more of the elements with the features, based on the designation received by the reception unit which is information instructing how to expand the language data into multiple sentences, as new training data for using the natural language model for a specific purpose. Equipped with, The aforementioned reception unit is When the aforementioned specific application is a classifier that identifies the intent of an utterance, the learning unit accepts a specification for increasing or decreasing the accuracy of the classifier estimated in the model learned by the learning unit. The aforementioned learning unit, Based on the specified increase or decrease in the accuracy of the classifier, the total number of expansions of the language data is adjusted, and the model is trained using sentences created according to the adjusted number of expansions. Learning device.

5. The aforementioned reception unit is Based on the user's specification of the accuracy of the classifier (up or down), the learning unit accepts the user's request to perform identification on the model it has learned, and after presenting the result of the identification, it accepts the user's request to further improve the accuracy of the classifier. The aforementioned learning unit, Based on the response indicating an improvement in the accuracy of the classifier, the model is further trained using the sentences created by expanding the language data. The learning device according to claim 4.

6. The aforementioned language data is This is language data written according to the BNF (Backus-Naur form). A learning device according to any one of claims 1 to 5.

7. Computers The system accepts a specification for a method of creating multiple sentences by expanding the language data, which is language data in which the elements constituting a sentence are described in a form that allows for substitution, and which includes features that are arbitrary language expressions that can be substituted for those elements, and stores this language data in a memory unit, and by substituting those elements with features. Based on the aforementioned designation, which is information instructing how the received language data should be expanded into multiple sentences, the language data is expanded by replacing one or more of the aforementioned features to satisfy the designation, and the resulting sentences are used as new training data for a natural language model to be used for a specific purpose, and the model is newly trained. It is a learning method, and furthermore, The system accepts the specification of the learning time when performing learning using the text created by expanding the aforementioned language data. Based on the specified learning time, the number of expansions of the language data is adjusted, and the model is trained using the sentences created according to the adjusted number of expansions. Learning methods.

8. Computers, A receiving unit that receives a specification for a method of creating multiple sentences by expanding the language data, which is language data in which the elements constituting a sentence are described in a form that allows for substitution of the elements, and which includes features that are arbitrary language expressions that can be substituted for the elements, and stores the language data in a memory unit, and by substituting the elements with features, A learning unit newly trains a natural language model using the text created by expanding the language data by replacing one or more of the elements with the features, based on the designation received by the reception unit which is information instructing how to expand the language data into multiple sentences, as new training data for using the natural language model for a specific purpose. It is a learning program designed to function as such. The aforementioned reception unit is The system accepts the specification of the learning time when performing learning using the text created by expanding the aforementioned language data. The aforementioned learning unit, Based on the specified learning time, the number of expansions of the language data is adjusted, and the model is trained using the sentences created according to the adjusted number of expansions. Learning program.

Images