JP2018028872A - Learning device, method for learning, program parameter, and learning program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase the accuracy of a label provided to a sentence.SOLUTION: A learning device according to the present application includes: a first learning section for allowing a learning unit and a restoration unit of a first recurrent neural network to learn the relation between a speech generation and a response in a predetermined conversation; a first generation section for generating a second recurrent neural network, using at least one of the learning unit and the restoration unit having learned by the action of the first learning section; and a second learning section for allowing the second recurrent neural network to learn the relation between a sentence and a label provided to the sentence.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a learning device, a learning method, a program parameter, and a learning program.

  2. Description of the Related Art Conventionally, a technique for labeling information using a technique called RNN (Recurrent neural network), deep learning, or the like is known. As an example of such a technique, a technique is known in which a learning device learns to output a label indicating the content or type of a sentence according to the content of the input sentence, and gives a label to the sentence. .

“Learning phrase representations using RNN encoder-decoder for statistical machine translation.” K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio. In Proceedings of the 2014 Confer-ence on EMNLP, 2014.

Socher, R .; Perelygin, A .; Wu, J .; Chuang, J .; Manning, CD; Ng, A .; and Potts, C. 2013.Recursive deep models for semantic compositionality over a sentiment treebank.In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP 2013), 1631-1642. Association for Computational Linguistics.

  However, in the prior art, the learning device may not be able to give an appropriate label.

  For example, if the learning device learns to output a label indicating the content or type of a sentence according to the content of the input sentence, it learns the relationship between the sentence and the label to be assigned to the sentence. Will be made to learn. However, if the correct data including the sentence and the label to be attached to the sentence cannot be prepared sufficiently, the relationship between the sentence and the label cannot be properly learned, and the learning device assigns an appropriate label. You may not be able to do it.

  The present application has been made in view of the above, and an object thereof is to improve the accuracy of a label applied to a sentence.

  The learning device according to the present application performs learning by a first learning unit that causes a learning device and a restorer included in the first recurrent neural network to learn a relationship between an utterance and a response in a predetermined dialogue, and the first learning unit performs learning. The relationship between the first generation unit that generates the second recurrent neural network using at least one of the learned learner and the restorer, and the sentence and the label given to the sentence is expressed as the second recurrent. And a second learning unit that causes the neural network to learn.

  According to one aspect of the embodiment, it is possible to improve the accuracy of a label attached to a sentence.

FIG. 1 is a diagram illustrating an example of a learning process executed by the learning device according to the embodiment. FIG. 2 is a diagram illustrating a configuration example of the learning device according to the embodiment. FIG. 3 is a diagram illustrating an example of information registered in the dialogue data database according to the embodiment. FIG. 4 is a diagram illustrating an example of information registered in the learning data database according to the embodiment. FIG. 5 is a first diagram illustrating variations of the RNN generated by the learning device according to the embodiment. FIG. 6 is a second diagram illustrating variations of the RNN generated by the learning device according to the embodiment. FIG. 7 is a flowchart for explaining an example of the flow of the learning process according to the embodiment. FIG. 8 is a diagram illustrating an example of a hardware configuration.

Hereinafter, a learning apparatus, a learning method, a program parameter, and a mode for carrying out a learning program according to the present application (hereinafter referred to as “embodiment”) will be described in detail with reference to the drawings. In addition, the learning apparatus, the learning method, the program parameter, and the learning program according to the present application are not limited by this embodiment. In the following embodiments, the same portions are denoted by the same reference numerals, and redundant description is omitted.
[Embodiment]

[1-1. Example of learning device)
First, an example of learning processing executed by the learning device will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of a learning process executed by the learning device according to the embodiment. In FIG. 1, the learning device 10 is an information processing device that performs a learning process described below, and is realized by, for example, a server device or a cloud system.

  More specifically, the learning device 10 can communicate with an arbitrary device such as the terminal device 100 used by the user via a predetermined network N such as the Internet. In addition, the learning device 10 can communicate with an input device 200 (see, for example, FIG. 2) for acquiring a dialogue between arbitrary users.

  The terminal device 100 and the input device 200 are realized by an information processing device such as a smart device such as a smartphone or a tablet, a desktop PC (Personal Computer), a notebook PC, or the like. The learning device 10 is realized by an arbitrary information processing device such as a server device or a cloud system.

[1-2. About the learning process)
Here, the learning apparatus 10 learns a model for giving attribute information indicating attributes such as the type and content of the sentence to the input sentence. Such attribute information is called a label, tag, or metadata, and is used when searching or selecting a sentence. In the following description, attribute information is referred to as a label.

  For example, when the learning apparatus 10 inputs a plurality of words included in a sentence in the order in which each word appears in the sentence, the recurrent neural network learns the characteristics of each word and the characteristics in the order in which each word appears. A network (hereinafter referred to as “RNN (Recurrent Neural Network)”) is learned. Here, when a plurality of pieces of information are sequentially input, the RNN encodes the characteristics of each information and the characteristics of the input order. And a learning device (encoder) that outputs encoded features and a restoration device (decoder) that restores new information based on the features output from the learning device. Then, a process of restoring a plurality of pieces of information input to the learning device in the input order from the features output by the learning device is executed.

  For example, the learning device 10 acquires a set of a predetermined sentence and a label to be given to the sentence as learning data used for RNN learning. In such a case, the learning device 10 learns the RNN to output a label to be given to the sentence when a word included in the sentence is input to the RNN. For example, the learning device 10 performs RNN learning using an arbitrary learning method used for learning such as deep learning or a neural network such as a back-propagation method.

  However, if the number of learning data cannot be acquired sufficiently, it is not possible to properly learn the characteristics of each word included in the sentence, the characteristics of the order in which the words appear, etc. There is a risk that the accuracy of the will decrease.

  Therefore, the learning device 10 executes the following learning process. First, the learning device 10 causes the learning device and the restoration device included in the first RNN to learn the relationship between utterances and responses in a predetermined dialogue. Subsequently, the learning device 10 generates the second RNN using at least one of the learning device or the restoring device in which learning has been performed. Then, the learning device 10 causes the second RNN to learn the relationship between the sentence and the label attached to the sentence. That is, the learning device 10 performs pre-training on the relationship between the sentence and the label using the utterance and the response included in the dialogue.

  Here, it is predicted that hidden commonality is included in the utterance and response included in the dialogue. For example, when there is an anime or movie based on a novel, whether the utterance and response included in the dialogue is a dialogue about a novel, a dialogue about an animation, a dialogue about a movie, etc. There is a hidden commonality. Further, it is predicted that a common context (context) exists in the utterance included in the dialogue and the response to the utterance. Such commonality is expected to be similar to the commonality between the sentence context and the label context. Then, if the relationship between the utterance and the response is learned in advance, it is considered that the commonality and co-occurrence between the sentence and the label can be reinforced.

  Therefore, the learning device 10 performs pre-training of the learning device and the restoration device included in the RNN using the utterance and the response. For example, when the utterance is input to the learning device, the learning device 10 performs pre-training of the learning device and the restoration device so that the restoration device outputs a response to the utterance. More specifically, the learning device 10 causes the learning device to learn the characteristics of each word along with the order in which each word included in the utterance appears in the utterance. Then, the learning device 10 uses a restorer to restore each word included in the response together with the order in which each word appears in the response from the feature that the learner has encoded the utterance, that is, the information output by the learner. Let them learn.

  Note that if the learning device 10 associates an utterance and a response with either a learning device or a restoring device, the learning device 10 needs to learn the utterance in association with the learning device and learn the response in association with the restoring device. There is no. For example, the learning device 10 may learn the response in association with the learning device, and learn the utterance in association with the decompressor. Specifically, the learning device 10 may learn the learning device and the restoration device so that when the response is input to the learning device, the restoration device outputs an utterance corresponding to the response. For example, the learning device 10 causes the learner to learn the characteristics of each word along with the order in which each word included in the response appears in the response. Then, the learning device 10 may cause the restoration device to learn to restore each word included in the utterance together with the order in which each word appears in the utterance from the features encoded by the learning device. Further, the learning device 10 may perform learning of a model that serves as a learner and restorer by sharing parameters between the learner and the restorer. For example, the learning apparatus 10 may perform learning of one RNN that operates as a learner / restorer.

[1-3. About generation of RNN]
Here, the learning device 10 may learn the relationship between the sentence and the label by using the RNN that has undergone pre-training, that is, the first RNN as it is. In addition, the learning device 10 may generate a new RNN using at least one of the learning device and the restoring device included in the first RNN, and let the generated RNN learn the relationship between the sentence and the label. Good.

  For example, the learning device 10 generates the second RNN using a learning device included in the first RNN and a newly generated restoration device. Here, the newly generated restorer can be any restorer as long as it is a restorer included in the first RNN. For example, a transfer coefficient for transmitting information between neurons is initialized ( For example, it may be a randomized restorer. In the second RNN generated in this way, it is considered that the hidden commonality feature included in the utterance and the response is reflected on the learning device side. In other words, the second RNN is considered to easily converge the transfer coefficients between the neurons so that an appropriate label is output when a sentence is input. For this reason, even if the amount of learning data does not exist sufficiently, the second RNN shows the relationship between the sentence and the label included in the learning data, that is, the hidden common feature of the sentence and the label. It is thought that it can learn quickly.

  Therefore, the learning device 10 learns decoding of the second RNN so that when the sentence of the learning data is input to the second RNN, the second RNN outputs a label to be given to the input sentence. For example, the learning device 10 connects a randomly initialized restorer and a first RNN learner, fixes the learner parameters, and updates the parameters on the restorer side. As a result, the learning device 10 can generate an RNN that gives an appropriate label to the sentence, and thus can improve the accuracy of the label given to the sentence.

  Here, when learning is performed in the third RNN, that is, when a label is assigned to a sentence other than the learning data, the learning apparatus 10 may use the second RNN as it is to further learn the whole, and uses the second RNN. It is also possible to generate a new third RNN and use the new third RNN at the time of measurement (hereinafter referred to as “measurement RNN”). For example, it is predicted that the reconstructor included in the second RNN has learned the characteristics of the relationship between the text and the label. More specifically, the reconstructor included in the second RNN assigns to the sentences in which the plurality of words appear in the order from the information indicating the characteristics of the plurality of words and the characteristics of the order in which each word appears. Learning is performed so that information indicating a label to be output is output. Therefore, the learning device 10 may use, as the measurement RNN, the third RNN obtained by further performing overall learning on the third RNN including the restorer included in the second RNN and the learner included in the first RNN. Further, the learning device 10 may use the third RNN that is further subjected to the overall learning using the learning device of the first RNN without passing through the second RNN as the measurement RNN. Also, the learning device 10 fixes the learning device parameters and updates only the restoring device parameters during the learning of the third RNN, and unfixes the learning device parameters after a certain amount of learning has progressed. Thus, learning (training) of the entire third RNN may be performed.

  And the learning apparatus 10 may provide a label with respect to the sentence used as the label provision object using 3rd RNN. For example, the learning device 10 inputs each word included in the sentence in the order in which each word appears in the input layer of the learning device included in the third RNN, and the information output from the output layer of the learning device, that is, the input Feature information indicating the features of each word included in the written sentence and the features of the order in which each word appears is input to the input layer of the decompressor. And the learning apparatus 10 determines the label which should be provided with respect to a sentence from the information which shows the label which the decompressor output based on the feature information.

[1-4. About variations of RNN]
Here, the learning device 10 may generate the measurement RNN using an arbitrary process as long as it uses a learning device or a restoration device that has been pre-trained using the speech and response included in the dialogue. . For example, the learning device 10 generates a second RNN from a learning device included in the first RNN and a new restoration device, and learns the second RNN using learning data. And the learning apparatus 10 is good also considering the 2nd RNN which learned as it is as a measurement RNN.

  In addition, the learning device 10 generates a second RNN using a learning device included in the first RNN and a new restoration device, and learns the second RNN using learning data. Subsequently, the learning device 10 generates a third RNN using a decompressor included in the second RNN and a learner included in the first RNN, and learns the third RNN using the learning data. Then, the learning device 10 may output the third RNN that has been learned using the learning data as the measurement RNN.

  Note that the learning device 10 may use the same learning data or may use different learning data when performing the second RNN learning and the third RNN learning. In addition to learning data that is a set of a sentence and a label to be assigned to the sentence, the learning device 10 uses a pair of an utterance and a response included in the dialogue as learning data, and learns the second RNN and the third RNN. You may go.

[1-5. (About text and labels)
Here, the learning apparatus 10 may perform learning of an RNN that gives an arbitrary label given to an arbitrary sentence. That is, the sentence to which the learning device 10 gives a label may be a sentence relating to a writing such as a novel, a newspaper article, a blog, a microblog, a criticism, a posting, or a record such as a medical record. The sentence may be a single sentence or a sentence composed of a plurality of sentences. For example, the learning device 10 may perform sentence learning or the like in units of books or works.

  Further, the label provided by the learning device 10 may be a word or a sentence indicating the content or attribute of the sentence, or may be a numerical value or a character string used for searching or the like.

[1-6. Example formula)
Next, an example of a mathematical expression indicating the learning process described above will be described. For example, when the state of the “j” -th node of the learning device at time “t” is described as “hj <t>”, the state of each node in the learning device can be expressed by the following equation (1). Here, the wavy “hj <t>” in the equation (1) can be expressed by the equation (2), and “zj” and “rj” can be expressed by the following equations (3) and (4). expressed. Here, “xt” indicates the “t” th input. Further, “W” and “U” in the equations (2) to (4) are parameters indicating the connection coefficient of the learning device.

  Then, the learning result by the learning device can be expressed by Expression (5).

  Here, the output of the learning device is transmitted to each node of the restoring device. The input to such a restorer node can be expressed by equation (6).

  Then, if the state of the “j” -th node of the decompressor at time “t” is “h′j <t>”, “h′j <t>” can be expressed by the following equation (7). it can. Here, “h′j <t>” with a wave in the equation (7) can be expressed by the following equation (8), and “z′j” and “r′j” 9) and the formula (10). In addition, “W ′” and “U ′” in Expression (8) to Expression (10) are parameters indicating the connection coefficient of the restorer.

  For example, in the learning device 10, the learning device learns a plurality of words included in a sentence, that is, the characteristics and order of the word group, and the restoration device uses the output of the learning device to obtain the plurality of words input to the learning device. “W”, “U”, “W ′”, and “U” in Expressions (1) to (10) so as to output information indicating labels to be attached to sentences appearing in the order of input. The value of “′” is learned.

[1-7. Example of learning process)
Next, an example of a learning process executed by the learning device 10 will be described with reference to FIG. First, the learning apparatus 10 accepts a dialogue such as a user's statement A or statement B as an input (step S1). For example, the input device 200 acquires a voice of a dialogue performed between a plurality of users using a voice acquisition device such as a microphone. Then, the input device 200 converts the dialogue into text data using an arbitrary voice recognition technique, and transmits the converted text data, that is, the text data of the dialogue to the learning device 10.

  More specifically, the input device 200 uses the text data of the utterance A by a certain user as utterance data (hereinafter sometimes referred to as “utterance data”), and other users' responses to the utterance A. Is generated as response data (hereinafter may be referred to as “response data”). The input device 200 transmits the speech data and the response data to the learning device 10 as dialogue data.

  In such a case, the learning device 10 associates the utterance and the response included in the input dialogue sentence with either the learning device or the restoration device of the first RNN, respectively, and hidden information inherent in the utterance and the response. That is, pre-training is performed on the hidden commonality inherent in speech and response (step S2). For example, the learning device 10 extracts an utterance and a response corresponding to the utterance from the text data of the dialog. Then, when the learning device 10 inputs the words included in the utterance in the order included in the utterance to the learning device EN01 included in the first RNN, the utterance input to the learning device EN01 from the output of the learning device EN01. The learning unit EN01 and the restoring unit DE01 perform learning so that the restoring unit DE01 outputs each word included in the response corresponding to the in the order included in the response. More specifically, the learning device 10 corrects the transfer coefficients of the learning device EN01 and the restoring device DE01.

  Next, the learning device 10 generates a second RNN that combines the pre-training learner EN01 and the new decompressor DE02, and from the sentence of the labeled sentence (that is, learning data), The second RNN is learned so as to generate a label (step S3). For example, when the learning device 10 inputs each word included in the sentence of the labeled sentence to the learning device EN01 included in the second RNN in the order included in the sentence, the learning apparatus 10 reads the labeled sentence from the output of the learning device EN01. The learning unit EN01 and the restoring unit DE02 perform learning so that the restoring unit DE02 outputs information indicating the label of.

  Here, it is assumed that the learning device 10 has acquired a sentence to be given a label from the terminal device 100 (step S4). In such a case, the learning device 10 executes processing at the time of measurement. More specifically, the learning device 10 uses the third RNN, which is a combination of the restoration device DE02 that has performed label learning using a labeled sentence, and the learning device EN01 that has undergone pre-training, as an RNN for measurement. The label given to the sentence acquired from 100 is determined (step S5). For example, the learning device 10 inputs each word included in the sentence acquired from the terminal device 100 to the learning device EN01 of the measurement RNN in the order in which each word appears. And the learning apparatus 10 determines the label which should be provided to a sentence based on the information which the decompressor DE02 of measurement RNN output. Thereafter, the learning device 10 outputs information indicating the label that is the determination result to the terminal device 100 (step S6).

[2. Configuration of learning device]
Hereinafter, an example of a functional configuration of the learning device 10 that realizes the above-described learning process will be described. FIG. 2 is a diagram illustrating a configuration example of the learning device according to the embodiment. As illustrated in FIG. 2, the learning device 10 includes a communication unit 20, a storage unit 30, and a control unit 40.

  The communication unit 20 is realized by, for example, a NIC (Network Interface Card). The communication unit 20 is connected to the network N by wire or wireless, and transmits and receives information to and from the terminal device 100 and the input device 200.

  The storage unit 30 is realized by, for example, a semiconductor memory device such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 30 also stores a dialogue data database 31 and a learning data database 32. Note that the storage unit 30 further stores, for example, information indicating the relationship between the node of the RNN to be learned and the transfer coefficient.

  In the dialogue data database 31, dialogue data used for pre-training is registered. For example, FIG. 3 is a diagram illustrating an example of information registered in the dialogue data database according to the embodiment. In the example illustrated in FIG. 3, information having items such as “dialog ID”, “utterance information”, “response information”, and the like is registered in the dialog data database 31. Note that information other than the information shown in FIG. 3, such as the attributes of the user who made an utterance or a response, may be registered in the dialog data database 31.

  Here, the “dialog ID (Identifier)” is information for identifying a dialog. “Speech information” is text data of an utterance. “Response information” is text data of a response to the associated utterance. In the example shown in FIG. 3, conceptual values such as “dialog # 1”, “utterance # 1”, “response # 1”, and the like have been described. Are registered in the dialogue data database 31.

  For example, in the example illustrated in FIG. 3, the conversation ID “dialog # 1”, the utterance information “utterance # 1”, and the response information “response # 1” are registered in association with each other. Such information indicates that the text data of the utterance included in the dialog indicated by the dialog ID “dialog # 1” is “utterance # 1”, and the text data of the response to this utterance is “response # 1”. .

  Returning to FIG. 2, the description will be continued. In the learning data database 32, learning data used for RNN learning, that is, labeled sentences are registered. For example, FIG. 4 is a diagram illustrating an example of information registered in the learning data database according to the embodiment. In the example illustrated in FIG. 4, information having items such as “learning data ID”, “text information”, and “label information” is registered in the learning data database 32. In addition to the information shown in FIG. 3, arbitrary information related to the labeled sentence may be registered in the learning data database 32.

  Here, the “learning data ID” is information for identifying a dialogue as learning data. “Sentence information” is a sentence of a labeled sentence, that is, text data. “Label information” is information indicating a label to be assigned to the associated text. In the example shown in FIG. 4, conceptual values such as “data # 1”, “sentence # 1”, “label # 1”, and the like are described. However, in practice, an identifier for identifying a labeled sentence, Text data or the like is registered in the learning data database 32.

  For example, in the example illustrated in FIG. 4, the learning data ID “data # 1”, the sentence information “sentence # 1”, and the label information “label # 1” are registered in association with each other. In such information, the text data of the sentence included in the labeled sentence indicated by the learning data ID “data # 1” is “sentence # 1”, and the label to be given to this sentence is “label # 1”. ".

  Returning to FIG. 2, the description will be continued. The control unit 40 is a controller. For example, various programs stored in a storage device inside the learning device 10 are stored in a RAM or the like by a processor such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). This is realized by being executed as a work area. The control unit 40 is a controller, and may be realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

  As illustrated in FIG. 2, the control unit 40 includes an acquisition unit 41, a first learning unit 42, a second learning unit 43, a generation unit 44, and a measurement unit 45.

  The acquisition unit 41 acquires an utterance in a predetermined dialogue and a response to the utterance. For example, when acquiring the text data of the dialog from the input device 200, the acquiring unit 41 analyzes the acquired text data and extracts the text data of the utterance and the text data of the response to the utterance. Then, the acquisition unit 41 registers the extracted utterance text data and response text data in the dialog data database 31 as dialog data. The acquisition unit 41 may extract an utterance and a response corresponding to the utterance using an arbitrary character string analysis technique such as morphological analysis.

  The first learning unit 42 causes the learning device and the restoration device included in the first RNN to learn the relationship between the utterance and the response in a predetermined dialogue. That is, the first learning unit 42 performs pre-training of the RNN using dialogue.

  For example, the first learning unit 42 acquires utterances and responses included in the dialogue from the dialogue data database 31. In addition, the first learning unit 42 prepares a first RNN to be pretrained, such as an initialized first RNN. Then, when the utterance is input to the first RNN learner, the first learning unit 42 learns the learner and the restorer so that the first RNN restorer outputs a response to the utterance. More specifically, the first learning unit 42 inputs each word included in the utterance to the learning device in the order in which the words appear in the utterance, and each word included in the response from each feature encoded by the learning device. The learner and the restorer are trained to restore the words in the order in which they appear in the response. That is, the first learning unit 42 causes the learning device and the restoration device to learn the characteristics of each word in the utterance and the characteristics of the order in which each word appears.

  Note that the first learning unit 42 may perform learning of the learning device and the restoration device so that when the response is input to the learning device, the restoration device outputs an utterance corresponding to the response. For example, the first learning unit 42 causes the learner to learn the features of each word together with the order in which each word included in the response appears in the utterance, and each feature included in the utterance from the features encoded by the learner. The restorer may be trained to restore words with the order in which each word appears in the response.

  The second learning unit 43 uses the labeled sentence that is the learning data registered in the learning data database 32 to generate the second RNN generated using at least one of the learning device and the restoring device included in the first RNN. Or, the third RNN is learned. An example of processing executed by the second learning unit 43 will be described later.

  The generation unit 44 generates the second RNN using at least one of the learning device or the restoration device learned by the first learning unit 42, that is, the learning device or the restoration device included in the first RNN. For example, the generation unit 44 generates the second RNN using a learning device and a new restorer included in the first RNN.

  When such a second RNN is generated, the second learning unit 43 causes the second RNN to learn the relationship between the sentence of the labeled sentence registered in the learning data database 32 and the label of the labeled sentence.

  In addition, when the second learning unit 43 performs the second RNN learning, the generation unit 44 performs the learning by the first RNN learning device that has been pretrained by the first learning unit 42 and the second learning unit 43. A third RNN is generated that includes a reconstructor of the performed second RNN. Then, the generation unit 44 registers the generated third RNN as the measurement RNN in the storage unit 30 or the like.

  Note that the second learning unit 43 may perform further learning using the learning data on the third RNN. For example, the second learning unit 43 causes the third RNN to learn the relationship between the sentence of the labeled sentence registered in the learning data database 32 and the label of the labeled sentence. The second learning unit 43 may register the learned third RNN as the measurement RNN in the storage unit 30 or the like.

  The measurement unit 45 executes measurement processing using the measurement RNN. For example, when the measurement unit 45 receives a sentence to be given a label from the terminal device 100, the measurement unit 45 uses the third RNN generated by the generation unit 44 as the measurement RNN to identify the label to be given to the received sentence. To do. For example, the measurement unit 45 inputs each word included in the accepted sentence to the learning RNN learning device in the order in which the words appear in the sentence, and accepts the word based on the information output from the measuring RNN decompressor. The label given to the given sentence is specified. Then, the measurement unit 45 transmits information indicating the identified label to the terminal device 100.

  Note that the measurement unit 45 may specify a label to be given to the accepted sentence, with the third RNN that has been further learned using the labeled sentence by the second learning unit 43 as the measurement RNN. Moreover, the measurement part 45 may specify the label provided to the received sentence by making 2nd RNN learned by the 2nd learning part 43 into measurement RNN.

[3. Variation of RNN]
Next, with reference to FIG. 5 and FIG. 6, variations of generation of the RNN used by the learning device 10 as the measurement RNN will be described. FIG. 5 is a first diagram illustrating variations of the RNN generated by the learning device according to the embodiment. FIG. 6 is a second diagram illustrating variations of the RNN generated by the learning device according to the embodiment.

  First, an example of a process for setting the second RNN as the measurement RNN will be described with reference to FIG. For example, the learning device 10 generates a first RNN including the learning device EN01 and the restoring device DE01, and performs pre-training using the utterance and the response (step S10). Subsequently, the learning device 10 takes out the learning device EN01 included in the first RNN (step S11), generates a second RNN using the learning device EN01 and the new restoration device DE02, and describes the sentence and label of the labeled sentence. Is used to learn the second RNN (step S12). Then, the learning device 10 uses the learned second RNN restorer DE02 (step S13) and the learned second RNN learner EN01 (step S14) as they are to obtain the measurement RNN, and the sentence Measurement processing for specifying the label is executed (step S15).

  Next, with reference to FIG. 6, an example of processing in which the third RNN that has performed learning using the labeled sentence is used as the measurement RNN will be described. For example, the learning device 10 generates a first RNN including the learning device EN01 and the restoring device DE01, and performs pre-training using the utterance and the response (step S20). Subsequently, the learning device 10 takes out the learning device EN01 included in the first RNN (step S21), generates a second RNN using the learning device EN01 and the new restoration device DE02, and describes the sentence and label of the labeled sentence. Is used to learn the second RNN (step S22).

  Further, the learning device 10 takes out the restoring device DE02 of the second RNN that has performed learning (Step S23), generates a third RNN that is combined with the learning device EN01 included in the first RNN that has undergone pre-training (Step S24), The learning of the third RNN is executed again using the labeled sentence (step S25). Then, the learning device 10 uses the learned third RNN learner DE02 (step S26) and the learned third RNN learner EN01 (step S27) as they are to obtain the measurement RNN, and the sentence. Measurement processing for specifying the label is executed (step S28).

[4. Example of flow of processing executed by learning device]
Next, an example of the flow of learning processing executed by the learning device 10 will be described with reference to FIG. FIG. 7 is a flowchart for explaining an example of the flow of the learning process according to the embodiment. First, the learning device 10 acquires a dialogue sentence (step S101), and extracts an utterance and a response from the dialogue sentence (step S102). Then, the learning device 10 performs pre-training with the learning device and the restoration device included in the first RNN so as to reproduce the response from the utterance (step S103).

  Subsequently, the learning device 10 generates the second RNN using the pre-trained learner and the new reconstructor (step S104), and generates information indicating the label from the sentence of the labeled sentence. The second RNN is learned (step S105). Then, the learning device 10 generates a third RNN using the learned second RNN restorer and the learner included in the pre-trained first RNN (step S106), and the generated third RNN is used as the measurement RNN. Are used to execute the measurement process (step S107), and the process is terminated.

[5. (Modification)
In the above, an example of the learning process by the learning device 10 has been described. However, the embodiment is not limited to this. Hereinafter, the variation of the learning process which the learning apparatus 10 performs is demonstrated.

[5-1. About RNN for measurement]
In the above-described example, the learning device 10 specifies a label to be given to a sentence to be given a label, using the generated measurement RNN. However, the embodiment is not limited to this. For example, the learning apparatus 10 may generate a program for assigning a label from a sentence using the generated measurement RNN and output the generated program. Such a program is executed by an information processing device other than the learning device 10, so that the label can be specified from the sentence using the measurement RNN generated by the learning device 10. That is, the measurement RNN generated by the learning device 10 may be used as a program parameter of a program that specifies a label from a sentence.

[5-2. Device configuration〕
In the example described above, the learning device 10 performs the learning process and the measurement process in the learning device 10. However, the embodiment is not limited to this. For example, the learning device 10 may execute only the learning process, and other devices may execute the measurement process. Further, the learning device 10 may store the dialogue data database 31 and the learning data database 32 in an external storage server.

[5-3. Others]
In addition, among the processes described in the above embodiment, all or part of the processes described as being automatically performed can be performed manually, or the processes described as being performed manually can be performed. All or a part can be automatically performed by a known method. In addition, the processing procedures, specific names, information including various data and parameters shown in the above text and drawings can be arbitrarily changed unless otherwise specified. For example, the various types of information illustrated in each drawing is not limited to the illustrated information.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured. For example, the first learning unit 42 and the second learning unit 43 illustrated in FIG. 2 may be integrated.

  In addition, the above-described embodiments can be appropriately combined within a range in which processing contents are not contradictory.

[6. program〕
Further, the learning device 10 according to the embodiment described above is realized by a computer 1000 having a configuration as shown in FIG. 8, for example. FIG. 8 is a diagram illustrating an example of a hardware configuration. The computer 1000 is connected to an output device 1010 and an input device 1020, and an arithmetic device 1030, a primary storage device 1040, a secondary storage device 1050, an output IF (Interface) 1060, an input IF 1070, and a network IF 1080 are connected via a bus 1090. Have

  The arithmetic device 1030 operates based on a program stored in the primary storage device 1040 and the secondary storage device 1050, a program read from the input device 1020, and the like, and executes various processes. The primary storage device 1040 is a memory device such as a RAM that temporarily stores data used by the arithmetic device 1030 for various arithmetic operations. The secondary storage device 1050 is a storage device in which data used for various calculations by the calculation device 1030 and various databases are registered, and is realized by a ROM (Read Only Memory), HDD, flash memory, or the like.

  The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various types of information such as a monitor and a printer. For example, USB (Universal Serial Bus), DVI (Digital Visual Interface), This is realized by a standard connector such as HDMI (registered trademark) (High Definition Multimedia Interface). The input IF 1070 is an interface for receiving information from various input devices 1020 such as a mouse, a keyboard, and a scanner, and is realized by, for example, a USB.

  The input device 1020 includes, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), and a PD (Phase change rewritable disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), and a tape. It may be a device that reads information from a medium, a magnetic recording medium, a semiconductor memory, or the like. The input device 1020 may be an external storage medium such as a USB memory.

  The network IF 1080 receives data from other devices via the network N and sends the data to the arithmetic device 1030, and transmits data generated by the arithmetic device 1030 to other devices via the network N.

  The arithmetic device 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic device 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040, and executes the loaded program.

  For example, when the computer 1000 functions as the learning device 10, the arithmetic device 1030 of the computer 1000 implements the function of the control unit 40 by executing a program loaded on the primary storage device 1040.

[7. effect〕
As described above, the learning device 10 causes the learning device and the restoration device included in the first RNN to learn the relationship between the utterance and the response in a predetermined dialogue. Then, the learning device 10 generates the second RNN using at least one of the learning device or the restoring device in which learning is performed, and the relationship between the sentence and the label attached to the sentence is determined as the second RNN. To learn. As described above, the learning apparatus 10 performs pre-training of the RNN using the hidden commonality included in the utterance and the response to the utterance, so that it is possible to improve the accuracy of the label attached to the sentence. In addition, when the learning data is abundant to some extent, it is possible to acquire an RNN with better classification accuracy (performance) than an RNN learned by a conventional learning method.

  The learning device 10 learns the learning device and the restoration device so that the restoration device outputs a response to the utterance when the utterance is input to the learning device. For example, the learning device 10 causes the learner to learn the characteristics of each word together with the order in which each word included in the utterance appears in the utterance, and each word included in the response is recorded from the feature encoded by the learner. The restorer is trained to restore with the order in which the words appear in the response.

  Further, the learning device 10 learns the learning device and the restoration device so that when the response is input to the learning device, the restoration device outputs the utterance corresponding to the response. For example, the learning device 10 causes the learner to learn the features of each word together with the order in which each word included in the response appears in the response, and each word included in the utterance is recorded from the features encoded by the learner. The restorer is trained to restore words with the order in which they appear in the utterance.

  As a result of the processing described above, the learning device 10 can perform pre-training of the RNN using the hidden commonality included in the utterance and the response to the utterance.

  In addition, the learning device 10 generates a second RNN using a learning device and a new restoration device included in the first RNN that has undergone pre-training. Further, the learning device 10 generates a third RNN including a learning device included in the first RNN that has been subjected to pre-training, and a restoring device that is included in the second RNN that has been subjected to learning using learning data. And the learning apparatus 10 specifies the label provided to the received sentence using 3rd RNN, when the sentence used as the provision object of a label is received. For this reason, the learning apparatus 10 can prevent the deterioration of the precision of the label provided to a sentence.

  Further, the learning device 10 generates a third RNN including a learning device included in the first RNN that has been subjected to pre-training, and a restoring device that is included in the second RNN that has been subjected to learning using learning data. Then, the learning device 10 causes the generated third RNN to learn the relationship between the sentence and the label given to the sentence, and the label given to the accepted sentence using the learned third RNN. Is identified. For this reason, the learning apparatus 10 can prevent the deterioration of the precision of the label provided to a sentence.

  Moreover, the learning apparatus 10 specifies the label provided to the received sentence using 2nd RNN, when the sentence used as the provision object of a label is received. For this reason, the learning apparatus 10 can prevent the deterioration of the precision of the label provided to a sentence.

  As described above, some of the embodiments of the present application have been described in detail with reference to the drawings. However, these are merely examples, and various modifications, including the aspects described in the disclosure section of the invention, based on the knowledge of those skilled in the art, It is possible to implement the present invention in other forms with improvements.

  Moreover, the above-mentioned “section (module, unit)” can be read as “means”, “circuit”, and the like. For example, the generation unit can be read as generation means or a generation circuit.

DESCRIPTION OF SYMBOLS 20 Communication part 30 Storage part 31 Dialogue data database 32 Learning data database 40 Control part 41 Acquisition part 42 1st learning part 43 2nd learning part 44 Generation part 45 Measuring part 100 Terminal apparatus 200 Input device

Claims (12)

A first learning unit that causes a learning device and a restoration device included in the first recurrent neural network to learn a relationship between an utterance and a response in a predetermined dialogue;
A first generator that generates a second recurrent neural network using at least one of a learner or a restorer that has been learned by the first learning unit;
A learning apparatus comprising: a second learning unit that causes the second recurrent neural network to learn a relationship between a sentence and a label attached to the sentence. The first learning unit performs learning of the learning device and the restoration device so that when the utterance is input to the learning device, the restoration device outputs a response to the utterance. Item 4. The learning device according to Item 1. The first learning unit causes the learner to learn the features of each word together with the order in which the words included in the utterance appear in the utterance, and includes the response from the features encoded by the learner. The learning apparatus according to claim 2, wherein the restorer is made to learn each word to be restored together with an order in which each word appears in the response. The first learning unit performs learning of the learning device and the restoring device so that when the response is input to the learning device, the restoring device outputs an utterance corresponding to the response. The learning apparatus according to claim 1. The first learning unit causes the learner to learn the characteristics of each word together with the order in which each word included in the response appears in the response, and includes the utterance from the features encoded by the learner. The learning apparatus according to claim 4, wherein the restorer is made to learn each word to be restored together with an order in which each word appears in the utterance. The said 1st production | generation part produces | generates the said 2nd recurrent neural network using the learning device and new restorer which were learned by the said 1st learning part. The one of Claims 1-5 characterized by the above-mentioned. The learning device according to any one of the above. A second generator for generating a third recurrent neural network including a learning device learned by the first learning unit and a restorer included in the second recurrent neural network learned by the second learning unit;
A specific unit for specifying a label to be given to the accepted sentence using the third recurrent neural network generated by the second generation unit when the sentence to be given a label is accepted; The learning apparatus according to claim 6. A second generator for generating a third recurrent neural network including a learning device learned by the first learning unit and a restorer included in the second recurrent neural network learned by the second learning unit;
A third learning unit that causes the third recurrent neural network generated by the second generation unit to learn the relationship between the sentence and the label attached to the sentence;
A specific unit that identifies a label to be given to the accepted sentence using the third recurrent neural network learned by the third learning part when the sentence to be given a label is accepted; The learning apparatus according to claim 6, further comprising: A specific unit that identifies a label to be given to the accepted sentence using the second recurrent neural network learned by the second learning part when the sentence to be given a label is accepted; The learning apparatus according to claim 6, further comprising: A learning method executed by a learning device,
A first learning step of causing a learning device and a restoration device included in the first recurrent neural network to learn a relationship between an utterance and a response in a predetermined dialogue;
A first generation step of generating a second recurrent neural network using at least one of a learner or a reconstructor learned in the first learning step;
A learning method comprising: a second learning step of causing the second recurrent neural network to learn a relationship between a sentence and a label attached to the sentence. A first learning step of causing a learning device and a restoration device included in the first recurrent neural network to learn a relationship between an utterance and a response in a predetermined dialogue;
A first generation step of generating a second recurrent neural network using at least one of a learner or a reconstructor learned in the first learning step;
A second learning step of causing the second recurrent neural network to learn a relationship between a sentence and a label attached to the sentence;
A second generation step of generating a third recurrent neural network for estimating a label to be given to the input sentence using the second recurrent neural network;
A program parameter including the third recurrent neural network generated by the learning method. A first learning procedure for causing a learning device and a restoration device included in the first recurrent neural network to learn a relationship between an utterance and a response in a predetermined dialogue;
A first generation procedure for generating a second recurrent neural network using at least one of a learner or a reconstructor trained by the first learning procedure;
A learning program for executing a second learning procedure for causing the second recurrent neural network to learn a relationship between a sentence and a label given to the sentence.

Images