JP2017059205A - Subject estimation system, subject estimation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a subject estimation system and the like capable of highly accurately estimating a subject of a dialog even if learning data is insufficient.SOLUTION: A subject estimation system comprises a convolution neural network 10 for estimating a subject label of a dialog. The convolution neural network 10 includes: a convolution layer 12 consisting of one or more topic dependent convolution layers for performing convolution calculation dependent on a topic with respect to input of a word string vector sequence corresponding to a dialog text which is transcription of the dialog and one topic independent convolution layer for performing convolution calculation independent of the topic; a pooling layer 13 for performing pooling processing with respect to output of the convolution layer 12; and a whole coupling layer 14 for performing whole coupling processing with respect to output of the pooling layer 13.SELECTED DRAWING: Figure 3

Description

  The present invention relates to a subject estimation system, a subject estimation method, and a program for estimating a subject of dialogue.

  There is a system that performs pattern recognition using a convolutional neural network (for example, Patent Document 1). Patent Document 1 discloses a general method of pattern recognition using a convolutional neural network.

  A method of applying a convolutional neural network to the field of natural language processing is also known (for example, Non-Patent Document 1). Non-Patent Document 1 discloses a method of classifying sentences using a convolutional neural network trained using a known data set.

US Patent Application Publication No. 2003/0174881

Yoon Kim, “Convolutional Neural Networks for Sentence Classification”, [online], [Search March 29, 2016], Internet <URL: http://arxiv.org/abs/1408.5882>

  However, in the sentence classification method using the above-described conventional technique, it is assumed that the convolutional neural network is learned using sufficient learning data, and the case where the learning data is insufficient is not taken into consideration.

  Therefore, even if the convolutional neural network disclosed in the above prior art is used, there is a problem that the task of estimating the subject of dialogue cannot be performed with high accuracy if the learning data is not sufficient.

  Therefore, the present invention focuses on the above-described problems, and provides a subject estimation system, a subject estimation method, and a program capable of estimating the subject of a conversation with higher accuracy even when learning data is not sufficient. For the purpose.

  In order to solve the above problems, a theme estimation system according to an aspect of the present invention includes a convolutional neural network, and is a theme estimation system for estimating a theme label of a dialog, the convolution neural network writing a dialog. One or more topic-dependent convolution layers that perform topic-dependent convolution operations on a word string vector sequence input corresponding to the generated conversation text, and one topic-independent convolution layer that performs convolution operations that do not depend on the topic A convolution layer, a pooling layer that performs a pooling process on the output of the convolution layer, and a total coupling layer that performs a total coupling process on the output of the pooling layer.

  These general or specific aspects may be realized by a system, a method, an integrated circuit, a computer program, or a recording medium, and are realized by any combination of the system, method, integrated circuit, computer program, and recording medium. May be.

  ADVANTAGE OF THE INVENTION According to this invention, even when learning data is not enough, the theme estimation system etc. which can estimate the theme of a dialog with higher precision are realizable.

The figure which shows the structure of the convolution neural network which the theme estimation system of the comparative example uses Diagram for explaining that the convolutional neural network of the comparative example does not support multi-label output The figure which shows the structure of the convolution neural network which the theme estimation system in Embodiment 1 utilizes FIG. 3 is a block diagram showing a functional configuration at the time of identification of the subject estimation system in the first embodiment. The figure which shows an example of the hardware constitutions of a computer required in order to perform a theme estimation system The flowchart which shows operation | movement of the theme estimation system at the time of identification shown in FIG. FIG. 3 is a block diagram showing a functional configuration during learning of the subject estimation system in the first embodiment. The flowchart which shows operation | movement of the theme estimation system at the time of learning shown in FIG. The figure which shows the experimental verification result of the theme estimation system in Embodiment 1 FIG. 7 is a block diagram showing a functional configuration during additional learning of the subject estimation system in the second embodiment. The flowchart which shows operation | movement of the theme estimation system at the time of the additional learning shown in FIG. The graph which shows the experimental verification result of the theme estimation system in Embodiment 2

(Knowledge that became the basis of the present invention)
The task of estimating the subject of dialogue assumed in the present disclosure is a task of estimating what is the subject of dialogue, which is a natural language exchange based on two parties, regardless of whether a person or a machine. The task estimates a subject in the dialogue by inputting a word string during the dialogue and taking a segment of the word row recognized as related to one topic as a unit.

  In a subject estimation system based on a convolutional neural network, learning of a convolutional neural network is performed using learning data, and evaluation is performed using evaluation data.

  In the above tasks assumed in this disclosure, a finite number of topics are defined, and a finite number of subjects that are desired to be output are also defined. Also, what the topic is is given to the subject estimation system together with the word string in the segment.

  However, as described above, when the convolutional neural network disclosed in the above-described prior art is used in the subject estimation system, it is assumed that the convolutional neural network is learned using sufficient learning data. The case where the learning data is not sufficient is not considered.

  For example, when data obtained by recording a dialogue between humans is used as learning data, there may be a small amount of dialogue and less learning data for a certain topic. Also, for a subject that is desired to be obtained as an output, it may happen that there is little learning data for a certain subject. Such a case where there is little learning data and there is a bias in the number has not been studied.

  Therefore, even if the convolutional neural network disclosed in the above prior art is used, if the learning data is not sufficient, there arises a problem that the task of estimating the subject of dialogue cannot be performed with high accuracy.

  In order to solve such a problem, a subject estimation system according to an aspect of the present invention includes a convolutional neural network, and is a subject estimation system for estimating a subject label of a dialogue, wherein the convolutional neural network includes: One or more topic-dependent convolutional layers that perform topic-dependent convolution operations on an input of a word string vector sequence corresponding to the dialog text that transcribes the conversation, and one topic non-conformation that performs convolution operations that do not depend on the topic A convolution layer including a dependent convolution layer; a pooling layer that performs a pooling process on an output of the convolution layer; and an all coupling layer that performs a total coupling process on an output of the pooling layer.

  With this configuration, it is possible to realize a theme estimation system that can estimate the theme of a conversation with higher accuracy even when the learning data is not sufficient. More specifically, according to this configuration, one topic-independent convolution layer that is learned for each topic with high accuracy when there is a large amount of learning data, and learning regardless of a topic with high accuracy when there is a small amount of learning data Since one or more topic-dependent convolutional layers are integrated in a higher hierarchy, high subject estimation performance can be obtained even when the amount of learning data is small.

  The convolutional neural network may estimate the subject label of the dialogue for the input by solving the input as a two-class classification problem.

  Further, in the convolutional neural network, learning dialogue text that transcribes a dialogue, wherein the dialogue time-series text is pre-divided into segments for each topic, and a topic label corresponding to each divided segment Is used as learning data to cause each of the one or more topic-dependent convolutional layers to perform a convolution operation that depends on the dependent topic for each dependent topic that is a dependent topic. The first weight may be learned, and the second weight may be learned so that the topic-independent convolution layer performs a convolution operation that does not depend on the dependent topic.

  Each of the one or more topic-dependent convolutional layers depends on the dependent topic by inputting a word string vector sequence related to the dependent topic among word string vector sequences corresponding to the learning dialogue text. The first weight is learned so as to perform a convolution operation, and the topic-independent convolution layer performs a convolution operation that does not depend on the dependent topic by inputting a word string vector sequence corresponding to the learning dialogue text. The second weight may be learned as described above.

  In order to solve the above problem, a theme estimation method according to an aspect of the present invention is a theme estimation method of a theme estimation system that includes a convolutional neural network and estimates a theme label of a dialog. A topic-dependent convolution processing step that performs a topic-dependent convolution operation on the input of a word string vector sequence corresponding to the dialogue text that has been transcribed, and a topic non-conformation that performs a topic-independent convolution operation on the input. A pooling processing step for performing a pooling process on the dependent convolution processing step, an output of the topic-dependent convolution processing step and an output of the topic-independent convolution processing step, and a fully combining process on the output of the pooling processing step And a full join processing step.

  Thereby, even when the learning data is not sufficient, it is possible to realize a theme estimation method that can estimate the theme of the dialogue with higher accuracy. More specifically, since the result of the topic-dependent convolution operation and the result of the topic-independent convolution operation are integrated at a later stage, high subject estimation performance can be obtained even when the amount of learning data is small.

  In the topic-dependent convolution processing step, the topic-independent convolution is performed by performing a convolution operation on the word string vector sequence and a first weight (A) that fires at a specific word indicating a dependent topic that is a dependent topic. In the processing step, a convolution operation is performed between the word string vector sequence and a second weight (Z) ignited by a word indicating a topic other than the dependent topic. In the pooling processing step, an output of the topic-dependent convolution processing step is performed. And the topic-independent convolution processing step output for extracting the maximum value in the time direction, and in the all connection processing step, weighted addition using the connection weight is performed on the output of the pooling processing step. Later, it is also possible to perform the full connection process by performing probability distribution.

  And an output step of estimating and outputting the subject label of the dialogue by comparing the probability distribution of the output of the pooling processing step with a threshold value.

  And an input step for inputting a word string vector sequence corresponding to the dialog text, wherein the input step further includes dialog text in which the dialog is transcribed, wherein the dialog is converted into text in time series. And a vectorizing step of calculating a word vector of each word string included in the dialog text by a predetermined method to obtain a word string vector string.

  Further, the conversational text for learning that transcribes the conversation to the convolutional neural network, the time-series text of the conversation being divided in advance into segments for each topic, and topics corresponding to the divided segments. A learning dialog text pre-assigned with a label is used as learning data to learn the first weight so as to perform a convolution operation depending on a dependent topic which is a topic on which the topic-dependent convolution processing step depends. And a second step of learning the second weight so that the topic-independent convolution processing step performs a convolution operation independent of the dependent topic using the learning dialogue text.

  In the first step, the first weight is learned using a word string vector string related to the dependent topic among word string vector strings corresponding to the learning dialogue text, and in the second step, the first weight is learned. The second weight may be learned using a word string vector string related to a topic other than the dependent topic among word string vector strings corresponding to the learning dialogue text.

  Further, when the number of word string vector sequences related to the first dependent topic among the word string vector sequences corresponding to the learning dialogue text is smaller than the number of word string vector sequences related to the second dependent topic, The first step and the second step may be performed using a dialogue text related to the first dependent topic obtained by searching the Web as semi-teacher data of the learning data.

  The present invention is not only realized as an apparatus, but also realized as an integrated circuit including processing means included in such an apparatus, or realized as a method using the processing means constituting the apparatus as a step. Can be realized as a program for causing a computer to execute, or as information, data, or a signal indicating the program. These programs, information, data, and signals may be distributed via a recording medium such as a CD-ROM or a communication medium such as the Internet.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Note that each of the embodiments described below shows a specific example of the present invention. Numerical values, shapes, components, steps, order of steps and the like shown in the following embodiments are merely examples, and are not intended to limit the present invention. In addition, among the constituent elements in the following embodiments, constituent elements that are not described in the independent claims indicating the highest concept are described as optional constituent elements. In all the embodiments, the contents can be combined.

(Embodiment 1)
In the present embodiment, a subject estimation system that includes a convolutional neural network and estimates a subject label of an interaction will be described. In the following, first, the structure of the convolutional neural network used by the subject estimation system of the comparative example will be described with reference to FIG. 1, and then the convolution neural network used by the subject estimation system according to the present embodiment will be used with reference to FIG. The network structure and the like will be described.

[Structure of subject estimation system of comparative example]
FIG. 1 is a diagram showing the structure of a convolutional neural network 80 used by the subject estimation system of the comparative example. The convolutional neural network 80 shown in FIG. 1 includes an input feature 81, a convolutional layer 82, a pooling layer 83, and a fully connected layer 84.

  The input feature 81 converts an input word string into a vector string by a predetermined method. The convolution layer 82 cuts out adjacent vector sequences of one to several words, and then performs a convolution operation using a learned weight matrix. The pooling layer 83 performs an operation for obtaining the maximum value in the time direction with respect to the output of the convolution layer 82. The total coupling layer 84 multiplies the output of the pooling layer 83 by the coupling weight for each output element of the total coupling layer 84, and finally performs probability distribution using the softmax function.

  When the convolutional neural network 80 in such a comparative example is used for the subject estimation task, the weight of the part highly related to the subject in the word string in the dialogue is increased by learning using the learning data, and a specific language A particular subject can be inferred if an expression is included.

  A frame indicated by a thick line in the input feature 81 in FIG. 1 is a time window. FIG. 1 shows two types, a time window for one word and a time window for two words. These time windows are shifted from the beginning to the end of the vector sequence obtained by converting the input word sequence along the time direction, and a convolution operation and nonlinear processing are performed for each time window to obtain an output value. An output element that holds these is represented by a square indicated by a thick line in the convolution layer 82.

  Note that the convolution operation using the convolutional neural network 80 is also referred to as filter processing. The number of output elements of the convolution layer 82 is determined by multiplying the total number of filters and the number of shifts of the time window. On the other hand, since the total coupling layer 84 has output elements corresponding to all the subject labels, the number of output elements of the total coupling layer 84 is determined by the number of all the subject labels. Therefore, in the subject estimation system using the convolutional neural network 80 of the comparative example, the problem (subject estimation) is solved as a multi-class classification problem.

  However, as described above, in the convolutional neural network 80 of the comparative example, when learning data is small, the learning data is pulled to the estimation accuracy of the subject, and the task of estimating the subject of the dialogue cannot be performed accurately. There is.

  Furthermore, the convolutional neural network 80 of the comparative example also has a problem that it does not support multi-label output (linguistic ambiguity). This will be described with reference to FIG.

  FIG. 2 is a diagram for explaining that the convolutional neural network 80 of the comparative example does not support multi-label output.

  2 (a) and 2 (b), a subject estimation system using a neural network 80 that convolves a dialogue sentence that is a different topic but has the same word “rate” in the word string being talked about is the subject. The estimated case is shown. More specifically, in FIG. 2A, the topic is “ACCOMMODATION” (“accommodation”), and when the word string “The room rate is twenty dollars.” During the conversation, The subject label “Pricerange” estimated by the subject estimation system using the network 80 is shown. FIG. 2B shows a convolutional neural network 80 of the comparative example when the topic is “ATTRACTION” (“attraction”) and the word string “So kids have to pay the same rate as adults.” The subject label “Pricerange” estimated by the subject estimation system that uses is shown.

  In FIG. 2B, although the same word “rate” is present in the word string being talked about, the topic label “Fee” is correct because it is a different topic from FIG. However, in FIG. 2B, the same subject label “Pricerange” as in FIG. 2A is estimated. As described above, in the subject estimation system using the convolutional neural network 80 of the comparative example, when the same word “rate” is present in the word string being talked about but the subject changes in context (when the language expression is ambiguous) ), There is also a problem that it can not cope.

[The structure of the subject estimation system of this embodiment]
FIG. 3 is a diagram showing the structure of a convolutional neural network used by the subject estimation system in the present embodiment.

  The convolutional neural network 10 shown in FIG. 3 includes an input feature 11, a convolutional layer 12, a pooling layer 13, and a fully connected layer 14. As will be described in detail later, the input feature 11 is coupled to a topic-dependent convolution layer and a topic-independent convolution layer that constitute the convolution layer 12 according to the topic. Further, the pooling layer 13 is coupled to a topic-dependent convolution layer and a topic-independent convolution layer according to the topic. Further, the entire coupling layer 14 is coupled to the pooling layer 13 according to the topic.

  The input feature 11 receives a word string and converts the input word string into a vector string by a predetermined method. The input feature 11 may be a vector string obtained by converting a word string to be input by a predetermined method. That is, as the input feature 11, a word string vector sequence corresponding to the dialog text that transcribes the dialog may be input. More specifically, the input feature 11 calculates, by a predetermined method, a vector of each word in a word string included in the dialogue text that is a dialogue text in which the dialogue is transcribed and the dialogue is converted into text in time series. The word string vector sequence obtained (converted) may be input.

  The input feature 11 includes a topic-dependent convolution layer or a topic-independent layer constituting the convolution layer 12 according to the topic of one to a few words of adjacent vector sequences extracted by the time window. Coupled to a convolution layer (described below).

  In the example shown in FIG. 3, first, a word string vector sequence 11 a corresponding to the dialogue text “if you take a dorm bed per... Then, after that, a word string vector string 11b corresponding to the dialogue text “if you want to buy a spec... Here, a frame indicated by a thick line in the input feature 11 in FIG. 3 is a time window. FIG. 3 shows two types of time windows for one word and two words, as in FIG.

  The convolution layer 12 includes one or more topic-dependent convolution layers that perform topic-dependent convolution operations on word string vector sequence inputs, and the topic-independent convolution operations on the word string vector sequence inputs. It consists of one topic-independent convolution layer that performs The topic-dependent convolution layer performs a convolution operation between the word string vector sequence and a first weight that fires at a specific word indicating the dependent topic. The topic-independent convolution layer performs a convolution operation between the word string vector sequence and a second weight that is fired with a word indicating a topic other than the dependent topic (independent of the dependent topic).

  Thus, the convolution layer 12 has a topic-dependent convolution layer associated with a topic and a topic-independent convolution layer not associated with the topic. The input feature 11 is then coupled to a topic-dependent or topic-independent convolution layer depending on the topic.

  In the present embodiment, for example, it is assumed that there are two types of topics, #a and #b. Of course, the topic is not limited to two types.

  As shown in FIG. 3, the convolution layer 12 does not depend on the topic-dependent convolution layer 12a that performs the convolution operation depending on the topic #a and the topic-dependent convolution layer 12b that performs the convolution operation depending on the topic #b. It consists of three parts, a topic-independent convolution layer 12z that performs convolution operations.

  More specifically, the topic-dependent convolutional layer 12a includes a vector sequence of 1 to several words that are cut out in the word sequence vector sequence 11a, and a learned weight matrix that fires at a specific word indicating the topic #a ( A convolution operation with the first weight) is performed. The topic-dependent convolutional layer 12b includes a vector sequence of 1 to several words that are cut out in the word sequence vector sequence 11b and a learned weight matrix (first weight) that fires at a specific word indicating the topic #b. Perform a convolution operation. The topic-independent convolutional layer 12z includes an adjacent vector sequence of 1 to several words extracted in the word sequence vector sequence 11a and an adjacent vector sequence of 1 to several words extracted in the word sequence vector sequence 11b, and topic #a. And a convolution operation with a learned weight matrix (second weight) that is ignited by a word indicating a topic other than the topic.

  The pooling layer 13 performs a pooling process on the output of the convolution layer 12. More specifically, the pooling layer 13 performs an operation for extracting the maximum value in the time direction from the output of the topic-dependent convolution layer and the output of the topic-independent convolution layer.

  In the example shown in FIG. 3, when the word string vector sequence 11a is input to the input feature 11, the pooling layer 13a is coupled to the topic-dependent convolution layer 12a and the topic-independent convolution layer 12z. The pooling layer 13a performs an operation for extracting the maximum value in the time direction from these outputs. When the word string vector sequence 11b is input to the input feature 11, the pooling layer 13b is coupled to the topic-dependent convolution layer 12b and the topic-independent convolution layer 12z. The pooling layer 13b performs an operation for extracting the maximum value in the time direction from these outputs.

  The total coupling layer 14 performs a total coupling process on the output of the pooling layer 13. More specifically, the total coupling layer 14 performs probability distribution on the output of the pooling layer 13 after performing weighted addition using coupling weights. In the present embodiment, the total coupling layer 14 adds and adds the coupling weight to the output of the pooling layer 13 for each output element, and finally generates a probability distribution using the softmax function.

  In the example shown in FIG. 3, when the word string vector string 11a is input to the input feature 11, the pooling layer 13a and the fully connected layer 14a are combined. The total coupling layer 14a multiplies the output of the pooling layer 13a by a coupling weight and adds the result, and finally, a probability distribution is made using a softmax function. When the word string vector string 11b is input to the input feature 11, the pooling layer 13b and the fully connected layer 14b are combined. The total coupling layer 14b multiplies the output of the pooling layer 13b by a coupling weight, and finally adds a probability distribution using a softmax function.

  Then, the total connection layer 14 estimates and outputs the subject label of the dialogue by comparing the probability distribution of the output of the pooling layer 13 with a threshold value.

  In this way, the convolutional neural network 10 estimates the subject label of the dialogue for the input by solving the input as a two-class classification problem.

  In other words, in the all coupling layer 84 of the comparative example described in FIG. 1, there are output elements corresponding to all the subject labels, and the problem has been solved (learned) as a multi-class classification problem. On the other hand, the subject estimation system using the convolutional neural network 10 in this embodiment has a topic-dependent convolution layer specialized for each subject (topic) and a topic-independent convolutional layer not specialized for a topic. Therefore, as shown in FIG. 3, the subject label pricerange can be solved (learned) as a two-class classification problem of pricerange or not pricerange (NOT pricerange). As a result, even if there is a bias in the number of learning data for each subject, the learning results for a small number of subjects are not affected by the learning results for a large number of subjects. Is improved.

(Learning of convolutional neural network 10)
Here, learning of the convolutional neural network 10 in the present embodiment will be described.

  In the present embodiment, the learning data (training data) is a learning dialogue text that transcribes a dialogue, and the dialogue time-series text is divided in advance into segments for each topic, and for each divided segment. The learning dialogue text in which the label of the topic corresponding to is assigned in advance is used. As the learning dialogue text, for example, a DSTC4 (Dialog State Tracking Challenge 4) data set may be used.

  In the convolutional neural network 10, the first weight is learned so that each of the one or more topic-dependent convolutional layers 12a and 12b performs a convolution operation depending on the topic for each dependent topic, and the topic-independent convolutional layer 12z. Then, the second weight is learned so that the convolution operation independent of the dependent topic is performed. Each of the one or more topic-dependent convolution layers receives a word string vector sequence related to the dependent topic among the word string vector sequences corresponding to the learning dialogue text, and performs a convolution operation depending on the dependent topic. The first weight is learned. The topic-independent convolution layer 12z receives the word string vector sequence corresponding to the learning dialogue text and learns the second weight so as to perform a convolution operation that does not depend on the dependent topic.

  In the convolutional neural network 10, learning of the convolution weights (first weight and second weight) of the convolution layer 12 and the connection weights of all the connection layers 14 is performed based on a difference (error) between a desired output and an actual output. . As a learning algorithm for performing learning based on a difference (error) between a desired output and an actual output, a stochastic gradient descent method (SGD) or the like is known. Since the learning algorithm may be a known algorithm, a description thereof is omitted here.

  Such learning processing is performed on each of the topic-dependent convolutional layer 12a, the topic-dependent convolutional layer 12b, and the topic-independent convolutional layer 12z constituting the convolutional layer 12. As a result, the topic-dependent convolutional layers 12a and 12b are connected to a linguistic expression specialized for the topic. However, since the amount of interactive sentences of topics depending on each of the learning data is limited, the amount of learning data is relatively small. Less. On the other hand, in the topic-independent convolutional layer 12z, the amount of learning data is relatively large because the topic-independent convolutional layer 12z is linked to language expression independent of the topic.

  Then, both (topic-dependent convolutional layers 12a and 12b and topic-independent convolutional layer 12z) are configured to be connected at the subsequent stage (pooling layer 13 and all coupling layer 14) as shown in FIG. Therefore, a balance is taken by learning. That is, the connection weight between the topic-dependent convolutional layers 12a and 12b and the topic-independent convolutional layer 12z is adjusted for a topic with a large amount of learning data and a topic with a small amount of learning data.

  As a result, when a pick with a lot of learning data is input as a conversation sentence, the output of the topic-dependent convolution layer is emphasized, and when a topic with a small amount of learning data is input as a conversation sentence, the topic-independent convolution layer is output. Since the output is emphasized, the difference in performance due to the difference in the amount of learning data is reduced. In this configuration, it is possible to raise the performance (estimation accuracy) for a topic with particularly little learning data.

  Furthermore, since the topic-dependent convolution layers 12a and 12b and the topic-independent convolution layer 12z are configured to be connected at an upper layer (the pooling layer 13 and the fully connected layer 14) as shown in FIG. The problem described in 2 does not occur. That is, the theme estimation system in the present embodiment including the convolutional neural network 10 can also estimate the multi-domain dialogue theme by the convolutional neutral network.

[Functional configuration of the subject estimation system of the present embodiment]
Next, the learning time and identification time of the subject estimation system in the above-described embodiment will be specifically described. Below, it demonstrates using the functional block diagram and operation | movement figure at the time of the learning and identification of this subject estimation system.

(At identification)
FIG. 4 is a block diagram showing a functional configuration during identification of the subject estimation system in the present embodiment. Here, the convolutional neural network 10 used by the subject estimation system is described as a functional block diagram.

  The subject estimation system in the present embodiment shown in FIG. 4 includes an input unit 110, a word vector sequence control unit 111, a topic-dependent convolutional layer calculation unit 121, a topic-independent convolutional layer calculation unit 123, a pooling layer calculation unit 131, and a full combination. A layer calculation unit 141 and an output unit 150 are provided. Further, the subject estimation system includes a storage unit 122 that stores weight A (first weight), a storage unit 124 that stores weight Z (second weight), and a storage unit 142 that stores weight f. Here, the input unit 110 and the word vector sequence control unit 111 are functional configurations of the input feature 11 described above. The topic-dependent convolutional layer calculation unit 121, the storage unit 122, the topic-independent convolutional layer calculation unit 123, and the storage unit 124 are functional configurations of the convolutional layer 12 described above. The pooling layer calculation unit 131 has the functional configuration of the pooling layer 13 described above, and the total coupling layer calculation unit 141 and the storage unit 142 have the functional configuration of the total coupling layer 14 described above.

  In the subject estimation system shown in FIG. 4, input information 50 composed of word string information 51 and topic information 52 is input. The input unit 110 receives a word string from the word string information 51 and simultaneously receives a topic from the topic information 52. The word vector sequence control unit 111 converts the word sequence into a vector sequence by a predetermined method. Here, as a method for converting a word string into a vector, various methods called bag-of-words and a method of compressing the dimension have been proposed. If these known methods are the above-mentioned predetermined methods, Good.

  The topic-dependent convolution layer calculation unit 121 performs a convolution operation using the word string vector sequence and the weight A (first weight). This weight A (first weight) exists by the total number of filters for each topic. The topic-independent convolution layer calculation unit 123 performs a convolution operation using the word string vector string and the weight Z (second weight). There are as many weights Z (second weights) as there are topic-independent filters.

  The pooling layer calculation unit 131 takes out the maximum value by looking at the outputs corresponding to the filters of the topic-dependent convolutional layer calculation unit 121 and the topic-independent convolutional layer calculation unit 123 in the time direction.

  The total coupling layer calculation unit 141 multiplies the output corresponding to the filter for each output element by adding the weight f, and finally generates a probability distribution by the softmax function. With the Softmax function, for example, when the output subjects are “pricerange” and “NOT pricerange” as in the example shown in FIG. 3, the output of the two elements is 0 or more and the sum is 1. Adjusted.

  The output unit 150 compares a threshold value (for example, 0.5) with the result of the total connected layer calculation unit 141, and outputs a subject whose probability distribution exceeds the threshold value.

  The subject estimation system is executed by a computer having a hardware configuration as shown in FIG. FIG. 5 is a diagram illustrating an example of a hardware configuration of a computer necessary for executing the subject estimation system.

  As shown in FIG. 5, the computer that executes the present subject estimation system includes a CPU 1001, a memory 1002, an external storage device 1003, a network interface 1004, an output device 1006, and an input device 1007. These are connected by a bus.

  All operations of the subject estimation system are performed by the CPU 1001, and values and programs that require updating of weights and the like are stored in the memory 1002. A large amount of data such as learning data is stored in the external storage device 1003. The network interface 1004 is used to access data on the Internet 1005 and capture learning data from the outside. Also, an output device 1006 and an input device 1007 are necessary as a user interface. The input device 1007 is a device serving as a user interface such as an input button, a touch pad, or a touch panel display, and accepts a user operation.

  FIG. 6 is a flowchart showing the operation of the subject estimation system at the time of identification shown in FIG.

  First, the input unit 110 receives an input word string and a topic label (S101). Next, the word vector string control unit 111 calculates each word of the input word string by a predetermined method to obtain a vector string of words (word strings) (S102). Next, the topic-dependent convolution layer calculation unit 121 performs a convolution operation (topic-dependent convolution operation) between the word vector string and the weight A (first weight) stored in the storage unit 122 (S103). Next, the topic-independent convolutional layer calculation unit 123 performs a convolution operation (topic-independent convolution operation) between the vector sequence of words and the weight Z (second weight) stored in the storage unit 124 (S104). Next, the pooling layer calculation unit 131 performs a pooling process for extracting the maximum value of each filter from the output of the topic-dependent convolutional layer calculation unit 121 and the output of the topic-independent convolutional layer calculation unit 123 (S105). Next, the total connection layer calculation unit 141 performs a total connection layer process on the output of the pooling layer calculation unit 131 (S106). More specifically, the total connected layer calculation unit 141 weights the output of the pooling layer calculation unit 131 with the weight f stored in the storage unit 122 and adds the weighted values. Then, after performing weighted addition for all output labels, probability distribution is performed. Finally, the output unit 150 determines the output label by comparing the probability distribution with respect to the output label and the threshold (S107).

  Note that when the theme estimation system shown in FIG. 4 performs multi-label output, the processes of S103 and S104 may be performed in parallel, and the results thereof may be integrated at a later stage. The same applies to the following learning.

(During learning)
FIG. 7 is a block diagram showing a functional configuration during learning of the subject estimation system in the present embodiment. Elements similar to those in FIG. 4 are denoted by the same reference numerals, and detailed description thereof is omitted. The subject estimation system at the time of learning shown in FIG. 7 is executed by a computer having a hardware configuration as shown in FIG. 5 described above, similarly to the subject estimation system at the time of identification shown in FIG.

  7 is different from the functional configuration diagram at the time of identification shown in FIG. 4 in the learning data 60, the error determination unit 160, and the weight update unit 161.

  The subject string estimation system at the time of learning receives word string information 61 and topic information 62 as learning data (training data). The learning data 60 stores the subject information 63 which is the subject information 63 corresponding to the word string information 51 and the topic information 52 input to the subject estimation system at the time of learning, and is desirable as an output. ing.

  The error determination unit 160 sets the probability distribution of the desired label obtained from the probability distribution for the subject label output from the output unit 150 and the subject information 63 as 1.0, and the probability of other labels as 0.0. Compare with the probability distribution and output the difference between the probability distributions as an error.

  Based on the error value output from the error determination unit 160 by a predetermined learning algorithm, the weight update unit 161 updates the update amounts of the weight A (first weight), the weight Z (second weight), and the weight f, respectively. Determine and perform those updates. Such updating of the weight is repeatedly executed over the entire learning data while changing the learning coefficient.

  FIG. 8 is a flowchart showing the operation of the subject estimation system during learning shown in FIG. Note that the processing of S201 to S207 is the same as the processing of S101 to S107 shown in FIG.

  The subject estimation system at the time of learning estimates the probability distribution for the subject label based on the input word string and the topic by the processing up to S207. Next, the error determination unit 160 obtains a desired subject from the subject information 63, sets a desired value (probability) of the probability distribution with respect to the subject label, and determines the difference between the set value and the estimated probability distribution with respect to the subject label as an error. Is calculated as (S208). Next, the weight update unit 161 uses the weight A (first weight) used in the topic-dependent convolution layer calculation unit 121 and the weight f related to the current topic and the topic-independent convolution layer used in the all-connection layer calculation unit 141. The weight Z (first weight) used in the calculation unit 123 is updated by a predetermined learning algorithm (S209).

  Note that these learnings are performed to determine whether or not a preset termination condition is satisfied (S210), and are repeated until the termination condition is satisfied. This termination determination is made on the condition that the error is not improved even if each weight is updated, or on the condition that the error is equal to or less than the threshold value.

[Effects]
As described above, the subject estimation system using the convolutional neural network 10 according to the present embodiment can estimate the subject of the conversation with higher accuracy even when the learning data is not sufficient. The subject estimation system can also estimate multi-domain interaction themes.

  More specifically, the convolution layer 12 of the convolutional neural network 10 includes a topic-dependent topic-dependent convolution layer and a topic-independent topic-independent convolution layer, merges them in the pooling layer, Balance the output from the topic-dependent convolutional layer with the output from the topic-independent convolutional layer. As a result, when a pick with a lot of learning data is input as a conversation sentence, the output of the topic-dependent convolution layer is emphasized, and when a topic with a small amount of learning data is input as a conversation sentence, the topic-independent convolution layer is output. Since the output is emphasized, the difference in performance due to the difference in the amount of learning data is reduced.

  Here, an experimental verification result is shown that the estimation accuracy of the subject estimation system using the convolutional neural network 10 of the present embodiment is improved as compared with the subject estimation system using the neural network of the comparative example. It explains using.

  FIG. 9 is a diagram showing an experimental verification result of the subject estimation system in the present embodiment. FIG. 9 shows the result of comparing the accuracy of the subject estimation of the subject estimation system of the present embodiment and the comparative example when using the Dialog State Tracking Challenge 4 (DSTC4) dialogue corpus. The DSTC4 dialogue corpus includes dialogue of five domains (Attraction, Accommodation, Food, Shopping, Transportation). In each domain, 54 types of themes such as Pricerange, Preference, and Exhibit can be estimated for the dialogue section. FIG. 9 shows the result when the dialogue subject label is “Pricerange”. . Further, (42/30) of “ACCOMMODATION (42/30)” shown in FIG. 9 means that the number of learning data is 42 and the number of test data is 30.

  The general model of the comparative example is a theme estimation system using the convolutional neural network 80 shown in FIG. 1, for example, and means a theme estimation system in which all topics are learned by one convolutional neural network 80. Further, the Topic-specific Model of the comparative example means a case where a subject estimation system corresponding to a domain is configured for each domain, such as a subject estimation system using a neural network that learns only the interaction of the domain of Attraction. That is, it is a case where the neural network of another subject estimation system is trained for each topic.

  On the other hand, the multi-topic model is a theme estimation system using the convolutional neural network 10 shown in FIG. 3 and means the theme estimation system in the present embodiment.

  As shown in FIG. 9, in the experimental results, the multi-topic model has a higher accuracy rate of each domain when the dialogue subject label is “Pricerange” than that of the comparative example. It can also be seen that the multi-topic model is improved over the two comparative examples with respect to the F value (overall) indicating the estimation accuracy.

  Note that the estimation accuracy of all conversation subject labels using the DSTC4 dialogue corpus was 48% for Multi-topic model, 43% for General Model, and 43% for Topic-specific Model. It can be seen that the estimation accuracy is improved over the two comparative examples.

(Embodiment 2)
In Embodiment 1, it has been described that the estimation accuracy of subject estimation is improved by configuring the convolution layer 12 with a topic-dependent topic-dependent convolution layer and a topic-independent topic-independent convolution layer. In the configuration of this convolution layer 12, as described above, there is a tendency that learning data for the topic-dependent convolution layer is reduced. In this embodiment, in order to compensate for the shortage of learning data, a functional configuration diagram and an operation diagram are used when the subject estimation system using the convolutional neural network 10 described in the first embodiment uses semi-supervised learning. I will explain.

  FIG. 10 is a block diagram showing a functional configuration during additional learning of the subject estimation system in the present embodiment. Elements similar to those in FIG. 7 are denoted by the same reference numerals, and detailed description thereof is omitted. The subject estimation system at the time of additional learning shown in FIG. 10 is executed by a computer having a hardware configuration as shown in FIG. 5 described above, similarly to the subject estimation system at the time of learning shown in FIG.

  Compared with the functional configuration diagram at the time of learning shown in FIG. 7, the external data acquisition unit 170 is added to the functional configuration diagram at the time of additional learning shown in FIG.

  When the number of word string vector sequences related to a certain dependent topic is less than the number of word string vector sequences related to another dependent topic among the word string vector sequences corresponding to the learning dialogue text, the external data acquisition unit 170 In this case, a dialog text related to a certain dependent topic obtained by searching the Web is acquired as semi-teacher data of learning data.

  More specifically, when the content of the dialog data relates to travel planning, for example, the external data acquisition unit 170 acquires text information related to travel as unsupervised learning data from an Internet travel review site. However, the correct word label of the subject information is not given to the information on the word-of-mouth site of the travel like the learning data (supervised learning data) using the known data set described above. The topic label is not given to the information on the travel word-of-mouth site.

  Therefore, the subject estimation system according to the present embodiment increases the supervised learning data by giving a correct answer label to the information of the travel word-of-mouth site that is such unsupervised learning data. Thereby, supervised learning data regarding a topic with a small amount can be increased. More specifically, the subject estimation system shown in FIG. 10 performs an operation at the time of identification (identification processing in S101 to S107 shown in FIG. 6) to estimate a subject. In addition, the subject estimation system shown in FIG. 10 sequentially inputs all the labels for the topic of information on the word-of-mouth travel site. Furthermore, the topic label and the subject label are assigned only to those obtained by the above-described method so that the estimated probability of the theme of the output of the total connected layer calculation unit 141 corresponding to the topic is larger than a preset threshold value. .

  Next, the subject estimation system shown in FIG. 10 uses the information on the word-of-mouth site of the trip to which the topic label and the subject label are assigned, and again performs the learning operation (the learning process in S201 to S210 shown in FIG. 8). The identification process for the travel review site information and the learning process again are repeated. In the learning process that repeats in this way (semi-supervised learning process), the threshold value may be initially set high and gradually reduced. Also, when obtaining text data from a word-of-mouth site, for example, if the title is an exhibition, it can be expected that the content is related to the subject label exhibit, so set a related phrase for each subject label, It is effective to add restrictions such as. In addition, it is effective to limit the topic label from the title. In other words, the external data acquisition unit 170 acquires external unsupervised data such as a review text of a travel review site, and excludes data unrelated to the conversation subject from the acquired unsupervised learning data by keywords, Supervised learning data corresponding to useful predetermined dialog themes can be obtained.

  FIG. 11 is a flowchart showing the operation of the subject estimation system during additional learning shown in FIG.

  First, supervised learning data is prepared (S301). For example, dialogue data between humans is recorded by voice and written manually. Or, chat by keyboard input and save the text. Furthermore, in the dialogue, an annotation is given manually to identify what is the topic from where to where. Crowd sourcing can be used for annotation. However, since these operations are costly, the learning data is often insufficient to learn.

  Next, the subject estimation system of the present embodiment performs the learning process of S201 to S210 using the obtained supervised learning data (S302).

  Next, the external data acquisition unit 170 acquires unsupervised learning data related to the topic or subject (S303). Specifically, as described above, the external data acquisition unit 170 does not simply acquire by narrowing down the granularity of a travel word-of-mouth site, but uses a vocabulary related to a topic or subject label to obtain a title or other information. Using the headline, the narrowed-down data is acquired as unsupervised learning data (word string).

  Next, the subject estimation system of the present embodiment estimates the topic label and the subject label of unsupervised learning data using the convolutional neural network 10 learned by the learning process of S302, and the estimated result is obtained without the teacher. A topic label and a subject label are assigned to the learning data (S304).

  Next, the subject estimation system of the present embodiment performs the learning process of S201 to S210 again using the unsupervised learning data to which the topic label and the subject label are assigned (S305). Since the weight (first weight, second weight) of the convolutional neural network 10 changes between the learning result in S302 and the learning result in S305, the topic label and the subject label estimated in S304 also change accordingly.

  Next, the subject estimation system according to the present embodiment performs an end determination based on the amount of change in weight update or the like (S306). If the end condition is not satisfied, S304 and S305 are repeated.

[Effects]
As described above, the subject estimation system using the convolutional neural network 10 according to the present embodiment performs the learning process by identifying unsupervised learning data and assigning a correct answer label in a pseudo manner even when the learning data is not sufficient. By repeating it, supervised learning data can be increased sufficiently. Thereby, the theme estimation system of the present embodiment can estimate the theme of the dialog with higher accuracy.

  Here, the subject estimation system using the convolutional neural network 10 of the present embodiment will be described using experimental verification results to improve estimation accuracy by repeating the learning process as described above.

  FIG. 12 is a graph showing experimental verification results of the subject estimation system in the present embodiment. FIG. 12 shows the effect of the semi-supervised learning process performed by the subject estimation system of the present embodiment when the DSTC4 dialogue corpus is used.

  In this experiment, unsupervised data acquired from the outside by the semi-supervised learning process performed by the subject estimation system of the present embodiment in the two-class classification of whether the dialogue sentence belongs to the Exhibit theme or not belongs to the teacher data (supervised Data) to the original supervised data. Here, the original number of supervised learning data is 762, and the added number of unsupervised learning data is 20 to 753.

  As shown in FIG. 12, it can be seen that the binary classification accuracy is improved by 3% at maximum by performing the semi-supervised learning process on the unsupervised data acquired from the outside and adding it as supervised learning data.

  As described above, the subject estimation system and the subject estimation method of the present invention have been described in the first embodiment and the second embodiment, but there is no particular limitation on the subject or apparatus in which each process is performed. It may be processed by a processor or the like (described below) embedded in a specific device located locally. Further, it may be processed by a cloud server or the like arranged at a location different from the local device.

  Note that the present invention further includes the following cases.

  (1) Specifically, the above apparatus is a computer system including a microprocessor, ROM, RAM, a hard disk unit, a display unit, a keyboard, a mouse, and the like. A computer program is stored in the RAM or hard disk unit. Each device achieves its functions by the microprocessor operating according to the computer program. Here, the computer program is configured by combining a plurality of instruction codes indicating instructions for the computer in order to achieve a predetermined function.

  (2) A part or all of the constituent elements constituting the above-described apparatus may be constituted by one system LSI (Large Scale Integration). The system LSI is an ultra-multifunctional LSI manufactured by integrating a plurality of components on a single chip, and specifically, a computer system including a microprocessor, ROM, RAM, and the like. . A computer program is stored in the RAM. The system LSI achieves its functions by the microprocessor operating according to the computer program.

  (3) A part or all of the constituent elements constituting the above-described device may be constituted by an IC card that can be attached to and detached from each device or a single module. The IC card or the module is a computer system including a microprocessor, a ROM, a RAM, and the like. The IC card or the module may include the super multifunctional LSI described above. The IC card or the module achieves its function by the microprocessor operating according to the computer program. This IC card or this module may have tamper resistance.

  (4) Further, the present invention may be the method described above. Further, the present invention may be a computer program that realizes these methods by a computer, or may be a digital signal composed of the computer program.

  (5) In addition, the present invention provides a computer-readable recording medium such as a flexible disk, hard disk, CD-ROM, MO, DVD, DVD-ROM, DVD-RAM, BD ( It may be recorded on a Blu-ray (registered trademark) Disc), a semiconductor memory, or the like. The digital signal may be recorded on these recording media.

  In the present invention, the computer program or the digital signal may be transmitted via an electric communication line, a wireless or wired communication line, a network represented by the Internet, a data broadcast, or the like.

  The present invention may be a computer system including a microprocessor and a memory, wherein the memory stores the computer program, and the microprocessor operates according to the computer program.

  In addition, the program or the digital signal is recorded on the recording medium and transferred, or the program or the digital signal is transferred via the network or the like and executed by another independent computer system. You may do that.

  INDUSTRIAL APPLICABILITY The present invention can be used for a subject estimation system and a subject estimation method that perform a task of estimating what a conversation subject is, which is a natural language exchange between two people, regardless of a human or a machine.

10, 80 Convolutional neural network 11, 81 Input features 11a, 11b Word sequence vector sequence 12, 82 Convolution layer 12a, 12b Topic-dependent convolution layer 12z Topic-independent convolution layer 13, 13a, 13b, 83 Pooling layer 14, 14a, 14b , 84 Fully connected layer 50 Input information 51, 61 Word string information 52, 62 Topic information 60 Learning data 63 Subject information 110 Input unit 111 Word vector string control unit 121 Topic-dependent convolutional layer calculation unit 122, 124, 142 Storage unit 123 Topic Independent convolutional layer calculation unit 131 Pooling layer calculation unit 141 Fully connected layer calculation unit 150 Output unit 160 Error determination unit 161 Weight update unit 170 External data acquisition unit 1001 CPU
1002 Memory 1003 External storage device 1004 Network interface 1005 Internet

Claims (12)

A subject estimation system for estimating a subject label of a dialog, comprising a convolutional neural network,
The convolutional neural network is:
One or more topic-dependent convolutional layers that perform topic-dependent convolution operations on an input of a word string vector sequence corresponding to the dialog text that transcribes the conversation, and one topic non-conformation that performs convolution operations that do not depend on the topic A convolution layer composed of a dependent convolution layer;
A pooling layer that performs a pooling process on the output of the convolution layer;
A total coupling layer that performs a total coupling process on the output of the pooling layer,
Thematic estimation system. The convolutional neural network estimates the subject label of the dialog for the input by solving the input as a two-class classification problem;
The subject estimation system according to claim 1. In the convolutional neural network, a dialogue text for learning that transcribes a dialogue, the time-series text of the dialogue is divided in advance into segments for each topic, and a label for a topic corresponding to each divided segment is obtained in advance. The first weight is used to cause each of the one or more topic-dependent convolutional layers to perform a convolution operation depending on the dependent topic for each dependent topic that is a dependent topic, using the given learning dialogue text as learning data. And learning a second weight so that the topic-independent convolution layer performs a convolution operation that does not depend on the dependent topic.
The subject estimation system according to claim 1. Each of the one or more topic-dependent convolutional layers receives a word string vector string related to the dependent topic from among word string vector strings corresponding to the learning dialogue text, thereby performing a convolution operation depending on the dependent topic. Learning the first weight to perform
The topic-independent convolution layer learns the second weight so as to perform a convolution operation independent of the dependent topic by inputting a word string vector sequence corresponding to the learning dialogue text.
The subject estimation system according to claim 3. A subject estimation method for a subject estimation system, comprising a convolutional neural network, for estimating a subject label of a dialogue, comprising:
A topic-dependent convolution processing step for performing a topic-dependent convolution operation on an input of a word string vector sequence corresponding to the dialog text that transcribes the dialog;
A topic-independent convolution processing step for performing a topic-independent convolution operation on the input;
A pooling processing step for performing a pooling process on the output of the topic-dependent convolution processing step and the output of the topic-independent convolution processing step;
A full join process step for performing a full join process on the output of the pooling process step,
Thematic estimation method. In the topic-dependent convolution processing step, a convolution operation is performed between the word string vector sequence and a first weight (A) that fires at a specific word indicating a dependent topic that is a dependent topic,
In the topic-independent convolution process step, a convolution operation is performed between the word string vector sequence and a second weight (Z) that fires at a word indicating a topic other than the dependent topic,
In the pooling processing step, an operation for extracting a maximum value in the time direction from the output of the topic-dependent convolution processing step and the output of the topic-independent convolution processing step is performed,
In the total connection processing step, after performing weighted addition using the connection weight on the output of the pooling processing step, the total connection processing is performed by performing probability distribution.
The subject estimation method according to claim 5. An output step of estimating and outputting a subject label of the dialogue by comparing a probability distribution of an output of the pooling processing step with a threshold value,
The subject estimation method according to claim 6. An input step of inputting a word string vector sequence corresponding to the dialogue text;
The input step includes
Further, an accepting step of accepting a dialog text that transcribes the dialog, wherein the dialog is converted into text in time series, and
A vectorization step of obtaining a word string vector sequence by calculating a vector of each word of the word sequence included in the dialogue text by a predetermined method,
The subject estimation method according to claim 5. further,
In the convolutional neural network, learning dialogue text in which dialogue is transcribed, the time-series text of the dialogue is divided in advance into segments for each topic, and a label for a topic corresponding to each divided segment is preliminarily obtained. A first step of learning the first weight so as to perform a convolution operation depending on a dependent topic that is a topic on which the topic-dependent convolution processing step depends, using the given learning dialogue text as learning data;
Using the learning dialogue text, the topic-independent convolution processing step includes a second step of learning the second weight so as to perform a convolution operation independent of the dependent topic.
The subject estimation method according to claim 6. In the first step, the first weight is learned using a word string vector string related to the dependent topic among word string vector strings corresponding to the learning dialogue text;
In the second step, the second weight is learned using a word string vector string related to a topic other than the dependent topic among word string vector strings corresponding to the learning dialogue text.
The subject estimation method according to claim 9. further,
When the number of word string vector sequences related to the first dependent topic is less than the number of word string vector sequences related to the second dependent topic among the word string vector sequences corresponding to the learning dialogue text, the Web is searched. Using the dialogue text related to the first dependent topic obtained as the semi-teacher data of the learning data, the first step and the second step are performed.
The subject estimation method according to claim 9 or 10. A computer readable program comprising a convolutional neural network for estimating a subject label for dialogue,
A topic-dependent convolution processing step for performing a topic-dependent convolution operation on an input of a word string vector sequence corresponding to the dialog text that transcribes the dialog;
A topic-independent convolution processing step for performing a topic-independent convolution operation on the input;
A pooling processing step for performing a pooling process on the output of the topic-dependent convolution processing step and the output of the topic-independent convolution processing step;
A full join process step for performing a full join process on the output of the pooling process step,
program.

Images