JP7517435B2 - Learning device, inference device, their methods, and programs - Google Patents

Description

The present invention relates to labeling technology.

In recent years, a technology called speech sequence labeling has been proposed for the purpose of understanding conversations and discourse, which takes a speech sequence as input and estimates a label representing the conversation or discourse scene for each utterance (for example, non-patent document 1).

For example, non-patent document 1 discloses a deep neural network (hereinafter referred to as a labeling network) that realizes speech sequence labeling by taking as input a speech text sequence obtained by speech recognition of a conversation between an operator and a customer at a contact center and estimating a label for each utterance as the response scene (opening, understanding the purpose, identity verification, response, or closing).

A large amount of labeled training data is required for learning a labeling network such as that in Non-Patent Document 1. However, it is difficult to collect a large amount of labeled training data in a new domain every time labeling is performed in that domain because the cost of labeling is enormous. Here, Non-Patent Document 2 proposes a method for realizing unsupervised domain adaptation that performs labeling in a new domain from labeled data (hereinafter, labeled training data) in a domain that has been previously applied (hereinafter, source domain) and unlabeled data (hereinafter, unlabeled training data) in a domain to which the new domain is to be applied (hereinafter, target domain).

R. Masumura, S. Yamada, T. Tanaka, A. Ando, H. Kamiyama, and Y. Aono, “Online call scene segmentation of contact center dialogues based on role aware hierarchical LSTM-RNNs,” Proceedings of the Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pp. 811-815, 2018. Y. Ganin and V. Lempitsky, “Unsupervised domain adaptation by backpropagation,” Proceedings of the 32nd International Conference on Machine Learning (ICML), vol. 37, pp. 1180-1189, 2015.

However, the method of Non-Patent Document 2 uses labeled training data of the source domain and unlabeled training data of the target domain to learn a labeling model that estimates a label corresponding to a single image belonging to the target domain. In other words, the method of Non-Patent Document 2 performs unsupervised domain adaptation for a simple classification problem of a single image, and no unsupervised domain adaptation method has been established for a complex classification problem that estimates a label sequence corresponding to a sequence of multiple pieces of information by taking into account the logical relationship between the series of information (for example, estimating a label sequence for each corresponding scene for a spoken text sequence).

The present invention has been made in consideration of these points, and aims to perform unsupervised domain adaptation of a labeling model that estimates a label sequence corresponding to multiple information sequences by taking into account the logical relationships between the sequences.

A labeling model including a logical relationship understanding means that receives an input information sequence, which is a sequence of multiple pieces of information having a logical relationship, obtains an intermediate feature sequence that takes into account the logical relationship of the input information sequence, and outputs the intermediate feature sequence; a labeling means that receives a first sequence based on the intermediate feature sequence, obtains an estimated label sequence of a label sequence corresponding to the input information sequence, and outputs the estimated label sequence; and a domain identification model that receives a second sequence based on the intermediate feature sequence, obtains estimated domain information of domain identification information indicating whether each piece of partial information included in the input information sequence belongs to a source domain or a target domain, and outputs the estimated domain information sequence. A learning device uses teacher data including labeled teacher data, which is a labeled learning information sequence belonging to the source domain, and unlabeled teacher data, which is an unlabeled learning information sequence belonging to the target domain, as the input information sequence, and learns the labeling model so that the estimation accuracy of the estimated label sequence is high and the estimation accuracy of the estimated domain information sequence is low, and performs adversarial learning to learn the domain identification model so that the estimation accuracy of the estimated domain information sequence is high, and obtains and outputs at least parameters of the labeling model.

This enables unsupervised domain adaptation of a labeling model that estimates a label sequence corresponding to multiple information sequences by taking into account the logical relationships between those sequences.

FIG. 1 is a block diagram illustrating a learning device according to the first embodiment. FIG. 2 is a block diagram illustrating a detailed configuration of the learning unit of the first embodiment. FIG. 3 is a conceptual diagram illustrating a network used in the learning process of the first embodiment. FIG. 4 is a block diagram illustrating the inference device of the first embodiment. FIG. 5 is a conceptual diagram illustrating the trained labeling network of the first embodiment. FIG. 6 is a block diagram illustrating the learning device of the second embodiment. FIG. 7 is a conceptual diagram illustrating a network used in the learning process of the second embodiment. FIG. 8 is a graph illustrating the experimental results. FIG. 9 is a block diagram illustrating the hardware configuration of the learning device and the inference device according to the embodiment.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In each embodiment, an example of unsupervised domain adaptation of a labeling model based on a deep neural network that takes a spoken text sequence (a sequence of multiple pieces of information having a logical relationship) as input and outputs a sequence of labels (label sequence) corresponding to a corresponding scene (e.g., opening, understanding the subject matter, identity verification, response, closing) (sequence labeling) is shown. However, these are only examples and do not limit the present invention. That is, the present invention can be used for unsupervised domain adaptation of a labeling model that estimates an arbitrary label sequence corresponding to an arbitrary sequence of multiple pieces of information by considering the logical relationship of the sequence of the information. Note that there is no limitation on the logical relationship of the sequence of multiple pieces of information, and it is sufficient that some relationship exists between the multiple pieces of information. Examples of logical relationships include context, dependency relationships between words, grammatical relationships in languages, and inter-frame relationships in audio and video, but these do not limit the present invention. In addition, the labeling model is not limited to a model based on a deep neural network, and any model such as a probability model or a classifier may be used as long as it estimates and outputs a label sequence corresponding to an input sequence of information.

[First embodiment]
A first embodiment of the present invention will be described.
<Functional configuration of the learning device 11 and learning process>
As illustrated in FIG. 1, the learning device 11 of the first embodiment includes a learning unit 11a and storage units 11b and 11c, and receives labeled teacher data of the source domain and unlabeled teacher data of the target domain as input, and obtains and outputs parameters (model parameters) of a labeling network of the target domain by learning. In the drawings, for the sake of simplicity, the source domain is represented as "SD", the target domain as "TD", and the network as "NW". The labeled teacher data of the source domain illustrated here is a labeled learning information sequence belonging to the source domain, and includes an utterance text sequence of the source domain (an input information sequence that is a sequence of multiple pieces of information having a logical relationship) and a correct label sequence corresponding to the utterance text sequence. The unlabeled teacher data of the target domain illustrated here is an unlabeled learning information sequence belonging to the target domain, and includes an utterance text sequence of the target domain but does not include a correct label sequence. Furthermore, a learning schedule, which is a schedule of a combination ratio of losses and the like to be described later, may be input to the learning device 11, and the learning device 11 may perform a learning process according to the learning schedule. The learning device 11 may also output parameters of a domain identification network for realizing unsupervised domain adaptation. As illustrated in FIG. 2, the learning unit 11a includes, for example, a control unit 11aa, a loss function calculation unit 11ab, a gradient inversion unit 11ac, and a parameter update unit 11ad. The learning unit 11a stores each piece of data obtained during the processing in the storage units 11b and 11c or a temporary memory (not shown). The learning unit 11a reads the data as necessary and uses it for each process.

<Network 100>
3 shows an example of the configuration of the network 100 used in the learning process by the learning device 11. The network 100 shown in FIG.

<Labeling Network 150>
The labeling network 150 illustrated in FIG. 3 receives (as input) a spoken text sequence T ₁ , ..., T _N (an input information sequence that is a sequence of a plurality of pieces of information having a logical relationship), obtains and outputs an estimated label sequence L ₁ , ..., L _N that is an estimated sequence of a sequence of labels corresponding to the spoken text sequence T ₁ , ..., T _N. Here, the spoken text sequence T ₁ , ..., T _N is a sequence of N spoken texts T _n . Here, n is an index corresponding to time, n=1, ..., N, for example, and N is an integer equal to or greater than 1, and generally N is an integer equal to or greater than 2. The spoken text T _n may be a word such as "I'm sorry" or "Yes," or may be a sentence including M(n) words T _n,1 , ..., T _n,M(n) such as "I'm sorry for the slow reply speed." Here, M(n) is an integer equal to or greater than 2. The estimated label sequence _L1 , ..., _LN illustrated here is a sequence of N estimated labels _Ln (where n = 1, ..., N). The estimated labels _Ln in this example correspond to the spoken text _Tn , and represent, for example, corresponding scenes of the spoken text _Tn (for example, opening, understanding the subject matter, identity verification, response, closing). The labeling network 150 illustrated here has a logical relationship understanding layer 110 (logical relationship understanding means) and a labeling layer 120 (labeling means).

<Logical Relationship Understanding Layer 110>
The logical relationship understanding layer 110 receives a spoken text sequence T ₁ , ..., T _N , obtains an intermediate feature sequence LF ₁ , ..., LF _N considering the context (logical relationship) of the spoken text sequence T ₁ , ..., T _N , and outputs the intermediate feature sequence LF ₁ , ..., LF _N. The intermediate feature sequence LF ₁ , ..., LF _N is a sequence of N intermediate features LF _n (where n=1, ..., N). The intermediate feature LF _n corresponds to the spoken text T _n . The logical relationship understanding layer 110 illustrated in FIG. 3 includes short-term context understanding networks 111-1, ..., 111-N (short-term logical relationship understanding means) and a long-term context understanding network 112 (long-term logical relationship understanding means). For example, the short-term context understanding networks 111-1, ..., 111-N are mutually identical networks (e.g., networks with the same parameters), and each short-term context understanding network 111-n represents a state corresponding to each n=1, ..., N (e.g., each time). The short-term context understanding networks 111-1, ..., 111-N illustrated here receive a spoken text sequence T ₁ , ..., T _N. Each short-term context understanding network 111-n (where n= ₁ , ..., N) that receives each spoken text T _n (each partial information included in the input information sequence) included in the spoken text sequence T 1 , ..., T _N obtains short-term intermediate features SF n (short-term intermediate features that consider the logical relationship of information in the partial information) that take into account the context of words in the received spoken text T _n (e.g., short-term context on a word-by-word basis), and outputs the short-term intermediate features _{SF n .} As a result, a series of short-term intermediate features SF ₁ , ..., SF _N are output from the short-term context understanding networks 111-1, ..., 111-N. Note that when the speech text T _n includes only one word, the context of the word in the speech text T _n depends only on the one word, and the SF _n obtained in this case is also a short-term intermediate feature that takes into account the context of the word. However, this does not limit the present invention. For example, each short-term context understanding network 111-n may be divided into a plurality of short-term context understanding networks 111-n1, ..., 111-nK' (where K' is an integer equal to or greater than 2). For example, k'=1, ..., K' is an index representing a layer of the short-term context understanding network, and each short-term context understanding network 111-nk' represents a network from the input layer to the k'th layer of the short-term context understanding network. In this case, a short-term intermediate feature SF _nk' of the k'th layer is output from each short-term context understanding network 111-nk'. The long-term context understanding network 112 illustrated here receives a sequence of multiple short-term intermediate features SF ₁ , ..., SF _N (short-term intermediate feature sequence), obtains intermediate feature sequences LF ₁ , ..., LF _N (long-term intermediate feature sequence considering logical relationships between multiple pieces of partial information included in an input information sequence) taking into account the context between multiple spoken texts T _n included in the spoken text sequence T ₁ , ..., T _N (for example, a long-term context of a sentence unit or a long-term context spanning multiple sentences), and outputs the intermediate feature sequence LF ₁ , ..., LF _N. However, this does not limit the present invention. For example, when short-term intermediate features SF _nk' of the k'th layer are output from each short-term context understanding network 111-nk' as described above, SF _nK' may be input to the long-term context understanding network 112 as SF _n , or multiple SF nK _' of SF _n1 , ..., SF _nK' may be input. In addition, the long-term context understanding network 112 may be divided into a plurality of long-term context understanding networks 112-1, ..., 112-K (where K is an integer of 2 or more). For example, k = 1, ..., K is an index representing a layer of the long-term context understanding network, and each long-term context understanding network 112-k represents a network from the input layer to the k-th layer of the long-term context understanding network. In this case, each long-term context understanding network 112-k (where k = 1, ..., K) may receive a series of a plurality of short-term intermediate features SF _n , and obtain and output intermediate features that consider the context between a plurality of speech texts T _n corresponding to the received series SF _n . In this case, K intermediate feature series {LF ₁₁ , ..., LF _N1 }, ..., {LF _1K , ..., LF _NK } are output by the long-term context understanding networks 112-1, ..., 112-K. LF _nk (where n=1, . . . , N, k=1, . . . , K) represents intermediate features corresponding to each n (eg, each time) in the kth layer output from each long-term context understanding network 112-k.

Here, the short-term context understanding network 111-n can be configured, for example, by combining an embedding layer that converts words into numerical values using a dictionary, a unidirectional LSTM (long-short term memory), a bidirectional LSTM, an attention mechanism, etc. (For example, see Non-Patent Document 1, etc.). The long-term context understanding network 112 can be configured, for example, by combining a unidirectional LSTM, a bidirectional LSTM, etc.

<Labeling layer 120>
The labeling layer 120 receives the intermediate feature sequence LF ₁ , ..., LF _N (first sequence based on the intermediate feature sequence), obtains an estimated label sequence L ₁ , ..., L _N corresponding to the spoken text sequence T ₁ , ..., T _N , and outputs the estimated label sequence L ₁ , ..., L _N. The labeling layer 120 illustrated in FIG. 3 includes label prediction networks 120-1, ..., 120-N. For example, the label prediction networks 120-1, ..., 120-N are identical networks (e.g., networks with identical parameters), and each label prediction network 120-n represents a state corresponding to each n=1, ..., N (e.g., each time). The label prediction network 120-n illustrated here receives the intermediate feature LF _n , obtains an estimated label L _n corresponding to the spoken text T _n , and outputs the estimated label L _n . The label prediction network 120-n can be configured, for example, by a fully connected neural network with a softmax function as an activation function. In addition, when K intermediate feature sequences {LF ₁₁ , ..., LF _N1 }, ..., {LF _1K , ..., LF _NK } are output from the long-term context understanding networks 112-1, ..., 112-K, the label prediction network 120-n receives, for example, LF _nK as the intermediate feature LF _n , obtains an estimated label L _n corresponding to the spoken text T _n , and outputs the estimated label L _n . However, the label prediction network 120-n may receive a plurality of intermediate features LF _nk from the intermediate feature sequence {LF ₁₁ , ..., LF _N1 }, ..., {LF _1K , ..., LF _NK }, obtains an estimated label L _n corresponding to the spoken text T _n , and outputs the estimated label L _n .

<<Domain Identification Model 130>>
The domain identification model 130 illustrated in FIG. 3 receives the intermediate feature sequence LF ₁ , ..., LF _N (a second sequence based on the intermediate feature sequence), obtains estimated domain information D n (where n= ₁ , ..., _N ) of domain identification information indicating whether each spoken text T _n (each piece of partial information included in the input information sequence) included in the spoken text sequence T 1 , ..., T N belongs to the source domain or the target _domain (whether each spoken text T _n belongs to the source domain or the target domain), and outputs the sequence D ₁ , ..., D _N of the estimated domain information. The domain identification model 130 illustrated here includes N domain identification networks 130-1, ..., 130-N. For example, the domain identification networks 130-1, ..., 130-N are identical networks (for example, networks with the same parameters), and each domain identification network 130-n represents a state corresponding to each n=1, ..., N (for example, each time). For example, each domain identification network 130-n (where n=1,...,N) receives intermediate features LF _n , obtains estimated domain information D _n , and outputs it. However, this does not limit the present invention. For example, instead of each domain identification network 130-n, there may be a plurality of domain identification networks 130-nk. For example, k=1,...,K is an index representing a layer of the long-term context understanding network, and each domain identification network 130-nk represents a network corresponding to each n (e.g., each time). In this case, each domain identification network 130-nk may receive intermediate features LF _nk (nε{1,...,N},kε{1,...,K}), and use them to obtain estimated domain information D _nk and output it. D _nk (where n=1,...,N,k=1,...,K) represents estimated domain information corresponding to each n (e.g., each time) output from each domain identification network 130-nk. The domain identification network 130-n (or the domain identification network 130-nk) can be configured, for example, by a fully connected neural network using a softmax function as an activation function.

<Learning process>
In the learning process, labeled teacher data of the source domain (a labeled training information sequence belonging to the source domain) and unlabeled teacher data of the target domain (unlabeled training information sequence belonging to the target domain) are input to the learning unit 11a of the learning device 11. The learning unit 11a performs adversarial learning on the above-mentioned network 100, using teacher data including the labeled teacher data of the source domain and the unlabeled teacher data of the target domain as the spoken text sequence _T1 , ..., _TN , to learn a labeling network 150 (labeling model) so that the estimation accuracy of the estimated label sequence _L1 , ..., _LN is high and the estimation accuracy of the estimated domain information sequence _D1 , ..., _DN is low, and to learn a domain discrimination model 130 so that the estimation accuracy of the estimated domain information sequence _D1 , ..., _DN is high. That is, the learning unit 11a _{performs adversarial learning between the labeling network 150 and the domain discrimination model 130 based on a loss function (hereinafter, label prediction loss) that represents an error between the estimated label sequence L 1 , ...} , _LN output from the labeling network 150 when the above-mentioned teacher data is input to the network 100 as the speech text sequence T ₁ , ..., T _N and the corresponding correct label sequence of labeled teacher data of the source domain, and a loss function (hereinafter, domain discrimination loss) that represents an error between the estimated domain information sequence D 1 , ..., D N output from the domain discrimination model 130 and the correct label sequence of estimated domain information identified from the labeled teacher data of the source domain and the unlabeled teacher data of the target domain. Note that the estimated label sequence L ₁ , ..., _LN output from the labeling network 150 when the unlabeled teacher data of the target domain is input to the network 100 is not used to calculate the label prediction loss.

The learning unit 11a performs this adversarial learning using, for example, the backpropagation method. In this case, a gradient inversion layer 141-n (where n=1, ..., N) is provided between the logical relationship understanding layer 110 and the domain identification model 130 (for example, between the long-term context understanding network 112-n and the domain identification network 130-n), and the gradient is inverted in the gradient inversion layer 141-n only during backpropagation. Here, by performing learning so as to reduce the label prediction loss, the estimation accuracy of the estimated label sequence L ₁ , ..., L _N in the labeling network 150 is improved. In addition, the gradient inversion layer 141-n inverts the gradient only during backpropagation of errors, and learning is performed to reduce domain discrimination loss, thereby learning the domain discrimination model 130 so as to increase the estimation accuracy of the estimated domain information sequence D ₁ , ..., D _N , and learning the logical relationship understanding layer 110 to obtain the intermediate feature sequence LF ₁ , ..., LF _N that reduces the estimation accuracy of the estimated domain information sequence D ₁ , _... , D N. This adversarial learning makes it possible to learn the labeling network 150 that can accurately estimate the estimated label sequence L ₁ , ..., L _N but can generate the intermediate feature sequence LF ₁ , ..., LF _N whose domain is not estimated by the domain discrimination model 130. As a result, the labeling network 150 can acquire the intermediate feature sequence LF ₁ , ..., LF _N that is effective for label prediction while suppressing the dependency on the domain, thereby realizing unsupervised domain adaptation.

This learning process can be realized by optimizing (minimizing) a loss function that is a linear combination of the label prediction loss and the domain classification loss. The combination ratio of the linear combination of the label prediction loss and the domain classification loss may be predetermined or may be specified by a learning schedule input to the learning unit 11a.

The learning unit 11a may perform the above-mentioned learning while changing the combination ratio of the label prediction loss and the domain discrimination loss according to the number of learning steps based on a learning schedule. For example, the learning unit 11a may perform learning using only the label prediction loss as a loss function in the early stages of learning, and as the number of learning steps increases, the proportion of the domain discrimination loss in the loss function gradually increases. Furthermore, the learning unit 11a may perform learning until the learning converges with a constant combination ratio, change the combination ratio, and perform learning until it converges again, repeatedly performing learning while changing the combination ratio based on the learning schedule.

The learning unit 11a may also prepare and learn a variety of domain identification models 130 and/or labeling networks 150 such as those exemplified above, and later select, from among the labeling networks 150 obtained by each learning, the labeling network 150 that provides the best estimation accuracy of the label sequence in the target domain.

The training process may be batch training, mini-batch training, or online training.

The learning unit 11a stores the parameters of the labeling network 150 obtained by the above-mentioned learning in the memory unit 11b, and stores the parameters of the domain identification model 130 (parameters of domain identification networks 130-1, ..., 130-) in the memory unit 11c. The learning device 11 outputs the parameters of the labeling network 150 stored in the memory unit 11b. The parameters of the labeling network 150 are used in the inference process described below. Normally, the parameters of the domain identification model 130 are not used in the inference process, and therefore do not need to be output from the learning device 11. However, the learning device 11 may output at least any of the parameters of the domain identification model 130 (parameters of domain identification networks 130-1, ..., 130-).

The above-mentioned learning process will be functionally illustrated with reference to FIG.
Step S11: Labeled teacher data of the source domain and unlabeled teacher data of the target domain are input to the control unit 11aa. The control unit 11aa generates teacher data including the labeled teacher data of the source domain and the unlabeled teacher data of the target domain. The control unit 11aa also initializes parameters of the network 100.
Step S12: The loss function calculation unit 11ab inputs the training data as spoken text sequences T ₁ , . . . , T _N to the network 100, obtains a label prediction loss and a domain discrimination loss, and obtains a loss function by linearly combining them.
Step S13: The parameter update unit 11ad backpropagates information based on the loss function according to the backpropagation method, and updates the parameters of the domain identification model 130 and the labeling layer 120.
Step S14: The gradient inversion unit 11ac inverts the gradient of the information based on the loss function backpropagated from the domain discrimination model 130 and backpropagates it to the logical relationship understanding layer 110. The information based on the loss function backpropagated from the labeling layer 120 is backpropagated to the logical relationship understanding layer 110 as it is.
Step S15: The parameter update unit 11ad updates the parameters of the logical relationship understanding layer 110 using the backpropagated information according to the backpropagation method.
Step S16: The control unit 11aa judges whether or not a termination condition (for example, a condition that the number of parameter updates has reached a predetermined number) has been satisfied. If the termination condition has not been satisfied, the control unit 11aa returns the process to step S12. On the other hand, if the termination condition has been satisfied, the control unit 11aa outputs the parameters of the labeling network 150. If necessary, the control unit 11aa may also output at least one of the parameters of the domain identification networks 130-1, ..., 130-N.

<Functional configuration and inference processing of the inference device 13>
4, the inference device 13 of the first embodiment includes an inference unit 13a and a storage unit 13b. The storage unit 13b stores parameters of the labeling network 150 obtained as described above.

Inference processing
In the inference process, an utterance text sequence for inference (input information sequence) is input to the inference unit 13a. The inference unit 13a applies the utterance text sequence for inference to a labeling network 150 (labeling model) specified by parameters stored in the storage unit 13b, obtains an estimated label sequence of a label sequence corresponding to the utterance text sequence for inference, and outputs the estimated label sequence. For example, in the case of the labeling network 150 illustrated in FIG. 5, the inference unit 13a inputs the utterance text sequence for inference T ₁ , ..., T _N to the logical relation understanding layer 110 to obtain an intermediate feature sequence LF ₁ , ..., LF _N corresponding to the utterance text sequence for inference T ₁ , ..., T _N. For example, the inference unit 13a inputs a spoken text sequence T ₁ , ..., T _N for inference to the short-term context understanding networks 111-1, ..., 111-N, respectively, to obtain a sequence of short-term intermediate features SF ₁ , ..., SF _N , and inputs the sequence of short-term intermediate features SF ₁ , ..., SF _N to the long-term context understanding network 112 to obtain an intermediate feature sequence LF ₁ , ..., LF _N. Furthermore, the inference unit 13a inputs the intermediate feature sequence LF ₁ , ..., LF _N to the labeling layer 120 to obtain and output an estimated label sequence L ₁ , ..., L _N corresponding to the spoken text sequence T ₁ , ..., T _N.

<Features of the First Embodiment>
In this embodiment, the labeling network 150 includes _a logical relationship understanding layer 110 that receives a spoken text sequence T ₁ , ..., _TN , obtains an intermediate feature sequence LF ₁ , ..., LF _N taking into account the context of the spoken text sequence T ₁ , ..., _TN , and outputs the intermediate feature sequence LF ₁ , ..., LF _N , and a labeling layer 120 that receives the intermediate feature sequence LF ₁ , ..., LF _N , obtains an estimated label sequence L ₁ , ..., LN corresponding to the spoken text sequence T ₁ , ..., _TN , and outputs the estimated label sequence L ₁ , ..., _LN . The labeling network 150 includes a logical relationship understanding layer 110 that receives a spoken text sequence T ₁ , ..., _TN , obtains estimated domain information D _n that is domain identification information indicating whether each spoken text T _n included in the spoken text sequence T ₁ , ..., _TN belongs to the source domain or the target domain, and outputs the estimated domain information sequence D ₁ , ..., D For the domain discrimination model 130 that outputs _{L N} , the learning device 11 performed adversarial learning using teacher data including labeled teacher data of the source domain and unlabeled teacher data of the target domain as the spoken text sequence T ₁ , ..., T _N to train the labeling network 150 so that the estimation accuracy of the estimated label sequence L ₁ , ..., L _N is high and the estimation accuracy of the estimated domain information sequence D ₁ , ..., D _N is low, and to train the domain discrimination model 130 so that the estimation accuracy of the estimated domain information sequence D ₁ , ..., D _N is high. This reduces the domain dependency of the labeling network 150, and as a result, unsupervised domain adaptation of the labeling network 150 that estimates the label sequence L ₁ , ..., L _N corresponding to the spoken text sequence T ₁ , ..., T N by taking into account the context of the spoken text sequence T 1 , ..., T _N becomes possible.

[Second embodiment]
In the second embodiment, a network for identifying a domain from the context between a plurality of spoken texts _Tn (a long-term intermediate feature sequence taking into account the logical relationship between a plurality of pieces of partial information included in an input information sequence) and a network for identifying a domain from short-term intermediate features _SFn (short-term intermediate features taking into account the logical relationship between information in a partial information) taking into account the context of words in the spoken _text Tn are simultaneously used for adversarial learning. This makes it possible to more efficiently remove domain dependency and learn a labeling network for a target domain with higher accuracy. The following description will focus on the differences from the first embodiment, and the same reference symbols will be used to simplify the description of matters common to the first embodiment.

<Functional configuration of the learning device 21 and learning process>
As illustrated in FIG. 6, the learning device 21 of the second embodiment has a learning unit 21a and storage units 11b, 21c, and 21d, and receives labeled teacher data of the source domain and unlabeled teacher data of the target domain as input, and obtains and outputs parameters (model parameters) of the labeling network of the target domain by learning. Furthermore, a learning schedule may be input to the learning device 21, and the learning device 21 may perform a learning process according to the learning schedule. The learning device 21 may also output parameters of a domain identification network for realizing unsupervised domain adaptation. As illustrated in FIG. 2, the learning unit 21a has, for example, a control unit 11aa, a loss function calculation unit 21ab, a gradient inversion unit 11ac, and a parameter update unit 21ad. The learning unit 21a stores each piece of data obtained during the processing in the storage units 11b, 21c, and 21d or a temporary memory not illustrated. The learning unit 21a reads the data as necessary and uses it for each process.

<Network 200>
Fig. 7 shows an example of the configuration of a network 200 used in the learning process by the learning device 21. The network 200 shown in Fig. 7 has a labeling network 150 (labeling model) and a domain discrimination model 230. The labeling network 150 is the same as that in the first embodiment, so a description thereof will be omitted, and the domain discrimination model 230 will be described below.

<<Domain Identification Model 230>>
7 includes short-term context domain identification networks 231-1, ..., 231-N (short-term logically related domain identification means) and a long-term context domain identification network 232 (long-term logically related domain identification means). For example, the short-term context domain identification networks 231-1, ..., 231-N are identical networks (e.g., networks with identical parameters), and each short-term context domain identification network 231-n represents a state corresponding to each n=1, ..., N (e.g., each time).

The long-term context domain identification network 232 receives the intermediate feature sequence LF ₁ , ..., LF _N (long-term intermediate feature sequence) output from the long-term context understanding network 112, obtains and outputs a sequence of estimated domain information LD ₁ , ..., LD _N. Here, each piece of estimated domain information LD _n (where n=1, ..., N) is estimated information of domain identification information indicating whether each spoken text T _n belongs to the source domain or the target _domain . Unlike the domain identification network 130-n of the first embodiment, the long-term context domain identification network 232 illustrated in FIG. 7 aims to remove the domain dependency of the context (logical relationship) across multiple spoken texts T _n , which are words or sentences _, from the labeling network 150 by continuously capturing the input short-term intermediate feature sequence SF ₁ , ..., SF _N (for example, by continuously capturing the short-term intermediate feature sequence SF 1 , ..., SF N in the time direction). However, this does not limit the present invention. For example, instead of the long-term context domain identification network 232, a plurality of long-term context domain identification networks 232-1, ..., 232-K (where K is an integer equal to or greater than 2) may be present. For example, k = 1, ..., K is an index representing a layer of the long-term context understanding network. In this case, each long-term context domain identification network 232-k (where k = 1, ..., K) may receive a long-term intermediate feature sequence LF _nk (n ∈ {1, ..., N}, k ∈ {1, ..., K}), and obtain and output estimated domain information LD _nk representing whether the speech text T _n corresponding to the received long-term intermediate feature sequence LF _nk belongs to the source domain or the target domain. LF _nk (where n = 1, ..., N, k = 1, ..., K) represents an intermediate feature corresponding to each n (for example, each time) output from each long-term context understanding network 112-k exemplified in the first embodiment. Even in this case, the domain dependency of the context across multiple spoken texts _Tn can be removed from the labeling network 150. Here, the long-term context domain identification network 232 can be configured by, for example, a combination of a unidirectional LSTM or bidirectional LSTM and a fully connected neural network using a softmax function as an activation function.

The short-term context domain identification networks 231-1, ..., 231-N receive the series of short-term intermediate features SF ₁ , ..., SF _N (a second series based on the intermediate feature series, a short-term intermediate feature series) output from the short-term context understanding networks 111-1, ..., 111-N (short-term logical relation understanding means), and obtain and output a series of estimated domain information SD ₁ , ..., SD _N. That is, each short-term context domain identification network 231-n (where n=1, ..., N) receives the short-term intermediate features SF _n output from the short-term context understanding network 111-n, obtains estimated domain information SD _n of domain identification information indicating whether each spoken text T _n belongs to the source domain or the target domain, and outputs the estimated domain information SD _n . Unlike the long-term context domain identification network 232, the short-term context domain identification network 231-n aims to efficiently remove the domain dependency of a single utterance text T _n such as a specific word or document having domain dependency by estimating whether the utterance text T _n belongs to the source domain or the target domain for _each short-term intermediate feature SF n. However, this does not limit the present invention. For example, instead of each short-term context domain identification network 231-n, there may be a plurality of short-term context domain identification networks 231-n1, ..., 232-nK' (where K' is an integer of 2 or more). For example, k' = 1, ..., K' is an index representing a layer of the short-term context domain identification network, and each short-term context domain identification network 231-nk' represents a network corresponding to each n (e.g., each time). In this case, each short-term context domain identification network 231-nk' (where k'=1,...,K') may receive short-term intermediate features SF _nk' (n∈{1,...,N},k'∈{1,...,K'}), obtain and output estimated domain information SD _nk' indicating whether the speech text T _n corresponding to the received short-term intermediate features SF _nk' belongs to the source domain or the target domain. Here, SF _nk' is a short-term intermediate feature corresponding to each n (e.g., each time) output from each short-term context understanding network 111-nk' exemplified in the first embodiment. Here, the short-term context domain identification network 231-n can be configured by a combination of fully connected neural networks, etc., with a softmax function as an activation function.

<Learning process>
In the learning process, labeled teacher data of the source domain (a labeled training information sequence belonging to the source domain) and unlabeled teacher data of the target domain (unlabeled training information sequence belonging to the target domain) are input to the learning unit 21a of the learning device 21. The learning unit 21a performs adversarial learning on the above-mentioned network 200, using teacher data including the labeled teacher data of the source domain and the unlabeled teacher data of the target domain as the spoken text sequence _T1 , ..., _TN , to learn the labeling network ₁₅₀ (labeling model) so that the estimation accuracy of the estimated label sequence _L1 , ..., _LN is high and the estimation accuracy of the estimated domain information sequence _LD1 , ..., _LDN and _SD1 , ..., _SDN is low, and to learn the domain discrimination model 230 so that the estimation accuracy of the estimated domain information sequence _LD1 , ..., _LDN and SD1, ..., _SDN is high. That is, the learning unit 21a calculates a loss function (hereinafter, label prediction _loss) that expresses _{the error between the estimated label sequence L 1 ,} ..., L _N output from the labeling network 150 and the corresponding correct label sequence of labeled teacher data of the source domain when the above-mentioned teacher data is input to the network 200 as the speech text sequence T 1 , ..., T N, a loss function (hereinafter, long-term context domain identification loss) that expresses the error between the sequence LD ₁ , ..., LD _N of estimated domain information output from the long-term context domain identification network 232 and the correct label sequence of estimated domain information identified from the labeled teacher data of the source domain and the unlabeled teacher data of the target domain, and a loss function (hereinafter, long-term context domain identification loss) that expresses the error between the sequence SD ₁ , ..., SD Adversarial learning is performed between the labeling network 150 and the domain discrimination model 230 based on _N and a loss function (hereinafter, short-term context domain discrimination loss) that represents the error between the correct label sequence of estimated domain information identified from labeled training data of the source domain and unlabeled training data of the target domain. Note that the estimated label sequence L ₁ , ..., L _N output from the labeling network 150 when unlabeled training data of the target domain is input to the network 200 is not used to calculate the label prediction loss.

The learning unit 21a performs this adversarial learning using, for example, an error backpropagation method. In this case, a gradient inversion layer 242-n (where n=1, ..., N) is provided between the long-term context understanding network 112 and the long-term context domain identification network 232, and a gradient inversion layer 241-n (where n=1, ..., N) is provided between the short-term context understanding network 111-n and the short-term context domain identification network 231-n, and the gradient is inverted in the gradient inversion layers 242-n and 241-n only during error backpropagation. Here, by performing learning so that the label prediction loss is reduced, the estimation accuracy of the estimated label sequence L ₁ , ..., L _N in the labeling network 150 is improved. Furthermore, by inverting the gradient inversion layer 242-n only during error backpropagation and performing training to reduce the long-term context domain discrimination loss, it is possible to train the long-term context domain discrimination network ₂₃₂ to increase the estimation accuracy of the estimated domain information series LD ₁ , ..., LD _N , and perform adversarial training to train the long-term context domain discrimination network 232 to obtain the intermediate feature series LF ₁ , ..., LF _N that reduces the estimation accuracy of the estimated domain information series LD 1 , ..., LD _N. Furthermore, by inverting the gradient inversion layer 241-n only during backpropagation and performing learning to reduce the short-term context domain discrimination loss, it is possible to train the short-term context domain discrimination networks 231-1, ..., 231-N to increase the estimation accuracy of the estimated domain information series SD ₁ , ..., SD _N , and perform adversarial learning to train the short-term context understanding networks 111-1, ..., 111-N to obtain the short-term intermediate feature series SF ₁ , ..., SF _N that reduces the estimation accuracy of the estimated domain information series SD ₁ , ..., SD _N. By these adversarial learnings, it is possible to train the labeling network 150 that can accurately estimate the estimated label series L ₁ , ..., L _N but can generate the intermediate feature series LF ₁ , ..., LF _N and the short-term intermediate feature series SF ₁ , ..., SF _N whose domain is not estimated by the domain discrimination model 230. This makes it possible for the labeling network 150 to acquire a sequence of intermediate features LF ₁ , ..., LF _N that are effective for predicting labels while suppressing dependency on the domain of context across multiple spoken texts T _n , and to acquire a sequence of short-term intermediate features SF ₁ , ..., SF _N that are effective for predicting labels while suppressing dependency on the domain on a per-speech text T _n basis, thereby realizing unsupervised domain adaptation with higher accuracy.

This learning process can be realized by optimizing (minimizing) a loss function that is a linear combination of the label prediction loss, the long-term context domain classification loss, and the short-term context domain classification loss. The combination ratio of the linear combination of the label prediction loss, the long-term context domain classification loss, and the short-term context domain classification loss may be predetermined or may be specified by a learning schedule input to the learning unit 21a.

The learning unit 21a may perform the above-mentioned learning while changing the combination ratio of the label prediction loss, the long-term context domain identification loss, and the short-term context domain identification loss according to the number of learning steps based on a learning schedule. For example, the learning unit 21a may perform learning using only the label prediction loss as a loss function in the early stages of learning, and as the number of learning steps increases, the ratio of the long-term context domain identification loss and the short-term context domain identification loss in the loss function gradually increases. Furthermore, the learning unit 21a may perform learning until the learning converges with a certain combination ratio, change the combination ratio, and perform learning until it converges again, while repeatedly changing the combination ratio based on the learning schedule.

The domain identification model 230 may have only one of the long-term context domain identification network 232 and the short-term context domain identification networks 231-1, ..., 231-N.

When the domain identification model 230 has only the long-term context domain identification network 232, the gradient inversion layers 241-1, ..., 241-N are omitted, and the learning process is performed based on a loss function that is a linear combination of the label prediction loss and the long-term context domain identification loss. In this case, the combination ratio of the linear combination may be predetermined or may be specified by a learning schedule input to the learning unit 21a. The learning unit 21a may also perform the above-mentioned learning while changing the combination ratio of the label prediction loss and the long-term context domain identification loss according to the number of learning steps based on the learning schedule. For example, the learning unit 21a may perform learning using only the label prediction loss as a loss function in the early stage of learning, and learning may be performed so that the proportion of the long-term context domain identification loss in the loss function gradually increases as the number of learning steps increases. Furthermore, the learning unit 21a may repeatedly perform learning such that the learning converges at a certain combination ratio, the combination ratio is changed, and learning is performed again until convergence occurs, while changing the combination ratio based on the learning schedule.

When the domain identification model 230 has only the short-term context domain identification networks 231-1, ..., 231-N, the gradient inversion layers 242-1, ..., 242-N are omitted, and the learning process is performed based on a loss function that is a linear combination of the label prediction loss and the short-term context domain identification loss. In this case, the combination ratio of the linear combination may be predetermined or may be specified by a learning schedule input to the learning unit 21a. The learning unit 21a may also perform the above-mentioned learning while changing the combination ratio of the label prediction loss and the short-term context domain identification loss according to the number of learning steps based on the learning schedule. For example, the learning unit 21a may perform learning using only the label prediction loss as a loss function in the early stage of learning, and may learn so that the proportion of the short-term context domain identification loss in the loss function gradually increases as the number of learning steps increases. Furthermore, the learning unit 21a may repeatedly perform learning such that the learning converges at a certain combination ratio, the combination ratio is changed, and learning is performed again until it converges, while changing the combination ratio based on the learning schedule.

Also, the domain identification model 230 may have only the long-term context domain identification network 232 and a part of the short-term context domain identification networks 231-1, ..., 231-N. That is, a part of the short-term context domain identification networks 231-1, ..., 231-N may be omitted. In this case, the estimated domain information SD _n corresponding to the omitted short-term context domain identification network 231-n and the correct label of the estimated domain information corresponding thereto are not used in the calculation of the short-term context domain identification loss.

The learning unit 21a may also prepare a plurality of various domain identification models 230 and/or labeling networks 150 as exemplified above, perform learning, and later select a labeling network 150 that has the best estimation accuracy of the label sequence by the labeling network 150 in the target domain from among the labeling networks 150 obtained by each learning. The plurality of domain identification models 230 prepared may include, for example, at least one of the domain identification model 230 including the long-term context domain identification network 232 and the short-term context domain identification network 231-1, ..., 231-N, the domain identification model 230 having only the long-term context domain identification network 232, the domain identification model 230 having only the short-term context domain identification network 231-1, ..., 231-N, and the domain identification model 230 of the first embodiment, as described above.

The training process may be batch training, mini-batch training, or online training.

The learning unit 21a stores the parameters of the labeling network 150 obtained by the above-mentioned learning in the memory unit 11b, stores the parameters of the short-term context domain identification networks 231-1, ..., 231-N in the memory unit 21c, and stores the parameters of the long-term context domain identification model 232 in the memory unit 21d. The learning device 21 outputs the parameters of the labeling network 150 stored in the memory unit 11b. The parameters of the labeling network 150 are used in the inference process. Normally, the parameters of the short-term context domain identification networks 231-1, ..., 231-N and the parameters of the long-term context domain identification model 232 are not used in the inference process, so they do not need to be output from the learning device 21. However, the learning device 21 may output at least one of the parameters of the short-term context domain identification networks 231-1, ..., 231-N and the parameters of the long-term context domain identification model 232.

The above-mentioned learning process will be functionally illustrated with reference to FIG.
Step S21: Labeled teacher data of the source domain and unlabeled teacher data of the target domain are input to the control unit 11aa. The control unit 11aa generates teacher data including the labeled teacher data of the source domain and the unlabeled teacher data of the target domain. The control unit 11aa also initializes parameters of the network 200.
Step S22: The loss function calculation unit 21ab inputs the teacher data as the spoken text sequence T ₁ , . . . , T _N to the network 200, and obtains the loss function as described above.
Step S23: The parameter update unit 21ad backpropagates information based on the loss function according to the backpropagation method, and updates the parameters of the domain identification model 230 and the labeling layer 120.
Step S24: The gradient inversion unit 11ac inverts the gradient of the information based on the loss function backpropagated from the domain discrimination model 230 and backpropagates it to the logical relationship understanding layer 110. The information based on the loss function backpropagated from the labeling layer 120 is backpropagated to the logical relationship understanding layer 110 as it is.
Step S25: The parameter update unit 21ad updates the parameters of the logical relationship understanding layer 110 using the backpropagated information according to the backpropagation method.
Step S26: The control unit 11aa judges whether or not the termination condition is satisfied. If the termination condition is not satisfied, the control unit 11aa returns the process to step S22. On the other hand, if the termination condition is satisfied, the control unit 11aa outputs the parameters of the labeling network 150. If necessary, the control unit 11aa may also output at least one of the parameters of the short-term context domain identification networks 231-1, ..., 231-N and the parameters of the long-term context domain identification model 232.

The functional configuration and inference processing of the inference device 13 in the second embodiment are the same as those in the first embodiment, so description is omitted.

<Features of the second embodiment>
In this embodiment, the labeling network 150 includes _a logical relationship understanding layer 110 that receives a spoken text sequence T ₁ , ..., _TN , obtains an intermediate feature sequence LF ₁ , ..., LF _N taking into account the context of the spoken text sequence T ₁ , ..., _TN , and outputs the intermediate feature sequence LF ₁ , ..., LF _N , and a labeling layer 120 that receives the intermediate feature sequence LF ₁ , ..., LF _N , obtains an estimated label sequence L ₁ , ..., LN corresponding to the spoken text sequence T ₁ , ..., _TN , and outputs the estimated label sequence L ₁ , ..., _LN . The labeling network 150 includes a logical relationship understanding layer 110 that receives a spoken text sequence T ₁ , ..., TN, obtains estimated domain information LD _n and _{SD n that} are domain identification information indicating whether each spoken text T _n included in the spoken text sequence T ₁ , ..., _TN belongs to a source domain or a target domain, and outputs the estimated domain information sequence LD ₁ , ..., LD For the domain discrimination model 230 that outputs label sequence _L1 , ..., _LDN and SD1, ..., _SDN , the learning device 21 performed adversarial learning using teacher data including labeled teacher data of the source domain and unlabeled teacher data of the target domain as the spoken text sequence _T1 , ..., _TDN to train the labeling network 150 so that the estimation accuracy of the estimated label sequence _L1 , ..., _LDN is high and the estimation accuracy of the estimated domain information sequences _LD1 , ..., _LDN and _SD1 , ..., _SDN is low, and to train the domain discrimination model 230 so that the estimation accuracy of the estimated domain information sequences _LD1 , ..., _LDN and _SD1 , ..., _SDN is high. This enables unsupervised domain adaptation of the labeling network 150 that estimates the label sequence _L1 , ..., _LN corresponding to the spoken text sequence _T1 , ..., _TDN in consideration of the context of the spoken text sequence.

In particular, in this embodiment, the domain identification model 230 includes at least one of a short-term context domain identification network 231-1, ..., 231-N that receives a sequence of short-term intermediate features SF ₁ , ..., SF _N output from the short-term context understanding network 111-1, ..., 111-N, obtains and outputs a sequence of estimated domain information SD ₁ , ..., SD _N , and a long-term context domain identification network 232 that receives a sequence of intermediate features LF ₁ , ..., LF _N output from the long-term context understanding network 112, obtains and outputs a sequence of estimated domain information LD ₁ , ..., LD _N. This makes it possible to efficiently remove at least one of the domain dependency of the spoken text T _n alone and the domain dependency of the context across the spoken text T _n from the labeling network 150. As a result, unsupervised domain adaptation of the labeling network 150 can be performed with higher accuracy.

Furthermore, since the domain discrimination model 230 includes at least the long-term context domain discrimination network 232, it is possible to efficiently remove domain dependency of context across multiple spoken texts _Tn from the labeling network 150. As a result, it is possible to perform unsupervised domain adaptation with high accuracy for the labeling network 150 that estimates a label sequence _L1 , ..., _LN corresponding to an utterance text sequence _T1 , ..., _TN by taking into account the context of the utterance text sequence.

Furthermore, since the domain discrimination model 230 includes both the short-term context domain discrimination networks 231-1, ..., 231-N and the long-term context domain discrimination network 232, it is possible to efficiently remove the domain dependency of a single utterance text _Tn and the domain dependency of a context across multiple utterance texts _Tn from the labeling network 150. In this case, unsupervised domain adaptation of the labeling network 150 can be performed with higher accuracy.

<Experimental Results>
The following are examples of experimental results of unsupervised domain adaptation performed according to the above-described embodiment. The experimental conditions are as follows.
(1) Classify each utterance text of the simulated data of the utterance text sequence into five classes of corresponding scenes, and estimate a label representing each corresponding scene.
(2) The five domains excluding the target domain (new domain) were treated as source domains (applied domains), and a labeling network was trained using only data from the source domain. The obtained labeling network and the labeling networks obtained in the first and second embodiments were trained, and the recognition performance (labeling accuracy rate) for the spoken text sequence of the target domain was verified using the obtained labeling network.
(3) The classification performance was verified for 60 phone call text sequences (60 calls x 6 domains = 360 phone call simulation data) belonging to six target domains (online shopping, ISP, securities, local government, mobile phone, and PC support).
(4) Each utterance text contains approximately 100 sentences.

Figure 8 shows an example of the experimental results. As shown in Figure 8, in both the first and second embodiments, unsupervised domain adaptation to the target domain is possible with high accuracy by using already existing labeled data in the source domain, even if labeled data in the target domain is not prepared. In particular, the method of the second embodiment enables unsupervised domain adaptation with higher accuracy, and the classification accuracy is improved by an average of 3.4% compared to the method of learning only with source domain data.

[Hardware configuration]
The learning devices 11, 21 and the inference device 13 in each embodiment are devices configured by a general-purpose or dedicated computer having a processor (hardware processor) such as a CPU (central processing unit) and memories such as a RAM (random-access memory) and a ROM (read-only memory) executing a predetermined program. This computer may have one processor and memory, or may have multiple processors and memories. This program may be installed in the computer, or may be recorded in a ROM or the like in advance. In addition, some or all of the processing units may be configured using electronic circuits that realize processing functions independently, rather than electronic circuits that realize functional configurations by reading programs like a CPU. In addition, an electronic circuit that configures one device may include multiple CPUs.

9 is a block diagram illustrating the hardware configuration of the learning device 11, 21 and the inference device 13 in each embodiment. As illustrated in FIG. 9, the learning device 11, 21 and the inference device 13 in this example have a CPU (Central Processing Unit) 10a, an input unit 10b, an output unit 10c, a RAM (Random Access Memory) 10d, a ROM (Read Only Memory) 10e, an auxiliary storage device 10f, and a bus 10g. The CPU 10a in this example has a control unit 10aa, a calculation unit 10ab, and a register 10ac, and executes various calculation processes according to various programs loaded into the register 10ac. The input unit 10b is an input terminal to which data is input, a keyboard, a mouse, a touch panel, etc. The output unit 10c is an output terminal to which data is output, a display, a LAN card controlled by the CPU 10a that has loaded a specific program, etc. The RAM 10d is a static random access memory (SRAM), a dynamic random access memory (DRAM), or the like, and has a program area 10da in which a predetermined program is stored and a data area 10db in which various data is stored. The auxiliary storage device 10f is, for example, a hard disk, a magneto-optical disc (MO), a semiconductor memory, or the like, and has a program area 10fa in which a predetermined program is stored and a data area 10fb in which various data is stored. The bus 10g connects the CPU 10a, the input unit 10b, the output unit 10c, the RAM 10d, the ROM 10e, and the auxiliary storage device 10f so that information can be exchanged. The CPU 10a writes the program stored in the program area 10fa of the auxiliary storage device 10f to the program area 10da of the RAM 10d according to the loaded OS (Operating System) program. Similarly, the CPU 10a writes various data stored in the data area 10fb of the auxiliary storage device 10f to the data area 10db of the RAM 10d. The addresses on the RAM 10d to which the programs and data are written are then stored in the register 10ac of the CPU 10a. The control unit 10aa of the CPU 10a sequentially reads out these addresses stored in the register 10ac, reads out the programs and data from the areas on the RAM 10d indicated by the read addresses, causes the calculation unit 10ab to sequentially execute the calculations indicated by the programs, and stores the results of the calculations in the register 10ac. With this configuration, the functional configurations of the learning devices 11 and 21 and the inference device 13 are realized.

The above-mentioned program can be recorded on a computer-readable recording medium. An example of a computer-readable recording medium is a non-transitory recording medium. Examples of such recording media include magnetic recording devices, optical disks, magneto-optical recording media, semiconductor memories, etc.

The distribution of this program is, for example, by selling, transferring, lending, etc., portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of a server computer and transferring the program from the server computer to other computers via a network. As described above, a computer that executes such a program, for example, first temporarily stores in its own storage device the program recorded in the portable recording medium or the program transferred from the server computer. Then, when executing the process, the computer reads the program stored in its own storage device and executes the process according to the read program. In addition, as another execution form of this program, the computer may read the program directly from the portable recording medium and execute the process according to the program, and further, each time the program is transferred from the server computer to this computer, the computer may execute the process according to the received program one by one. In addition, the server computer may not transfer the program to this computer, but may execute the above-mentioned process by a so-called ASP (Application Service Provider) type service that realizes the processing function only by issuing an execution instruction and obtaining the result. In this embodiment, the program includes information used for processing by an electronic computer and equivalent to a program (such as data that is not a direct instruction to a computer but has the nature of defining computer processing).

The device may be configured using not only the CPU 10a but also a GPU (Graphics Processing Unit). In each embodiment, the device is configured by executing a specific program on a computer, but at least a part of the processing contents may be realized by hardware.

Note that the present invention is not limited to the above-described embodiments. For example, in each embodiment, the unlabeled training data of the target domain includes a speech text sequence of the target domain but does not include a correct label sequence. However, at least a portion of the unlabeled training data of the target domain may include a correct label sequence. In this case, the correct label sequence of the unlabeled training data of the target domain may or may not be used to train the networks 100 and 200.

Also, as described above, in the above-mentioned embodiment, for the sake of clarity, the case is illustrated in which the sequence of multiple pieces of information having a logical relationship is a spoken text sequence, and the label sequence is a sequence of labels representing the corresponding scenes of each utterance (for example, opening, understanding the subject, identity verification, response, closing). However, this is only an example, and other information sequences, such as a sentence sequence, a programming language sequence, a voice signal sequence, or a video signal sequence, may be used as the sequence of multiple pieces of information having a logical relationship. Also, other label sequences, such as a label sequence representing a situation or action, a label sequence representing a place or time, a label sequence representing a part of speech, or a label sequence representing program content, may be used as the label sequence. Also, each model, such as the labeling model, may not be a model based on a deep neural network, but may be other models based on a probability model, a classifier, or the like. Also, the logical relationship understanding layer 110 in each embodiment receives the spoken text sequence T ₁ , ..., T _N , obtains and outputs the intermediate feature sequence LF ₁ , ..., LF _N that takes into account the context (logical relationship) of the spoken text sequence T ₁ , ..., T _N. However, the logical relationship understanding layer 110 may receive the spoken text sequence T ₁ , ..., T _N and obtain and output a sequence consisting of less than N or more than N intermediate features. Also, the labeling layer 120 of each embodiment receives the intermediate feature sequence LF ₁ , ..., LF _N (first sequence based on the intermediate feature sequence) and obtains and outputs an estimated label sequence L ₁ , ..., L _N corresponding to the spoken text sequence T ₁ , ..., T _N. However, the labeling layer 120 may receive the intermediate feature sequence LF ₁ , ..., LF _N (first sequence based on the intermediate feature sequence) and obtain and output a sequence of less than N or more than N estimated labels. Also, the domain identification model 130, 230 of each embodiment receives the intermediate feature sequence LF ₁ , ..., LF _N (second sequence based on the intermediate feature sequence) and obtains and outputs N pieces of estimated domain information. However, the domain discrimination model 130, 230 of the embodiment may receive the intermediate feature sequence LF ₁ , . . . , LF _N (a second sequence based on the intermediate feature sequence) and obtain and output less than N or more than N pieces of estimated domain information.

In addition, the various processes described above may not only be executed in chronological order as described, but may also be executed in parallel or individually depending on the processing capacity of the device executing the processes or as necessary. Needless to say, other modifications may be made as appropriate without departing from the spirit of the present invention.

The present invention makes it possible to efficiently remove the domain dependency of intermediate features, for example, for sequence labeling problems that take complex contexts into account. In particular, by efficiently removing the domain dependency of intermediate features for each of short-term and long-term logical relationships (contexts) as exemplified in the second embodiment, it is possible to learn a labeling network that is more adaptable to the target domain, and improve the labeling accuracy in the target domain.

Unsupervised domain adaptation technology has been considered for image recognition in the past, but this invention is the first to apply it to the problem of estimating a label sequence corresponding to multiple information sequences by considering the logical relationships between the sequences, such as in language processing. This unsupervised domain adaptation technology can significantly reduce the cost of labeling, which has been a barrier to the expansion of the contact center business industry, for example.

In particular, in the method illustrated in the second embodiment, for example, a domain identification network for a spoken text unit (e.g., a call unit) can be designed as a mechanism across sentence boundaries using unidirectional or bidirectional LSTM. This makes it possible to capture domain dependency, such as a specific conversation flow that depends on the industry of a contact center, and to efficiently remove it from the labeling network, thereby improving the estimation accuracy of labels in the target domain.

In addition, in the method illustrated in the second embodiment, for example, a domain identification network for each spoken text (for example, for each call) can be designed as a mechanism that does not cross the boundaries of spoken text. This makes it possible to capture domain dependency caused by specific words that depend on the industry of contact centers, for example, and to efficiently remove them from the labeling network, thereby improving the estimation accuracy of labels in the target domain.

11, 21 Learning devices 11a, 21a Learning unit 13 Inference device 13a Inference unit 100, 200 Network 110 Logical relationship understanding layer 111-1, ..., 111-N Short-term context understanding network 112 Long-term context understanding network 120 Labeling layer 120-1, ..., 120-N Label prediction network 130, 230 Domain identification model 130-1, ..., 130-N Domain identification network 231-1, ..., 231-N Short-term context domain identification network 232 Long-term context domain identification network

Claims (5)

a logical relationship understanding means for receiving an input information sequence, which is a sequence of a plurality of pieces of information having a logical relationship, and outputting an intermediate feature sequence taking into consideration the logical relationship of the input information sequence;
a labeling means for receiving a first sequence based on the intermediate feature sequence and outputting an estimated label sequence of a label sequence corresponding to the input information sequence;
A labeling model including
a domain discrimination model that receives a second sequence based on the intermediate feature sequence and outputs a sequence of estimated domain information indicating whether each piece of information included in the input information sequence belongs to a source domain or a target domain;
Whereas,
a learning unit that uses, as the input information sequence, teacher data including labeled teacher data which is a labeled learning information sequence belonging to a source domain and unlabeled teacher data which is an unlabeled learning information sequence belonging to a target domain, learns the labeling model so that the estimation accuracy of the estimated label sequence is high and the estimation accuracy of the estimated domain information sequence is low, and performs adversarial learning to learn the domain discrimination model so that the estimation accuracy of the estimated domain information sequence is high, and obtains and outputs parameters of the labeling model,
The logical relationship understanding means includes:
a plurality of short-term logical relationship understanding means for receiving each piece of information included in the input information sequence and outputting short-term intermediate features taking into account logical relationships within the received information;
a long-term logical relationship understanding means for receiving a short-term intermediate feature sequence consisting of a plurality of short-term intermediate features, and outputting a long-term intermediate feature sequence taking into consideration a logical relationship between a plurality of pieces of information included in the input information sequence,
The labeling means comprises:
receiving the long-term intermediate feature sequence as the first sequence, and outputting an estimated label sequence of a label sequence corresponding to the information sequence;
The domain identification model is
a short-term logical relation domain identification means for receiving the short-term intermediate feature sequence as the second sequence and outputting the sequence of estimated domain information; a logical relationship understanding means for receiving an input information sequence, which is a sequence of a plurality of pieces of information having a logical relationship, and outputting an intermediate feature sequence taking into consideration the logical relationship of the input information sequence;
a labeling means for receiving a first sequence based on the intermediate feature sequence and outputting an estimated label sequence of a label sequence corresponding to the input information sequence;
A labeling model including
a domain discrimination model that receives a second sequence based on the intermediate feature sequence and outputs a sequence of estimated domain information indicating whether each piece of information included in the input information sequence belongs to a source domain or a target domain;
Whereas,
a learning unit that uses, as the input information sequence, teacher data including labeled teacher data which is a labeled learning information sequence belonging to a source domain and unlabeled teacher data which is an unlabeled learning information sequence belonging to a target domain, learns the labeling model so that the estimation accuracy of the estimated label sequence is high and the estimation accuracy of the estimated domain information sequence is low, and performs adversarial learning to learn the domain discrimination model so that the estimation accuracy of the estimated domain information sequence is high, and obtains and outputs parameters of the labeling model,
The logical relationship understanding means includes:
a plurality of short-term logical relationship understanding means for receiving each piece of information included in the input information sequence and outputting short-term intermediate features taking into account logical relationships within the received information;
a long-term logical relationship understanding means for receiving a short-term intermediate feature sequence consisting of a plurality of short-term intermediate features, and outputting a long-term intermediate feature sequence taking into consideration a logical relationship between a plurality of pieces of information included in the input information sequence,
The labeling means comprises:
receiving the long-term intermediate feature sequence as the first sequence, and outputting an estimated label sequence of a label sequence corresponding to the information sequence;
The domain identification model is
A learning device comprising: a short-term logical relational domain identification means for receiving the short-term intermediate feature sequence as the second sequence and outputting the estimated domain information sequence; and a long-term logical relational domain identification means for receiving the long-term intermediate feature sequence as the second sequence and outputting the estimated domain information sequence. An inference device having an inference unit that applies an input information sequence for inference to the labeling model specified by the parameters obtained by the learning device of claim 1 or 2, obtains an estimated label sequence of a label sequence corresponding to the input information sequence for inference, and outputs the estimated label sequence. A learning method for the learning device of claim 1 or 2, or an inference method for the inference device of claim 3. A program for causing a computer to function as the learning device of claim 1 or 2 or the inference device of claim 3.

Images