JP7099254B2 - Learning methods, learning programs and learning devices - Google Patents

Description

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

Machine learning such as neural networks may be used for automatic summarization to generate summarization from documents such as newspapers, websites, and electric bulletin boards. For example, a model in which an RNN (Recurrent Neural Networks) encoder that vectorizes an input sentence and an RNN decoder that repeatedly predicts words in the summary sentence by referring to the vector of the input sentence is used to generate the summary sentence.

As an example of the method of learning such a model, there is a method of calculating the loss used for updating the model parameters for each word of the reference summary which is the correct summary sentence corresponding to the input sentence of the training sample. For example, at the time of model learning, the RNN decoder inputs the vector of the input sentence, the correct word one hour ago, and the number of remaining characters until the RNN decoder outputs the EOS called the sentence end symbol, and outputs the EOS. Repeatedly calculate the probability distribution of words for each time until. "EOS" here is an abbreviation for "End Of Sentence". The loss is calculated by comparing the probability distribution of the word calculated for each time with the correct word at the time. For example, the probability distribution of the word calculated at the first time is compared with the first word in the word string included in the reference summary. Also, the probability distribution of the word calculated at the second time is compared with the second word from the beginning of the reference summary.

When the above model learning is performed, the word order of the words is satisfied even if the sentence meaning is the same between the summary sentence output by the RNN decoder and the reference summary of the correct answer, while the limitation on the number of words in the summary sentence is satisfied to some extent. If is different, it is an evaluation that causes a loss.

For this reason, an index called ROUGE may be used to evaluate an automatically generated summary. The term "ROUGE" as used herein refers to an index indicating the degree of overlap of the word N-gram between the reference summary of the correct answer and the summary sentence output by the summary sentence generation system in which the model is incorporated. A technique called MRT (Minimum Risk Training) that tunes the parameters of the model of the RNN encoder and the RNN decoder based on such ROUGE has also been proposed.

Japanese Unexamined Patent Publication No. 2016-62181 Japanese Unexamined Patent Publication No. 2013-167985 JP-A-2015-170224 Japanese Unexamined Patent Publication No. 2014-123219

Ayana, Shiqi Shen, Yu Zhao, Zhiyuan Liu, Maosong Sun “Neural Headline Generation with Sentence-wise Optimization” Submitted on 7 Apr 2016

However, in the above technique, all the abstracts whose word order is different from the reference abstract of the correct answer are highly evaluated, so that a model for generating the abstract with low readability may be learned.

That is, in the above MRT, even if the summary sentence has a different word order from the reference summary of the correct answer, a high ROUGE value is calculated if the degree of word duplication is high. Then, the abstract sentence having a high ROUGE value may include a abstract sentence having a non-grammatical expression by exchanging the word order with the reference abstract of the correct answer. In some cases, a model that generates a summary sentence with low readability may be learned, partly because the parameters of the model are updated based on the summary sentence having a non-grammatical expression.

In one aspect, it is an object of the present invention to provide learning methods, learning programs and learning devices that can suppress the learning of models that generate less readable summaries.

In one aspect, it is a learning method that performs machine learning of a model that generates a summary sentence from an input sentence, by acquiring the input sentence and the summary sentence of the correct answer and replacing the word order of the words included in the summary sentence of the correct answer. A pseudo-sentence in which a non-grammatical expression is simulated is generated, and the pseudo-sentence is generated from the input sentence by the model. The computer executes a process of updating the parameters of the model based on the generation probability of the summary sentence of the correct answer generated from the input sentence.

It is possible to suppress the training of a model that produces a summary sentence with low readability.

FIG. 1 is a block diagram showing a functional configuration of the learning device according to the first embodiment. FIG. 2 is a diagram showing an example of a use case of the article summarization tool. FIG. 3 is a diagram showing an example of an input sentence. FIG. 4A is a diagram showing an example of a reference summary. FIG. 4B is a diagram showing an example of a system summary. FIG. 4C is a diagram showing an example of a system summary. FIG. 5 is a diagram showing an example of the processing content of MRT. FIG. 6 is a diagram showing an example of the generation probability and the ROUGE value. FIG. 7A is a diagram showing an example of a reference summary. FIG. 7B is a diagram showing an example of a system summary. FIG. 7C is a diagram showing an example of a system summary. FIG. 7D is a diagram showing an example of a system summary. FIG. 8 is a diagram showing an example of how to update the parameters of the model. FIG. 9 is a diagram showing an example of the first model learning. FIG. 10 is a diagram showing an example of the first model learning. FIG. 11 is a diagram showing an example of the first model learning. FIG. 12 is a diagram showing an example of input / output to the model in the first system. FIG. 13 is a diagram showing an example of a method for calculating the degree of overlap. FIG. 14 is a diagram showing an example of a method for calculating the degree of overlap with an error. FIG. 15 is a diagram showing an example of a method for calculating the degree of overlap with an error. FIG. 16 is a diagram showing an example of input / output to the model in the second system. FIG. 17 is a flowchart showing the procedure of the learning process according to the first embodiment. FIG. 18 is a flowchart showing the procedure of the first loss calculation process according to the first embodiment. FIG. 19 is a flowchart showing a procedure of the second loss calculation process according to the first embodiment. FIG. 20 is a diagram showing a hardware configuration example of a computer that executes the learning program according to the first and second embodiments.

The learning method, learning program and learning device according to the present application will be described below with reference to the attached drawings. It should be noted that this embodiment does not limit the disclosed technique. Then, each embodiment can be appropriately combined as long as the processing contents do not contradict each other.

FIG. 1 is a block diagram showing a functional configuration of the learning device according to the first embodiment. The learning device 1 shown in FIG. 1 provides a learning service that accepts various articles such as newspapers, electric bulletin boards, and websites as input sentences and executes learning of a model that generates a summary sentence.

As one embodiment, the learning device 1 can be implemented by installing a learning program that realizes the above learning service as package software or online software on an arbitrary computer. By causing the computer to execute the above learning program in this way, the computer can function as the learning device 1. The computer referred to here may be any information processing device. For example, in addition to desktop-type or notebook-type personal computers and workstations, mobile communication terminals such as smartphones and mobile phones, tablet terminals, wearable terminals, and the like are included in this category. Further, the learning device 1 may be implemented as a server device in which the terminal device used by the user is a client and the above learning service is provided to the client. In this case, the learning device 1 accepts a request for model learning in which learning data including a plurality of learning samples or identification information capable of calling the learning data via a network or a storage medium is input. Then, the learning device 1 is implemented as a server device that provides a learning service that outputs a learning execution result of model learning for the learning data received in the model learning request. In this case, the learning device 1 may be implemented on-premises as a server that provides the learning service, or may be implemented as a cloud that provides the learning service by outsourcing.

[Example of trained model use case]
The trained model learned by the above learning service can be implemented as an article summarization tool that accepts the original text of an article such as a newspaper article, an electric bulletin board, or a website as an input sentence and generates the summary sentence.

Here, the above-mentioned article summarization tool can be incorporated as one function of an application whose user is a media business operator who operates various media such as newspapers, electric bulletin boards, and websites, as one aspect.

At this time, the above application may be implemented as stand-alone software executed by a terminal device used by a person concerned with a media company, for example, an editor or the like. In addition, among the functions provided by the above applications, front-end functions such as inputting the original text and displaying the summary text are provided by terminal devices such as reporters and editors, and back-end functions such as the generation of the summary text are provided. The function may be provided as a Web service.

FIG. 2 is a diagram showing an example of a use case of the article summarization tool. FIG. 2 shows an example of the transition of the article summary screen 20 displayed on the terminal device used by a person concerned with the media business.

The upper part of FIG. 2 shows an article summary screen 20 in an initial state in which inputs for various items are not set. For example, the article summary screen 20 includes GUI (Graphical User Interface) components such as an original text input area 21, a summary display area 22, a pull-down menu 23, a summary button 24, and a clear button 25. Of these, the original text input area 21 corresponds to an area for inputting the original text such as an article. Further, the summary display area 22 corresponds to an area for displaying the summary text corresponding to the original text input in the original text input area 21. Further, the pull-down menu 23 corresponds to an example of a GUI component that specifies the maximum number of characters in the summary sentence. Further, the summary button 24 corresponds to an example of a GUI component that accepts the execution of a command that generates a summary sentence corresponding to the original text input in the original text input area 21. Further, the clear button 25 corresponds to an example of a GUI component that clears the text of the original text input in the original text input area 21.

As shown in FIG. 2, in the original text input area 21 of the article summary screen 20, text input can be accepted via an input device such as a keyboard (not shown). In addition to accepting text input via the input device in this way, in the original text input area 21, text can be imported from a document file created by an application such as word processing software.

By inputting the original text in the original text input area 21 in this way, the article summary screen 20 transitions from the state shown in the upper part of FIG. 2 to the state shown in the middle part of FIG. 2 (step S1). .. For example, when the text of the original text is input to the original text input area 21, it is possible to accept the execution of the command to generate the summary text via the operation for the summary button 24. It is also possible to clear the text input in the original text input area 21 via the operation for the clear button 25. In addition, via the pull-down menu 23, it is also possible to accept the designation of the maximum number of characters desired by a person concerned with the media business from among the plurality of maximum number of characters. Here, as an example of a scene in which a bulletin board bulletin board is generated as a summary from the original text of a newspaper or news article, an example in which 80 characters corresponding to an example of the maximum number of characters that can be displayed on the electric bulletin board is specified is shown. There is. This is just an example, and when a headline is generated from an article in a newspaper or a website, the maximum number of characters corresponding to the headline can be selected.

When the summary button 24 is operated with the original text input in the original text input area 21, the article summary screen 20 is shown in the lower part of FIG. 2 from the state shown in the middle part of FIG. Transition to the state (step S2). In this case, the text of the original text input in the original text input area 21 is input to the trained model as an input sentence, and the summary sentence is generated. The generation of this summary may be performed on the terminal device of the media operator's party, or may be executed on the back end server device. As a result, as shown in the lower part of FIG. 2, the summary sentence generated by the trained model is displayed in the summary display area 22 of the article summary screen 20.

In this way, the text of the summary sentence displayed in the summary display area 22 of the article summary screen 20 can be edited via an input device (not shown) or the like.

By providing the above-mentioned article summarization tool, it is possible to reduce the work of article summarization performed by reporters, editors, and the like. That is, the work of article summarization is the most in the process of delivering news to the media, such as "selection of delivered articles", "sending to media editing system", "article summarization", "headline creation", "review", etc. There is an aspect that the effort is large. For example, when article summarization is done manually, it is necessary to select important information from the entire article and reconstruct the text. For this reason, the technical significance of automating or semi-automating the work of article summarization is high.

Here, as an example, the use case where the article summarization tool is used by the people concerned with the media business is taken as an example, but the article summarization tool is used by the viewer who receives the article distribution from the media business. It doesn't matter if you do. For example, the article summarization tool can be used as a function to read out the summary sentence instead of reading out the whole article by a smart speaker or the like.

[One aspect of RNN model learning issues]
As explained in the background technique section above, when calculating the loss used to update the model parameters for each word of the correct reference summary corresponding to the input sentence of the training sample, the word order is different from the reference summary, but the meaning of the sentence is. The evaluation of similar summaries may be underestimated.

Such a failure example of model learning will be described with reference to FIGS. 3 and 4A to 4C. FIG. 3 is a diagram showing an example of an input sentence. FIG. 4A is a diagram showing an example of a reference summary. 4B and 4C are diagrams showing an example of a system summary. In the following, the correct summary sentence included in the training sample may be referred to as a "reference summary", and the summary sentence generated by the model from the input sentence may be referred to as a "system summary".

Here, as an example, a case where a pair of the input sentence 30 shown in FIG. 3 and the reference summary 40 shown in FIG. 4A is input as a learning sample during model learning will be given as an example. At this time, when the system summary 40B shown in FIG. 4B and the system summary 40C shown in FIG. 4C are generated from the input sentence 30 by the model to which the RNN (Recurrent Neural Networks) encoder and the RNN decoder are connected, the following evaluation is performed. Will be done.

That is, between the reference summary 40 shown in FIG. 4A and the system summary 40B shown in FIG. 4B, the words match at each position from the beginning to the end. As an example, FIGS. 4A and 4B show the fifth word from the beginning of the reference summary 40 and the system summary 40B in bold. For example, when the word located at the fifth position from the beginning of the system summary 40B is predicted, as shown in FIG. 4B, the probability of the word "AI" in the probability distribution of the words of the input sentence 30 output by the RNN decoder. Is the best. The word of the reference summary 40 located at the fifth position from the beginning is also "AI" as shown in FIG. 4A. When the words in the system summary 40B at the positions corresponding to the positions of the words match for each word included in the reference summary 40 in this way, the loss is "0".

On the other hand, the reference summary 40 shown in FIG. 4A and the system summary 40C shown in FIG. 4C have the same meaning, but the word order of the eighth word from the beginning is different between the reference summary 40 and the system summary 40C. As an example, FIGS. 4A and 4C show the fifth word from the beginning of the reference summary 40 and the system summary 40C in bold. For example, when the word located at the fifth position from the beginning of the system summary 40C is predicted, as shown in FIG. 4C, the probability of the word "call center" in the probability distribution of the words of the input sentence 30 output by the RNN decoder. Is the best. On the other hand, the word of the reference summary 40 located at the fifth position from the beginning is "AI" as shown in FIG. 4A. When the word order is changed between the reference summary 40 and the system summary 40C and the word arrangement is different, even if the system summary 40C has the same meaning as the reference summary 40, a loss occurs.

From these things, different evaluations will be made between the system summary 40B and the system summary 40C. However, the meanings of the system summary 40B and the system summary 40C are the same. Therefore, from the aspect of summarization, it cannot be said that it is appropriate unless the same evaluation is made, and the system summarization 40C is underestimated as compared with the system summarization 40B.

[Current MRT]
As described above, a technique called MRT (Minimum Risk Training) has been proposed from the aspect of suppressing underestimation of system summaries having a different word order from reference summaries during model learning. For example, in MRT, the parameters of the RNN encoder and RNN decoder models are tuned based on the ROUGE that represents the N-gram overlap of words between the correct reference summary and the system summary.

FIG. 5 is a diagram showing an example of the processing content of MRT. As shown in FIG. 5, in the model learning of the RNN encoder and the RNN decoder, a pair of the input sentence x and the reference summary y of the correct answer is used as a learning sample. Of these input sentences x and the reference summary y of the correct answer, the input sentence x is input to the model.

When the input sentence x is input in this way, the RNN decoder of the model having the parameter θ outputs a plurality of system summaries _y'1 to y according to the probability distribution of the words output at each time from the beginning to the EOS (End of Sentence). ′ ₃ is sampled.

For example, at each time from the beginning to EOS, the probability is calculated for each word registered in the model dictionary, that is, for each word that appears in the input sentence in the entire learning data including a plurality of learning samples. The above sampling can be realized by extracting words at each time according to the probability distribution of the words at each time obtained by such a calculation. Here, for convenience of explanation, an example in which _three system summaries _y'1 to y'3 are sampled is given, but any number of system summaries y'may be sampled.

Then, in MRT, the generation probability that the system summary y'is generated from the input sentence x and the word n- between the reference summary y and the system summary _{y'for each} system summary y'1 to y'3. The ROUGE value representing the degree of overlap of the gram is calculated. Then, in MRT, the loss L _MRT (θ) is calculated according to the following equation (1) from the generation probabilities of the system summaries _y'1 to _y'3 and the ROUGE value.

Here, "P (y'| x; θ)" in the above equation (1) indicates the probability that the system summary y'is generated from the input sentence x when the parameter of the model is θ. Further, "D" in the above equation (1) refers to learning data which is a set of learning samples including an input sentence x and a reference summary y. Further, "S" in the above equation (1) refers to a set of system summaries generated from the input sentence x when the parameter of the model is θ. Further, "Δ (y', y)" in the above equation (1) refers to the degree of word duplication calculated between the system summary y'and the reference summary y, and here, as an example, ROUGE or the like is used. It is assumed that the negative gain is calculated as the ROUGE value by using the function.

The MRT then updates the model parameter θ based on the loss L _MRT . For example, MRT obtains the gradient, that is, ∂L _MRT (θ) / ∂θ _i by partially differentiating L _MRT (θ) with respect to θ _i , and updates the model parameter θ _i , that is, θ _i ← θ + (∂). L _MRT (θ) / ∂θ _i ) is calculated.

By updating the model parameter θ _i based on the loss _LMRT (θ) in this way, the probability of generating a system summary with a high ROUGE value is increased, while the probability of generating a system summary with a low ROUGE value is decreased. Learning is realized.

The change in the loss _LMRT (θ) before and after the parameter update using this ROUGE value will be described with reference to FIG. FIG. 6 is a diagram showing an example of the generation probability and the ROUGE value. The upper table of FIG. 6 shows the generation probability that the model having the parameter θ _t in the model learning in the t-th round generates the system summary y'from the input sentence x, and the ROUGE value between the reference summary and the system summary y'. Is shown. The light hatched portion shown in the table of FIG. 6 indicates that it is calculated by the calculation formula of the generation probability of the system summary y'included in the above formula (1), while the dark hatched portion shown in the table of FIG. It is shown that the hatched part is calculated by the ROUGE function included in the above equation (1).

For example, assuming that the generation probabilities and ROUGE values of the system summaries _y'1 to _y'3 generated from the input sentence x by the model having the parameter θ _t are the values shown in the upper table of FIG. 6, _LMRT (θ). _t ) can be calculated as follows. That is, the loss L _MRT (θ _t ) is the ROUGE value of the system summary y ′ ₁ generation probability and the ROUGE value, the system summary y ′ ₂ generation probability and the ROUGE value, and the system summary y ′ ₃ generation probability and the ROUGE value. It can be calculated from the sum of the values. That is, the loss L _MRT (θ _t ) is calculated as 0.2 × (−0.3) +0.6 × (−0.1) +0.2 × (−0.6) to be −0.24. It is calculated.

The system summary y ′ ₁ to y ′ ₃ generation probability and ROUGE value generated from the input sentence x by the model whose parameters are updated from θ _t to θ _{t + 1} based on such loss L _MRT (θ _t ) are shown in FIG. It is assumed that it is as shown in the lower table.

On the other hand, in the lower table shown in FIG. 6, the generation probability that the model having the parameter θ _{t + 1} in the model learning in the t + 1 round generates the system summary y ′ from the input sentence x, and the reference summary and the system summary y ′. The ROUGE value between is shown. In this case as well, the loss _LMRT (θ _{t + 1} ) is the generation probability and ROUGE value of the system summary y ′ ₁ , the generation probability and the ROUGE value of the system summary y ′ ₂ , and the generation probability of the system summary y ′ ₃ . And can be obtained from the sum of the ROUGE values. That is, the loss L _MRT (θ _{t + 1} ) is −0.46 by the calculation of 0.3 × (−0.3) +0.1 × (−0.1) +0.6 × (−0.6). It is calculated.

By updating the model parameters from θ _t to θ _{t + 1} in this way, model learning that reduces the loss L _MRT (θ _{t + 1} ) in the t + 1 round rather than the loss L _MRT (θ _t ) in the t round is realized. You can see that it has been done.

[One aspect of the current MRT issues]
However, as explained in the background technology section above, when the word order difference is irrelevant and the model parameters are updated according to the degree of word duplication, all system summaries with different word orders from the correct reference summary are used. The ROUGE value of is highly evaluated. Therefore, even if non-grammatical expressions are included in the system summary, which has a different word order from the correct reference summary, the loss of the system summary is underestimated and the model parameters are learned. As a result, a model that produces a poorly readable system summary may be trained.

Such a failure example of model learning will be described with reference to FIGS. 7A to 7D. FIG. 7A is a diagram showing an example of a reference summary. 7B-7D are diagrams showing an example of a system summary. Here, as an example, a case where a pair of the input sentence 30 shown in FIG. 3 and the reference summary 70 shown in FIG. 7A is input as a training sample is given as an example when training the model. At this time, when the system summaries 70B to 70D having the same ROUGE values shown in FIGS. 7B to 7D are generated from the input sentence 30 by the model to which the RNN encoder and the RNN decoder are connected, the following evaluation is performed. ..

That is, the word order and the set of words match between the reference summary 70 shown in FIG. 7A and the system summary 70B shown in FIG. 7B. Since the set of words matches between the reference summary 70 and the system summary 70B in this way, the loss is "0". Further, the word order is different between the reference summary 70 shown in FIG. 7A and the system summary 70C shown in FIG. 7C, but the set of words is the same. Since the set of words matches between the reference summary 70 and the system summary 70C in this way, the loss is "0". Further, the word order is different between the reference summary 70 shown in FIG. 7A and the system summary 70D shown in FIG. 7D, but the set of words is the same. Since the set of words matches between the reference summary 70 and the system summary 70D in this way, the loss is "0". In this way, the same evaluation is made between the system summary 70B and the system summary 70D having the same ROUGE value.

However, the system summary 70D contains non-grammatical expressions not found in the system summary 70B or system summary 70C. For example, the correct usage is that the case particle "de" is used for "chat", such as "... in chat ..." shown in system summary 70B and system summary 70C. Nevertheless, in "... chat is ..." shown in the system summary 70D, the case particle "ga" is used for "chat", and there is a grammatical error. Further, due in part to the grammatical error, in the system summary 70D, the modified part of "chat" modifies the modified part of "automatically respond", which is an erroneous dependency.

As described above, in the current MRT, if the ROUGE values are at the same level, system summaries 70B and system summaries 70C that do not include non-grammatical expressions, and non-grammatical expressions and erroneous dependencies are included. The same evaluation will be made with the system summary 70D. That is, when the system summary 70D containing a non-grammatical expression is included in the system summary during model training, the negative gain of the ROUGE value of the system summary 70D is the negative gain of the ROUGE value of the system summary 70B and the system summary 70C. Works as well as. Thus, the model is updated based on the loss overworked by the negative gain of the ROUGE value of the system summary 70D containing non-grammatical expressions, resulting in the learning of a model that produces a less readable summary. It may end up.

[One aspect of problem-solving approach]
Therefore, the learning device 1 according to the present embodiment replaces the word order of the words included in the reference summary of the correct answer to generate a pseudo-sentence in which the non-grammatical expression is simulated, and the model generates the pseudo-sentence. Update the model parameters so that the model is more likely to generate a reference summary than the probability.

FIG. 8 is a diagram showing an example of how to update the parameters of the model. As shown in FIG. 8, in the model learning of the RNN encoder and the RNN decoder, a pair of the input sentence x and the reference summary y of the correct answer is used as a learning sample as in the MRT shown in FIG.

Of these input sentences x and the reference summary y of the correct answer, the input sentence x is input to the model. When the input sentence x is input in this way, the learning device 1 has a plurality of system summaries _y'1 to y according to the probability distribution of words output by the RNN decoder of the model having the parameter θ at each time from the beginning to the EOS. ′ ₃ is sampled.

Then, the learning device ₁ has a generation probability that the system summary y'is generated from the input sentence x for each system summary _y'1 to y'3, and the word between the reference summary y and the system summary y'. The ROUGE value representing the degree of overlap of n-gram is calculated. Then, the learning device 1 calculates the loss L _MRT (θ) according to the above equation (1) from the generation probabilities of the system summaries _y'1 to _y'3 and the ROUGE value.

As described above, also in this embodiment, the process until the loss _LMRT (θ) is calculated from the generation probability of the system summary y'and the ROUGE value is the same as the above MRT, but the loss _LMRT (θ) itself. Is not used as a loss.

That is, the present embodiment is different in that it defines a loss improved from the above MRT. For example, in this embodiment, the term of the loss _Lord (θ) that penalizes the pseudo-sentence z including the non-grammatical expression together with the term of the loss _LMRT (θ) based on the generation probability of the system summary y'and the ROUGE value. The loss L (θ) to which is added is defined as the following equation (2). In addition, "α" in the following equation (2) is a weighting coefficient, and for example, any value of 0 to 1 can be adopted.

Here, the above loss _Lord (θ) is calculated by the following equation (3). “D” in the following equation (3) refers to training data which is a set of training samples including an input sentence x and a reference summary y. Further, "S'(y)" in the following equation (3) refers to a set of pseudo sentences z generated from the reference summary y of the correct answer. Further, "p (z | x; θ)" in the following equation (3) indicates the probability that a pseudo sentence z is generated from the input sentence x when the parameter of the model is θ. Further, "p (y | x; θ)" in the following equation (3) indicates the probability that the correct reference summary y is generated from the input sentence x.

For example, the learning device 1 is a set S'of pseudo-sentences z ₁ to z ₃ in which a non-grammatical expression is simulated by exchanging the word order of the words included in the reference summary y of the correct answer. (Y) is generated. At this time, by sampling the pseudo-sentence z by changing the word order of the words without changing the number of words included in the reference summary y of the correct answer, the ROUGE value calculated with the reference summary y is "1". It is possible to generate a pseudo sentence z that becomes. Here, for convenience of explanation, an example in which three pseudo-sentences z ₁ to z ₃ are sampled is given, but any number of pseudo-sentences z may be sampled.

Further, the learning device 1 calculates a generation probability p (y | x; θ) in which the reference summary y is generated from the input sentence x, and the pseudo sentence z is generated from the input sentence x for each pseudo sentence z. The generation probability p (z | x; θ) is calculated. For example, in the example of FIG. 8, the generation probability p (y | x; θ) of the reference summary y is calculated as “0.2”. Further, the generation probability p (z ₁ | x; θ) of the pseudo sentence z ₁ is calculated as “0.3”. Further, the generation probability p (z ₂ | x; θ) of the pseudo sentence z ₂ is calculated as “0.4”. Further, the generation probability p (z ₃ | x; θ) of the pseudo sentence z ₃ is calculated as “0.1”.

Based on the calculation result of such a generation probability, a calculation example of the loss _Lord (θ) will be described. For example, in the case of the pseudo-sentence z ₁ in the set S'(y) defined in Σ, the generation probability of the pseudo-sentence z ₁ (p (z ₁ | x; θ) = 0.3) and the generation of the reference summary y. The probability (p (y | x; θ) = 0.2) is compared. In this case, the generation probability of the pseudo sentence z ₁ is larger than the generation probability of the reference summary y. Therefore, in the above equation (3), the difference between the generation probability of the pseudo sentence z ₁ and the generation probability of the reference summary y, that is, p (z ₁ | x; θ) −p (y | x; θ) = 0. 1 is positive. As a result, p (z ₁ | x; θ) -p (y | x; θ) = 0.1 is selected by the max function.

Further, in the case of the pseudo sentence z ₂ , the generation probability of the pseudo sentence z ₂ (p (z ₂ | x; θ) = 0.4) and the generation probability of the reference summary y (p (y | x; θ) = 0. 2) is compared. In this case, the generation probability of the pseudo sentence z ₂ is larger than the generation probability of the reference summary y. Therefore, in the above equation (3), the difference between the generation probability of the pseudo sentence z ₂ and the generation probability of the reference summary y, that is, p (z ₂ | x; θ) -p (y | x; θ) = 0. 2 is positive. Also in this case, p (z ₂ | x; θ) −p (y | x; θ) = 0.2 is selected by the max function.

Further, in the case of the pseudo sentence z ₃ , the generation probability of the pseudo sentence z ₃ (p (z ₃ | x; θ) = 0.1) and the generation probability of the reference summary y (p (y | x; θ) = 0. 2) is compared. In this case, the generation probability of the pseudo sentence z ₃ is smaller than the generation probability of the reference summary y. Therefore, in the above equation (3), the difference between the generation probability of the pseudo sentence z ₃ and the generation probability of the reference summary y, that is, p (z ₃ | x; θ) -p (y | x; θ) = −0. .1 is negative. As a result, 0 is selected by the max function.

The loss _Lord (θ) can be calculated as 0.3 (= 0.1 + 0.2 + 0) by summing up the losses calculated for each of the elements of the pseudo sentences z ₁ to z ₃ .

As described above, in this embodiment, the following model learning can be realized by calculating the loss L (θ) based on the loss _Lord (θ) in addition to the loss L _MRT (θ). For example, while improving the ROUGE value by the term of loss L _MRT (θ), the model parameters are updated so that the probability of generating the reference summary y exceeds the probability of generating the pseudo-sentence z by the term of loss _Lord (θ). can do.

For this reason, the degree of duplication of the reference summary and the word is high, and while giving the effect of increasing the probability of generating the system summary having a different word order from the reference summary, it is a non-grammatical expression even in the summary sentence with the high degree of duplication of the reference summary and the word. It is possible to give a reaction that imposes a penalty on the generation of a pseudo-sentence containing. Therefore, it is possible to update the parameters that increase the probability of generating a system summary that does not include non-grammatical expressions even in a summary sentence with a high degree of duplication of a reference summary and a word.

Therefore, according to the learning device 1 according to the present embodiment, it is possible to suppress the learning of a model that generates a summary sentence having low readability.

[Functional configuration of learning device 1]
Next, an example of the functional configuration of the learning device 1 according to this embodiment will be described. As shown in FIG. 1, the learning device 1 includes a learning data storage unit 2, a first model storage unit 3, a first learning unit 5, a second model storage unit 8, and a second learning unit. Has 10 and. In addition to the functional units shown in FIG. 1, the learning device 1 may have various functional units of a known computer, for example, various functional units such as various input devices and voice output devices.

The functional units such as the first learning unit 5 and the second learning unit 10 shown in FIG. 1 are virtually realized by the following hardware processor as an example. Examples of such processors include DLUs (Deep Learning Units), GPGPUs (General-Purpose computing on Graphics Processing Units), GPU clusters, and the like. In addition, a CPU (Central Processing Unit) or an MPU (Micro Processing Unit) may be used. For example, when the processor deploys the learning program as a process on a memory such as a RAM (Random Access Memory), the functional unit is virtually realized. Here, DLU, GPGPU, GPU cluster, CPU, and MPU are exemplified as an example of the processor, but the above-mentioned functional unit may be realized by any processor regardless of general-purpose type or specialized type. .. In addition, the above-mentioned functional unit does not prevent it from being realized by hard-wired logic such as ASIC (Application Specific Integrated Circuit) and FPGA (Field Programmable Gate Array).

Further, functional units such as the learning data storage unit 2, the first model storage unit 3, and the second model storage unit 8 shown in FIG. 1 include an HDD (Hard Disk Drive), an optical disk, an SSD (Solid State Drive), and the like. Storage device can be adopted. The storage device does not necessarily have to be an auxiliary storage device, and various semiconductor memory elements such as RAM, EPPROM, and flash memory can also be adopted.

Here, in FIG. 1, from the aspect of improving the learning speed of the model in the second learning unit 10, the parameters after the preprocessing after the first learning unit 5 is made to execute the preprocessing for learning the parameters of the model. Is used to illustrate a case where the second learning unit 10 is made to execute the above model learning. This is just an example, and the preprocessing by the first learning unit 5 does not necessarily have to be performed. For example, the preprocessing by the first learning unit 5 may be skipped, and the second learning unit 10 may be made to execute the above model learning using the initial parameters. In the following, the model learning that is the preprocessing executed by the first learning unit 5 is referred to as "first model learning", and the above model learning executed by the second learning unit 10 is referred to as "first model learning". It may be described as "second model learning".

The learning data storage unit 2 is a storage unit that stores learning data. Here, the learning data includes, as an example, D learning samples, so-called learning cases. Further, the training sample includes a pair of the input sentence x and the reference summary y. Note that FIG. 1 illustrates a case where the same learning data is used for the first learning unit 5 and the second learning unit 10 as an example, but the first learning unit 5 and the second learning unit 5 are illustrated. It is also possible that training data different among 10 are used for model training.

The first model storage unit 3 and the second model storage unit 8 are both storage units that store information about the model.

As one embodiment, the following information is stored in the first model storage unit 3 and the second model storage unit 8. For example, the layer structure of the model such as neurons and synapses of each layer of the input layer, the hidden layer and the output layer forming a neural network to which the RNN encoder and the RNN decoder are connected, and the model parameters such as the weight and bias of each layer are set. The included model information is stored. Here, before the model learning is executed by the first learning unit 5, the first model storage unit 3 stores parameters initially set by random numbers as model parameters. Further, at the stage after the model learning is executed by the first learning unit 5, the parameters of the model learned by the first learning unit 5 are stored in the first model storage unit 3. Further, at the stage after the model learning is executed by the second learning unit 10, the parameters of the model learned by the second learning unit 10 are stored in the second model storage unit 8.

The first learning unit 5 is a processing unit that executes the first model learning that is the above-mentioned preprocessing. Here, as an example of the first model learning, a case where model learning called log-likelihood optimization is executed will be illustrated.

As shown in FIG. 1, the first learning unit 5 has an input control unit 5I, a model execution unit 6, and an update unit 7.

The input control unit 5I is a processing unit that controls the input to the model.

As one embodiment, the input control unit 5I controls the input of data to the model of the neural network to which the RNN encoder and the RNN decoder are connected for each training sample included in the training data.

Specifically, the input control unit 5I initializes the value of the loop counter d that counts the training sample. Subsequently, the input control unit 5I acquires a learning sample corresponding to the loop counter d among the D learning samples stored in the learning data storage unit 2. After that, the input control unit 5I increments the loop counter d, and repeatedly executes the process of acquiring the training sample from the training data storage unit 2 until the value of the loop counter d becomes equal to the total number D of the training samples. Here, an example of acquiring the learning data stored in the storage inside the learning device 1 has been given, but the learning data can be obtained from an external computer connected via a network, for example, a file server, a removable medium, or the like. It does not matter if it is acquired.

Each time the learning sample is acquired in this way, the input control unit 5I inputs the input sentence x included in the learning sample to the RNN encoder 6A. As a result, a vector in which the word string of the input sentence x is vectorized, a so-called intermediate representation, is output from the RNN encoder 6A to the RNN decoder 6B. At the same time or before and after this, the input control unit 5I sets the value of the register holding the number of remaining characters until the RNN decoder 6B outputs the EOS called the sentence end symbol to a predetermined upper limit number of characters, for example, user input or user setting. Initialize to a value. The details of the subsequent input to the RNN decoder 6B, the output from the RNN data, and the update of the model parameters using the same will be described later.

The model execution unit 6 is a processing unit that executes a model of a neural network to which the RNN encoder 6A and the RNN decoder 6B are connected.

As one aspect, the model execution unit 6 has M units corresponding to the number of words M of the input sentence of the learning sample input by the input control unit 5I according to the model information stored in the first model storage unit 3. Deploy LSTM (Long Short-Term Memory) on the work area. As a result, M LSTMs are made to function as RNN encoders 6A. In this RNN encoder 6A, according to the input control by the input control unit 5I, the mth word from the beginning of the input sentence is input to the LSTM corresponding to the mth word in order from the first word of the input sentence of the learning sample. At the same time, the output of the LSTM corresponding to the m-1st word is input to the LSTM corresponding to the mth word. By repeating such input from the LSTM corresponding to the first word to the LSTM corresponding to the Mth word at the end, a vector of the input sentence of the learning sample, a so-called intermediate representation, can be obtained. The intermediate representation of the input sentence generated by the RNN encoder 6A in this way is input to the RNN decoder 6B.

As a further aspect, the model execution unit 6 has N LSTMs corresponding to the number of words N of the correct reference summary input by the input control unit 5I according to the model information stored in the first model storage unit 3. On the work area. As a result, N RSTMs are made to function as the RNN decoder 6B. In these RNN decoders 6B, an intermediate expression of the input sentence of the learning sample is input from the RNN encoder 6A according to the input control of the input control unit 5I, and the EOS tag is input from the input control unit 5I to each of N LSTMs. The number of characters remaining until output is entered. By operating N LSTMs according to these inputs, the RNN decoder 6B outputs a word probability distribution for each N LSMTs. The "probability distribution of words" referred to here refers to the distribution of probabilities calculated for each word appearing in the input sentence in the entire learning sample.

The update unit 7 is a processing unit that updates the parameters of the model.

As one embodiment, when the probability distribution of words is output from the nth LSTM of the RNN decoder 6B, the update unit 7 sets the word having the maximum probability in the probability distribution as the nth word from the beginning of the system summary. Generate. After that, when the nth word of the system summary is generated, the update unit 7 loses from the nth word among the words included in the correct reference summary and the nth word generated as the system summary. calculate. In this way, the loss is calculated for each of the N LSTMs of the RNN decoder 6B. Then, the update unit 7 calculates the parameters for updating the models of the RNN encoder 6A and the RNN decoder 6B by executing the optimization of the log-likelihood based on the loss of each LSTM. Then, the update unit 7 updates the parameters of the model stored in the first model storage unit 3 to the parameters obtained by optimizing the log-likelihood. The update of this parameter can be repeatedly executed over all the training samples, and the training data D can also be repeatedly executed over a predetermined number of epochs.

The processing contents of the input control unit 5I, the model execution unit 6, and the update unit 7 will be described with reference to FIGS. 9 to 11. 9 to 11 are diagrams showing an example of the first model learning. 9 to 11 show a case where the input control unit 5I acquires a pair of the input sentence 30 shown in FIG. 3 and the reference summary 70 shown in FIG. 7A as a learning sample.

As shown in FIG. 9, the model execution unit 6 vectorizes the word string included in the input sentence 30 acquired by the input control unit 5I. That is, the model execution unit 6 expands M LSTM6a-1 to 6a-M corresponding to the number of words M of the input sentence 30 in the work area used by the model execution unit 6. As a result, M RSTM6a-1 to 6a-M function as the RNN encoder 6A. Then, the input control unit 5I inputs the word of the input sentence 30 into the LSTM6a corresponding to the position of the word in order from the first word included in the input sentence 30, and inputs the output of the previous LSTM6a. By repeating such input from LSTM6a-1 corresponding to the first word "our company" to LSTM6a-M corresponding to the last word ".", The vector of the input sentence 30 is obtained. The vector of the input sentence 30 generated by the RNN encoder 6A in this way is input to the RNN decoder 6B.

After that, the model execution unit 6 inputs the vector of the input sentence 30, the correct word one hour ago, the number of remaining characters until the RNN decoder 6B outputs the EOS called the sentence end symbol, and the like, and every time until the EOS is output. Repeatedly calculate the probability distribution of words.

For example, at the first time of calculating the probability distribution of the word to be collated with the first word of the reference summary 70, the operation shown in FIG. 9 is performed. That is, as shown in FIG. 9, the input control unit 5I outputs the LSTM6a-M and the initial symbol called BOS (Begin Of Sentence) with respect to the LSTM6b-1 expanded in the work area used by the model execution unit 6. And input the number of characters "37" of the reference summary 70 as the number of remaining characters. As a result, the probability distribution of the word at the first time (t = 1) is output by LSTM6b-1. As a result, the update unit 7 calculates the loss from the probability distribution of the word at the first time and the correct word "call center" at the first time. In this case, the smaller the loss is calculated as the probability of the correct word "call center" at the first time is closer to 1 and the probability of the other words is closer to 0.

Further, at the second time for calculating the probability distribution of the word to be collated with the second word from the beginning of the reference summary 70, the operation shown in FIG. 10 is performed. That is, as shown in FIG. 10, the input control unit 5I inputs the output of LSTM6b-1 and the correct word "call center" one hour before to LSTM6b-2, and one hour from the number of remaining characters at the first hour. The number of characters "30" obtained by subtracting the number of characters of the correct word of the eye is input as the number of remaining characters at the second time. As a result, the probability distribution of the word at the second time (t = 2) is output by LSTM6b-2. As a result, the update unit 7 calculates the loss from the probability distribution of the word at the second time and the correct word "no" at the second time. In this case, the smaller the loss is calculated as the probability of the correct word "no" at the second time is closer to 1 and the probability of the other words is closer to 0.

Further, at the third time when the probability distribution of the word to be collated with the third word from the beginning of the reference summary 70 is calculated, the operation shown in FIG. 11 is performed. That is, as shown in FIG. 11, the input control unit 5I inputs the output of LSTM6b-2 and the correct word "no" one hour before to LSTM6b-3, and two hours from the number of remaining characters at the second time. The number of characters "29" obtained by subtracting the number of characters of the correct word of the eye is input as the number of remaining characters at the third time. As a result, the probability distribution of the word at the third time (t = 3) is output by LSTM6b-3. As a result, the update unit 7 calculates the loss from the probability distribution of the word at the third time and the correct word "inquiry" at the third time. In this case, the smaller the loss is calculated as the probability of the correct word "inquiry" at the third time is closer to 1 and the probability of the other words is closer to 0.

By repeatedly executing such processing until the sentence end symbol "EOS" is output from LSTM6b, the update unit 7 calculates the loss for each word of the reference summary 70. Further, a process of calculating the loss for each word of the reference summary is executed for all the training samples included in the training data. When the loss for each word of the reference summary is calculated for all the training samples included in the training data in this way, the updater 7 maximizes the objective function _Lt shown in the following equation (4) with respect to the parameter θ. "Optimization of log-likelihood" is executed as the first model learning. Here, the probability "p (y | x; θ)" in the following equation (4) is obtained by the total product of losses at each time as shown in the following equation (5). Note that "D" in the following equation (4) refers to a set of learning samples including the input sentence x and the reference summary y. Further, "t" of "y _<t " in the following formula (5) indicates the position of a word in the reference summary, for example, the first word of the reference summary is represented by y ₁ , and the second word is y. It is represented by ₂ , ..., The last word is represented by y _t-1 .

After that, the update unit 7 updates the parameters of the model stored in the first model storage unit 3 to the parameters θ obtained by optimizing the log-likelihood. This update of the parameter θ can be repeated a predetermined number of times for the training data D. The parameters of the model stored in the first model storage unit 3 in this way are used by the second learning unit 10.

Returning to the description of FIG. 1, the second learning unit 10 is a processing unit that executes the above-mentioned second model learning. As shown in FIG. 1, the second learning unit 10 includes an input control unit 10I, a model execution unit 11, a summary generation unit 12, a first probability calculation unit 13, an overlap degree calculation unit 14, and a second. It has a loss calculation unit 15, a pseudo sentence generation unit 16, a second probability calculation unit 17, a second loss calculation unit 18, and an update unit 19.

The input control unit 10I is a processing unit that controls the input to the model.

As one embodiment, the input control unit 10I controls the input of data to the model of the neural network to which the RNN encoder 11A and the RNN decoder 11B are connected for each training sample included in the training data.

Specifically, the input control unit 10I initializes the value of the loop counter d that counts the training sample. Subsequently, the input control unit 10I acquires a learning sample corresponding to the loop counter d among the D learning samples stored in the learning data storage unit 2. After that, the input control unit 10I increments the loop counter d, and repeatedly executes the process of acquiring the training sample from the training data storage unit 2 until the value of the loop counter d becomes equal to the total number D of the training samples. Here, an example of acquiring the learning data stored in the storage inside the learning device 1 has been given, but the learning data can be obtained from an external computer connected via a network, for example, a file server, a removable medium, or the like. It does not matter if it is acquired.

Each time the learning sample is acquired in this way, the input control unit 10I inputs the input sentence x included in the learning sample to the RNN encoder 11A. As a result, a vector in which the word string of the input sentence x is vectorized, a so-called intermediate representation, is output from the RNN encoder 11A to the RNN decoder 11B. At the same time or before and after this, the input control unit 10I sets the value of the register holding the number of remaining characters until the RNN decoder 11B outputs the EOS called the sentence end symbol to a predetermined upper limit number of characters, for example, user input or user setting. Initialize to a value. The details of the subsequent input to the RNN decoder 11B, the output from the RNN data, and the update of the model parameters using the same will be described later.

The model execution unit 11 is a processing unit that executes a model of a neural network to which the RNN encoder 11A and the RNN decoder 11B are connected.

As one aspect, the model execution unit 11 has M units corresponding to the number of words M of the input sentence of the learning sample input by the input control unit 10I according to the model information stored in the first model storage unit 3. Deploy the LSTM on the work area. As a result, M LSTMs are made to function as RNN encoders 11A. In this RNN encoder 11A, according to the input control by the input control unit 10I, the mth word from the beginning of the input sentence is input to the LSTM corresponding to the mth word in order from the first word of the input sentence of the learning sample. At the same time, the output of the LSTM corresponding to the m-1st word is input to the LSTM corresponding to the mth word. By repeating such input from the LSTM corresponding to the first word to the LSTM corresponding to the Mth word at the end, a vector of the input sentence of the learning sample, a so-called intermediate representation, can be obtained. The intermediate representation of the input sentence generated by the RNN encoder 11A in this way is input to the RNN decoder 11B.

As a further aspect, the model execution unit 11 displays K LSTMs corresponding to each time on the work area until the sentence end symbol "EOS" is output according to the model information stored in the first model storage unit 3. Expand to. As a result, K RSTMs are made to function as the RNN decoder 11B. In these RNN decoders 11B, an intermediate expression of the input sentence of the learning sample is input from the RNN encoder 11A according to the input control of the input control unit 10I, and the EOS tag is input from the input control unit 10I to each K LSTM. The number of characters remaining until output is entered. By operating K LSTMs according to these inputs, the RNN decoder 11B outputs a word probability distribution for each K LSMTs.

In addition to the input control unit 10I and the model execution unit 11, the second learning unit 10 has the above-mentioned loss L _MRT (θ) from the aspect of calculating the loss L (θ) used by the update unit 19 for updating the parameters of the model. ) Can be classified into a first system for calculating as a first loss and a second system for calculating the loss _Lord (θ) as a second loss.

Of these, the first system includes a summary generation unit 12 that generates a system summary, a first probability calculation unit 13 that calculates the generation probability of the system summary, and a duplication that calculates the multiplicity of the system summary and the reference summary. A degree calculation unit 14 and a first loss calculation unit 15 for calculating the first loss described above are included.

Hereinafter, the processing contents in the first system of the second model learning will be described with reference to FIG. 12. FIG. 12 is a diagram showing an example of input / output to the model in the first system. FIG. 12 shows a case where the input control unit 10I acquires a pair of the input sentence 30 shown in FIG. 3 and the reference summary 70 shown in FIG. 7A as a learning sample.

As shown in FIG. 12, the model execution unit 11 vectorizes the word string included in the input sentence 30 acquired by the input control unit 10I, similarly to the model execution unit 6 described above. That is, the model execution unit 11 expands M LSTM11a-1 to 11a-M corresponding to the number of words M of the input sentence 30 in the work area used by the model execution unit 11. These M LSTM11a-1 to 11a-n function as the RNN encoder 11A. Then, the input control unit 10I inputs the word of the input sentence 30 into the LSTM11a corresponding to the position of the word in order from the first word included in the input sentence 30, and inputs the output of the previous LSTM11a. By repeating such input from LSTM11a-1 corresponding to the first word "our company" to LSTM11a-M corresponding to the last word ".", The vector of the input sentence 30 is obtained. The vector of the input sentence 30 generated by the RNN encoder 11A in this way is input to the RNN decoder 11B.

After that, the model execution unit 11 inputs the vector of the input sentence 30, the word predicted one time ago, the number of characters remaining until the RNN decoder 11B outputs the EOS, and the like, and the word is output at each time until the EOS is output. Calculate the probability distribution repeatedly.

Here, in the second model learning, unlike the first model learning, the word generated one hour before is not the correct word one hour before each time of the RNN decoder 11B, but the word generated one hour before is input by the input control unit 10I. Entered. Further, in the second model learning, the loss of the system summarization with respect to the reference summarization is not calculated for each time of the RNN decoder 6B as in the first model learning. That is, in the second model learning, as shown in FIG. 12, a system summary is generated by repeatedly outputting words based on the probability distribution of words from LSTM11b corresponding to each time until the EOS tag is output. After that, the loss of the system summary is calculated.

For example, at the first time of predicting the first word of the system summary, the input control unit 10I outputs the LSTM11a-M and the initial symbol to the LSTM11b-1 expanded in the work area used by the model execution unit 11. Enter the number of characters "37" of the reference summary 70 together with "BOS" as the number of remaining characters. Here, as an example of the maximum number of characters, the case where the number of characters in the reference summary is adopted is illustrated, but the number of characters may be limited to be shorter than the number of characters in the reference summary, or may be limited to the number of characters longer than the number of characters in the reference summary. You can also. As a result, the probability distribution of the word at the first time (t = 1) is output by LSTM11b-1. Based on the probability distribution of this word, the summary generator 12 extracts the first word of the system summary. For example, the summary generation unit 12 can execute a lottery according to the probability distribution of words and extract the winning words by the lottery. In addition, the summary generation unit 12 randomly samples one word from the words belonging to a predetermined number having a high probability, for example, the top five. Here, in the example shown in FIG. 12, the processing after the second time will be described by taking as an example the case where "call center" is randomly sampled as the first word of the system summary.

Subsequently, at the second time when the second word from the beginning of the system summary is predicted, the input control unit 10I gives the LSTM11b-2 the output of the LSTM11b-1 and the prediction result "call center" one hour before. The number of characters "30" obtained by subtracting the number of characters of the word predicted at the first time from the number of remaining characters at the time is input as the number of characters remaining at the second time. As a result, the probability distribution of the word at the second time (t = 2) is output by LSTM11b-2. By executing the lottery of words based on the probability distribution of the words, the summary generation unit 12 samples the words won in the lottery.

After that, the summary generation unit 12 samples the words of the system summary at each time until EOS is output by LSTM11b-K. By generating system summaries by such sampling, the summarization generation unit 12 can generate a predetermined number, for example, S system summaries y'for one input sentence. When S system summaries are generated in this way, the first probability calculation unit 13 is based on the probabilities of the words generated at each time of the system summaries y'for each S system summaries y'. The generation probability p (y'| x, θ) in which the system summary y'is generated from the input sentence x is calculated.

Here, in the second model learning, the first loss _LMRT (θ) is calculated according to the above equation (1). That is, the first loss _LMRT (θ) is the system summary and reference summary calculated by the multiplicity calculation unit 14, which will be described later, in addition to the generation probability of the system summary calculated by the first probability calculation unit 13. Calculated based on the degree of duplication of words between.

As shown in FIG. 12, the multiplicity Δ (y', y) used for calculating the first loss is not necessarily calculated using all the words included in the system summary. That is, the multiplicity calculation unit 14 has, for each of the S system summaries generated by the summary generation unit 12, between the reference summary and the sentence within the upper limit of the number of characters in the system summary, for example, the number of characters of the reference summary. Calculate the degree of word duplication. Thereby, the word of the part exceeding the upper limit number of characters in the system summary, that is, the hatched part shown in FIG. 12 can be excluded from the calculation target of the degree of duplication.

For example, as shown in the following equation (6), the duplication degree calculation unit 14 includes a trim function that cuts out the word corresponding to the n-byte character string corresponding to the maximum number of characters from the beginning of the character string of the system summary. The degree of duplication of n-gram can be calculated according to the function.

FIG. 13 is a diagram showing an example of a method for calculating the degree of overlap. FIG. 13 shows an example in which the multiplicity Δ (y ′, y) is calculated according to the above equation (6). As shown in FIG. 13, in the system summary y', the _first word y'1, the _second word y'2 from the beginning, ..., The k-1th word from the beginning y'k-1, and the first word y'k _-1 . The _k -th word y'k, ..., And the last word y' _{| y'|} are included. On the other hand, the reference summary y includes the first word y ₁ , the second word y ₂ , ..., And the last word y _{| y |} . In this case, a word having a number of bytes corresponding to the reference summary y from the system summary y'by trim (y', byte (y)), that is, the first word y'1, the _second word y'2 from the _first , ... -The _k -1th word y'k-1 from the beginning is cut off. Then, the trim (trim (y', byte (y), y) cut out from the first word y'1 of the system summary y'to the k-1th word y'k _{-1 by} ROUGE (trim (y', byte (y)), y). The degree of word duplication between y', byte (y)) and the reference summary y is calculated. By calculating the multiplicity Δ (y ′, y) according to the above equation (6) in this way, the words from the kth to the end of the system summary y ′ exceeding the upper limit, that is, the words y ′ _k to the word. y' _{| y'|} can be excluded from the calculation target of the multiplicity. As a result, the word from the kth to the end of the system summary y'that exceeds the maximum number of characters, that is, the word _y'k to the word y' _{| y'|} contains a word that overlaps with the reference summary y. Therefore, it is possible to prevent the system summary y'from being overestimated.

In addition to limiting the calculation target of the degree of duplication to words within the maximum number of characters in the system summary, as shown in the following equation (7), the duplication degree calculation unit 14 is insufficient for the maximum number of characters in the system summary. It is also possible to calculate the length or the length exceeding the upper limit of the number of characters in the system summary as an error of giving a penalty to the degree of duplication. In addition, "C" shown in the following formula (7) refers to hyperparameters set by the developer or user of the above learning program.

FIG. 14 is a diagram showing an example of a method for calculating the degree of overlap with an error. FIG. 14 shows an example in which the multiplicity Δ (y ′, y) with an error is calculated according to the above equation (7). In the example shown in FIG. 14, as in the example shown in FIG. 13, the first word y'1 to k- ₁ of the system summary y'by ROUGE (trim (y', byte (y), y)). The degree of overlap between the trim (y', byte (y)) cut out to the word y'k _-1 and the reference summary y is calculated. Further, according to the above equation (7), the absolute value of the difference in length between the system summary and the reference summary, for example | byte (y')-byte (y) |, is given to the multiplicity as an error. .. For example, in the example of FIG. 14, since the length of the system summary is larger than that of the reference summary, the length byte (y')-byte (y) exceeding the maximum number of characters is added to the multiplicity. As a result, the multiplicity Δ (y ′, y) with an error is calculated. In this way, an error | byte (y')-byte (y) | is added to the overlap degree calculated by ROUGE according to the above equation (7) to calculate the overlap degree Δ (y', y) with an error. do. As a result, the loss of the system summary that does not reach the maximum number of characters and the system summary that exceeds the maximum number of characters increases, and as a result, it is possible to realize model learning that enhances the evaluation of the system summary whose number of characters matches the maximum number of characters.

Further, the multiplicity calculation unit 14 does not necessarily have to calculate the error to be given to the multiplicity even in the system summary that does not necessarily reach the upper limit of the number of characters. For example, the multiplicity calculation unit 14 can calculate the length exceeding the upper limit number of characters of the system summary as an error by narrowing down the case where the system summary exceeds the upper limit number of characters according to the following equation (8).

FIG. 15 is a diagram showing an example of a method for calculating the degree of overlap with an error. FIG. 15 shows an example in which the multiplicity Δ (y ′, y) with an error is calculated according to the above equation (8). In the example shown in FIG. 15, as in the example shown in FIG. 13, the first word y'1 to k- ₁ of the system summary y'by ROUGE (trim (y', byte (y), y)). The degree of overlap between the trim (y', byte (y)) cut out to the word y'k _-1 and the reference summary y is calculated. Further, when the system summary exceeds the maximum number of characters, max (0, byte (y')-byte (y)) adds the length byte (y')-byte (y) exceeding the maximum number of characters to the degree of duplication. By doing so, the multiplicity Δ (y ′, y) with an error is calculated. On the other hand, when the system summary does not reach the upper limit of characters, "0" is selected by max (0, byte (y')-byte (y)), so that no error is given to the degree of duplication and the degree of duplication is increased. It is calculated as Δ (y ′, y) as it is. As a result, the loss of the system summary exceeding the maximum number of characters increases without increasing the loss of the system summary that does not reach the maximum number of characters, and as a result, model learning that enhances the evaluation of the system summary within the maximum number of characters can be realized.

After the multiplicity Δ (y ′, y) with such an error is calculated, the first loss calculation unit 15 is used for each predetermined number generated by the summary generation unit 12, for example, S system summaries. The first loss is calculated from the calculation result of the generation probability that the system summary is generated from the input statement and the multiplicity Δ (y ′, y) with an error calculated by the multiplicity calculation unit 14. Further, the first loss calculation unit 15 relates to a set S (x, θ) of S system summaries y'by executing a calculation calculated for each S system summaries and summing up the first losses. The sum of the first losses is calculated.

Returning to the description of FIG. 1, the second system includes a pseudo-sentence generation unit 16 that generates a pseudo-sentence, a second probability calculation unit 17 that calculates a reference summary generation probability and a pseudo-sentence generation probability, and the above. The second loss calculation unit 18 for calculating the second loss of the above is included.

For example, the pseudo-sentence generation unit 16 has a set S'(y) of pseudo-sentences z in which a non-grammatical expression is simulated by exchanging the word order of the words included in the reference summary y of the correct answer. ) Is generated. At this time, the pseudo-sentence generation unit 16 calculates the pseudo-sentence z by changing the word order of the words without changing the number of words included in the correct reference summary y. It is possible to generate a pseudo-sentence z whose ROUGE value is "1".

Here, when the second loss is calculated, the configuration of the RNN encoder 11A, the input to the RNN encoder 11A, and the output from the RNN encoder 11A are not different from the case where the first loss is calculated. On the other hand, when the second loss is calculated, the configuration of the RNN encoder 11A, the input to the RNN encoder 11A and the output from the RNN encoder 11A are different from the case where the first loss is calculated.

For example, when the generation probability of the pseudo sentence z used for calculating the second loss is calculated, the model execution unit 11 inputs by the input control unit 10I according to the model information stored in the first model storage unit 3. Expand J LSTMs corresponding to the number of words J of the pseudo sentence z to be performed on the work area. As a result, J RSTMs are made to function as the RNN decoder 11B. An intermediate expression of the input sentence x of the learning sample is input from the RNN encoder 11A to these RNN decoders 11B according to the input control of the input control unit 10I, and the tags of the input control units 10I to EOS are input for each J LSTM. The number of characters remaining until is output is input. Further, the word of the pseudo sentence z one hour ago is input to the J LSTMs of the RNN decoder 11B according to the input control of the input control unit 10I. By operating the J LSTMs according to these inputs, the RNN decoder 11B outputs the word probabilities of the pseudo sentence z at each time for each J LSMT. In this way, based on the probability of the word at each time of the pseudo sentence z output by each LSMT of the RNN decoder 11B, the second probability calculation unit 17 generates the generation probability p in which the pseudo sentence z is generated from the input sentence x ( z | x; θ) is calculated.

Hereinafter, the processing contents in the second system of the second model learning will be described with reference to FIG. FIG. 16 is a diagram showing an example of input / output to the model in the second system. In FIG. 16, the input sentence 30 shown in FIG. 3 is input to the RNN encoder 11A by the input control unit 10I, and the word at each time of the pseudo sentence z, which is the same sentence as the system summary shown in FIG. 7D, is the RNN decoder. An example of inputting to 11B is shown. Since the configuration of the RNN encoder 11A, the input to the RNN encoder 11A, and the output from the RNN encoder 11A are the same as those shown in FIG. 12, the description of the RNN decoder 11B will be started.

As shown in FIG. 16, the model execution unit 11 inputs the vector of the input sentence 30, the word at each time of the pseudo sentence z, the number of characters remaining until the RNN decoder 11B outputs the EOS, and the like, and outputs the EOS. Repeatedly calculate the probability distribution of words at each time.

Here, when the generation probability of the pseudo sentence z is calculated, unlike the case where the system summary is generated, it is included in the pseudo sentence z instead of the word generated one hour before in the LSTM11b at each time of the RNN decoder 11B. The word of the pseudo sentence z one time before is input by the input control unit 10I.

For example, at the first time, the input control unit 10I has the output of the LSTM11a-M and the number of characters of the reference summary 70 together with the initial symbol "BOS" for the LSTM11b-1 expanded in the work area used by the model execution unit 11. Enter "37" as the number of remaining characters. Here, as an example of the maximum number of characters, the case where the number of characters in the reference summary is adopted is illustrated, but the number of characters may be limited to be shorter than the number of characters in the reference summary, or may be limited to the number of characters longer than the number of characters in the reference summary. You can also. As a result, the probability distribution of the word at the first time (t = 1) is output by LSTM11b-1. At this time, the second probability calculation unit 17 stores the probability corresponding to the word "AI" at the beginning of the pseudo sentence z in the work area (not shown) in the probability distribution of the word at the first time.

Subsequently, at the second time, the input control unit 10I sends the RSTM11b-2 the output of the SSTM11b-1 and the word "AI" of the pseudo-sentence z one hour before, and the first hour from the number of remaining characters in the first hour. The number of characters "35" obtained by subtracting the number of characters of the word "AI" of the pseudo sentence z of the above is input as the number of remaining characters at the second time. As a result, the probability distribution of the word at the second time (t = 2) is output by LSTM11b-2. At this time, the second probability calculation unit 17 stores the probability corresponding to the second word "no" from the beginning of the pseudo sentence z in the work area (not shown) in the probability distribution of the word at the second time.

After such input to the RNN decoder 11B is repeated until the J-2 time, the input control unit 10I outputs the LSTM11b-J-2 to the LSTM11b-J-1 at the J-1 time. And the number of characters "5" obtained by subtracting the number of characters of the word "sale" of the pseudo sentence z at the J-2 time from the number of remaining characters at the first time together with the word "sale" of the pseudo sentence z one hour before is the J-1 time. Enter as the number of characters remaining in the eye. As a result, the probability distribution of words at the J-1 time (t = J-1) is output by LSTM11b-J-1. At this time, the second probability calculation unit 17 stores the probability corresponding to the J-1st word "inquiry" from the beginning of the pseudo sentence z in the work area (not shown) in the probability distribution of the word at the J-1th time. do.

Finally, at the J time, the input control unit 10I tells LSTM11b-J from the number of remaining characters at the first time together with the output of LSTM11b-J-1 and the word "inquiry" of the pseudo sentence z one hour before. -Enter the number of characters "0" obtained by subtracting the number of characters of the word "inquiry" of the pseudo sentence z at the time 1 as the number of remaining characters at the J time. As a result, the probability distribution of the word at the J time (t = J) is output by LSTM11b-J. At this time, the second probability calculation unit 17 stores the probability corresponding to the Jth word "EOS" from the beginning of the pseudo sentence z in the work area (not shown) in the probability distribution of the word at the J time.

Based on the probability of the word at each time of the pseudo sentence z stored in the work area in this way, the second probability calculation unit 17 generates a generation probability p (z | x) in which the pseudo sentence z is generated from the input sentence x. ; Θ) is calculated. Thereby, the generation probability of the pseudo sentence z can be obtained for each pseudo sentence z.

Even when the generation probability of the reference summary y used for calculating the second loss is calculated, the generation probability of the reference summary y can be calculated in the same manner as in the case of calculating the generation probability of the pseudo sentence z. That is, the model execution unit 11 puts I LSTMs corresponding to the number of words I of the reference summary y input by the input control unit 10I on the work area according to the model information stored in the first model storage unit 3. Expand to. As a result, I RSTMs are made to function as the RNN decoder 11B. An intermediate expression of the input sentence x of the learning sample is input from the RNN encoder 11A to these RNN decoders 11B according to the input control of the input control unit 10I, and the tags of the input control units 10I to EOS are input for each I LSTM. The number of characters remaining until is output is input. Further, the word of the reference summary y one hour ago is input to the I LSTMs of the RNN decoder 11B according to the input control of the input control unit 10I. By operating the I LSTMs according to these inputs, the RNN decoder 11B outputs the word probabilities at each time of the reference summary y for each I LSMT. In this way, based on the probability of the word at each time of the reference summary y output by each LSMT of the RNN decoder 11B, the second probability calculation unit 17 generates the generation probability p in which the reference summary y is generated from the input sentence x ( y | x; θ) is calculated.

After the generation probability of the pseudo sentence z is calculated in this way, the second loss calculation unit 18 compares the generation probability of the pseudo sentence z and the generation probability of the reference summary y. At this time, when the generation probability of the pseudo sentence z is larger than the generation probability of the reference summary y, the second loss calculation unit 18 determines the difference between the generation probability of the pseudo sentence z ₁ and the generation probability of the reference summary y, that is, p ( z | x; θ) −p (y | x; θ) is calculated as the second loss. On the other hand, when the generation probability of the pseudo sentence z is not larger than the generation probability of the reference summary y, the second loss calculation unit 18 calculates a predetermined set value, for example, a value of zero or more as the second loss. After that, the second loss calculation unit 15 executes a calculation calculated for each pseudo-sentence z and sums the second losses, so that the second loss calculation unit 15 relates to a set S'(y) of S'pseudo-sentences z. Calculate the sum of losses.

As described above, for all the training samples included in the training data, the process of calculating the sum of the first losses for the S system summaries and the sum of the second losses for the S'pseudo-sentence z is repeatedly executed. Will be done. When the sum of the first loss and the sum of the second losses are calculated for all the training samples included in the training data in this way, the update unit 19 performs the objective function L (θ) shown in the above equation (2). ) Is minimized Update the model parameter to the model parameter θ. The parameters of the model updated in this way are stored in the second model storage unit 8. This update of the parameter θ can be repeated a predetermined number of times for the training data D. As a result, the model information stored in the second model storage unit 8 can be provided as a generation model of the summary sentence.

[Processing flow]
FIG. 17 is a flowchart showing the procedure of the learning process according to the first embodiment. The flowchart of the learning process shown in FIG. 17 is a diagrammatic representation of the procedure of the second model learning executed by the second learning unit 10. FIG. 17 shows, as an example, a flowchart of an example in which the degree of duplication with an error is calculated according to the above equation (8). For example, from the aspect of improving the learning speed of the model in the second learning unit 10, the model learned by the first learning unit 5 after executing the first model learning by the first learning unit 5 as a preprocessing. The learning process shown in FIG. 17 can be started using the parameters of.

As shown in FIG. 17, the processes of steps S101 to S103 are executed for each of the D training samples included in the training data. That is, the input control unit 10I acquires one of the learning samples included in the learning data stored in the learning data storage unit 2 (step S101).

By inputting the learning sample acquired in step S101 into the first system in this way, the first loss calculation process is executed (step S102).

(1) First Loss Calculation Process FIG. 18 is a flowchart showing a procedure of the first loss calculation process according to the first embodiment. This process corresponds to the process of step S102 described above. As shown in FIG. 18, the summary generator 12 samples the words at each time based on the probability distribution of the words output from the RNN decoder, so that the S of the input sentence x of the learning sample acquired in step S101 is satisfied. Generate system summaries y'(step S301). Then, the first probability calculation unit 13 calculates the generation probability of the S system summaries y'generated in step S301 (step S302).

After that, the processes of the following steps S303 to the following steps S306 are executed for each of the S system summaries y'generated in the step S301. That is, the multiplicity calculation unit 14E cuts out a word having an upper limit number of characters, for example, a number of bytes corresponding to the reference summary y from the system summary y'according to the trim (y', byte (y)) shown in the above equation (8). (Step S303).

Then, the multiplicity calculation unit 14 has trim (y', byte (y)) cut out in step S303 according to ROUGE (trim (y', byte (y), y) shown in the above equation (8). )) And the multiplicity of words with the reference summary y (step S304).

Further, the overlap degree calculation unit 14 has a length byte (y) for which the system summary y'exceeds the upper limit number of characters according to max (0, byte (y')-byte (y)) shown in the above equation (8). ′) -Byte (y) is calculated as an error (step S305). If the system summary does not reach the upper limit of characters, "0" is selected by max (0, byte (y')-byte (y)), so the error given to the degree of duplication is calculated as "0". To.

By adding the error calculated in step S305 to the overlap degree calculated in step S304, the overlap degree Δ (y ′, y) with an error is derived.

After that, the first loss calculation unit 15 calculates the first loss from the calculation result of the probability for the system summary y'calculated in step S302 and the multiplicity Δ (y', y) with an error (the first loss calculation unit 15). Step S306).

When the first loss is calculated for each of the S system summaries y'generated in step S301, the first loss calculation unit 15 sums up the first losses calculated for each of the S system summaries. By executing the calculation to be performed, the sum of the first losses corresponding to the set S (x, θ) of the system summary y'is calculated (step S307), and the process of step S102 shown in FIG. 17 is terminated. ..

Returning to the description of FIG. 17, the learning sample acquired in step S101 is input to the second system, so that the second loss calculation process is executed (step S103).

(2) Second Loss Calculation Process FIG. 19 is a flowchart showing a procedure of the second loss calculation process according to the first embodiment. This process corresponds to the process of step S103 described above. As shown in FIG. 19, the pseudo-sentence generation unit 16 replaces the word order of the words included in the reference summary y of the correct answer to the pseudo-sentence z in which a non-grammatical expression is simulated. The set S'(y) is generated (step S501).

After that, the processes of the following steps S502 to the following steps S505 are executed for each S'pseudo-sentence z generated in step S501. That is, the second probability calculation unit 17 calculates the generation probability p (z | x; θ) in which the pseudo sentence z is generated from the input sentence x (step S502). Then, the second loss calculation unit 18 compares the generation probability of the pseudo sentence z calculated in step S502 and the generation probability of the reference summary y (step S503).

Here, when the generation probability of the pseudo sentence z is larger than the generation probability of the reference summary y (step S503Yes), the second loss calculation unit 18 executes the following processing. That is, the second loss calculation unit 18 has _a difference between the generation probability of the pseudo sentence z1 and the generation probability of the reference summary y according to the above equation (3), that is, p (z | x; θ) -p (y | x; θ) is calculated as the second loss (step S504).

On the other hand, when the generation probability of the pseudo sentence z is not larger than the generation probability of the reference summary y (step S503No), the second loss calculation unit 18 has a predetermined set value, for example, zero or more according to the above equation (3). Is calculated as the second loss (step S505).

After that, when the second loss is calculated for each S'pseudo-sentence z generated in step S501, the second loss calculation unit 18 executes the following processing. That is, the second loss calculation unit 18 corresponds to the set S'(x, θ) of the pseudo-sentence z by executing the calculation of summing the second losses calculated for each S'pseudo-sentence. The sum of the second losses to be performed is calculated (step S506), and the process of step S103 shown in FIG. 17 is terminated.

Then, for all the training samples included in the training data, the sum of the first losses corresponding to the set S (x, θ) of the system summary y'and the set S'(x, θ) of the pseudo-sentence z correspond. When the sum of the second losses is calculated, the update unit 19 has the minimum objective function L (θ) in which the parameters of the model stored in the second model storage unit 8 are shown in the above equation (2). The parameter θ of the model to be converted is updated (step S104), and the process is terminated.

[One aspect of the effect]
As described above, the learning device 1 according to the present embodiment replaces the word order of the words included in the reference summary of the correct answer to generate a pseudo sentence in which the non-grammatical expression is simulated, and the model is simulated. Update the model parameters so that the model is more likely to generate a reference summary than it is to generate a statement. For this reason, the degree of duplication of the reference summary and the word is high, and while giving the effect of increasing the probability of generating the system summary having a different word order from the reference summary, it is a non-grammatical expression even in the summary sentence with the high degree of duplication of the reference summary and the word. It is possible to give a reaction that imposes a penalty on the generation of a pseudo-sentence containing. Therefore, it is possible to update the parameters that increase the probability of generating a system summary that does not include non-grammatical expressions even in a summary sentence with a high degree of duplication of a reference summary and a word. Therefore, according to the learning device 1 according to the present embodiment, it is possible to suppress the learning of a model that generates a summary sentence having low readability.

Although the embodiments relating to the disclosed apparatus have been described so far, the present invention may be implemented in various different forms other than the above-described embodiments. Therefore, another embodiment included in the present invention will be described below.

[Distributed and integrated]
Further, each component of each of the illustrated devices does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of them may be functionally or physically distributed / physically in any unit according to various loads and usage conditions. Can be integrated and configured. For example, the first learning unit 5 or the second learning unit 10 may be connected via a network as an external device of the learning device 1. Further, another device may have the first learning unit 5 or the second learning unit 10, respectively, and may realize the function of the learning device 1 by being connected to a network and cooperating with each other. Further, another device has all or a part of the learning data storage unit 2, the first model storage unit 3, or the second model storage unit 8, respectively, and is connected to a network to cooperate with each other to achieve the above learning. The function of the device 1 may be realized.

[Learning program]
Further, the various processes described in the above embodiment can be realized by executing a program prepared in advance on a computer such as a personal computer or a workstation. Therefore, in the following, an example of a computer that executes a learning program having the same function as that of the above embodiment will be described with reference to FIG. 20.

FIG. 20 is a diagram showing a hardware configuration example of a computer that executes the learning program according to the first and second embodiments. As shown in FIG. 20, the computer 100 includes an operation unit 110a, a speaker 110b, a camera 110c, a display 120, and a communication unit 130. Further, the computer 100 has a CPU 150, a ROM 160, an HDD 170, and a RAM 180. Each of these 110 to 180 parts is connected via the bus 140.

As shown in FIG. 20, the HDD 170 stores a learning program 170a that exhibits the same function as the second learning unit 10 shown in the first embodiment. The learning program 170a may be integrated or separated as in the case of each component of the second learning unit 10 shown in FIG. That is, not all the data shown in the first embodiment may be stored in the HDD 170, and the data used for processing may be stored in the HDD 170.

Under such an environment, the CPU 150 reads the learning program 170a from the HDD 170 and deploys it to the RAM 180. As a result, the learning program 170a functions as a learning process 180a, as shown in FIG. The learning process 180a expands various data read from the HDD 170 into an area allocated to the learning process 180a in the storage area of the RAM 180, and executes various processes using the expanded various data. For example, as an example of the process executed by the learning process 180a, the process shown in FIGS. 17 to 19 is included. In the CPU 150, not all the processing units shown in the first embodiment need to operate, and it is sufficient that the processing units corresponding to the processes to be executed are virtually realized.

The learning program 170a may not necessarily be stored in the HDD 170 or the ROM 160 from the beginning. For example, the learning program 170a is stored in a "portable physical medium" such as a flexible disk inserted into the computer 100, a so-called FD, a CD-ROM, a DVD disk, a magneto-optical disk, or an IC card. Then, the computer 100 may acquire and execute the learning program 170a from these portable physical media. Further, the learning program 170a is stored in another computer or server device connected to the computer 100 via a public line, the Internet, a LAN, a WAN, or the like, and the computer 100 acquires and executes the learning program 170a from these. You may try to do it.

The following additional notes will be further disclosed with respect to the embodiments including the above embodiments.

(Appendix 1) This is a learning method for machine learning of a model that generates a summary sentence from an input sentence.
Get the input sentence and the summary sentence of the correct answer,
By exchanging the word order of the words included in the summary sentence of the correct answer, a pseudo sentence in which the non-grammatical expression is simulated is generated.
Based on the generation probability of the pseudo-sentence in which the pseudo-sentence is generated from the input sentence by the model, and the generation probability of the summary sentence of the correct answer in which the summary sentence of the correct answer is generated from the input sentence by the model. Update the parameters of the model,
A learning method characterized by a computer performing processing.

(Appendix 2) The learning method according to Appendix 1, wherein the updating process updates the parameters of the model so that the generation probability of the summary sentence of the correct answer is higher than the generation probability of the pseudo sentence. ..

(Appendix 3) In the update process, when the generation probability of the pseudo sentence is higher than the generation probability of the summary sentence of the correct answer, the difference between the generation probability of the pseudo sentence and the generation probability of the summary sentence of the correct answer is lost. When the parameters of the model are updated by adding and the generation probability of the pseudo sentence is not higher than the generation probability of the summary sentence of the correct answer, the difference between the generation probability of the pseudo sentence and the generation probability of the summary sentence of the correct answer is calculated. The learning method according to Appendix 2, wherein the parameters of the model are updated without adding to the loss.

(Appendix 4) For each of a plurality of summary sentences generated by inputting the input sentence into the model, the generation probability of the summary sentence in which the summary sentence is generated from the input sentence by the model is calculated.
The computer further executes a process of calculating the degree of duplication of words in the summary sentence and the correct answer summary sentence for each of the plurality of summary sentences.
The updating process includes the probability of generating the summary sentence calculated for each of the plurality of summary sentences, the degree of duplication of words calculated for each of the plurality of summary sentences, the probability of generating the pseudo sentence, and the correct answer. The learning method according to Appendix 1, wherein the parameters of the model are updated based on the generation probability of the summary sentence.

(Appendix 5) The learning method according to Appendix 1, wherein the generation process generates the pseudo sentence by changing the word order of the words without changing the number of words included in the summary sentence of the correct answer. ..

(Appendix 6) A learning program that executes machine learning of a model that generates a summary sentence from an input sentence.
Get the input sentence and the summary sentence of the correct answer,
By exchanging the word order of the words included in the summary sentence of the correct answer, a pseudo sentence in which the non-grammatical expression is simulated is generated.
Based on the generation probability of the pseudo-sentence in which the pseudo-sentence is generated from the input sentence by the model, and the generation probability of the summary sentence of the correct answer in which the summary sentence of the correct answer is generated from the input sentence by the model. Update the parameters of the model,
A learning program characterized by having a computer perform processing.

(Appendix 7) The learning program according to Appendix 6, wherein the updating process updates the parameters of the model so that the generation probability of the summary sentence of the correct answer is higher than the generation probability of the pseudo sentence. ..

(Appendix 8) In the update process, when the generation probability of the pseudo sentence is higher than the generation probability of the summary sentence of the correct answer, the difference between the generation probability of the pseudo sentence and the generation probability of the summary sentence of the correct answer is lost. When the parameters of the model are updated by adding and the generation probability of the pseudo sentence is not higher than the generation probability of the summary sentence of the correct answer, the difference between the generation probability of the pseudo sentence and the generation probability of the summary sentence of the correct answer is calculated. The learning program according to Appendix 7, wherein the parameters of the model are updated without adding to the loss.

(Appendix 9) For each of a plurality of summary sentences generated by inputting the input sentence into the model, the generation probability of the summary sentence in which the summary sentence is generated from the input sentence by the model is calculated.
For each of the plurality of summary sentences, the computer is further executed to calculate the degree of duplication of words in the summary sentence and the correct summary sentence.
The updating process includes the probability of generating the summary sentence calculated for each of the plurality of summary sentences, the degree of duplication of words calculated for each of the plurality of summary sentences, the probability of generating the pseudo sentence, and the correct answer. The learning program according to Appendix 6, characterized in that the parameters of the model are updated based on the probability of generating a summary sentence.

(Appendix 10) The learning program according to Appendix 6, wherein the generated process generates the pseudo sentence by changing the word order of the words without changing the number of words included in the summary sentence of the correct answer. ..

(Appendix 11) A learning device that performs machine learning of a model that generates a summary sentence from an input sentence.
The acquisition unit that acquires the input sentence and the summary sentence of the correct answer,
A pseudo-sentence generator that generates a pseudo-sentence in which non-grammatical expressions are simulated by exchanging the word order of the words included in the correct summary sentence.
Based on the generation probability of the pseudo-sentence in which the pseudo-sentence is generated from the input sentence by the model, and the generation probability of the summary sentence of the correct answer in which the summary sentence of the correct answer is generated from the input sentence by the model. An update unit that updates the parameters of the model,
A learning device characterized by having.

(Supplementary Note 12) The learning device according to Supplementary Note 11, wherein the updating unit updates the parameters of the model so that the generation probability of the summary sentence of the correct answer is higher than the generation probability of the pseudo sentence.

(Appendix 13) When the generation probability of the pseudo sentence is higher than the generation probability of the summary sentence of the correct answer, the update unit adds the difference between the generation probability of the pseudo sentence and the generation probability of the summary sentence of the correct answer to the loss. Then, when the parameters of the model are updated and the probability of generating the pseudo-sentence is not higher than the probability of generating the summary of the correct answer, the difference between the probability of generating the pseudo-sentence and the probability of generating the summary of the correct answer is lost. The learning apparatus according to Appendix 12, wherein the parameters of the model are updated without adding to.

(Appendix 14) Probability calculation for calculating the generation probability of the summary sentence generated from the input sentence by the model for each of a plurality of summary sentences generated by inputting the input sentence into the model. Department and
Each of the plurality of abstract sentences further has a multiplicity calculation unit for calculating the degree of duplication of words in the abstract and the correct abstract.
The update unit includes the probability of generating the summary sentence calculated for each of the plurality of summary sentences, the degree of duplication of words calculated for each of the plurality of summary sentences, the probability of generating the pseudo sentence, and the summary of the correct answer. The learning device according to Appendix 11, characterized in that the parameters of the model are updated based on the probability of sentence generation.

(Appendix 15) The learning according to Appendix 11, wherein the pseudo-sentence generation unit generates the pseudo-sentence by changing the word order of the words without changing the number of words included in the summary sentence of the correct answer. Device.

1 Learning device 2 Learning data storage unit 3 First model storage unit 5 First learning unit 5I Input control unit 6 Model execution unit 7 Update unit 8 Second model storage unit 10 Second learning unit 10I Input control unit 11 Model execution unit 12 Summary generation unit 13 First probability calculation unit 14 Duplicate degree calculation unit 15 First loss calculation unit 16 Pseudo sentence generation unit 17 Second probability calculation unit 18 Second loss calculation unit 19 Update unit

Claims (7)

It is a learning method that performs machine learning of a model that generates a summary sentence from an input sentence.
Get the input sentence and the summary sentence of the correct answer,
By exchanging the word order of the words included in the summary sentence of the correct answer, a pseudo sentence in which the non-grammatical expression is simulated is generated.
Based on the generation probability of the pseudo-sentence in which the pseudo-sentence is generated from the input sentence by the model, and the generation probability of the summary sentence of the correct answer in which the summary sentence of the correct answer is generated from the input sentence by the model. Update the parameters of the model,
A learning method characterized by a computer performing processing. The learning method according to claim 1, wherein the updating process updates the parameters of the model so that the generation probability of the summary sentence of the correct answer is higher than the generation probability of the pseudo sentence. In the updating process, when the generation probability of the pseudo sentence is higher than the generation probability of the summary sentence of the correct answer, the difference between the generation probability of the pseudo sentence and the generation probability of the summary sentence of the correct answer is added to the loss. Update the parameters of the model, and if the probability of generating the pseudo-sentence is not higher than the probability of generating the summary of the correct answer, add the difference between the probability of generating the pseudo-sentence and the probability of generating the correct summary to the loss. The learning method according to claim 2, wherein the parameters of the model are updated without any problems. For each of the plurality of summary sentences generated by inputting the input sentence into the model, the generation probability of the summary sentence in which the summary sentence is generated from the input sentence by the model is calculated.
The computer further executes a process of calculating the degree of duplication of words in the summary sentence and the correct answer summary sentence for each of the plurality of summary sentences.
The updating process includes the probability of generating the summary sentence calculated for each of the plurality of summary sentences, the degree of duplication of words calculated for each of the plurality of summary sentences, the probability of generating the pseudo sentence, and the correct answer. The learning method according to any one of claims 1 to 3, wherein the parameters of the model are updated based on the generation probability of the summary sentence. The process of generating is one of claims 1 to 4, characterized in that the pseudo sentence is generated by changing the word order of the words without changing the number of words included in the summary sentence of the correct answer. The learning method described. It is a learning program that executes machine learning of a model that generates a summary sentence from an input sentence.
Get the input sentence and the summary sentence of the correct answer,
By exchanging the word order of the words included in the summary sentence of the correct answer, a pseudo sentence in which the non-grammatical expression is simulated is generated.
Based on the generation probability of the pseudo-sentence in which the pseudo-sentence is generated from the input sentence by the model, and the generation probability of the summary sentence of the correct answer in which the summary sentence of the correct answer is generated from the input sentence by the model. Update the parameters of the model,
A learning program characterized by having a computer perform processing. It is a learning device that performs machine learning of a model that generates a summary sentence from an input sentence.
The acquisition unit that acquires the input sentence and the summary sentence of the correct answer,
A pseudo-sentence generator that generates a pseudo-sentence in which non-grammatical expressions are simulated by exchanging the word order of the words included in the correct summary sentence.
Based on the generation probability of the pseudo-sentence in which the pseudo-sentence is generated from the input sentence by the model, and the generation probability of the summary sentence of the correct answer in which the summary sentence of the correct answer is generated from the input sentence by the model. An update unit that updates the parameters of the model,
A learning device characterized by having.

Images