JP2020071737A - Learning method, learning program and learning device - Google Patents

Abstract

To provide a learning method capable of suppressing learning of models that generate less readable summary.SOLUTION: The learning method of machine learning of a model that generates a summary from an input sentence, causes a computer to execute a series of processing including: acquiring an input sentence and a correct summary; generating a pseudo sentence in which a non-grammatical expression is simulated by changing the word order of the words included in the correct summary; and updating model parameters on the basis of the probability of generation of a pseudo sentence from the input sentence by model and the probability of generation of a correct summary by the model from the input sentence.SELECTED DRAWING: Figure 1

Description

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

  Machine learning such as a neural network may be used for automatic summarization for generating a summary from a document such as a newspaper, a website, or an electronic bulletin board. For example, a model in which an RNN (Recurrent Neural Networks) encoder that vectorizes an input sentence and an RNN decoder that repeats prediction of words in the abstract sentence by referring to the vector of the input sentence is connected is used for generating the abstract sentence.

  As an example of a method of learning such a model, there is a method of calculating a loss used for updating a parameter of a model for each word of a reference summary which is a correct summary sentence corresponding to an input sentence of a learning sample. For example, during model learning, the RNN decoder inputs the vector of the input sentence, the correct word one time before, and the number of characters remaining until the RNN decoder outputs EOS called a sentence end symbol, and outputs the EOS. Iteratively calculates the probability distribution of words at each time. “EOS” here is an abbreviation for “End Of Sentence”. In this way, the loss is calculated by comparing the probability distribution of words 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 number of words in the summary sentence is satisfied to some extent, but the word order of the words is different even if the summary sentence output by the RNN decoder and the correct reference summary have the same meaning. If is different, it is evaluated that a loss occurs.

  From this, an index called ROUGE may be used for evaluation of automatically generated summary sentences. The term "ROUGE" used herein refers to an index indicating the degree of N-gram overlap of words between the correct reference summary and the summary output by the summary generation system in which the model is incorporated. A technique called MRT (Minimum Risk Training) for tuning the parameters of the model of the RNN encoder and the RNN decoder based on such ROUGE has also been proposed.

JP, 2016-62181, A JP, 2013-167985, A JP, 2005-170224, A JP, 2014-123219, A

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 summary sentences whose word order is different from that of the correct reference summary are highly evaluated, so that a model that generates a summary sentence with low readability may be learned.

  That is, in the above MRT, a high ROUGE value is calculated if the degree of word overlap is high even in a summary sentence having a word order different from that of the correct reference summary. In addition, a summary sentence having a high ROUGE value may include a summary sentence having a non-grammatical expression by changing the word order from that of the correct reference summary. In some cases, the model parameters are updated based on the summary sentence having a non-grammatical expression in this way, and a model that generates a summary sentence with low readability may be learned.

  In one aspect, an object of the present invention is to provide a learning method, a learning program, and a learning device that can suppress the learning of a model that generates a low-readable summary sentence.

  In one aspect, a learning method for performing machine learning of a model that generates a summary sentence from an input sentence, wherein an input sentence and a summary sentence of a correct answer are acquired, and the word order of words included in the summary sentence of the correct answer is changed. A pseudo-sentence in which a non-grammatical expression is pseudo-reproduced is generated, the generation probability of the pseudo-sentence in which the pseudo-sentence is generated from the input sentence by the model, and the correct summary sentence by the model. The computer executes a process of updating the parameters of the model based on the generation probability of the correct summary sentence generated from the input sentence.

  It is possible to suppress learning of a model that generates a summary sentence with low readability.

FIG. 1 is a block diagram of the 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 summarizing tool. FIG. 3 is a diagram showing an example of an input sentence. FIG. 4A is a diagram illustrating an example of the 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 contents of MRT. FIG. 6 is a diagram illustrating 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 a method for updating model parameters. FIG. 9 is a diagram illustrating an example of the first model learning. FIG. 10 is a diagram illustrating an example of the first model learning. FIG. 11 is a diagram illustrating an example of the first model learning. FIG. 12 is a diagram showing an example of input / output to / from the model in the first system. FIG. 13 is a diagram showing an example of a method of calculating the degree of overlap. FIG. 14 is a diagram showing an example of a method of calculating the degree of overlap with an error. FIG. 15 is a diagram illustrating an example of a method of calculating the degree of overlap with an error. FIG. 16 is a diagram showing an example of input / output to / from the model in the second system. FIG. 17 is a flowchart illustrating the procedure of the learning process according to the first embodiment. FIG. 18 is a flowchart illustrating the procedure of the first loss calculation process according to the first embodiment. FIG. 19 is a flowchart illustrating the procedure of the second loss calculation process according to the first embodiment. FIG. 20 is a diagram illustrating a hardware configuration example of a computer that executes a learning program according to the first and second embodiments.

  A learning method, a learning program, and a learning device according to the present application will be described below with reference to the accompanying drawings. Note that this embodiment does not limit the disclosed technology. Then, the respective embodiments can be appropriately combined within the range in which the processing contents do not contradict each other.

  FIG. 1 is a block diagram of the 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, electronic bulletin boards, websites, and the like as input sentences and learns a model for generating a summary sentence.

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

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

  Here, the above-mentioned article summarizing tool can be incorporated as one function of an application in which a media business operator who manages various media such as newspapers, electronic bulletin boards, and websites is a user, as one aspect.

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

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

  In the upper part of FIG. 2, an article summary screen 20 in an initial state in which inputs for various items are not set is shown. 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 an original text such as an article. Further, the summary display area 22 corresponds to an area for displaying a summary sentence corresponding to the original sentence input to the original sentence input area 21. Further, the pull-down menu 23 corresponds to an example of a GUI component that specifies the maximum number of characters of the summary sentence. The summary button 24 corresponds to an example of a GUI component that receives execution of a command that generates a summary sentence corresponding to the original sentence input in the original sentence input area 21. 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 receiving text input via the input device in this way, in the original text input area 21, text can be imported from a file of a document 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 for generating the abstract text through the operation on the abstract button 24. In addition, the text entered in the original text input area 21 can be cleared by operating the clear button 25. In addition, via the pull-down menu 23, it is also possible to accept the specification of the upper limit number of characters desired by a person involved in the media business from a plurality of upper limit number of characters. Here, as an example of a scene in which the bulletin board of the electronic bulletin board is generated as a summary sentence from the original text of a newspaper or a news article, an example in which 80 characters corresponding to an example of the maximum number of characters that can be displayed on the electronic bulletin board is designated is shown. There is. This is just an example, and when a headline is generated from an article on a newspaper or a website, the maximum number of characters corresponding to the headline can be selected.

  Then, when the summary button 24 is operated while the original text is input in the original text input area 21, the article summary screen 20 is changed from the state shown in the middle of FIG. 2 to the lower of FIG. To the closed state (step S2). In this case, the text of the original text input to the original text input area 21 is input as an input text to the learned model to generate the summary text. The generation of this summary may be executed on the terminal device of a person involved in the media business, or may be executed on the back-end server device. As a result, as shown in the lower part of FIG. 2, the summary display area 22 of the article summary screen 20 displays the summary sentence generated by the learned model.

  As described above, the text of the summary sentence displayed in the summary display area 22 of the article summary screen 20 can be edited through an input device (not shown) or the like.

  By providing the article summarization tool as described above, it becomes possible to reduce the work of article summarization performed by a reporter, an editor, or the like. In other words, the work of summarizing articles is the most important in the process of distributing news to the media, such as "selection of distribution articles", "send to media editing system", "summary of articles", "creation of headlines", and "review". There is an aspect that it is labor intensive. For example, when the article summarization is performed manually, it is necessary to select important information from the entire article and reconstruct the sentence. From this, the technical significance that the work of article summarization is automated or semi-automated is high.

  In addition, here, as an example, the use case in which the article summarization tool is used by a person involved in 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. For example, an article summarization tool can be used as a function of reading out the summary text instead of reading out the entire text of the article using a smart speaker or the like.

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

  Such a model learning failure case 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 illustrating an example of the reference summary. 4B and 4C are diagrams showing an example of the system summary. In the following, a correct summary sentence included in the learning sample may be referred to as a “reference summary”, and a summary sentence generated from the input sentence by the model may be referred to as a “system summary”.

  Here, as an example, a case where the 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 described as an example. At this time, when the system summary 40B shown in FIG. 4B or the system summary 40C shown in FIG. 4C is generated from the input sentence 30 by the model in which the RNN (Recurrent Neural Networks) encoder and the RNN decoder are connected, the following evaluation is performed. Done.

  That is, in the reference summary 40 shown in FIG. 4A and the system summary 40B shown in FIG. 4B, words match at each position from the beginning to the end. In FIGS. 4A and 4B, as an example, the word located at the fifth position from the beginning of the reference summary 40 and the system summary 40B is shown in bold type. For example, when the fifth word 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 calculated. Will be the best. The word in the reference summary 40 located at the fifth position from the beginning is also “AI”, as shown in FIG. 4A. In this way, if each word included in the reference summary 40 matches the word in the system summary 40B at the position corresponding to the position of the word, 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 8th word from the beginning differs between the reference summary 40 and the system summary 40C. In FIGS. 4A and 4C, as an example, the fifth word from the beginning of the reference summary 40 and the system summary 40C is shown in bold type. 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. Will be the best. On the other hand, the word of the reference summary 40 located fifth from the beginning is “AI” as shown in FIG. 4A. In this way, when the word order is changed between the reference summary 40 and the system summary 40C so that the word arrangement is different, a loss occurs even if the system summary 40C has the same meaning as the reference summary 40.

  These results in different evaluations being made between system summary 40B and system summary 40C. However, the texts of the system summary 40B and the system summary 40C are the same. Therefore, from the aspect of summary, it cannot be said that it is appropriate unless the same evaluation is made, and the system summary 40C is underestimated compared to the system summary 40B.

[Current MRT]
As described above, a technique called MRT (Minimum Risk Training) has been proposed from the viewpoint of suppressing underestimation of a system summary having a word order different from that of the reference summary during model learning. For example, in MRT, the parameters of the model of the RNN encoder and the RNN decoder are tuned based on ROUGE, which 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 contents of MRT. As shown in FIG. 5, a pair of an input sentence x and a correct reference summary y is used as a learning sample for model learning of the RNN encoder and the RNN decoder. The input sentence x of the input sentence x and the correct reference summary y 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 words output at each time from the beginning to 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 a word at each time according to the probability distribution of the word at each time obtained by such calculation. Here, for convenience of description, an example in which _three system summaries y ′ _{1 to} y ′ ₃ are sampled has been described, but an arbitrary number of system summaries y ′ may be sampled.

Then, in MRT, for each system summary y ′ _{1 to} y ′ ₃ , 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 ′. A ROUGE value indicating the degree of overlap of the gram is calculated. Then, in MRT, the loss L _MRT (θ) is calculated from the generation probabilities of the system summaries y ′ _{1 to} y ′ ₃ and the ROUGE value according to the following equation (1).

  Here, “P (y ′ | x; θ)” in the above equation (1) indicates the probability that the system summary y ′ will be generated from the input sentence x, where θ is the parameter of the model. Further, “D” in the above equation (1) indicates learning data that is a set of learning samples including the input sentence x and the reference summary y. Further, “S” in the above equation (1) indicates a set of system summaries generated from the input sentence x, where θ is a parameter of the model. Further, “Δ (y ′, y)” in the above equation (1) indicates the degree of word overlap 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 a gradient, that is, ∂L _MRT (θ) / ∂θ _i by partially differentiating L _MRT (θ) with θ _i , and updates the model parameter θ _i , that is, θ _i ← θ + (∂ Calculate L _MRT (θ) / ∂θ _i ).

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

The change in the loss L _MRT (θ) before and after the parameter update using the ROUGE value will be described with reference to FIG. FIG. 6 is a diagram illustrating an example of the generation probability and the ROUGE value. In the table in the upper part of FIG. 6, the generation probability that the model having the parameter θ _t generates the system summary y ′ from the input sentence x in the model learning at the t-th round, and the ROUGE value between the reference summary and the system summary y ′. And are shown. The thin hatched areas shown in the table of FIG. 6 indicate that they are calculated by the formula for calculating the generation probability of the system summary y ′ included in the above equation (1), while the dark hatched areas shown in the table of FIG. The hatched portion indicates that it is calculated by the ROUGE function included in the above equation (1).

For example, if the generation probability and the ROUGE value 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, L _MRT (θ _{t 1} ) can be calculated as follows. That is, the loss L _MRT (θ _t ) is the ROUGE value of the generation probability of the system summary y ′ ₁ , and the ROUGE value of the generation probability of the system summary y ′ ₂ and the ROUGE value of the system summary y ′ _3. It can be calculated from the sum of the values. That is, the loss L _MRT (θ _t ) is −0.24 by the calculation of 0.2 × (−0.3) + 0.6 × (−0.1) + 0.2 × (−0.6). Is calculated.

The generation probabilities and ROUGE values of the system summaries y ′ _{1 to} y ′ ₃ generated from the input sentence x by the model in which the parameter is updated from θ _t to θ _{t + 1} based on the loss L _MRT (θ _t ) are shown in FIG. It is assumed that the table in the lower row is.

On the other hand, in the lower table shown in FIG. 6, in the model learning of the t + 1th round, the model having the parameter θ _{t + 1} generates the system summary y ′ from the input sentence x, the reference summary and the system summary y ′. And the ROUGE values between. Also in this case, the loss L _MRT (θ _{t + 1} ) is equal to the generation probability and the 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 . It can be calculated from the sum of and and the ROUGE value. 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). Is calculated.

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

[One aspect of current MRT issues]
However, as described in the section of the background art above, when the parameter of the model is updated according to the degree of overlap of words without regard to the difference in the word order as in MRT, all system summaries whose word order is different from the correct reference summary The ROUGE value of is highly evaluated. Therefore, even if a non-grammatical expression is included in the system summary whose word order differs from that of the correct reference summary, the parameters of the model are learned by underestimating the loss of the system summary. As a result, a model that produces a less readable system summary may be trained.

  Such a model learning failure case will be described with reference to FIGS. 7A to 7D. FIG. 7A is a diagram showing an example of a reference summary. 7B to 7D are diagrams showing an example of the 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 learning sample when learning the model will be described. 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 according to the model in which the RNN encoder and the RNN decoder are connected, the following evaluation is performed. ..

  In other words, the reference summary 70 shown in FIG. 7A and the system summary 70B shown in FIG. 7B have the same word order and the same set of words. In this way, since the set of words match between the reference summary 70 and the system summary 70B, the loss is “0”. Further, the reference summary 70 shown in FIG. 7A and the system summary 70C shown in FIG. 7C have different word orders, but the sets of words match. In this way, since the set of words match between the reference summary 70 and the system summary 70C, the loss becomes “0”. Further, the reference summaries 70 shown in FIG. 7A and the system summaries 70D shown in FIG. 7D also have the same word set, although the word order is different. In this way, since the set of words match between the reference summary 70 and the system summary 70D, the loss is “0”. In this way, the same evaluation is made among the system summaries 70B to 70D having the same ROUGE value.

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

  As described above, in the current MRT, if the ROUGE value is at the same level, the system summary 70B or the system summary 70C that does not include the non-grammatical expression and the non-grammatical expression or the incorrect dependency are included. The same evaluation will be made with the system summary 70D. That is, when the system summary 70D including a non-grammatical expression is included in the system summary during model learning, the negative gain of the ROUGE value of the system summary 70D and the system summary 70D of the system summary 70D is negative. Works the same as. In this way, the model is updated based on the loss caused by the excessive effect of the negative gain of the ROUGE value of the system summary 70D including the non-grammatical expression, and as a result, the model that produces the unreadable summary sentence is learned. It may end up.

[One aspect of approach to problem solving]
Therefore, the learning device 1 according to the present embodiment changes the word order of the words included in the correct reference summary to generate a pseudo sentence in which a non-grammatical expression is pseudo reproduced, and the model generates a pseudo sentence. Update the model parameters such that the model has a higher probability of producing a reference summary than the probability.

  FIG. 8 is a diagram showing an example of a method for updating model parameters. As shown in FIG. 8, in the model learning of the RNN encoder and the RNN decoder, as in the MRT shown in FIG. 5, a pair of the input sentence x and the correct reference summary y is used as a learning sample.

The input sentence x of the input sentence x and the correct reference summary y is input to the model. When the input sentence x is input in this way, the learning device 1 determines 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 EOS. ' ₃ is sampled.

Then, the learning device 1 generates, for each system summary y ′ _{1 to} y ′ ₃ , the generation probability that the system summary y ′ is generated from the input sentence x and the words between the reference summary y and the system summary y ′. A ROUGE value representing the degree of overlap of n-gram is calculated. Then, the learning device 1 calculates the loss L _MRT (θ) from the generation probabilities of the system summaries y ′ _{1 to} y ′ ₃ and the ROUGE value according to the above equation (1).

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

That is, the present embodiment is different in that the improved loss is defined from the above MRT. For example, in the present embodiment, the term of loss L _order (θ) that penalizes the pseudo sentence z including the non-grammatical expression together with the term of loss L _MRT (θ) based on the generation probability of the system summary y ′ and the ROUGE value. The loss L (θ) added with is defined by the following equation (2). Note that “α” in the following formula (2) is a weighting coefficient, and for example, an arbitrary value of 0 to 1 can be adopted.

Here, the loss L _order (θ) is calculated by the following equation (3). “D” in the following Expression (3) indicates learning data which is a set of learning samples including the input sentence x and the reference summary y. Further, “S ′ (y)” in the following expression (3) indicates a set of pseudo sentences z generated from the correct reference summary y. Further, “p (z | x; θ)” in the following expression (3) indicates the probability that the 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 Expression (3) indicates the probability that the correct reference summary y is generated from the input sentence x.

For example, the learning device 1 replaces the correct reference summary y with the word order of the words included in the reference summary y so that a set S ′ of pseudo sentences z _{1 to} z ₃ in which a non-grammatical expression is pseudo reproduced. (Y) is generated. At this time, by changing the word order of the words and sampling the pseudo-sentence z without changing the number of words included in the correct reference summary y, the ROUGE value calculated with the reference summary y is “1”. It is possible to generate a pseudo sentence z. Here, for convenience of description, an example in which _three pseudo sentences z _{1 to} z ₃ are sampled has been described, but an arbitrary number of pseudo sentences z may be sampled.

Further, the learning device 1 calculates the generation probability p (y | x; θ) that 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”. The generation probability p (z ₁ | x; θ) of the pseudo sentence z ₁ is calculated as “0.3”. Furthermore, 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”.

A calculation example of the loss L _order (θ) will be described based on the calculation result of the generation probability. For example, in the case of the pseudo sentences _{z 1} of the defined set S to sigma '(y), the pseudo-sentence generation probability _{z 1 (p (z 1 |} x; θ) = 0.3) generating 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 formula (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.

Also, in the case of the pseudo sentences _{z 2,} generation probability of the pseudo sentences _{z 2 (p (z 2 |} x; θ) = 0.4) and the reference Summary y probability generation of (p (y | x; θ ) = 0. 2) is compared with. 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 formula (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.

Also, in the case of the pseudo sentences _{z 3,} generation probability of the pseudo sentences _{z 3 (p (z 3 |} x; θ) = 0.1) and the reference Summary y probability generation of (p (y | x; θ ) = 0. 2) is compared with. 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 formula (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 L _order (θ) can be calculated as 0.3 (= 0.1 + 0.2 + 0) by summing up the losses calculated for each element of these pseudo sentences z _{1 to} z ₃ .

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

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

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

[Functional configuration of learning device 1]
Next, an example of a functional configuration of the learning device 1 according to the present 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. 10 and. The learning device 1 may have various functional units of a known computer other than the functional units shown in FIG. 1, such as functional units such as various input devices and audio 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 processors as an example. Examples of such processors include DLUs (Deep Learning Units) and GPGPUs (General-Purpose computing on Graphics Processing Units), as well as GPU clusters. In addition, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or the like may be used. For example, the processor develops the learning program as a process on a memory such as a RAM (Random Access Memory), so that the functional unit is virtually realized. Here, the DLU, GPGPU, GPU cluster, CPU, and MPU are illustrated as an example of the processor, but the functional unit may be realized by any processor regardless of general-purpose type or specialized type. .. In addition to the above, the above-mentioned functional unit is not prevented from being realized by hard-wired logic such as ASIC (Application Specific Integrated Circuit) and FPGA (Field Programmable Gate Array).

  Further, the 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 a HDD (Hard Disk Drive), an optical disk, an SSD (Solid State Drive), and the like. The storage device of can be adopted. The storage device does not necessarily have to be the auxiliary storage device, and various semiconductor memory elements such as RAM, EPPROM, and flash memory can be adopted.

  Here, in FIG. 1, from the aspect of improving the learning speed of the model in the second learning unit 10, after the pre-processing for learning the parameters of the model is executed by the first learning unit 5, the parameters after the pre-processing are performed. The case where the second learning unit 10 is caused to execute the above model learning by using This is merely an example, and the preprocessing by the first learning unit 5 may not necessarily be performed. For example, the pre-processing by the first learning unit 5 may be skipped and the second learning unit 10 may be made to execute the model learning using the initial parameters. Hereinafter, the model learning that is the preprocessing executed by the first learning unit 5 will be referred to as “first model learning”, and the model learning executed by the second learning unit 10 will be described. 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, for example, D learning samples, so-called learning cases. Further, the learning sample includes a pair of input sentence x and reference summary y. Note that FIG. 1 illustrates, as an example, a case where the same learning data is used for the first learning unit 5 and the second learning unit 10, but the first learning unit 5 and the second learning unit are used. It does not matter if different learning data among 10 are used for model learning.

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

  As one embodiment, the following information is stored in the first model storage unit 3 and the second model storage unit 8. For example, starting with the layer structure of a model such as neurons and synapses in each layer of an input layer, a hidden layer and an output layer forming a neural network to which an RNN encoder and an RNN decoder are connected, model parameters such as weight and bias of each layer are set. The model information including is stored. Here, at the stage before the model learning is executed by the first learning unit 5, the first model storage unit 3 stores the parameters initialized by random numbers as the model parameters. Further, at the stage after the model learning is performed 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-described preprocessing. Here, as an example of the first model learning, a case where model learning called optimization of log likelihood is executed is illustrated.

  As shown in FIG. 1, the first learning unit 5 includes an input control unit 5I, a model executing unit 6, and an updating 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 learning sample included in the learning data.

  Specifically, the input control unit 5I initializes the value of the loop counter d that counts the learning samples. Subsequently, the input control unit 5I acquires the learning sample corresponding to the loop counter d from 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 learning sample from the learning data storage unit 2 until the value of the loop counter d becomes equal to the total number D of learning samples. Here, an example in which the learning data saved in the storage inside the learning device 1 is acquired has been described, but the learning data is acquired from an external computer connected via a network, such as a file server, removable media, or the like. It does not matter if it is acquired.

  Each time a 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 obtained by vectorizing the word string of the input sentence x, that is, a so-called intermediate representation is output from the RNN encoder 6A to the RNN decoder 6B. Simultaneously with or before this, the input control unit 5I sets the value of the register that holds the number of remaining characters until the RNN decoder 6B outputs EOS, which is called a sentence end symbol, to a predetermined upper limit number of characters, such as 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 updating 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 the neural network to which the RNN encoder 6A and the RNN decoder 6B are connected.

  As one aspect, the model execution unit 6 uses M model numbers corresponding to the number M of words in 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. Expand LSTM (Long Short-Term Memory) on the work area. As a result, the M LSTMs function as the RNN encoder 6A. In the RNN encoder 6A, the mth word from the beginning of the input sentence of the learning sample is input to the LSTM corresponding to the mth word in order from the beginning word of the input sentence of the learning sample according to the input control by the input control unit 5I. At the same time, the output of the LSTM corresponding to the m−1th 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, that is, a so-called intermediate representation is obtained. The intermediate representation of the input sentence thus generated by the RNN encoder 6A is input to the RNN decoder 6B.

  As a further aspect, the model execution unit 6 according to the model information stored in the first model storage unit 3 has N LSTMs corresponding to the number N of words in the correct reference summary input by the input control unit 5I. On the work area. As a result, the N LSTMs function as the RNN decoder 6B. According to the input control of the input control unit 5I, these RNN decoders 6B receive the intermediate representation of the input sentence of the learning sample from the RNN encoder 6A, and input tags of the EOS from the input control unit 5I for every N LSTMs. The number of characters remaining until output is input. By operating the N LSTMs according to these inputs, the RNN decoder 6B outputs the probability distribution of words for each of the N LSMTs. The “word probability distribution” mentioned here refers to a probability distribution calculated for each word that appears in the input sentence in the entire learning sample.

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

  As one embodiment, when the probability distribution of words is output from the n-th LSTM of the RNN decoder 6B, the updating unit 7 determines the word with the highest probability in the probability distribution as the n-th word from the beginning of the system summary. To generate. After that, when the n-th word of the system summary is generated, the updating unit 7 loses a loss from the n-th word included in the correct reference summary and the n-th 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 updating unit 7 calculates the parameters for updating the models of the RNN encoder 6A and the RNN decoder 6B by executing the log-likelihood optimization based on the loss of each LSTM. Then, the updating 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. This parameter update can be repeatedly executed for all learning samples and also for the learning data D over a predetermined number of epochs.

  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 illustrating an example of the first model learning. 9 to 11 show a case where the input control unit 5I acquires the 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 develops M LSTMs 6a-1 to 6a-M corresponding to the number M of words of the input sentence 30 in the work area used by the model execution unit 6. This causes the M LSTMs 6a-1 to 6a-M to function as the RNN encoder 6A. Then, the input control unit 5I inputs the words of the input sentence 30 in order from the first word included in the input sentence 30 to the LSTM 6a corresponding to the position of the word, and inputs the output of the previous LSTM 6a. 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 thus generated by the RNN encoder 6A 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 before, and the number of remaining characters until the RNN decoder 6B outputs EOS called a sentence end symbol. Iteratively calculate the probability distribution of words.

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

  Further, at the second time when the probability distribution of the word to be matched with the second word from the beginning of the reference summary 70 is calculated, the operation shown in FIG. 10 is performed. That is, as shown in FIG. 10, the input control unit 5I inputs the output of the LSTM6b-1 and the correct word "call center" one hour before to the LSTM6b-2, and at the same time from the remaining number of characters at the first time to the first time. The number of letters "30", which is the number of letters of the correct word for the eyes, is input as the remaining number of letters at the second time. As a result, the LSTM6b-2 outputs the probability distribution of words at the second time (t = 2). As a result, the updating 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, a smaller loss is calculated as the probability of the correct word "no" at the second time is closer to 1 and the probabilities of the other words are closer to 0.

  Further, at the third time when the probability distribution of the word matching 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 the LSTM6b-2 and the correct word "no" one hour before to the LSTM6b-3 and the number of remaining characters at the second time from the second time. The number of letters "29", which is the number of letters of the correct word for the eyes, is input as the remaining number of letters at the third time. As a result, the LSTM6b-3 outputs the word probability distribution at the third time (t = 3). As a result, the updating 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, a smaller loss is calculated as the probability of the correct word “inquiry” at the third time is closer to 1 and the probabilities of the other words are closer to 0.

The updating unit 7 calculates the loss for each word of the reference summary 70 by repeatedly performing such processing until the end-of-sentence symbol “EOS” is output from the LSTM 6b. Further, the process of calculating the loss for each word of the reference summary is executed for all the learning samples included in the learning data. When the loss for each word of the reference summary is calculated for all learning samples included in the learning data in this way, the updating unit 7 maximizes the objective function L _t shown in the following Expression (4) for the parameter θ. The "logarithmic likelihood optimization" is executed as the first model learning. Here, the probability “p (y | x; θ)” in the following formula (4) is obtained by the total product of the losses at each time, as shown in the following formula (5). It should be noted that “D” in the following formula (4) indicates a set of learning samples including the input sentence x and the reference summary y. In addition, “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. ₂ , the last word is represented by y _t-1 .

  After that, the updating unit 7 updates the parameters of the model stored in the first model storage unit 3 to the parameter θ obtained by optimizing the log likelihood. This update of the parameter θ can be repeated for the learning data D a predetermined number of times. In this way, the model parameters stored in the first model storage unit 3 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-described 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, a duplication degree calculation unit 14, and a The first loss calculation unit 15, the pseudo sentence generation unit 16, the second probability calculation unit 17, the second loss calculation unit 18, and the update unit 19 are included.

  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 learning sample included in the learning data.

  Specifically, the input control unit 10I initializes the value of the loop counter d that counts the learning samples. Subsequently, the input control unit 10I acquires the learning sample corresponding to the loop counter d from 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 learning sample from the learning data storage unit 2 until the value of the loop counter d becomes equal to the total number D of learning samples. Here, an example in which the learning data saved in the storage inside the learning device 1 is acquired has been described, but the learning data is acquired from an external computer connected via a network, such as a file server, removable media, or the like. It does not matter if it is acquired.

  Every time a 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, that is, a so-called intermediate expression is output from the RNN encoder 11A to the RNN decoder 11B. Simultaneously with or before or after this, the input control unit 10I sets the value of the register, which holds the number of remaining characters until the ENN, which is called a sentence end symbol, is output to the RNN decoder 11B to a predetermined upper limit number of characters, such as 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 uses M model numbers corresponding to the number M of words in 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, the M LSTMs function as the RNN encoder 11A. In the RNN encoder 11A, the m-th word from the head of the input sentence is input to the LSTM corresponding to the m-th word in order from the head word of the input sentence of the learning sample according to the input control by the input control unit 10I. At the same time, the output of the LSTM corresponding to the m−1th 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, that is, a so-called intermediate representation is obtained. The intermediate representation of the input sentence thus generated by the RNN encoder 11A is input to the RNN decoder 11B.

  As a further aspect, the model execution unit 11 follows the model information stored in the first model storage unit 3 and outputs K LSTMs corresponding to each time on the work area until the end-of-sentence symbol “EOS” is output. Expand to. As a result, the K LSTMs function as the RNN decoder 11B. According to the input control of the input control unit 10I, the RNN encoder 11A inputs the intermediate representation of the input sentence of the learning sample to these RNN decoders 11B, and the input control unit 10I outputs an EOS tag for each K LSTMs. The number of characters remaining until output is input. By operating the K LSTMs according to these inputs, the RNN decoder 11B outputs the probability distribution of words for each of the K LSMTs.

In addition to the input control unit 10I and the model executing unit 11, the second learning unit 10 calculates the loss L (θ) used by the updating unit 19 to update the model parameters from the aspect of the above loss L _MRT (θ). ) Can be classified into a first system that calculates the first loss, and a second system that calculates the loss L _order (θ) as the second loss.

  Among them, in the first system, 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 an overlap that calculates the degree of overlap between the system summary and the reference summary. The degree calculation unit 14 and the first loss calculation unit 15 that calculates the first loss are included.

  The processing contents of the first system of the second model learning will be described below with reference to FIG. FIG. 12 is a diagram showing an example of input / output to / from 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 develops M LSTMs 11a-1 to 11a-M corresponding to the number M of words of the input sentence 30 in the work area used by the model execution unit 11. These M LSTMs 11a-1 to 11a-n are caused to function as the RNN encoder 11A. Then, the input control unit 10I inputs the words of the input sentence 30 in order from the first word included in the input sentence 30 to the LSTM 11a corresponding to the position of the word, and inputs the output of the previous LSTM 11a. The vector of the input sentence 30 is obtained by repeating such input from the LSTM11a-1 corresponding to the first word "our company" to the LSTM11a-M corresponding to the last word ".". The vector of the input sentence 30 thus generated by the RNN encoder 11A 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 before one time, and the number of remaining characters until the RNN decoder 11B outputs EOS, etc., and inputs the word at each time until EOS is output. Repeatedly calculate the probability distribution.

  Here, in the second model learning, unlike the first model learning, the word generated one hour before, rather than the correct word one hour before at each time of the RNN decoder 11B, is generated by the input control unit 10I. Is entered. Further, in the second model learning, the system summary loss with respect to the reference summary 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, the system summary is generated by repeatedly outputting the words based on the probability distribution of the words from the LSTM 11b corresponding to each time until the EOS tag is output. Loss of the system summary is calculated.

  For example, at the first time when the first word of the system summary is predicted, the input control unit 10I outputs the output of the LSTM 11a-M and the initial symbol to the LSTM 11b-1 expanded in the work area used by the model execution unit 11. The number of characters "37" in the reference summary 70 is input as the number of remaining characters together with "BOS". Here, as an example of the upper limit number of characters, the case where the number of characters in the reference summary is adopted has been illustrated, but the number of characters may be limited to a number shorter than the number of characters in the reference summary, or limited to a number of characters longer than the number of characters in the reference summary. You can also As a result, the LSTM 11b-1 outputs the probability distribution of words at the first time (t = 1). Based on the probability distribution of this word, the summary generation unit 12 extracts the first word of the system summary. For example, the summary generation unit 12 can perform a lottery according to the probability distribution of words and extract the words won by the lottery. In addition to this, the summary generation unit 12 randomly samples one word from the words belonging to the top predetermined number of probabilities, for example, the top five. Here, in the example illustrated in FIG. 12, the processing after the second time will be described by way of example only when “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 outputs 1 to the LSTM 11b-2 together with the output of the LSTM 11b-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 remaining characters at the second time. As a result, the LSTM 11b-2 outputs the probability distribution of words at the second time (t = 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 abstract generation unit 12 samples the words of the system abstract at each time until EOS is output by the LSTM 11b-K. By generating a system summary by such sampling, the summary generation unit 12 can generate a predetermined number, for example, S system summaries y ′ for one input sentence. When S number of system summaries are generated in this way, the first probability calculation unit 13 determines the probability of the word generated at each time of the system summarization y ′ for each of the S number of system summaries y ′. The generation probability p (y ′ | x, θ) of generating the system summary y ′ from the input sentence x is calculated.

Here, in the second model learning, the first loss L _MRT (θ) is calculated according to the above equation (1). That is, the first loss L _MRT (θ) is calculated in addition to the system summary generation probability calculated by the first probability calculation unit 13 as well as the system summary and reference summary calculated by the redundancy calculation unit 14 described later. It is calculated based on the degree of overlap between words.

  Thus, the degree of overlap Δ (y ′, y) used for calculating the first loss is not always calculated using all the words included in the system summary, as shown in FIG. That is, the degree-of-duplication calculation unit 14 sets, for each of the S system summaries generated by the summarization generator 12, between the reference summarization targeting sentences within the upper limit number of characters of the system summarization, for example, the number of characters of the reference summarization. Calculate the word overlap. As a result, the words in the part of the system summary that exceed the maximum number of characters, that is, the hatched part shown in FIG. 12, can be excluded from the calculation targets of the degree of overlap.

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

FIG. 13 is a diagram showing an example of a method of calculating the degree of overlap. FIG. 13 shows an example in which the degree of overlap 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 y ′ _{k−1 from} the beginning, the beginning From the kth word y ′ _k , ..., And the last word y ′ _{| y ′ |} . On the other hand, the reference summary y includes the _first word y ₁ , the second word y ₂ from the beginning, ..., And the last word y _{| y |} . In this case, a word of the 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 beginning, ... ·, k-1-th word _{y 'k-1} is cut off from the beginning. On top of that, ROUGE (trim (y ', byte (y)), y) , the system summarized y' cut out _{'from 1} k-1-th word _y' beginning of a word y to _k-1 trim ( The degree of word overlap between y ′, byte (y)) and the reference summary y is calculated. By calculating the degree of overlap Δ (y ′, y) according to the above equation (6) in this way, the words from the k-th to the end of the system summary y ′ that exceeds the upper limit number of characters, that is, word y ′ _k to word It is possible to exclude y ′ _{| y ′ |} from the calculation target of the overlap rate. As a result, the words from the kth to the end of the system summary y ′ exceeding the upper limit number of characters, that is, the words y ′ _k to y ′ _{| y ′ |} include words that overlap with the reference summary y. Thus, it is possible to prevent the system summary y ′ from being overestimated.

  Thus, in addition to limiting the calculation target of the degree of duplication to the words within the upper limit number of characters of the system summary, the degree of duplication calculation unit 14 does not exceed the upper limit number of characters of the system summary as shown in the following equation (7). The length or the length exceeding the upper limit number of characters of the system summary can be calculated as an error to be added to the degree of duplication as a penalty. In addition, "C" shown in the following formula (7) indicates a hyperparameter set by the developer or the user of the above learning program.

FIG. 14 is a diagram showing an example of a method of calculating the degree of overlap with an error. FIG. 14 shows an example in which the degree of overlap Δ (y ′, y) with an error is calculated according to the above equation (7). Also 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 ′ is determined by ROUGE (trim (y ′, byte (y)), y). The word duplication degree 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 added to the degree of overlap 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) that exceeds the upper limit number of characters is added to the degree of duplication. Thus, the multiplicity Δ (y ′, y) with an error is calculated. Thus, the error | byte (y ′) − byte (y) | is added to the degree of overlap calculated by ROUGE according to the above equation (7) to calculate the degree of overlap Δ (y ′, y) with an error. To do. As a result, the loss of system summaries that do not reach the maximum number of characters and system summaries that exceed the maximum number of characters increases, and as a result, model learning that enhances the evaluation of system summaries whose number of characters matches the maximum number of characters can be realized.

  Further, the degree-of-overlap calculation unit 14 may not necessarily calculate the error to be added to the degree of overlap even for system summaries that do not necessarily have the maximum number of characters. For example, the degree-of-overlap calculation unit 14 can calculate the length of the amount exceeding the upper limit number of characters of the system summary as an error by narrowing down when the system summary exceeds the upper limit number of characters according to the following formula (8).

FIG. 15 is a diagram illustrating an example of a method of calculating the degree of overlap with an error. FIG. 15 shows an example in which the degree of overlap Δ (y ′, y) with an error is calculated according to the above equation (8). Also in the example shown in FIG. 15, similarly to the example shown in FIG. 13, the _first word y ′ ₁ to k−1 of the system summary y ′ is determined by ROUGE (trim (y ′, byte (y)), y). The word duplication degree 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) that exceeds the maximum number of characters to the degree of duplication. By doing so, the degree of overlap Δ (y ′, y) with an error is calculated. On the other hand, when the system summary does not reach the upper limit of the number of characters, “0” is selected by max (0, byte (y ′) − byte (y)), so that no error is added to the degree of duplication and the degree of duplication is It is calculated as Δ (y ′, y) as it is. As a result, the loss of the system summary that is less than the upper limit number of characters is not increased, and the loss of the system summary that exceeds the upper limit number of characters is increased.

  After such a degree of overlap Δ (y ′, y) with an error is calculated, the first loss calculation unit 15 determines, for each predetermined number of system summaries 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 in which the system summary is generated from the input sentence and the multiplicity Δ (y ′, y) with the error calculated by the multiplicity calculation unit 14. Furthermore, the first loss calculation unit 15 executes the calculation for summing the first losses calculated for each of the S system summaries, thereby relating to the set S (x, θ) of the S system summaries y ′. Calculate the sum of the first losses.

  Returning to the description of FIG. 1, in the second system, a pseudo sentence generation unit 16 that generates a pseudo sentence, a second probability calculation unit 17 that calculates the reference summary generation probability and the pseudo sentence generation probability, and And a second loss calculating section 18 for calculating the second loss of.

  For example, the pseudo-sentence generator 16 replaces the correct reference summary y with the word order of the words included in the reference summary y to reproduce a set of pseudo sentences z in which a non-grammatical expression is pseudo-reproduced. ) Is generated. At this time, the pseudo-sentence generation unit 16 performs calculation with the reference summary y by changing the word order of the words and sampling the pseudo-sentence z without changing the number of words included in the correct reference summary y. It is possible to generate a pseudo sentence z having a ROUGE value of "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 the same as those when 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 those when 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 the input by the input control unit 10I according to the model information stored in the first model storage unit 3. The J LSTMs corresponding to the number J of words of the pseudo sentence z are expanded on the work area. This causes the J LSTMs to function as the RNN decoder 11B. In accordance with the input control of the input control unit 10I, these RNN decoders 11B receive the intermediate representation of the input sentence x of the learning sample from the RNN encoder 11A, and the input control unit 10I outputs the tag of the EOS for each J LSTMs. The remaining number of characters until is output is input. Furthermore, the words of the pseudo sentence z one hour before are input to the J LSTMs of the RNN decoder 11B according to the input control of the input control unit 10I. By operating J LSTMs according to these inputs, the RNN decoder 11B outputs the probability of the word at each time of the pseudo sentence z for each J LSMT. As described above, based on the probability of the word in the pseudo sentence z output by each LSMT of the RNN decoder 11B at each time, the second probability calculator 17 causes the generation probability p (that the pseudo sentence z is generated from the input sentence x to be p ( z | x; θ) is calculated.

  The processing contents of the second system of the second model learning will be described below with reference to FIG. 16. FIG. 16 is a diagram showing an example of input / output to / from the model in the second system. In FIG. 16, the input control unit 10I inputs the input sentence 30 shown in FIG. 3 to the RNN encoder 11A, 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 RNN decoder. An example of input to 11B is shown. Note that 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 in the example shown in FIG. 12, and therefore the description of the RNN decoder 11B will be started.

  As illustrated in FIG. 16, the model execution unit 11 receives the vector of the input sentence 30, the word at each time of the pseudo sentence z, the number of remaining characters until the RNN decoder 11B outputs EOS, and the like until the EOS is output. The probability distribution of words is calculated repeatedly at each time.

  Here, unlike the case where the system summary is generated, when the generation probability of the pseudo-sentence z is calculated, the pseudo-sentence z is not included in the pseudo-sentence z, but is included in the pseudo-sentence z, not the word generated one time before in the LSTM 11b of each time of the RNN decoder 11B. The word of the pseudo sentence z one hour before is input from the input control unit 10I.

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

  Then, at the second time, the input control unit 10I outputs the first time from the remaining number of characters at the first time to the LSTM11b-2 together with the output of the LSTM11b-1 and the word "AI" of the pseudo sentence z one hour before. The number of characters “35” obtained by subtracting the number of characters of the word “AI” of the pseudo sentence z is input as the number of remaining characters at the second time. As a result, the LSTM 11b-2 outputs the probability distribution of words at the second time (t = 2). At this time, the second probability calculation unit 17 saves the probability corresponding to the second word “no” from the beginning of the pseudo sentence z in the probability distribution of the words at the second time in a work area (not shown).

  After the 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 word number "5" obtained by subtracting the word number "sales" of the pseudo sentence z at 1 hour before from the remaining character number at the first time, and the word number "5" of the word "sale" of the pseudo sentence z at time J-2. Enter as the number of remaining characters in the eye. As a result, the LSTM 11b-J-1 outputs the probability distribution of words at the J-1 time (t = J-1). At this time, the second probability calculation unit 17 saves the probability corresponding to the J−1th word “inquiry” from the beginning of the pseudo-sentence z in the work area (not shown) in the probability distribution of the words at the J−1th time. To do.

  Finally, at the J-th time, the input control unit 10I outputs to the LSTM 11b-J the output of the LSTM 11b-J-1 and the word "inquiry" of the pseudo-sentence z one hour before, from the remaining number of characters at the 1st time to the J-number. The number of characters "0" obtained by subtracting the number of characters of the word "inquiry" of the pseudo sentence z at time -1 is input as the number of remaining characters at time J. As a result, the LSTM 11b-J outputs the probability distribution of words at the J-th time (t = 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 probability distribution of the words at the Jth time in a work area (not shown).

  As described above, the second probability calculation unit 17 uses the probability of the pseudo sentence z stored in the work area at each time to generate the pseudo sentence z 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 to calculate the second loss is calculated, the generation probability of the reference summary y can be calculated in the same manner as when calculating the generation probability of the pseudo sentence z. That is, the model execution unit 11 sets, on the work area, I LSTMs corresponding to the word number I of the reference summary y input by the input control unit 10I according to the model information stored in the first model storage unit 3. Expand to. As a result, the I LSTMs function as the RNN decoder 11B. In accordance with the input control of the input control unit 10I, the RNN decoder 11B receives the intermediate representation of the input sentence x of the learning sample from the RNN encoder 11A, and the input control unit 10I outputs the EOS tag for each I LSTM. The remaining number of characters until is output is input. Further, the word of the reference summary y one hour before is input to the I LSTMs of the RNN decoder 11B according to the input control of the input control unit 10I. By operating I LSTMs according to these inputs, the RNN decoder 11B outputs the probability of the word at each time of the reference summary y for each I LSMT. As described above, based on the word probabilities at each time of the reference summary y output by each LSMT of the RNN decoder 11B, the second probability calculation unit 17 causes the generation probability p (that 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 causes 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 higher 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 the calculation for summing the second losses calculated for each pseudo-sentence z, and thereby the second loss regarding the set S ′ (y) of S ′ pseudo-sentences z. Calculate the sum of the losses.

  As described above, the process of calculating the sum of the first loss for S system summaries and the sum of the second loss for S ′ pseudo sentences z is repeatedly executed for all learning samples included in the learning data. To be done. When the sum of the first loss and the sum of the second loss are calculated for all the learning samples included in the learning data in this way, the updating unit 19 causes the objective function L (θ ) Updates the model parameters to the model parameter θ that is minimized. The model parameters updated in this way are stored in the second model storage unit 8. This updating of the parameter θ can be repeated for the learning data D a predetermined number of times. As a result, the model information stored in the second model storage unit 8 can be provided as a summary sentence generation model.

[Process flow]
FIG. 17 is a flowchart illustrating the procedure of the learning process according to the first embodiment. The flowchart of the learning process shown in FIG. 17 is a schematic representation of the procedure of the second model learning executed by the second learning unit 10. FIG. 17 shows, as an example only, a flowchart of an example in which the degree of overlap with 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 the first model learning by the first learning unit 5 is executed as preprocessing. It is possible to start the learning process shown in FIG.

  As shown in FIG. 17, the processes of steps S101 to S103 are executed for each D learning samples included in the learning 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 to the first system in this manner, the first loss calculation process is executed (step S102).

(1) First Loss Calculation Processing FIG. 18 is a flowchart showing the procedure of the first loss calculation processing according to the first embodiment. This process corresponds to the process of step S102 described above. As illustrated in FIG. 18, the abstraction generation unit 12 samples the words for each time based on the probability distribution of the words output from the RNN decoder, thereby performing the S for the input sentence x of the learning sample acquired in step S101. Individual system summaries y'are generated (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 processing of the following step S303 to the following step S306 is executed for each of the S system summaries y'generated in step S301. That is, the degree-of-overlap calculation unit 14E cuts out a word having the maximum number of characters, for example, the number of bytes corresponding to the reference summary y, from the system summary y ′ according to trim (y ′, byte (y)) shown in the above equation (8). (Step S303).

  Then, the degree-of-overlap calculation unit 14 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 reference summary y, the degree of word overlap is calculated (step S304).

  Further, the degree-of-overlap 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 is less than the upper limit number of characters, “0” is selected by max (0, byte (y ′) − byte (y)), and thus the error to be added to the degree of duplication is calculated as “0”. It

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

  After that, the first loss calculating unit 15 calculates the first loss from the calculation result of the probability for the system summary y ′ calculated in step S302 and the overlapping degree Δ (y ′, y) with an error (( 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 totals the first losses calculated for each of the S system summaries. Is executed to calculate the first loss sum corresponding to the set S (x, θ) of the system summary y ′ (step S307), and the process of step S102 shown in FIG. 17 is ended. ..

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

(2) Second Loss Calculation Processing FIG. 19 is a flowchart showing the procedure of the second loss calculation processing according to the first embodiment. This process corresponds to the process of step S103 described above. As illustrated in FIG. 19, the pseudo-sentence generating unit 16 replaces the correct reference summary y with the word order of the words included in the reference summary y to reproduce the pseudo-sentence z of the pseudo-sentence z. A set S '(y) is generated (step S501).

  After that, the processing of the following step S502 to the following step S505 is executed for each of the S'pseudo sentences z generated in step S501. That is, the second probability calculator 17 calculates the generation probability p (z | x; θ) of generating the pseudo sentence z 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 (Yes in step S503), the second loss calculation unit 18 executes the following process. That is, the second loss calculating unit 18 calculates 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 (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 (No in step S503), the second loss calculation unit 18 sets a predetermined set value according to the above equation (3), for example, zero or more. Is calculated as the second loss (step S505).

  After that, when the second loss is calculated for each of the S ′ pseudo sentences z generated in step S501, the second loss calculation unit 18 executes the following processing. That is, the second loss calculating unit 18 executes the calculation for summing the second losses calculated for each of S ′ pseudo sentences, thereby corresponding to the set S ′ (x, θ) of the pseudo sentences z. The sum of the second losses to be calculated is calculated (step S506), and the process of step S103 shown in FIG. 17 is ended.

  After that, for all learning samples included in the learning data, the sum of the first loss corresponding to the set S (x, θ) of the system summary y ′ and the set S ′ (x, θ) of the pseudo sentence z are associated. When the sum of the second losses is calculated, the updating unit 19 determines the model parameter stored in the second model storage unit 8 as the minimum objective function L (θ) shown in the above equation (2). The parameter θ of the model to be converted is updated (step S104), and the process ends.

[One side of effect]
As described above, the learning device 1 according to the present exemplary embodiment generates the pseudo sentence in which the non-grammatical expressions are pseudo reproduced by changing the word order of the words included in the correct reference summary, and the model is simulated. Update the model parameters so that the model has a higher probability of generating a reference summary than a sentence generating probability. For this reason, the non-grammatical expression is provided even in a summary sentence having a high degree of overlap between the reference summary and the word, while having the effect of increasing the probability of generating a system summary having a high degree of overlap between the reference summary and the word and a word order different from that of the reference summary. A reaction that imposes a penalty on the generation of pseudo sentences including Therefore, it is possible to update the parameters that increase the generation probability of a system summary that does not include a non-grammatical expression even in a summary sentence with a high degree of word duplication. Therefore, according to the learning device 1 according to the present embodiment, it is possible to suppress the learning of the model that generates the abstract sentence having low readability.

  Although the embodiments of the disclosed device have been described so far, the present invention may be implemented in various different forms other than the embodiments described above. Therefore, other embodiments included in the present invention will be described below.

Distributed and integrated
In addition, each component of each illustrated device may not necessarily be physically configured as illustrated. That is, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part of the device may be functionally or physically distributed / arranged in arbitrary units according to various loads or usage conditions. It can be integrated and configured. For example, the first learning unit 5 or the second learning unit 10 may be connected as an external device of the learning device 1 via a network. Further, the functions of the learning device 1 may be realized by having the first learning unit 5 or the second learning unit 10 in another device, respectively, and connecting the devices by network to cooperate with each other. Further, another device has all or 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 perform the learning described above. The function of the device 1 may be realized.

[Learning program]
The various processes described in the above embodiments can be realized by executing a prepared program on a computer such as a personal computer or a workstation. Therefore, an example of a computer that executes a learning program having the same functions as those in the above-described embodiment will be described below with reference to FIG.

  FIG. 20 is a diagram illustrating a hardware configuration example of a computer that executes a 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 units 110 to 180 is connected via a bus 140.

  As shown in FIG. 20, the HDD 170 stores a learning program 170a having the same function as that of the second learning unit 10 described in the first embodiment. This learning program 170a may be integrated or separated similarly to each component of the second learning unit 10 shown in FIG. That is, the HDD 170 does not necessarily need to store all the data described in the first embodiment, and the data used for the processing may be stored in the HDD 170.

  Under such an environment, the CPU 150 reads out the learning program 170a from the HDD 170 and loads it on 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 in the 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, the processing shown in FIGS. 17 to 19 is included as an example of the processing executed by the learning process 180a. In the CPU 150, not all the processing units shown in the above-described first embodiment need to operate, and the processing unit corresponding to the processing to be executed may be virtually realized.

  The learning program 170a does not necessarily have to 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, an IC card, or the like. 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 a server device connected to the computer 100 via a public line, the Internet, a LAN, a WAN, etc., and the computer 100 acquires and executes the learning program 170a from these. You may do so.

  The following supplementary notes will be further disclosed regarding the embodiments including the above-described examples.

(Supplementary Note 1) A learning method for performing machine learning of a model for generating a summary sentence from an input sentence,
Get the input sentence and the summary of the correct answer,
Generating a pseudo sentence in which a non-grammatical expression is pseudo reproduced by exchanging the word order of the words included in the correct answer 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 correct summary sentence in which the correct answer summary sentence is generated from the input sentence by the model Updating the parameters of the model,
A learning method characterized in that a computer executes the processing.

(Supplementary note 2) The learning method according to Supplementary note 1, wherein in the updating process, the parameters of the model are updated so that the probability of generation of the correct summary sentence is higher than the probability of generation of the pseudo sentence. ..

(Supplementary Note 3) In the updating process, if the pseudo-sentence generation probability is higher than the correct-answer summary generation probability, the difference between the pseudo-sentence generation probability and the correct-answer summary sentence generation probability is lost. If the probability of generation 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 added. The learning method according to appendix 2, wherein the parameters of the model are updated without being added to the loss.

(Supplementary Note 4) For each of a plurality of summary sentences generated by inputting the input sentence into the model, a 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 further executes a process of calculating the degree of overlap of words in the summary sentence and the correct answer summary sentence,
The updating process includes the generation probability of the summary sentence calculated for each of the plurality of summary sentences, the degree of word duplication calculated for each of the plurality of summary sentences, the generation probability of 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.

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

(Supplementary note 6) A learning program for executing machine learning of a model for generating a summary sentence from an input sentence,
Get the input sentence and the summary of the correct answer,
Generating a pseudo sentence in which a non-grammatical expression is pseudo reproduced by exchanging the word order of the words included in the correct answer 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 correct summary sentence in which the correct answer summary sentence is generated from the input sentence by the model Updating the parameters of the model,
A learning program characterized by causing a computer to execute processing.

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

(Supplementary Note 8) In the updating process, if the pseudo-sentence generation probability is higher than the correct-answer summary generation probability, the difference between the pseudo-sentence generation probability and the correct-answer summary sentence generation probability is lost. If the probability of generation 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 added. The learning program according to appendix 7, wherein the parameters of the model are updated without being added to the loss.

(Supplementary Note 9) For each of a plurality of summary sentences generated by inputting the input sentence into the model, a 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, further causes the computer to perform a process of calculating the degree of overlap of words in the summary sentence and the correct answer summary sentence,
The updating process includes the generation probability of the summary sentence calculated for each of the plurality of summary sentences, the degree of word duplication calculated for each of the plurality of summary sentences, the generation probability of the pseudo sentence, and the correct answer. 7. The learning program according to appendix 6, wherein the parameters of the model are updated based on the summary sentence generation probability.

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

(Supplementary note 11) A learning device for performing machine learning of a model for generating a summary sentence from an input sentence,
An acquisition unit that acquires the input sentence and the summary of the correct answer,
A pseudo-sentence generating unit that generates a pseudo-sentence in which a non-grammatical expression is pseudo-reproduced by changing the word order of the words included in the correct answer 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 correct answer summary sentence in which the correct answer summary sentence is generated from the input sentence by the model An updating unit for updating the parameters of the model,
A learning device having:

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

(Supplementary Note 13) The updating unit adds the difference between the generation probability of the pseudo sentence and the generation probability of the correct summary sentence to the loss when the generation probability of the pseudo sentence is higher than the generation probability of the correct summary sentence. Then, the model parameters are updated, and if the pseudo-sentence generation probability is not higher than the correct-answer summary sentence generation probability, the difference between the pseudo-sentence generation probability and the correct-answer summary sentence generation probability is lost. 13. The learning device according to appendix 12, wherein the parameters of the model are updated without being added to.

(Supplementary Note 14) Probability calculation for calculating, for each of a plurality of summary sentences generated by inputting the input sentence into the model, a generation probability of the summary sentence generated by the model from the input sentence. Department,
For each of the plurality of summary sentences, further has a multiplicity calculation unit that calculates the multiplicity of words of the summary sentence and the correct summary sentence,
The update unit, the generation probability of the summary sentence calculated for each of the plurality of summary sentences, the degree of word duplication calculated for each of the plurality of summary sentences, the generation probability of the pseudo sentence and the summary of the correct answer. The learning device according to appendix 11, wherein the parameters of the model are updated based on a sentence generation probability.

(Supplementary note 15) The learning according to supplementary note 11, wherein the pseudo-sentence generating unit generates the pseudo-sentence by changing the word order of words without changing the number of words included in the correct answer summary sentence. apparatus.

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 Duplication 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)

A learning method for performing machine learning of a model for generating a summary sentence from an input sentence,
Get the input sentence and the summary of the correct answer,
Generating a pseudo sentence in which a non-grammatical expression is pseudo reproduced by exchanging the word order of the words included in the correct answer 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 correct summary sentence in which the correct answer summary sentence is generated from the input sentence by the model Updating the parameters of the model,
A learning method characterized in that a computer executes the processing.   The learning method according to claim 1, wherein in the updating process, the parameters of the model are updated so that the probability of generating the correct summary sentence is higher than the probability of generating the pseudo sentence.   In the updating process, if the pseudo-sentence generation probability is higher than the correct-answer summary generation probability, the difference between the pseudo-sentence generation probability and the correct-answer summary sentence generation probability is added to the loss. If the generation probability of the pseudo-sentence is not higher than the generation probability of the correct summary sentence, the model parameter is updated, and the difference between the pseudo sentence generation probability and the correct summary sentence generation probability is added to the loss. The learning method according to claim 2, wherein the parameters of the model are updated without being updated. 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 further executes a process of calculating the degree of overlap of words in the summary sentence and the correct answer summary sentence,
The updating process includes the generation probability of the summary sentence calculated for each of the plurality of summary sentences, the degree of word duplication calculated for each of the plurality of summary sentences, the generation probability of the pseudo sentence, and the correct answer. The learning method according to claim 1, wherein the parameters of the model are updated based on the generation probability of the summary sentence.   5. The generating process, wherein the pseudo sentence is generated by changing a word order of words without changing the number of words included in the correct answer summary sentence. The learning method described. 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 of the correct answer,
Generating a pseudo sentence in which a non-grammatical expression is pseudo reproduced by exchanging the word order of the words included in the correct answer 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 correct summary sentence in which the correct answer summary sentence is generated from the input sentence by the model Updating the parameters of the model,
A learning program characterized by causing a computer to execute processing. A learning device for performing machine learning of a model for generating a summary sentence from an input sentence,
An acquisition unit that acquires the input sentence and the summary of the correct answer,
A pseudo-sentence generating unit that generates a pseudo-sentence in which a non-grammatical expression is pseudo-reproduced by changing the word order of words included in the correct answer 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 correct summary sentence in which the correct answer summary sentence is generated from the input sentence by the model An updating unit for updating the parameters of the model,
A learning device having:

Images