JP2020071668A - Summary generation method and summary generation program - Google Patents

Abstract

To provide a summary generation method capable of generating a summary including unknown words not found in a model dictionary.SOLUTION: The summary generation method causes a computer to execute a series of processing that, when there is a proper expression that includes a common word between an input sentence and a summary output by a model in which the input sentence is input, a string of the first unique expression that includes the common words in the summary is replaced with a string of the second specific expression that includes common words in the input sentence.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a summary generation method and a summary generation program.

  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 the generation of the words of the abstract sentence by referring to the vector of the input sentence is connected is used for generating the abstract sentence.

JP, 2014-225185, A JP, 2017-27168, A JP, 2005-196513, A

  However, in the above technique, the vocabulary of the model dictionary is limited to the words that appear in the learning data, so there is a limit that unknown words that are not in the model dictionary cannot be generated as the words of the summary sentence.

  In one aspect, it is an object of the present invention to provide a summary generation method and a summary generation program that can generate a summary sentence including unknown words that are not in the model dictionary.

  In the summary generation method according to one aspect, when a specific expression including a common word exists between an input sentence and a summary sentence output by a model to which the input sentence is input, the common word is included in the summary sentence. The computer executes a process of replacing the character string of the first proper expression with the character string of the second proper expression including the common word in the input sentence.

  It is possible to generate summary sentences including unknown words that are not in the model dictionary.

FIG. 1 is a block diagram illustrating the functional configuration of each device included in the system 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 a method of replacing a unique expression. FIG. 4A is a diagram showing an example of a learning input sentence. FIG. 4B is a diagram showing an example of the correct answer summary sentence. FIG. 5 is a diagram illustrating an example of model learning. FIG. 6 is a diagram showing an example of a model dictionary. FIG. 7 is a diagram illustrating an example of the input sentence and the summary sentence. FIG. 8 is a flowchart illustrating the procedure of the summary generation process according to the first embodiment. FIG. 9 is a diagram illustrating an application example of the unique expression detection method. 10. FIG. 10 is a flowchart illustrating the procedure of the summary generation process according to the second embodiment. FIG. 11 is a diagram illustrating a hardware configuration example of a computer that executes the abstract generation program according to the first and second embodiments.

  A summary generation method and a summary generation program 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.

[System configuration]
FIG. 1 is a block diagram illustrating the functional configuration of each device included in the system according to the first embodiment. In the system 1 shown in FIG. 1, a machine learning service that performs model machine learning using learning data including a learning input sentence and a correct answer summary sentence, and a summary that generates a summary sentence from the input sentence using a learned model. Generation services are provided.

  As shown in FIG. 1, the system 1 includes a learning device 10 and a generation device 30. Upon receiving the model learned by the learning device 10, the generation device 30 generates a result for the given data.

  The learning device 10 corresponds to an example of a computer that provides the above machine learning service. When arranging the learning device 10 and the generation device 30 in another computer, the model is passed via network communication.

  As an embodiment, the learning device 10 can be implemented by installing a learning program that realizes the above machine learning service as package software or online software in an arbitrary computer. The computer can be caused to function as the learning device 10 by causing the computer to execute the learning program thus installed.

  By way of example only, the learning device 10 can be implemented as a server device that accommodates the generation device 30 as a client and provides the client with the above machine learning service. In this case, the learning device 10 may be implemented on-premises as a server that provides the machine learning service, or may be implemented as a cloud that provides the machine learning service by outsourcing.

  For example, the learning device 10 inputs learning data including a plurality of learning samples or identification information that can call the learning data via a network or a storage medium, and outputs the learning result of the model to the generation device 30. At this time, the learning device 10 can provide the parameters of the model of the neural network to which the RNN encoder and the RNN decoder are connected, as an example. In addition, the learning device 10 can also provide an application program in which the generation of a summary sentence realized by using a learned model is incorporated as a function. For example, to provide an application program that generates an article headline as a summary sentence from the original text of various articles such as newspapers, electronic bulletin boards, and websites, or generates a preliminary article from the original text of the article as a summary sentence. You can

  In addition, the above-described machine learning service providing form is merely an example, and the machine learning service may be provided in a form other than the above-described examples. For example, the learning program itself for realizing the above machine learning service may be provided as package software or online software, or a computer in which the above learning program is introduced may be provided.

  The generation device 30 corresponds to an example of a computer that provides the above summary generation service.

  As one embodiment, the generation device 30 can be implemented by installing a digest generation program that realizes the digest generation service described above as package software or online software in an arbitrary computer. The computer can be caused to function as the generation device 30 by causing the computer to execute the installed abstract generation program.

  As just one example, the above-described summary generation service is provided as one of the web service tools provided to media companies operating various media such as newspapers, electronic bulletin boards, and websites, for example, as an "article summary tool". can do. In this case, among the functions provided as the Web service, front-end functions such as inputting an original sentence and displaying a summary sentence are installed in a terminal device such as a reporter or an editor, and generate a summary sentence. The back-end function may be installed in the generation device 30.

[Example of use case of article summarization tool]
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. For example, the work of summarizing articles is a relative process in the process of distributing news to the media, such as "selecting distribution articles", "sending to media editing system", "summary of articles", "creation of headlines", and "review". There is an aspect that the labor is large. 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.

  Further, here, as an example, the generation device 30 is implemented as a computer that provides the above-described summary generation service, but the present invention is not limited to this. For example, the above-described learned model-embedded abstract generation program may be implemented as a stand-alone application program that is executed by an arbitrary computer, for example, a terminal device such as a reporter or an editor.

  Furthermore, here, the example in which the machine learning service and the summary generation service are executed by different business entities has been described, but these two services may be provided by the same business operator. In this case, the learning program and the abstract generating program may be executed by the same computer or computer system.

[One aspect of the issue]
As described in the background art section above, the vocabulary of the model dictionary is limited to the words that appear in the learning data, so there is a limit that unknown words that are not in the model dictionary cannot be generated as words of the summary sentence.

[One aspect of approach to problem solving]
Therefore, the generation device 30 according to the present embodiment adopts an approach of compensating an unknown word that is not in the model dictionary with a proper expression that appears in the input sentence. In other words, a corpus is sometimes used as learning data for machine learning of a model, but even a corpus with a large number of vocabularies includes proper nouns and unique expressions that include numerical expressions such as time expressions, numbers, and percentages. It is difficult to cover expressions (Named Entity). As described above, while it is difficult to increase the number of learning samples from the aspect of reducing unknown words in the model, it can be considered that most of the unknown words are likely to be proper expressions if the viewpoint is changed.

  From this, under the motivation of supplementing unknown words that are not in the model dictionary with the proper expressions that appear in the input sentence, if there is a proper expression that includes a common word between the input sentence and the summary sentence, Replace the unique expression string with the unique string of the input sentence. Hereinafter, a proper expression including a common word in the summary sentence will be referred to as a “first proper expression”, and a proper expression including a common word in the input sentence will be referred to as a “second proper expression”. May be listed.

  FIG. 3 is a diagram showing an example of a method of replacing a unique expression. FIG. 3 shows an example in which the summary sentence 40Y is generated by inputting the input sentence 40G into the learned model. As shown in FIG. 3, a specific expression is extracted from the input sentence 40G and the summary sentence 40Y.

  After the unique expressions are extracted in this way, the replacement target of the unique expressions is narrowed down to the unique expressions that include common words between the input sentence and the summary sentence. For example, as shown in bold in the relevant portion of the summary sentence 40Y, the proper expression "Yoshida Gomu" including the common word "Yoshida" is narrowed down as the first proper expression from the summary sentence 40Y. On the other hand, as shown in bold in the relevant part of the input sentence 40G, the proper expression "Yoshida ○○" including the common word "Yoshida" is narrowed down as the second proper expression.

  At this stage, it can be considered that the two correspond to each other, and the first proper expression "Yoshida Gomu" can be replaced with the second proper expression "Yoshida ○○". Can also be added.

  For example, if a person's name is taken as an example, relatives such as parents, children, and relatives may appear in the text of the article over a plurality of names. In such a case, an input sentence and a summary sentence do not necessarily have a corresponding relationship because they are unique expressions including common words.

  Therefore, it is possible to add a condition that the character strings adjacent to the specific expression are similar between the summary sentence 40Y and the input sentence 40G. For example, from the aspect that the modification part and the modified part are modified from the preceding phrase to the succeeding phrase, if the similarity between the character strings adjacent to each other before the unique expression is a predetermined threshold value, Can be adopted.

  Here, as an example of the numerical value, it is assumed that it is determined whether or not the similarity of 14 characters adjacent to each other before the unique expression is 80% or more. As underlined in the summary sentence 41Y and the input sentence 40G in FIG. 3, the adjacent 14 characters before the proper expressions of “Yoshida Gomu” and “Yoshida ○○” are both “instructor of disaster prevention classroom at the same school”. Served. As described above, since 14 characters match among 14 characters, the similarity is 100%, which exceeds 80% of the threshold value.

  From this, it can be estimated that the proper expression "Yoshida Gomu" of the summary sentence 40Y is likely to correspond to the proper expression "Yoshida ○○" of the input sentence 40G. In this case, the character string including the proper expression "Yoshida Gomu" in the summary sentence 40Y is replaced with the character string including the proper expression "Yoshida ○○" in the input sentence 40G. That is, the character string 42Y "Yoshida XX who was a lecturer in the disaster prevention classroom at the school" shown in black and white in reverse in FIG. 3 is the character string 42G "Italic disaster prevention at the school" shown in italics in FIG. It was replaced by Yoshida ○○, who was a lecturer in the classroom.

  As a result, the proper expression of the unknown word "○○" in the learned model was corrected using the input sentence 40G. 43Y "Yoshida ○○ who was a lecturer in the disaster prevention classroom at the school in November 2015 Issue is generated.

  As described above, according to the generation device 30 according to the present embodiment, it becomes possible to generate a summary sentence including an unknown word that is not in the model dictionary.

[Configuration of learning device 10]
As shown in FIG. 1, the learning device 10 includes a learning data storage unit 11, a model storage unit 12, an input control unit 13, an encoder execution unit 14, a decoder execution unit 15, a generation unit 16, and a calculation unit. 17 and an updating unit 18. The learning device 10 may have various functional units of a known computer other than the functional units shown in FIG. 1, such as various input devices and audio output devices.

  The functional units such as the input control unit 13, the encoder execution unit 14, the decoder execution unit 15, the generation unit 16, the calculation unit 17, and the update unit 18 illustrated in FIG. 1 are virtually realized by the following hardware processors as examples only. To be done. Examples of such a processor include a DLU (Deep Learning Unit), a GPGPU (General-Purpose computing on Graphics Processing Units), and a GPU cluster. In addition, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), and the like are included. That is, the functional unit is virtually realized by the processor developing the learning program as a process on a memory such as a RAM (Random Access Memory). 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).

  A storage device such as an HDD (Hard Disk Drive), an optical disk, or an SSD (Solid State Drive) can be adopted as the functional units such as the learning data storage unit 11 and the model storage unit 12 shown in FIG. 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.

  The learning data storage unit 11 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 an input sentence used for model learning and a correct summary sentence. Hereinafter, the former may be referred to as a “learning input sentence” from the aspect of identifying the label of the input sentence input at the time of learning the model and generating the summary sentence. Further, the former may be referred to as a “correct answer summary” from the aspect of identifying a label of a summary sentence referred to as a correct answer and a summary sentence generated from an input sentence when learning a model.

  The model storage unit 12 is a storage unit that stores information about the model.

  As an embodiment, the model storage unit 12 includes a layer structure of models such as neurons and synapses in each layer of an input layer, a hidden layer, and an output layer that form a model of a neural network to which an RNN encoder and an RNN decoder are connected. , Model information including model parameters such as weight and bias of each layer is stored. Here, in the stage before the model learning is executed, the model storage unit 12 stores the parameters initialized by random numbers as the model parameters. Further, at the stage after the model learning is executed, the model storage unit 12 stores the parameters of the learned model.

  The input control unit 13 is a processing unit that controls the input to the model.

  As one embodiment, the input control unit 13 activates the process when a model learning request is received. When the process is activated in this way, the input control unit 13 performs initial setting for model learning. For example, the input control unit 13 sets the number of characters of the correct summary sentence, and sets the number of characters designated by the user as the upper limit number of characters of the summary sentence generated by the model.

  After the initial value to be input to the RNN decoder is set in this way, the input control unit 13 starts input 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 13 initializes the value of the loop counter d that counts the learning samples. Subsequently, the input control unit 13 acquires the learning sample corresponding to the loop counter d from the D learning samples stored in the learning data storage unit 11. After that, the input control unit 13 increments the loop counter d, and repeatedly executes the process of acquiring the learning sample from the learning data storage unit 11 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 stored in the storage inside the learning device 10 is acquired has been described, but the learning data can be acquired from a removable medium or the like in addition to an external computer connected via a network, such as a file server. It does not matter if it is acquired.

  Each time a learning sample is acquired in this way, the input control unit 13 inputs the learning input sentence included in the learning sample to the RNN encoder. As a result, a vector obtained by vectorizing the word string of the learning input sentence, that is, a so-called intermediate expression is output from the RNN encoder to the RNN decoder. Simultaneously with or before or after this, the input control unit 13 sets the upper limit number of characters set by the input control unit 13 to the value of the register holding the number of remaining characters until the RNN decoder outputs EOS (End Of Sentence) called end-of-sentence symbol. Initialize to. The details of the subsequent input to the RNN decoder, the output from the RNN data, and the updating of the model parameters using the same will be described later.

  The encoder execution unit 14 is a processing unit that executes an RNN encoder.

  As an embodiment, the encoder execution unit 14 according to the model information stored in the model storage unit 12 corresponds to the number M of words in the learning input sentence input by the input control unit 13, and the M LSTMs (Long Short). -Term Memory) is expanded on the work area. This causes the M LSTMs to function as an RNN encoder. In this RNN encoder, the mth word from the beginning of the learning input sentence is input to the LSTM corresponding to the mth word in order from the beginning word of the learning input sentence according to the input control by the input control unit 13. 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 learning input sentence, that is, a so-called intermediate representation is obtained. The intermediate representation of the learning input sentence thus generated by the RNN encoder is input to the RNN decoder.

  The decoder execution unit 15 is a processing unit that executes an RNN decoder.

  As one embodiment, the decoder execution unit 15 sets N LSTMs corresponding to the number N of words of the correct answer summary sentence input by the input control unit 13 on the work area according to the model information stored in the model storage unit 12. Expand to. This causes the N LSTMs to function as an RNN decoder. Under the control of the input control unit 13, these RNN decoders receive the intermediate representation of the learning input sentence from the RNN encoder and output the EOS tag from the input control unit 13 for each N LSTMs. The remaining number of characters is entered. By operating the N LSTMs according to these inputs, the RNN decoder outputs the probability distribution of words for each of the N LSMTs to the generation unit 16. The “word probability distribution” mentioned here refers to a distribution of probabilities calculated for each word that appears in the learning input sentence in the entire learning sample.

  The generation unit 16 is a processing unit that generates words of a summary sentence.

  As one embodiment, when the probability distribution of words is output from the n-th LSTM of the RNN decoder, the generation unit 16 generates the word with the highest probability in the probability distribution as the n-th word from the beginning of the summary sentence. To do.

  The calculation unit 17 is a processing unit that calculates the loss for each word generated by the model.

  As one embodiment, when the generation unit 16 generates the nth word of the summary sentence, the calculation unit 17 calculates the nth word of the words included in the correct answer summary sentence and the nth word generated by the generation unit 16. The loss is calculated from the second word and.

  The update unit 18 is a processing unit that updates the parameters of the model.

  As an embodiment, when the loss is calculated for each of the N LSTMs of the RNN decoder, the updating unit 18 optimizes the log-likelihood based on the loss of each LSTM to determine the model of the RNN decoder. Calculate the parameters to update. Then, the updating unit 18 updates the parameters of the model stored in the model storage unit 12 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.

[Specific example of model learning]
Hereinafter, with reference to FIGS. 4 to 6, a specific example of a dictionary constructed by model learning will be described while explaining a specific example of model learning.

  4A and 4B exemplify an example of a learning sample input to the model. FIG. 4A is a diagram showing an example of a learning input sentence. FIG. 4B is a diagram showing an example of the correct answer summary sentence. FIG. 5 is a diagram illustrating an example of model learning. Of these, FIG. 5 shows model learning in a situation where the learning input sentence 60A shown in FIG. 4A and the learning sample d1 of the correct answer summary sentence 60B shown in FIG. 4B are input to the model.

  As shown in FIG. 5, when the correct answer summary sentence 60B shown in FIG. 4B is input, 14 LSTMs corresponding to the word number “14” of the correct answer summary sentence 60B are developed in the work area. Hereinafter, each of the LSTMs developed in the work area corresponding to each word in order from the first word of the correct answer summary sentence 60B is identified as “LSTM15A to LSTM15N”.

  For example, at the first time when the probability distribution of the word to be matched with the first word of the correct answer summary sentence 60B is calculated, according to the control of the input control unit 13, the learning input sentence shown in FIG. 4A from the last LSTM of the RNN encoder. The intermediate representation of 60A is input to the leading LSTM 15A. At the same time, in the leading LSTM 15A, the number of characters of the correct answer sentence "23" is set as the initial value of the number of remaining characters until the ENN is output to the RNN decoder, together with the beginning symbol called BOS (Begin Of Sentence) from the input control unit 13. Is entered. Thereby, the LSTM 15A calculates the probability distribution of the words at the first time (t = 1) by calculating the probability for each word that appears in the learning input sentence in the entire learning sample, and at the same time, calculates the learning input sentence 60A. Update the intermediate representation. Then, the LSTM 15A outputs the probability distribution of the word at the first time to the generation unit 16 and outputs the intermediate expression updated at the first time to the LSTM 15B in the next stage.

  When the probability distribution of words at the first time is output in this way, the generation unit 16 generates the word with the highest probability in the probability distribution, “Yoshida” in this example, as the first word of the summary sentence, The first word of the summary sentence generated at the first time is output to the calculation unit 17. Then, the calculation unit 17 calculates the loss at the first time from the first word “Yoshida” corresponding to the first time among the words included in the correct answer summary sentence 60B and the word “Yoshida” generated at the first time. calculate. In this case, a smaller loss is calculated as the probability of the correct word “Yoshida” at the first time is closer to 1 and the probabilities of the other words are closer to 0. After that, the input control unit 13 subtracts the number of characters “2” of the word “Yoshida” generated at the first time from the initial value “23” of the number of remaining characters held in the register to thereby set the value of the number of remaining characters to “ 21 ".

  Next, at the second time when the probability distribution of the word that matches the second word from the beginning of the correct answer summary sentence 60B is calculated, the intermediate expression updated at the first time is input from the LSTM 15A to the LSTM 15B. At the same time, under the control of the input control unit 13, the correct word "Yoshida" at the first time of the second time, that is, at the first time is input to the LSTM15B, and the number of remaining characters "21" held in the register is "21". Is input. With this, the LSTM 15B calculates the probability distribution of the words at the second time (t = 2) by calculating the probability for each word that appears in the learning input sentence in the entire learning sample, and at the same time, calculates the learning input sentence 60A. Update the intermediate representation. Then, the LSTM 15B outputs the probability distribution of the word at the second time to the generation unit 16, and outputs the intermediate expression updated at the second time to the LSTM 15C in the next stage.

  When the probability distribution of words at the second time is output in this way, the generation unit 16 determines that the word having the highest probability in the probability distribution, “rubber” in this example, is the second word from the beginning of the summary sentence. The second word from the beginning of the summary sentence generated at the second time is output to the calculation unit 17. Then, the calculating unit 17 determines the second time from the word “rubber” which is the second from the beginning corresponding to the second time and the word “rubber” generated at the second time among the words included in the correct answer summary sentence 60B. Calculate the loss at. In this case, a smaller loss is calculated as the probability of the correct word “rubber” at the second time is closer to 1 and the probabilities of the other words are closer to 0. After that, the input control unit 13 subtracts the number of characters “2” of the word “rubber” generated at the second time from the value of the number of remaining characters “21” held in the register to obtain the value of the remaining number of characters “19”. Update to.

  The above process is repeatedly executed until the 14th time when the end-of-sentence symbol “EOS” is output from the LSTM 15N. After that, the updating unit 18 performs optimization of the log-likelihood based on the loss at the 1st time to the 14th time to calculate parameters for updating the model of the RNN decoder, and then stores the parameters in the model storage unit 12. The parameters of the stored model are updated to the parameters obtained by the optimization of the log likelihood.

  Thus, when the learning data D including the learning sample d1 of the learning input sentence 60A shown in FIG. 4A and the correct answer summary sentence 60B shown in FIG. 4B is used for model learning, as an example, the learned model is The dictionary shown in FIG. 6 is constructed. FIG. 6 is a diagram showing an example of a model dictionary. In FIG. 6, as one aspect, a part of the words of the learning input sentence 60A included in the learning sample d1 is extracted and shown, but the words that appear in the entire learning data D are the words of the summary sentence. It is made into a dictionary as a target for calculating the probability of generation. For example, as shown in FIG. 6, in the dictionary of learned models, “reporter”, “meeting”, “new product”, “ha”, “as is”, “off-the-shelf”, “cost”, “attention” , "Yoshida", "Rubber", "Performance", etc. are included. In the learned model equipped with such a dictionary of words, as shown in FIG. 5, when an input sentence including the proper expression “Yoshida” is input, the word “rubber” is generated next to the word “Yoshida”. The model learning that increases the generation probability is realized by the learning sample d1.

[Configuration of Generator 30]
As illustrated in FIG. 1, the generation device 30 includes an input control unit 31, an encoder execution unit 32, a decoder execution unit 33, a generation unit 34, an extraction unit 35, a detection unit 36, and a replacement unit 37. Have. Note that the generation device 30 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 input control unit 31, the encoder execution unit 32, the decoder execution unit 33, the generation unit 34, the extraction unit 35, the detection unit 36, and the replacement unit 37 illustrated in FIG. Virtually realized. Examples of such a processor include DLU, GPGPU, and GPU cluster. In addition to these, a CPU, an MPU, and the like are included. That is, the above-mentioned functional unit is virtually realized by the processor developing the above-mentioned summary generation program as a process on a memory such as a RAM. 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, the above-mentioned functional unit does not prevent that it is realized by hard-wired logic such as ASIC or FPGA.

  Note that, here, as an example, an example in which a summary generation program in which the functions corresponding to the above summary generation service are packaged is executed is described. However, a program module is used in units such as the above-mentioned unique expression replacement function. It can be executed or referenced by a library.

  The input control unit 31 is a processing unit that controls input to the model.

  As one embodiment, the input control unit 31 activates a process when receiving a request to generate a summary sentence. When the process is activated in this way, the input control unit 31 accepts the input sentence to be the target of the abstract sentence generation and the designation of the upper limit number of characters of the abstract sentence to be generated in the learned model. After that, the input control unit 31 inputs the input sentence to the RNN encoder. As a result, a vector obtained by vectorizing the word string of the input sentence, that is, a so-called intermediate representation is output from the RNN encoder to the RNN decoder. Simultaneously with or before or after this, the input control unit 31 initializes the value of the register that holds the number of remaining characters until the RNN decoder outputs EOS called a sentence end symbol to the specified upper limit number of characters. The details of the subsequent input to the RNN decoder, the output from the RNN data, and the generation of the summary sentence using the same will be described later.

  The encoder execution unit 32 is a processing unit that executes an RNN encoder.

  As one embodiment, the encoder execution unit 32 outputs K LSTMs corresponding to the number k of words of the input sentence input by the input control unit 31 according to the model information of the learned model stored in the model storage unit 12. By expanding on the work area, these K LSTMs function as an RNN encoder. In this RNN encoder, under the input control of the input control unit 31, the kth word from the beginning of the input sentence is input to the LSTM corresponding to the kth word in order from the beginning word of the input sentence, and k The output of the LSTM corresponding to the -1st word is input to the LSTM corresponding to the kth word. By repeating such input from the LSTM corresponding to the head word to the LSTM corresponding to the Kth word at the end, a vector of the input sentence, that is, a so-called intermediate representation is obtained. The intermediate representation of the input sentence thus generated by the RNN encoder is input to the RNN decoder.

  The decoder execution unit 33 is a processing unit that executes an RNN decoder.

  As an embodiment, the decoder execution unit 33 expands the LSTM on the work area according to the model information of the learned model stored in the model storage unit 12 until the tag of the EOS is output, and thus the tag of the EOS is extracted. Until L is output, the L LSTMs expanded are made to function as an RNN decoder. Under the control of the input control unit 31, these RNN decoders receive the intermediate representation of the input sentence from the RNN encoder, and the number of remaining characters until the input control unit 31 outputs an EOS tag for each L LSTMs. Is entered. By operating the L LSTMs according to these inputs, the RNN decoder outputs a probability distribution of words to the generation unit 34 for each of the L LSMTs.

  The generation unit 34 is a processing unit that generates a summary sentence.

  As one embodiment, when the probability distribution of words is output from the l-th LSTM of the RNN decoder, the generation unit 34 generates the word having the highest probability in the probability distribution as the l-th word from the beginning of the summary sentence. To do. After that, when the EOS tag is output from the Lth LSTM of the RNN decoder, the generation unit 34 generates a summary sentence by combining words generated in order from the first LSTM to the Lth LSTM, and generates the summary sentence. The extracted summary sentence is output to the extraction unit 35.

[Specific example of summary generation]
Hereinafter, a specific example of the summary sentence generation using the learned model will be described with reference to FIG. 7.

  FIG. 7 is a diagram illustrating an example of the input sentence and the summary sentence. FIG. 7 shows an example in which the input sentence 40G1 is input to the learned model equipped with the word dictionary shown in FIG. The input sentence 40G1 corresponds to the entire sentence of the input sentence 40G shown in FIG. As shown in FIG. 7, when the input sentence 40G1 is input to the learned model, the learned model outputs the summary sentence 40Y. Here, attention is paid to a proper expression including the word “Yoshida” that is common between the input sentence 40G1 and the summary sentence 40Y. Then, the portion of the summary sentence 40Y corresponding to the proper expression "Yoshida ○○" included in the input sentence 40G1 is output as "Yoshida Gomu", which indicates that the input sentence 40G1 deviates from its original purpose. This is partly because it is an unknown word that does not have the proper expression "○○" in the word dictionary shown in FIG. Then, as shown in FIG. 5, in the learned model, when the input sentence including the proper expression “Yoshida” is input, the generation probability that the word “rubber” is generated next to the word “Yoshida” becomes high. Such model learning is realized by the learning sample d1. As a result, the portion of the summary sentence 40Y that should be output as "Yoshida ○○" is erroneously output as "Yoshida rubber".

  In the following, as one aspect, even if the summary sentence 40Y in which the unknown word is incorrect is output from the trained model, the unique expression replacement function that can generate the summary sentence including the unknown word “○○” is used. The description of each corresponding functional unit will be continued.

  The extraction unit 35 is a processing unit that extracts a unique expression.

  As one embodiment, the extraction unit 35 executes the morphological analysis of the text of each of the input sentence and the summary sentence. Using the result of this morphological analysis, the extraction unit 35 does not correspond to a label related to the unique expression NE corresponding to the position of each character included in the text, for example, the label “Y” corresponding to the unique expression or the unique expression. A labeling process for giving a label “N” is executed. Hereinafter, a label related to the unique expression NE may be referred to as “NE label”. An arbitrary entity extraction engine can be used for this labeling process, and open source software may be used. As a result, a unique expression is extracted for each character of the input sentence and the summary sentence.

  The detection unit 36 is a processing unit that detects a first proper expression and a second proper expression that correspond between the summary sentence and the input sentence.

  As one embodiment, the detection unit 36 determines the first unique expression and the second unique expression depending on whether the similarity of the entire set of unique expressions between the summary sentence and the input sentence is equal to or more than a predetermined threshold. To detect. As an example, the detection unit 36 detects, for each of the summary sentence and the input sentence, the character string in the section from when the NE label of the summary sentence or the input sentence changes from “N” to “Y” until it returns to “N”. Extract as a whole set of unique expressions. Then, the detection unit 36 selects one whole set of unique expressions of the summary sentence. Subsequently, the detection unit 36 sets, for each set of unique expressions of the input sentence, the similarity between the set of unique expressions of the input sentence and the set of unique expressions of the summary sentence being selected to a predetermined threshold L. It is determined whether or not the above. As a result, the entire set of eigenexpressions of the summary sentence and the overall set of eigenexpressions of the input sentence having the similarity of not less than the threshold L are detected as the first and second eigenexpressions.

  As an example, the Jacquard coefficient can be used for the similarity between the unique expressions. For example, the Jacquard coefficient can be calculated by the following equation (1). A cosine measure can also be used for the similarity. For example, the cosine measure can be calculated by the following formula (2). Here, “| YNE [y] ∧GNE [h] |” in the following two expressions indicates the common number of characters in the general set of the two unique expressions. Further, “| YNE [y] |” and “| GNE [h] |” indicate the number of characters in the entire set of unique expressions of the summary sentence or the input sentence. Further, “max (A, B)” in the following expression (1) indicates a function that returns the maximum value of A and B as a return value. Further, “sqrt (A)” in the following formula (2) indicates the square root of A.

| YNE [y] ∧ GNE [h] | / max (| YNE [y] |, | GNE [h] |) (1)
| YNE [y] ∧ GNE [h] | / (sqrt (| YNE [y] |) sqrt | GNE [h] |)) (2)

  For example, when YNE is “Yoshida rubber” and GNE is “Yoshida ○○”, | YNE [y] ∧GNE [h] | is two characters of “Yoshida”. Therefore, the jacquard coefficient can be calculated as "1/2" by calculating 2 / max (4,4). Further, the cosine measure can be calculated as “1” by calculating 2 / (√2 × √2). For example, by setting the threshold value L to a value such as "0.3" or "0.5", "Yoshida Gomu" is detected as the first proper expression, and "Yoshida ○○" is detected as the second. It can be detected as a proper expression.

  If there is a difference in the number of characters between the two sets of unique expressions, a subset with the same number of characters as the one with the smaller number of characters is extracted from the set with the larger number of characters, and the extracted subset of the unique expressions is extracted. It is also possible to calculate the degree of similarity with the entire set of unique expressions having a smaller number of characters for each. In this case, the combination having the highest similarity among the degrees of similarity exceeding the threshold L may be detected as the first proper expression and the second proper expression. Further, here, the case where the calculation of the degree of similarity is performed by using the character as the minimum unit is illustrated, but the calculation may be performed by using the word as the minimum unit.

  The replacement unit 37 is a processing unit that replaces the character string of the first proper expression with the character string of the second proper expression.

  As one embodiment, when the detection unit 36 extracts the first specific expression and the second specific expression, the replacement unit 37 makes the character strings adjacent to each specific expression similar between the summary sentence and the input sentence. Or not. For example, from the aspect that the modification part and the modified part are modified from the preceding clause to the succeeding clause, the replacing part 37 has a predetermined length N adjacent before the first proper expression and the second proper expression. It is determined whether the similarity of the character string is a predetermined threshold value C or not. At this time, when the similarity of the character string of the predetermined length N adjacent before each unique expression is equal to or more than the threshold value C, the first unique expression of the summary sentence and the second unique expression of the input sentence have a correspondence relationship. The possibility can be estimated. In this case, the replacement unit 37 replaces the character string including the first proper expression in the summary sentence with the character string including the second proper expression in the input sentence.

  By the processing of the extraction unit 35, the detection unit 36, and the replacement unit 37, the case described with reference to FIG. 7, that is, the summary sentence 40Y in which the error “rubber” is present in the unknown word “XX” is output from the learned model. Even if it is done, it is possible to generate a summary including the unknown word "○○".

  That is, as described with reference to FIG. 3, first, the specific expression is extracted from the input sentence 40G and the summary sentence 40Y. Then, from the summary sentence 40Y, the proper expression "Yoshida Gomu" including the common word "Yoshida" is narrowed down as the first proper expression. On the other hand, as shown in bold in the relevant part of the input sentence 40G, the proper expression "Yoshida ○○" including the common word "Yoshida" is narrowed down as the second proper expression. Furthermore, since the two specific expressions “Yoshida Gomu” and “Yoshida ○○” are adjacent to each other and the adjacent character string of a predetermined length N “I served as a lecturer in the disaster prevention class at the same school” matches, the similarity is judged to be It can be confirmed that both the card coefficient and the cosine measure are 1 which is equal to or more than the threshold value C. Therefore, it can be estimated that the proper expression "Yoshida Gomu" of the summary sentence 40Y is likely to correspond to the proper expression "Yoshida ○○" of the input sentence 40G. After such confirmation, the character string 42Y “Yoshida ○○ who served as a lecturer in the disaster prevention classroom at the school” shown in black and white in reverse video in FIG. 3 is shown in italics in FIG. It is replaced with the character string 42G "Yoshida ○○ who served as a lecturer in the disaster prevention classroom at the school".

  Therefore, the modified summary sentence 43Y in which the proper expression of the unknown word "○○" of the learned model was corrected by using the input sentence 40G "Pointed out by Yoshida ○○ who was a lecturer in the disaster prevention classroom at the school in November 2015. Can be generated and output to a predetermined output destination such as a terminal device connected to the generation device 30.

[Process flow]
FIG. 8 is a flowchart illustrating the procedure of the summary generation process according to the first embodiment. This processing is executed after the preprocessing for extracting the unique expression from the summary sentence and the input sentence is executed. Here, the parameters used in the summary generation process shown in FIG. 8 will be described. For example, “g” indicates a variable that indicates the position of the proper expression GNE of the input sentence G, and “0” is set as the initial value, for example. Further, “y” indicates a variable that indicates the position of the proper expression YNE of the summary sentence Y. The counters for these variables are held in registers or the like (not shown).

  As shown in FIG. 8, the detection unit 36 initializes the position y of the proper expression YNE of the summary sentence Y to "1" and increments the position g of the proper expression GNE of the input sentence G (step S101). Subsequently, the detection unit 36 determines whether the degree of similarity between the g-th unique expression GNE [g] of the input sentence G and the y-th unique expression YNE [y] of the summary sentence Y is equal to or greater than a threshold L. judge.

  At this time, when the similarity between GNE [g] and YNE [y] is equal to or greater than the threshold L (Yes in step S102), YNE [y] is detected as the first proper expression and GNE [g] is the second. It is detected as a proper expression. In this case, the replacement unit 37 determines whether or not the similarity between the character strings of the predetermined length N adjacent before GNE [g] and YNE [y] is the threshold value C or more (step S103).

  Here, when the similarity of the character string of the predetermined length N adjacent before GNE [g] and YNE [y] is more than threshold value C (step S103 Yes), GNE [g] and YNE [y] have a corresponding relationship. It can be estimated that In this case, the replacement unit 37 replaces the y-th unique expression YNE [y] in the summary sentence Y with the g-th unique expression GNE [g] in the input sentence G (step S104), and proceeds to the processing of step S105. Transition.

  On the other hand, when the degree of similarity between GNE [g] and YNE [y] is not greater than or equal to the threshold value L, or the degree of similarity between character strings of a predetermined length N that are adjacent before GNE [g] and YNE [y] is not greater than or equal to the threshold value C. In that case (No in step S102 or step S103), the process in step S104 is skipped and the process proceeds to step S105.

  Then, the detection unit 36 increments the position y of the unique expression YNE of the summary sentence Y (step S105). After that, until the position y of the unique expression YNE of the summary sentence Y becomes equal to the number of unique expressions | YNE | of the summary sentence Y (step S106 Yes), the processes from step S102 to step S105 are repeatedly executed. ..

  When the position y of the unique expression YNE of the summary sentence Y becomes the number of unique expressions | YNE | of the summary sentence Y (No in step S106), the position g of the unique expression GNE of the input sentence G is the number of unique expressions of the input sentence G. Until it becomes equal to | GNE | (step S107 Yes), the processes from step S101 to step S106 are repeatedly executed.

  After that, when the position g of the unique expression GNE of the input sentence G becomes equal to the number | GNE | of unique expressions of the input sentence G (No at step S107), the process ends.

[One side of effect]
As described above, the generation device 30 according to the present embodiment, when there is a specific expression including a common word between the input sentence and the summary sentence, converts the character string of the unique expression of the summary sentence into the unique expression of the input sentence. Replace with the string. Therefore, according to the generation device 30 according to the present embodiment, it is possible to generate an abstract sentence including an unknown word that is not in the model dictionary.

  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.

[Application example of unique expression detection method]
In the above-described first embodiment, an example of detecting the first proper expression and the second proper expression by using the proper expressions included in the summary sentence and the input sentence as a search key has been described, but the first proper characteristic is detected by another method. The expression and the second proper expression can also be detected. For example, the generation device 30 determines whether the word at the end or the beginning of the matching character string and the word adjacent to the end or the beginning of the matching character string that are sequentially searched from the beginning or the end of the input sentence and the summary sentence are unique expressions. It is also possible to detect the first proper expression and the second proper expression depending on whether or not.

  FIG. 9 is a diagram illustrating an application example of the unique expression detection method. The upper part of FIG. 9 shows the position g and the NE tag in the input sentence 40G for each character included in the input sentence 40G shown in FIG. On the other hand, in the lower part of FIG. 9, the position y and the NE tag in the summary sentence 40Y are shown for each character included in the summary sentence 40Y shown in FIG.

  For example, when the search is sequentially performed from the beginning of the input sentence 40G and the summary sentence 40Y, the following is performed. That is, as shown in FIG. 9, the character string from position g = 38 to position g = 45 in the input sentence 40G matches the character string from position y = 15 to position y = 23 in the summary sentence 40Y. On the other hand, the character "o" at the position g = 46 in the input sentence 40G and the character "go" at the position y = 24 in the summary sentence 40Y do not match. As a result, the matching character string “Yoshida who acted as a lecturer” surrounded by a frame 81 shown in FIG. 9 is obtained as a search result.

  When the matching character string is obtained, the word "Yoshida" at the end of the matching character string and the word adjacent to the end of the matching character string, that is, the word "○○" of the input sentence 40G surrounded by the frame 82 and It is determined whether or not the word “rubber” in the summary sentence 40Y is a proper expression. Since the NE tags of the word “Yoshida”, the word “○○” and the word “rubber” are all “Y”, it can be confirmed that they are unique expressions. In this case, the character string “Yoshida Gomu” including the word “Yoshida” at the end of the matching character string and the word “rubber” adjacent to the end of the matching character string in the summary sentence 40Y is the matching character string of the input sentence 40G. The word "Yoshida" at the end and the character string "Yoshida ○○" including the word "○○" adjacent to the end of the matching character string are replaced. That is, the character string "Yoshida rubber" surrounded by the broken line frame 83Y shown in FIG. 9 is replaced with the character string "Yoshida XX" surrounded by the solid line frame 83G in FIG.

  FIG. 10 shows an example of an algorithm of such a method of replacing a proper expression. 10. FIG. 10 is a flowchart illustrating the procedure of the summary generation process according to the second embodiment. This processing is also executed after the preprocessing for extracting the unique expression from the summary sentence and the input sentence is executed. Here, the parameters used in the summary generation process shown in FIG. 10 will be described. For example, “g” refers to a variable that indicates the position of the proper expression GNE of the input sentence G, and for example, “1” is set as the initial value. Further, “y” indicates a variable indicating the position of the proper expression YNE of the summary sentence Y. Further, “c” indicates the number of characters in the matching character string. Further, “h” indicates an index for searching the matching character string. Counters for these variables are held in a register or the like (not shown).

  As shown in FIG. 10, the detection unit 36 initializes the position y of the proper expression YNE of the summary sentence Y to “1”, initializes the number of characters c of the matching character string to “0”, and searches for the matching character string. The index h for use is initialized to "g" (step S301).

  Subsequently, the detection unit 36 determines whether or not the h-th character in the input sentence G and the y-th character in the summary sentence Y are the same character, that is, G [h] == Y [y]. It is determined whether or not (step S302). If G [h] == Y [y] is not satisfied (No in step S302), the process proceeds to step S308.

  At this time, if G [h] == Y [y] (Yes in step S302), the detection unit 36 determines the position y of the unique expression YNE of the summary sentence Y, the number of characters c of the matching character string, and the matching character string for searching. The index h of is incremented (step S303).

  Then, the detection unit 36 determines whether or not the number of characters c of the matching character string is smaller than a threshold value L to be compared with the matching character string, that is, whether or not the number of characters c of the matching character string <threshold L (step S304). ). When the number of characters c of the matching character string is smaller than the threshold value L (Yes in step S304), the process returns to step S302.

  On the other hand, when the number of characters c of the matching character string is not smaller than the threshold L (No in step S304), the detection unit 36 determines whether the h-th character in the input sentence G and the y-th character in the summary sentence Y are different characters. , Ie G [h]! = Y [y] is determined (step S305). G [h]! = Y [y] is not satisfied (No in step S305), the process returns to step S303.

  At this time, G [h]! = Y [y] (step S305 Yes), the detection unit 36 determines the h-1th character in the input sentence G, the hth character in the input sentence G, and the yth character in the summary sentence Y. It is determined whether the NE label is the label “Y” corresponding to the unique expression (step S306). When the NE labels of the three characters are not "Y" (No in step S306), the process proceeds to step S308.

  Here, when the NE label of the h−1th character in the input sentence G, the hth character in the input sentence G, and the yth character in the summary sentence Y is the label “Y” (step S306 Yes), Perform the following processing. That is, the replacement unit 37 replaces the character string of the unique expression including the character Y [y] at the position y in the summary sentence Y with the unique expression including the character G [h-1] at the position h-1 in the input sentence G. The character string is replaced (step S307).

  Then, the detection unit 36 increments the position g of the input sentence G (step S308). Then, until the position g of the input sentence G reaches the number of characters | G | of the input sentence G (step S309 Yes), the processes from step S301 to step S308 are repeatedly executed. Finally, when the position g of the input sentence G reaches the number of characters | G | of the input sentence G (No in step S309), the process ends.

  With such processing, it is possible to generate a summary sentence including unknown words that are not in the model dictionary, as in the summary generation processing shown in FIG.

[Application example of applicable conditions]
In the first and second embodiments, the similarity of the character strings adjacent to the proper expression between the summary sentence 40Y and the input sentence 40G is set as the application condition for the replacement of the first proper expression and the second proper expression. Although an example has been given, other weighting conditions may be set. For example, in the proper expression extraction, classes such as a person name, a place name, an organization name, and a numerical expression can be obtained for each character or morpheme in addition to the NE label, and this can be used for setting the application condition. For example, the class is common between the first proper expression and the second proper expression, and further, all the classes into which the characters or morphemes included in the first proper expression and the second proper expression are classified. The same condition can be set as the application condition.

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 input control unit 31, the encoder execution unit 32, the decoder execution unit 33, the generation unit 34, the extraction unit 35, the detection unit 36, or the replacement unit 37 may be connected as an external device of the generation device 30 via a network. .. As an example, the input control unit 31, the encoder execution unit 32, the decoder execution unit 33, and the generation unit 34 may be provided in an external device of the generation device 30. In addition, another device has the input control unit 31, the encoder execution unit 32, the decoder execution unit 33, the generation unit 34, the extraction unit 35, the detection unit 36, or the replacement unit 37, and they are network-connected and cooperate with each other. The functions of the generation device 30 described above may be realized. As an example, a device including the input control unit 31, the encoder execution unit 32, the decoder execution unit 33, and the generation unit 34 and a device including the extraction unit 35, the detection unit 36, and the replacement unit 37 cooperate with each other via a network. May be

[Summary generator]
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, in the following, an example of a computer that executes a summary generation program having the same functions as those in the above embodiments will be described with reference to FIG. 11.

  FIG. 11 is a diagram illustrating a hardware configuration example of a computer that executes the abstract generation program according to the first and second embodiments. As shown in FIG. 11, 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. 11, the HDD 170 includes the input control unit 31, the encoder execution unit 32, the decoder execution unit 33, the generation unit 34, the extraction unit 35, the detection unit 36, and the replacement unit 37 shown in the first embodiment. A summary generation program 170a having the same function is stored. This summary generation program 170a is integrated like the components of the input control unit 31, the encoder execution unit 32, the decoder execution unit 33, the generation unit 34, the extraction unit 35, the detection unit 36, and the replacement unit 37 shown in FIG. Or you may separate. 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 the digest generation program 170a from the HDD 170 and loads it on the RAM 180. As a result, the abstract generation program 170a functions as an abstract generation process 180a, as shown in FIG. The abstract generation process 180a expands various data read from the HDD 170 in the area allocated to the abstract generation 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. 8 and 10 is included as an example of the processing executed by the abstract generation 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.

  It should be noted that the above-mentioned summary generation program 170a does not necessarily have to be stored in the HDD 170 or the ROM 160 from the beginning. For example, the abstract generation program 170a is stored in a “portable physical medium” such as a flexible disk inserted into the computer 100, a so-called FD, CD-ROM, DVD disk, magneto-optical disk, IC card, or the like. Then, the computer 100 may acquire the summary generation program 170a from these portable physical media and execute it. Further, the summary generation program 170a is stored in another computer or a server device connected to the computer 100 via a public line, the Internet, a LAN (Local Area Network), a WAN, etc., and the computer 100 generates a digest from these. The program 170a may be acquired and executed.

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

(Supplementary Note 1) When there is a proper expression including a common word between an input sentence and a summary sentence output by the model into which the input sentence is input, a first proper characteristic including the common word in the summary sentence. Replacing the character string of the expression with a character string of the second proper expression including the common word in the input sentence,
A method for generating an abstract, characterized in that a computer executes the processing.

(Supplementary Note 2) The first proper expression and the second proper expression are detected depending on whether or not the similarity of the entire set of proper expressions between the input sentence and the summary sentence is equal to or more than a predetermined threshold value. 2. The abstract generating method according to appendix 1, wherein the computer further executes the process.

(Supplementary Note 3) The replacing process is performed by combining a first adjacent character string that is adjacent before or after the first proper expression and a second adjacent character string that is adjacent before or after the second proper expression. 3. The abstract generating method according to appendix 2, wherein the character string of the first proper expression is replaced with the character string of the second proper expression on condition that the similarity is equal to or higher than a predetermined threshold.

(Supplementary Note 4) Whether the word at the end or the beginning of the matching character string searched from the beginning or the end of the input sentence and the summary sentence and the word adjacent to the ending or the beginning of the matching character string are unique expressions. The method according to claim 1, wherein the computer further executes a process of detecting the first proper expression and the second proper expression depending on whether or not the first proper expression and the second proper expression are detected.

(Supplementary Note 5) In the replacing process, the character string of the first unique expression is replaced with the character string of the second unique expression on the condition that the length of the matching character string is equal to or greater than a predetermined threshold value. The abstract generation method according to appendix 4, characterized in that

(Supplementary Note 6) When there is a proper expression including a common word between the input sentence and the summary sentence output by the model into which the input sentence is input, the first proper characteristic including the common word in the summary sentence. Replacing the character string of the expression with a character string of the second proper expression including the common word in the input sentence,
A summary generation program characterized by causing a computer to execute a process.

(Supplementary Note 7) The first specific expression and the second specific expression are detected depending on whether or not the similarity of the entire set of specific expressions between the input sentence and the summary sentence is equal to or more than a predetermined threshold value. 7. The abstract generating program according to appendix 6, further causing the computer to execute processing.

(Supplementary Note 8) The replacing process is performed by combining a first adjacent character string that is adjacent before or after the first proper expression and a second adjacent character string that is adjacent before or after the second proper expression. 8. The abstract generation program according to appendix 7, wherein the character string of the first proper expression is replaced with the character string of the second proper expression on condition that the similarity is equal to or higher than a predetermined threshold value.

(Supplementary Note 9) Whether the word at the end or the beginning of the matching character string searched from the beginning or the end of the input sentence and the summary sentence, and the word adjacent to the ending or the beginning of the matching character string are unique expressions. 7. The abstract generation program according to appendix 6, further causing the computer to further execute a process of detecting the first proper expression and the second proper expression depending on whether or not the computer.

(Supplementary Note 10) In the replacing process, the character string of the first unique expression is replaced with the character string of the second unique expression on the condition that the length of the matching character string is equal to or larger than a predetermined threshold value. The summary generation program according to appendix 9, characterized in that.

(Supplementary Note 11) When there is a proper expression including a common word between the input sentence and the summary sentence output by the model to which the input sentence is input, the first proper characteristic including the common word in the summary sentence. A replacement unit that replaces the character string of the expression with the character string of the second proper expression including the common word in the input sentence,
An abstract generating device comprising:

(Supplementary Note 12) The first proper expression and the second proper expression are detected depending on whether or not the similarity of the entire set of proper expressions between the input sentence and the summary sentence is equal to or more than a predetermined threshold value. The abstract generating device according to appendix 11, further comprising a detecting unit.

(Supplementary note 13) The replacement unit resembles a first adjacent character string that is adjacent before or after the first proper expression and a second adjacent character string that is adjacent before or after the second proper expression. 13. The abstract generating apparatus according to appendix 12, wherein the character string of the first proper expression is replaced with the character string of the second proper expression on condition that the degree is equal to or higher than a predetermined threshold.

(Supplementary Note 14) Whether the word at the end or the beginning of the matching character string retrieved from the beginning or the end of the input sentence and the summary sentence, and the word adjacent to the ending or the beginning of the matching character string are unique expressions. 12. The abstract generating apparatus according to appendix 11, further comprising a detection unit that detects the first proper expression and the second proper expression depending on whether or not the first proper expression is present.

(Supplementary Note 15) The replacement unit replaces the character string of the first unique expression with the character string of the second unique expression on the condition that the length of the matching character string is equal to or greater than a predetermined threshold. 15. The abstract generating apparatus according to supplementary note 14.

10 learning device 11 learning data storage unit 12 model storage unit 13 input control unit 14 encoder execution unit 15 decoder execution unit 16 generation unit 17 calculation unit 18 update unit 30 generation device 31 input control unit 32 encoder execution unit 33 decoder execution unit 34 generation Part 35 Extraction part 36 Detection part 37 Substitution part

Claims (6)

When there is a proper expression including a common word between the input sentence and the summary sentence output by the model to which the input sentence is input, a character string of a first proper expression including the common word in the summary sentence. Is replaced with a character string of a second proper expression including the common word in the input sentence,
A method for generating an abstract, characterized in that a computer executes the processing.   The computer performs the process of detecting the first unique expression and the second unique expression depending on whether or not the similarity of the entire set of unique expressions between the input sentence and the summary sentence is equal to or more than a predetermined threshold value. The method according to claim 1, further comprising:   In the replacing process, the similarity between a first adjacent character string adjacent before or after the first proper expression and a second adjacent character string adjacent before or after the second proper expression is predetermined. 3. The abstract generation method according to claim 2, wherein the character string of the first unique expression is replaced with the character string of the second unique expression on the condition that the character string is equal to or more than the threshold value.   Depending on whether the word at the end or the beginning of the matching character string searched in order from the beginning or the end of the input sentence and the summary sentence and the word adjacent to the end or the beginning of the matching character string are unique expressions, The method according to claim 1, wherein the computer further executes a process of detecting the first proper expression and the second proper expression.   The replacing process replaces the character string of the first specific expression with the character string of the second specific expression on condition that the length of the matching character string is equal to or greater than a predetermined threshold. The summary generation method according to claim 4. When there is a proper expression including a common word between the input sentence and the summary sentence output by the model to which the input sentence is input, a character string of a first proper expression including the common word in the summary sentence. Is replaced with a character string of a second proper expression including the common word in the input sentence,
A summary generation program characterized by causing a computer to execute a process.

Images