Natural language processing device

Abstract

This natural language processing device comprises: an element information extraction unit that, for each information type, uses an element learning model corresponding to the information type to extract element information from each of a higher-level document and a lower-level document; and an entailment relation inference unit that, for each information type, uses an entailment relation learning model corresponding to the information type to determine entailment relation between element information extracted from the higher-level document and element information extracted from the lower-level document, and on the basis of a determination result, outputs a result of inference of the entailment relation between the higher-level document and lower-level document.

Description

Natural Language Processing Unit

[0001] The present invention relates to a natural language processing device.

[0002] When there is a higher-level document that lists multiple requirements and a lower-level document that lists the means to satisfy each of the requirements listed in the higher-level document, the correspondence between the requirements written in the higher-level document and the locations in the lower-level document that list the means to satisfy those requirements will be confirmed.

[0003] Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova: "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", Proceedings of NAACL HLT 2019, pages 4171-4186. Pretrained Japanese BERT models released.<URL: https: / / www.nlp.ecei.tohoku.ac.jp / news-release / 3284 / >

[0004] When searching for a part of a lower-level document that corresponds to a requirement stated in a higher-level document, one possible method is to use machine learning to match sentences within a document (see, for example, Non-Patent Document 1 and Non-Patent Document 2).

[0005] However, in the conventional method, when comparing a higher-level document with a lower-level document, characteristics such as word order, style, grammar, phrasing, or writing style adversely affect the judgment results.

[0006] The present invention has been made in light of the above circumstances, and its purpose is to provide a device that searches for parts of lower-level documents that correspond to requirements stated in higher-level documents, taking into account the knowledge structure of the application field.

[0007] According to one aspect of the present invention, a natural language processing device includes an element information extraction unit and an entailment inference unit. The element information extraction unit extracts element information from each of a higher-level document and a lower-level document using an element learning model corresponding to the information type for each information type. The entailment inference unit determines, for each information type, an entailment relationship between the element information extracted from the higher-level document and the element information extracted from the lower-level document using the entailment learning model corresponding to the information type, and outputs an inferred entailment relationship between the higher-level document and the lower-level document based on the determination result.

[0008] According to the present invention, it is possible to provide a natural language processing device that searches for a portion corresponding to a requirement described in a higher-level document from a lower-level document, taking into account the knowledge structure of the application field.

[0009] FIG. 1 is a block diagram showing an example of a functional configuration of a natural language processing apparatus according to an embodiment. FIG. 2 is a block diagram showing an example of element training data. FIG. 3 is a diagram showing an example of generation of an element learning model in a domain knowledge learning unit included in the natural language processing apparatus according to an embodiment. FIG. 4 is a diagram showing an example of generation of an entailment relationship learning model in an entailment relationship learning unit included in the natural language processing apparatus according to an embodiment. FIG. 5 is a diagram showing an example of extraction of element information in an element information extraction unit included in the natural language processing apparatus according to an embodiment. FIG. 6 is a diagram showing a first example of entailment relationship inference in an entailment relationship inference unit included in the natural language processing apparatus according to an embodiment. FIG. 7 is a diagram showing a second example of entailment relationship inference in an entailment relationship inference unit included in the natural language processing apparatus according to an embodiment. FIG. 8 is a diagram showing a third example of entailment relationship inference in an entailment relationship inference unit included in the natural language processing apparatus according to an embodiment. FIG. 9 is a block diagram showing an example of a hardware configuration of a natural language processing apparatus according to an embodiment.

[0010] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The natural language processing apparatus according to the embodiments is an apparatus that performs natural language processing using machine learning technology. In the following description, components having the same functions and configurations will be given the same reference numerals.

[0011] (Functional Configuration) First, an example of the functional configuration of a natural language processing apparatus 10 according to an embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the functional configuration of a natural language processing apparatus 10 according to an embodiment. Below, a case will be described in which the natural language processing apparatus 10 infers an implication relationship between a higher-level document and a lower-level document relating to information security. Note that the higher-level document and the lower-level document processed by the natural language processing apparatus 10 are not limited to the information security field.

[0012] As shown in FIG. 1, the natural language processing apparatus 10 includes a domain knowledge learning unit 11, an entailment learning unit 12, an element information extraction unit 13, and an entailment inference unit 14.

[0013] As a pre-learning step, the field knowledge learning unit 11 generates an element learning model 201 for each information type (classification) in an application field (for example, the information security field) using a machine learning algorithm that inputs element training data 101.

[0014] More specifically, for example, element teacher data 101 corresponding to each information type is input to the domain knowledge learning unit 11. In other words, different element teacher data 101 are input to the domain knowledge learning unit 11 for each information type. The element teacher data 101 is teacher data (pre-learning data) for each information type used in machine learning performed by the domain knowledge learning unit 11. The element teacher data 101 includes one or more text data. Note that two or more information types are sufficient. For each information type, the domain knowledge learning unit 11 learns the characteristics of words (including sentences expressing unique expressions) related to the information type that occur before, after, or around technical terms (e.g., technical terms related to information security) using a machine learning algorithm that inputs the element teacher data 101 corresponding to the information type. Then, the domain knowledge learning unit 11 generates an element learning model 201 for learning the characteristics of words for each information type. Therefore, the domain knowledge learning unit 11 generates a different element learning model 201 for each information type. The domain knowledge learning unit 11 transmits each element learning model 201 to the element information extraction unit 13 .

[0015] As a pre-learning step, the entailment learning unit 12 generates an entailment learning model 202 for each information type by a machine learning algorithm that uses the entailment training data 102 as an input.

[0016] More specifically, the entailment relationship learning unit 12 receives, for each information type, entailment relationship teacher data 102 corresponding to the information type. In other words, different entailment relationship teacher data 102 are input to the entailment relationship learning unit 12 for each information type. The entailment relationship teacher data 102 is teacher data for each information type used in machine learning performed by the entailment relationship learning unit 12. The entailment relationship teacher data 102 includes one or more text data. For each information type, the entailment relationship learning unit 12 learns entailment relationships between words related to the information type using a machine learning algorithm that receives as input the entailment relationship teacher data 102 corresponding to the information type, and generates an entailment relationship learning model 202 for determining entailment relationships between words related to the information type. Therefore, the entailment relationship learning unit 12 generates a different entailment relationship learning model 202 for each information type. The entailment relationship learning unit 12 transmits each entailment relationship learning model 202 to the entailment relationship inference unit 14.

[0017] For each information type, the element information extraction unit 13 uses a machine learning algorithm that uses an element learning model 201 corresponding to the information type to extract words related to the information type that are located before, after, or around technical terms from each of the higher-level document 103 and the lower-level document 104. Hereinafter, the words extracted by the element information extraction unit 13 will be referred to as "element information."

[0018] More specifically, the element information extraction unit 13 receives the upper-level document 103 and one or more lower-level documents 104 as input. The upper-level document 103 and the lower-level documents 104 are text data. For each information type, the element information extraction unit 13 uses an element learning model 201 corresponding to the information type to measure the feature amount of the element learning model 201 for words related to the information type that are located before, after, or around the technical term. The element information extraction unit 13 then extracts, as element information, words whose feature amount is equal to or greater than a judgment value. The element information extraction unit 13 transmits the element information extracted from the upper-level document 103 and the element information extracted from the lower-level documents 104 to the implication inference unit 14.

[0019] The entailment inference unit 14 determines, for each information type, an entailment relationship between element information extracted from the higher-level document 103 and element information extracted from the lower-level document 104, using a machine learning algorithm that uses an entailment learning model 202 corresponding to the information type. Based on the determination result, the entailment inference unit 14 generates and outputs inference result data 105 that indicates an inference result as to whether or not there is a correspondence between the higher-level document and the lower-level document. The inference result data 105 is, for example, a text file.

[0020] The machine learning model used in the domain knowledge learning unit 11, the entailment learning unit 12, the element information extraction unit 13, and the entailment inference unit 14 may be, for example, BERT (Bidirectional Encoder Representations from Transformers).

[0021] (Element Teacher Data) Next, an example of the element teacher data 101 will be described with reference to Fig. 2. Fig. 2 is a block diagram showing an example of the element teacher data 101.

[0022] 2, for example, five information types in the field of information security are set: "access rights," "management information classification," "prohibition / use," "management location," and "storage medium." In this case, the element teacher data 101 includes access rights element teacher data 101a, management information classification element teacher data 101b, prohibition / use element teacher data 101c, management location element teacher data 101d, and media type element teacher data 101e. Each of the access rights element teacher data 101a, management information classification element teacher data 101b, prohibition / use element teacher data 101c, management location element teacher data 101d, and media type element teacher data 101e includes one or more text data.

[0023] (Generation of Element Learning Model) Next, an example of generation of the element learning model 201 will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of generation of the element learning model 201 in the field knowledge learning unit 11.

[0024] As shown in FIG. 3, the domain knowledge learning unit 11 generates, for each information type, an element learning model 201 corresponding to the information type.

[0025] More specifically, for example, access right element teacher data 101a is input to the domain knowledge learning unit 11. In this case, the domain knowledge learning unit 11 learns access right-related words W1 that occur before, after, or in the vicinity of technical terms using a machine learning algorithm that uses the access right element teacher data 101a as input. Then, the domain knowledge learning unit 11 generates an access right element learning model 201a that learns the characteristics of the words W1.

[0026] For example, management information classification element training data 101b is input to the domain knowledge learning unit 11. In this case, the domain knowledge learning unit 11 learns a word W2 related to the management information classification that occurs before, after, or in the vicinity of the technical term using a machine learning algorithm that uses the management information classification element training data 101b as input. Then, the domain knowledge learning unit 11 generates a management information classification element learning model 201b that learns the characteristics of the word W2.

[0027] For example, prohibited and causative element teacher data 101c is input to the domain knowledge learning unit 11. In this case, the domain knowledge learning unit 11 learns prohibited and causative words W3 that occur before, after, or in the vicinity of the technical term using a machine learning algorithm that uses the prohibited and causative element teacher data 101c as input. Then, the domain knowledge learning unit 11 generates a prohibited and causative element learning model 201c that learns the characteristics of word W3.

[0028] For example, management location element teacher data 101d is input to the domain knowledge learning unit 11. In this case, the domain knowledge learning unit 11 learns a word W4 related to a management location that occurs before, after, or in the vicinity of a technical term using a machine learning algorithm that uses the management location element teacher data 101d as input. Then, the domain knowledge learning unit 11 generates a management location element learning model 201d that learns the characteristics of the word W4.

[0029] For example, media type element teacher data 101e is input to the domain knowledge learning unit 11. In this case, the domain knowledge learning unit 11 learns a word W5 related to a storage medium that occurs before, after, or in the vicinity of the technical term using a machine learning algorithm that uses the media type element teacher data 101e as input. Then, the domain knowledge learning unit 11 generates a storage medium element learning model 201e that learns the characteristics of word W5.

[0030] (Generation of entailment learning model) Next, an example of generation of the entailment learning model 202 will be described with reference to Fig. 4. Fig. 4 is a diagram showing an example of generation of the entailment learning model 202 in the entailment learning unit 12.

[0031] As shown in FIG. 4, the entailment learning unit 12 generates an entailment learning model 202 corresponding to each information type.

[0032] More specifically, for example, access right entailment relationship teacher data 102a is input to the entailment relationship learning unit 12. In this case, the entailment relationship learning unit 12 learns the entailment relationship of the word W1 regarding the access right by a machine learning algorithm using the access right entailment relationship teacher data 102a as input. Then, the entailment relationship learning unit 12 generates an access right entailment relationship learning model 202a that learns the entailment relationship of the word W1.

[0033] For example, management information class entailment relation training data 102b is input to the entailment relation learning unit 12. In this case, the entailment relation learning unit 12 learns the entailment relation of word W2 related to the management information class by a machine learning algorithm using the management information class entailment relation training data 102b as input. Then, the entailment relation learning unit 12 generates a management information class entailment relation learning model 202b that learns the entailment relation of word W2.

[0034] For example, the forbidden / causative entailment relation training data 102c is input to the entailment relation learning unit 12. In this case, the entailment relation learning unit 12 learns the forbidden / causative entailment relation of word W3 by a machine learning algorithm using the forbidden / causative entailment relation training data 102c as input. Then, the entailment relation learning unit 12 generates a forbidden / causative entailment relation learning model 202c that learns the entailment relation of word W3.

[0035] For example, management location entailment relationship teacher data 102d is input to the entailment relationship learning unit 12. In this case, the entailment relationship learning unit 12 learns the entailment relationship of word W4 related to management location by a machine learning algorithm that uses the management location entailment relationship teacher data 102d as input. Then, the entailment relationship learning unit 12 generates a management location entailment relationship learning model 202d that learns the entailment relationship of word W4.

[0036] For example, storage medium entailment relation training data 102e is input to the entailment relation learning unit 12. In this case, the entailment relation learning unit 12 learns the entailment relation of word W5 related to the storage medium by a machine learning algorithm that uses the storage medium entailment relation training data 102e as input. Then, the entailment relation learning unit 12 generates a storage medium entailment relation learning model 202e that learns the entailment relation of word W5.

[0037] (Extraction of Element Information) Next, an example of extraction of element information will be described with reference to Fig. 5. Fig. 5 is a diagram showing an example of extraction of element information by the element information extraction unit 13. Fig. 5 shows a case where element information is extracted from a higher-level document 103. The element information extraction unit 13 executes a similar process when extracting element information from a lower-level document.

[0038] 5, the upper document 103 is input to the element information extraction unit 13. The element information extraction unit 13 extracts element information from the upper document 103 for each information type by a machine learning algorithm using an element learning model 201 corresponding to the information type.

[0039] More specifically, the element information extraction unit 13 uses a machine learning algorithm using the access right element learning model 201a to measure the feature amounts of words W1 related to access rights that are located before, after, or around technical terms included in the higher-level document 103. Then, the element information extraction unit 13 extracts, as element information E1, words W1 whose feature amounts are equal to or greater than a predetermined judgment value.

[0040] The element information extraction unit 13 uses a machine learning algorithm using the management information classification element learning model 201b to measure the feature amount of words W2 related to the management information classification that are located before, after, or around technical terms included in the higher-level document 103. Then, the element information extraction unit 13 extracts words W2 whose feature amount is equal to or greater than a predetermined judgment value as element information E2.

[0041] The element information extraction unit 13 uses a machine learning algorithm that uses the prohibited / causative element learning model 201c to measure the feature amounts of prohibited / causative words W3 that occur before, after, or around technical terms included in the higher-level document 103. Then, the element information extraction unit 13 extracts, as element information E3, words W3 whose feature amounts are equal to or greater than a predetermined threshold value.

[0042] The element information extraction unit 13 uses a machine learning algorithm using the management location element learning model 201d to measure the feature amount of a word W4 related to a management location that appears before, after, or around a technical term included in the upper document 103. Then, the element information extraction unit 13 extracts, as element information E4, a word W4 whose feature amount is equal to or greater than a predetermined judgment value.

[0043] The element information extraction unit 13 uses a machine learning algorithm using the storage medium element learning model 201e to measure the feature amounts of words W5 related to storage media that are located before, after, or around technical terms included in the higher-level document 103. Then, the element information extraction unit 13 extracts words W5 whose feature amounts are equal to or greater than a predetermined judgment value as element information E5.

[0044] The determination values ​​of the feature amounts used to extract the element information E1 to E5 may be set to the same value for each information type, or different values ​​may be set.

[0045] Similarly, the element information extraction unit 13 extracts element information E1 to E5 from the lower-level document 104.

[0046] The element information extraction unit 13 transmits the element information E1 to E5 extracted from the higher-level document 103 and the element information E1 to E5 extracted from the lower-level document 104 to the implication inference unit .

[0047] (Inference of an implication relationship between a higher-level document and a lower-level document) Next, three examples of inference of an implication relationship between a higher-level document and a lower-level document will be shown.

[0048] First, a first example of inference of an implication relationship will be described with reference to Fig. 6. Fig. 6 is a diagram showing a first example of inference of an implication relationship in the implication relationship inference unit 14.

[0049] 6, the entailment inference unit 14 compares, for each information type, the element information extracted from the higher-level document 103 with the element information extracted from the lower-level document 104. That is, the entailment inference unit 14 in this example determines the entailment relationship of the element information for each information type.

[0050] More specifically, the implication relationship inference unit 14 determines the implication relationship between the element information E1 extracted from the upper document 103 and the element information E1 extracted from the lower document 104 using a machine learning algorithm that uses the access right implication relationship learning model 202a.

[0051] The implication relationship inference unit 14 determines the implication relationship between the element information E2 extracted from the upper document 103 and the element information E2 extracted from the lower document 104 using a machine learning algorithm that uses the management information classification implication relationship learning model 202b.

[0052] The entailment inference unit 14 determines the entailment relationship between the element information E3 extracted from the higher-level document 103 and the element information E3 extracted from the lower-level document 104 using a machine learning algorithm that uses the prohibitive / causative entailment learning model 202c.

[0053] The implication inference unit 14 determines the implication relationship between the element information E4 extracted from the higher-level document 103 and the element information E4 extracted from the lower-level document 104 using a machine learning algorithm that uses the management location implication learning model 202d.

[0054] The implication inference unit 14 determines the implication between the element information E5 extracted from the higher-level document 103 and the element information E5 extracted from the lower-level document 104 using a machine learning algorithm that uses the storage medium implication learning model 202e.

[0055] The implication inference unit 14 infers the implication between the higher-level document 103 and the lower-level document 104 based on the determination result of the implication for each information type, and generates inference result data 105 .

[0056] Next, a second example of inference of an implication relationship will be described with reference to Fig. 7. Fig. 7 is a diagram showing a second example of inference of an implication relationship in the implication relationship inference unit 14.

[0057] As shown in FIG. 7, the implication inference unit 14 in this example determines the implication between the element information extracted from the upper document 103 and the element information extracted from the lower document 104 based on the combination information of the element information.

[0058] More specifically, the implication inference unit 14 generates combination information of element information E1 to E6 extracted from the higher-level document 103 based on the schema of the higher-level document 103. For example, element information E1 is located at the top, and element information E2 to E5 are linked in order downwards. Element information E6 is linked to the lower-level element of element information E2. Similarly, the implication inference unit 14 generates combination information of element information E1 to E5 extracted from the lower-level document 104 based on the schema of the lower-level document 104. For example, element information E1 is located at the top, and element information E2 to E5 are linked in order downwards. The implication inference unit 14 compares the combination information of element information E1 to E6 extracted from the higher-level document 103 with the combination information of element information E1 to E5 extracted from the lower-level document 104, and infers an implication relationship between the higher-level document 103 and the lower-level document 104 based on the comparison result, thereby generating inference result data 105. The implication inference unit 14 may execute the first and second examples of inference of an implication relationship. That is, the implication inference unit 14 may infer an implication relationship between the higher-level document 103 and the lower-level document 104 based on the implication relationship of the element information for each information type and the comparison result of the combination information, and generate inference result data 105.

[0059] Next, a third example of inference of an implication relationship will be described with reference to Fig. 8. Fig. 8 is a diagram showing a third example of inference of an implication relationship in the implication relationship inference unit 14. In this example, a case will be described in which a lower-level document 104 corresponding to a requirement stated in a higher-level document 103 is extracted based on the reference relationships of a plurality of lower-level documents 104.

[0060] 8, the entailment inference unit 14 generates context data from the context of the higher-level document 103. The entailment inference unit 14 searches for element information serving as a search key from the multiple lower-level documents 104 in accordance with the context data, thereby inferring an entailment relationship that takes into account the context of the higher-level document 103 and the reference relationships between the multiple lower-level documents 104.

[0061] More specifically, for example, the upper document 103 includes a sentence of condition 1. The sentence of condition 1 then includes a sentence of condition 2 and a sentence of condition 3. The sentence of condition 2 then includes requirement 1, and the document of condition 3 includes requirement 2. Based on the above relationships, the implication inference unit 14 generates context data of condition 1, condition 2, condition 3, requirement 1, and requirement 2 shown in FIG. 8 . For example, in this example, each of condition 1, condition 2, condition 3, requirement 1, and requirement 2 corresponds to an information type.

[0062] First, a case will be described in which the implication inference unit 14 extracts the lower-level document 104 corresponding to requirement 1 of the higher-level document 103. Based on the reference relationships among the lower-level documents 104a to 104j, the implication inference unit 14 extracts the lower-level document 104b, which includes element information of condition 1, from among the lower-level documents 104a, 104b, and 104c. The implication inference unit 14 confirms that the element information of condition 1 in the higher-level document 103 and the element information of condition 1 in the lower-level document 104b are in an implication relationship. The implication inference unit 14 confirms that the lower-level document 104b includes element information of condition 2, and that the element information of condition 2 in the higher-level document 103 and the element information of condition 2 in the lower-level document 104b are in an implication relationship. Next, the implication inference unit 14 extracts the lower-level document 104d, which includes element information of condition 2, from the lower-level documents 104d and 104e, which are referenced by the lower-level document 104b. The implication inference unit 14 confirms that the lower-level document 104d includes element information of requirement 1, and that there is an implication relationship between the element information of requirement 1 in the higher-level document 103 and the element information of requirement 1 in the lower-level document 104d. Next, the implication inference unit 14 extracts the lower-level document 104g that includes element information of requirement 1 from the lower-level documents 104f, 104g, and 104h that are referenced by the lower-level document 104d. The implication inference unit 14 confirms that there is an implication relationship between the element information of requirement 1 in the higher-level document 103 and the element information of requirement 1 in the lower-level document 104g.

[0063] Next, a case will be described in which the implication inference unit 14 extracts the lower-level document 104 corresponding to requirement 2 of the higher-level document 103. Based on the reference relationships among the lower-level documents 104a to 104j, the implication inference unit 14 extracts the lower-level document 104b, which includes element information of condition 1, from among the lower-level documents 104a, 104b, and 104c. The implication inference unit 14 confirms that the element information of condition 1 in the higher-level document 103 and the element information of condition 1 in the lower-level document 104b are in an implication relationship. The implication inference unit 14 confirms that the lower-level document 104b includes element information of condition 3, and that the element information of condition 3 in the higher-level document 103 and the element information of condition 3 in the lower-level document 104b are in an implication relationship. Next, the implication inference unit 14 extracts the lower-level document 104e, which includes element information of condition 3, from the lower-level documents 104d and 104e, which are referenced by the lower-level document 104b. The implication inference unit 14 confirms that the lower-level document 104e includes element information of requirement 2, and that there is an implication relationship between the element information of requirement 2 in the upper-level document 103 and the element information of requirement 2 in the lower-level document 104e. Next, the implication inference unit 14 extracts the lower-level document 104j that includes element information of requirement 2 from the lower-level documents 104i and 104j that are referenced by the lower-level document 104e. The implication inference unit 14 confirms that there is an implication relationship between the element information of requirement 2 in the upper-level document 103 and the element information of requirement 2 in the lower-level document 104j.

[0064] (Hardware Configuration) Next, an example of the hardware configuration of the natural language processing apparatus 10 will be described with reference to Fig. 9. Fig. 9 is a block diagram showing an example of the hardware configuration of the natural language processing apparatus 10. Here, an example will be described in which the natural language processing apparatus 10 is configured by a computer 20.

[0065] As shown in FIG. 9, the computer 20 includes a processor 21 , a read-only memory (ROM) 22 , a random access memory (RAM) 23 , an auxiliary storage device 24 , and an input / output interface 25 .

[0066] The processor 21, the ROM 22, the RAM 23, the auxiliary storage device 24, and the input / output interface 25 are electrically connected to one another via a bus 26. The processor 21, the ROM 22, the RAM 23, the auxiliary storage device 24, and the input / output interface 25 transmit and receive data or control signals via the bus 26.

[0067] The processor 21 is configured by a general-purpose hardware processor including, for example, a CPU (Central Processing Unit) and a GPU (Graphical Processing Unit), etc. The processor 21 controls the entire computer 20. That is, the processor 21 controls the ROM 22, the RAM 23, the auxiliary storage device 24, and the input / output interface 25.

[0068] The ROM 22 is a non-volatile memory that constitutes part of the main storage device. The ROM 22 is, for example, an erasable programmable read-only memory (EPROM). The ROM 22 is a non-transitory storage medium that stores firmware, programs, and the like. For example, the processor 21 loads firmware from the ROM 22 into the RAM 23 and executes it.

[0069] The RAM 23 is a volatile memory that constitutes part of the main storage device. The RAM 23 is a dynamic random access memory (DRAM) or a static random access memory (SRAM), etc. The RAM 23 temporarily stores programs used in processing by the processor 21 and data used to execute the programs. The processor 21 executes the programs in the RAM 23 to perform operations on the data in the RAM 23 and store the results of the operations in the RAM 23.

[0070] The auxiliary storage device 24 is configured with non-volatile memory such as a hard disk drive (HDD) or a solid state drive (SSD). The auxiliary storage device 24 non-temporarily stores programs to be executed by the processor 21 and data required for executing the programs. The processor 21 loads the programs and data in the auxiliary storage device 24 into the RAM 23 and executes the programs to perform various functions.

[0071] The input / output interface 25 is connected to an external input device 31, an output device 32, etc., and enables input of information from the input device 31 and output of information to the output device 32. For example, the input / output interface 25 may be a wired interface or a wireless interface. The wired interface includes a port to which a device is connected, etc. The wireless interface includes Bluetooth (registered trademark), Wi-Fi (registered trademark), etc.

[0072] The input device 31 may include a keyboard, a mouse, a touch panel, a receiving device, a disk drive, etc. The input device 31 is not limited to these and may include any other input device. The output device 32 may include a display, a transmitting device, a disk drive, etc. The output device 32 is not limited to these and may include any other output device. The input device 31 and the output device 32 may be configured as an input / output device 33 that has the functions of both the input device 31 and the output device 32.

[0073] The element training data 101 , the entailment training data 102 , the upper document 103 , and the lower document 104 are input to the natural language processing apparatus 10 via the input device 31 .

[0074] The program non-temporarily stored in the auxiliary storage device 24 is provided to the computer 20, for example, via a storage medium 34 on which the program is non-temporarily recorded and which can be read by the computer 20. Such a storage medium 34 is called a non-temporarily computer-readable storage medium. Non-temporarily computer-readable storage media include disks such as flexible disks, optical disks (CD-ROM, CD-R, DVD-ROM, DVD-R, etc.), and magneto-optical disks (MO, etc.), as well as semiconductor memories.

[0075] The programs non-temporarily stored in the auxiliary storage device 24 include a natural language processing program. The natural language processing program is a program that causes the computer 20 to execute at least some of the functions of the components of the natural language processing device 10, namely, the domain knowledge learning unit 11, the entailment learning unit 12, the element information extraction unit 13, and the entailment inference unit 14.

[0076] If the storage medium 34 is a disk, the program non-temporarily stored in the auxiliary storage device 24 is read into the auxiliary storage device 24 via a disk drive, which is the input device 31, and the input / output interface 25, or if the storage medium 34 is a semiconductor memory, via a port, which is the input / output interface 25, and non-temporarily stored therein. Alternatively, the program may be stored in a server on a network, downloaded from the server, and non-temporarily stored in the auxiliary storage device 24.

[0077] When the computer 20 starts up, the processor 21 executes a program in the ROM 22 and loads and starts an OS (Operating System) into the RAM 23. Under control of the OS, the processor 21 monitors input instructions, connections to external devices, and the like. Under control of the OS, the processor 21 also sets up a program area and a data area in the RAM 23. In response to an input instruction to start the natural language processing device 10, the processor 21 loads a natural language processing program from the auxiliary storage device 24 into the program area of ​​the RAM 23, and loads data used in executing the natural language processing program from the auxiliary storage device 24 into the data area of ​​the RAM 23. The processor 21 calculates data in the data area in accordance with the natural language processing program and writes the calculation results to the data area. Through these operations, the processor 21, RAM 23, auxiliary storage device 24, input / output interface 25, and bus 26 work together to execute at least some of the functions of the components of the natural language processing device 10.

[0078] (Effects of the Embodiment) With the configuration according to the embodiment, the natural language processing apparatus 10 can generate, for each information type, an element learning model 201 that learns the characteristics of words before, after, or around a technical term. The natural language processing apparatus 10 can generate, for each information type, an entailment learning model 202 that learns the entailment relationships of words before, after, or around a technical term. The natural language processing apparatus 10 can extract element information for each information type from each of the higher-level document 103 and the lower-level document 104. The natural language processing apparatus 10 can determine, for each information type, the entailment relationships between the element information extracted from the higher-level document 103 and the element information extracted from the lower-level document 104. This enables the natural language processing apparatus 10 to make determinations based on the knowledge structure of the application field when searching the lower-level document 104 for a portion corresponding to a requirement described in the higher-level document 103. In other words, when comparing the higher-level document and the lower-level document, the natural language processing apparatus 10 can eliminate the influence of the order of words in the sentence, writing style, wording, etc. Therefore, the natural language processing apparatus 10 can improve the accuracy of the inference result for inferring the correspondence between the higher-level document 103 and the lower-level document 104 .

[0079] Furthermore, with the configuration according to this embodiment, the natural language processing apparatus 10 can generate combination information of element information included in the higher-level document 103 and combination information of element information included in the lower-level document 104. The natural language processing apparatus 10 compares the combination information of element information included in the higher-level document 103 with element information extracted from the combination information of element information included in the lower-level document 104, and can infer an implication relationship between the higher-level document 103 and the lower-level document 104 based on the comparison result, and generate inference result data 105.

[0080] Furthermore, with the configuration according to this embodiment, the natural language processing apparatus 10 can generate context data from the context of the upper document 103. The entailment inference unit 14 can infer entailment taking into account the context of the upper document 103 and the referential relationships of the lower documents 104 by searching for element information serving as a search key from the lower documents 104 that are in a referential relationship according to the context data.

[0081] The present invention is not limited to the above-described embodiments. Various modifications are possible in the implementation stage without departing from the spirit of the invention. Furthermore, the embodiments may be implemented in appropriate combinations, in which case the combined effects can be obtained. Furthermore, the above-described embodiments include various inventions, and various inventions can be extracted by combining selected elements from the disclosed elements. For example, if the problem can be solved and the desired effect can be obtained even if some elements are deleted from all elements shown in the embodiments, the configuration from which these elements are deleted can be extracted as an invention.

[0082] DESCRIPTION OF SYMBOLS 10...Natural language processing device 11...Domain knowledge learning unit 12...Entailment relationship learning unit 13...Element information extraction unit 14...Entailment relationship inference unit 20...Computer 21...Processor 22...ROM 23...RAM 24...Auxiliary storage device 25...Input / output interface 26...Bus 31...Input device 32...Output device 33...Input / output device 34...Storage medium 101...Element teacher data 101a...Access right element teacher data 101b...Management information category element teacher data 101c...Prohibition / causative element teacher data 101d...Management location element teacher data 101e...Media type element teacher data 102...Entailment relationship teacher data 102a...Access right entailment relationship teacher data 102b...Management information category entailment relationship teacher data 102c...Prohibition / causative entailment relationship teacher data 102d...Management location entailment relationship teacher data 102e...Storage medium entailment relationship teacher data 103...Higher-level document 104, 104a to 104j...Lower-level documents 105...Inference result data 201...Element learning model 201a...Access right element learning model 201b...Management information classification element learning model 201c...Prohibition / causative element learning model 201d...Management location element learning model 201e...Storage medium element learning model 202...Entailment relationship learning model 202a...Access right entailment relationship learning model 202b...Management information classification entailment relationship learning model 202c...Prohibition / causative entailment relationship learning model 202d...Management location entailment relationship learning model 202e...Storage medium entailment relationship learning model

Claims

1. An element information extraction unit that extracts element information from each of a superior document and an inferior document using an element learning model corresponding to the information type for each information type; and an implication relationship inference unit that determines an implication relationship between the element information extracted from the superior document and the element information extracted from the inferior document using an implication relationship learning model corresponding to the information type for each information type, and outputs a result of inferring the implication relationship between the superior document and the inferior document based on the determination result. A natural language processing device comprising:

2. The natural language processing device according to claim 1, further comprising a domain knowledge learning unit that generates an element learning model corresponding to the information type by a machine learning algorithm that takes as input the element teacher data corresponding to the information type for each information type.

3. The natural language processing device according to claim 1, further comprising an implication relationship learning unit that generates an implication relationship learning model corresponding to the information type by a machine learning algorithm that takes as input the implication relationship teacher data corresponding to the information type for each information type.

4. The implication relationship inference unit according to claim 1 outputs a result of inferring the implication relationship between the superior document and the inferior document based on a comparison between combination information of the element information extracted from the superior document and combination information of the element information extracted from the inferior document.

Images