Document retrieval system

Abstract

A document retrieval system considering the concept of document retrieval is provided. The document retrieval system (100) includes an input unit (101), a first processing unit (102), a storage unit (105), a second processing unit (103), and an output unit (104). The input unit (101) has the function of inputting a first document (20), the first processing unit (102) has the function of forming a first graph structure (21) using the first document (20), the storage unit (105) has the function of storing a second graph structure (11), the second processing unit (103) has the function of calculating the similarity between the first graph structure (21) and the second graph structure (11), and the output unit (104) has the function of providing information. The first processing unit (102) has the function of dividing the first document (20) into multiple symbols. The nodes and edges of the first graph structure (21) have labels, and the labels are composed of multiple symbols.

Description

Technical Field

[0001] One aspect of the present invention relates to a document retrieval system. Furthermore, another aspect of the present invention relates to a document retrieval method. Background Technology

[0002] Various document retrieval technologies have been proposed. Among existing document retrieval methods, word (string) retrieval is the primary method. For example, web pages utilize page ranking algorithms, while the patent field utilizes category dictionaries. Furthermore, there are methods that obtain a set of words and use Jaccard coefficients, Dice coefficients, Simpson coefficients, etc., to represent document similarity. Additionally, there are methods that use TF-IDF, Bag of Words (BoW), Doc2Vec, etc., to vectorize documents and compare cosine similarity. Moreover, there are methods that evaluate the similarity of text strings using Hamming distance, Levenstein distance, Jaro-Winkler distance, etc., to retrieve desired documents. Furthermore, Patent Document 1 discloses a language processing device that calculates the distance between string structures by converting the constituent units of a text into string structures, and compares them from the perspective of text similarity.

[0003] [Preliminary Technology Documents]

[0004] [Patent Literature]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 2005-258624 Summary of the Invention

[0006] The technical problem that the invention aims to solve

[0007] To retrieve documents across various fields, more accurate document retrieval methods are needed. For example, patent documents (specifications, patent application scopes, etc.) or contracts often use similar words. Therefore, it is important not only to consider the words used in the document, but also to employ retrieval techniques that take into account the document's concepts.

[0008] Therefore, one objective of this invention is to provide a document retrieval system that takes into account the concept of a document. Furthermore, one objective of this invention is to provide a document retrieval method that takes into account the concept of a document.

[0009] Note that the description of these objectives does not preclude the existence of other objectives. One aspect of the invention does not require achieving all of the above objectives. Objectives other than those described above can be readily understood and extracted from the description, drawings, claims, etc.

[0010] means of solving technical problems

[0011] One aspect of the present invention is a document retrieval system comprising an input unit, a first processing unit, a storage unit, a second processing unit, and an output unit. The input unit has the function of inputting a first document; the first processing unit has the function of forming a first graph structure using the first document; the storage unit has the function of storing a second graph structure; the second processing unit has the function of calculating the similarity between the first graph structure and the second graph structure; the output unit has the function of providing information; and the first processing unit has the function of segmenting the first document into multiple tokens. The nodes and edges of the first graph structure have labels, and the labels are composed of multiple tokens.

[0012] In the aforementioned document retrieval system, preferably, the first processing unit has the function of assigning word classes to symbols.

[0013] Furthermore, in the aforementioned document retrieval system, it is preferable that the first processing unit has the function of performing dependency analysis, and that the first processing unit has the function of linking partial tokens based on the results of dependency analysis.

[0014] Furthermore, in the aforementioned document retrieval system, it is preferable that the first processing unit has the function of replacing symbols including representative words or hypernyms with representative words or hypernyms.

[0015] Furthermore, in the aforementioned document retrieval system, it is preferable that the first processing unit uses the second document to form a second graph structure.

[0016] Furthermore, in the aforementioned document retrieval system, it is preferable that, when the labels of the edges of the graph structure have antonyms, the first processing unit has the function of generating a new graph structure by reversing the direction of the edges of the graph structure and replacing the labels of the edges with antonyms.

[0017] Furthermore, in the aforementioned document retrieval system, it is preferable that the second processing unit has the following function: by vectorizing the first graph structure and the second graph structure, it evaluates the vector similarity between the vectorized first graph structure and the vectorized second graph structure.

[0018] Furthermore, in the aforementioned document retrieval system, it is preferable that the second processing unit has the function of vectorizing the first graph structure and the second graph structure using the Weisfeiler-Lehman kernel.

[0019] Furthermore, in the aforementioned document retrieval system, it is preferable that the first processing unit has the function of linking the second symbol and the first symbol when the part of speech assigned to the first symbol is a noun and the part of speech assigned to the second symbol located directly before the first symbol is an adjective.

[0020] Furthermore, in the aforementioned document retrieval system, it is preferable that the first processing unit has the function of linking the third and fourth tokens when both the word class assigned to the third token and the word class assigned to the fourth token located immediately after the third token are nouns.

[0021] Invention Effects

[0022] According to one aspect of the present invention, a document retrieval system that takes into account the concept of a document can be provided. Furthermore, according to one aspect of the present invention, a document retrieval method that takes into account the concept of a document can be provided.

[0023] By analyzing each sentence of a document to obtain its conceptual graph structure and calculating the similarity of these graph structures, documents that are conceptually similar can be retrieved. Furthermore, by combining this method with existing retrieval methods, the accuracy of ranking and other aspects can be improved.

[0024] Note that the effects of one aspect of the present invention are not limited to those described above. The described effects do not preclude the existence of other effects. Furthermore, other effects refer to effects other than those described above, which will be described in the following description. Those skilled in the art can derive and appropriately extract effects other than those described above from the description, drawings, etc. Moreover, one aspect of the present invention has at least one of the described effects and / or other effects. Therefore, one aspect of the present invention may sometimes lack the described effects, depending on the circumstances. Attached Figure Description

[0025] Figure 1 This is a diagram illustrating an example of a document retrieval system.

[0026] Figure 2 This is a flowchart illustrating an example of a document retrieval method.

[0027] Figures 3A to 3C This is a graph showing the results obtained in each step.

[0028] Figures 4A to 4C This is a graph showing the results obtained in each step.

[0029] Figures 5A to 5D This is a graph showing the results obtained in each step.

[0030] Figures 6A to 6C This is a graph showing the results obtained in each step.

[0031] Figure 7 This is a diagram illustrating an example of the hardware of a document retrieval system.

[0032] Figure 8 This is a diagram illustrating an example of the hardware of a document retrieval system. Detailed Implementation

[0033] The embodiments will be described in detail with reference to the accompanying drawings. Note that the present invention is not limited to the following description, and those skilled in the art will readily understand that its methods and details can be varied in many ways without departing from the spirit and scope of the invention. Therefore, the present invention should not be construed as being limited only to the contents described in the embodiments shown below.

[0034] Furthermore, in the structure of the invention described below, the same symbols are used in different figures to represent the same parts or parts having the same function, and repeated descriptions are omitted. Additionally, when representing parts having the same function, the same shading lines are sometimes used without additional symbols.

[0035] Furthermore, for ease of understanding, the positions, sizes, and extents of the constituent elements shown in the accompanying drawings may not always reflect their actual positions, sizes, and extents. Therefore, the disclosed invention is not necessarily limited to the positions, sizes, and extents shown in the accompanying drawings.

[0036] Furthermore, the ordinal numbers such as "first," "second," and "third" used in this specification are for the convenience of identifying the constituent elements, and are not intended to limit the quantity.

[0037] (Implementation Method 1)

[0038] In this embodiment, refer to Figures 1 to 4C This invention describes a document retrieval system and method.

[0039] Document Retrieval System

[0040] Figure 1 This is a diagram illustrating the structure of the document retrieval system 100. That is to say, Figure 1 An example of the structure of a document retrieval system as one aspect of the present invention is shown.

[0041] The document retrieval system 100 can be installed in an information processing device such as a personal computer used by the user. Alternatively, a processing unit of the document retrieval system 100 can be installed in a server so that the document retrieval system 100 can be used by accessing the processing unit from a client PC via a network.

[0042] like Figure 1 As shown, the document retrieval system 100 includes an input unit 101, a graph structure forming unit 102, a similarity calculation unit 103, an output unit 104, and a storage unit 105. Furthermore, the aforementioned processing unit includes the graph structure forming unit 102 and the similarity calculation unit 103.

[0043] Input unit 101 inputs document 20. Document 20 is a document specified by the user for retrieval. Document 20 can be text data, audio data, or image data. Input unit 101 can be an input device such as a keyboard, mouse, touch sensor, microphone, scanner, or camera.

[0044] The document retrieval system 100 may also have the function of converting audio data into text data. For example, the graph structure forming unit 102 may have this function. In addition, the document retrieval system 100 may also include an audio-to-text conversion unit with this function.

[0045] The document retrieval system 100 may also have optical character recognition (OCR) functionality. This allows it to recognize the text contained in image data to generate text data. For example, the image structure forming unit 102 may have this functionality. Furthermore, the document retrieval system 100 may also include a character recognition unit with this functionality.

[0046] Storage unit 105 stores documents 10_1 to 10_n (n is an integer of 2 or more). Documents 10_1 to 10_n are comparison documents of document 20. Hereinafter, documents 10_1 to 10_n are sometimes collectively referred to as multiple documents 10. Multiple documents 10 are stored in storage unit 105 via input unit 101, storage medium, communication, etc.

[0047] The multiple documents 10 stored in the storage unit 105 are preferably text data. For example, by converting audio data or image data into text data, the data size can be reduced, thereby reducing the load on the storage unit 105.

[0048] Furthermore, the storage unit 105 stores graph structures 11_1 to 11_n. Graph structures 11_1 to 11_n correspond to documents 10_1 to 10_n, respectively. Additionally, graph structures 11_1 to 11_n are formed in the graph structure forming unit 102 using documents 10_1 to 10_n, respectively. Hereinafter, graph structures 11_1 to 11_n are sometimes collectively referred to as a plurality of graph structures 11.

[0049] It is preferable to assign the same ID to document 10_i (where i is an integer greater than or equal to 1 and less than or equal to n) and graph structure 11_i. This allows document 10_i and graph structure 11_i to be associated. By pre-forming graph structures 11_1 to 11_n, the time required to retrieve documents can be shortened.

[0050] Furthermore, the storage unit 105 can store the document 20. Additionally, it can also store the graph structure 21 corresponding to the document 20. Furthermore, the graph structure 21 is formed in the graph structure forming unit 102 using the document 20.

[0051] The graph structure forming unit 102 has the function of forming a graph structure using a document. Therefore, the graph structure forming unit 102 preferably has the functions of performing morpheme analysis, dependency analysis, abstraction, and forming a graph structure. In addition, the graph structure forming unit 102 has the function of referring to the concept dictionary 112. By referring to the concept dictionary 112, a graph structure corresponding to the document is formed in the graph structure forming unit 102. The document refers to document 20 and multiple documents 10.

[0052] The preferred graph structure is a directed graph. A directed graph is a graph consisting of nodes and directed edges. Furthermore, a more preferred graph structure is a directed graph in which nodes and edges are labeled. By using a labeled directed graph structure, similarity and retrieval accuracy can be improved.

[0053] In addition, although Figure 1 The diagram shows the structure of the concept dictionary 112 in a device different from the document retrieval system 100, but it is not limited thereto. The concept dictionary 112 may also be included in the document retrieval system 100.

[0054] Furthermore, the functions for morpheme analysis and dependency analysis can also be located in a device different from the document retrieval system 100. In this case, the document retrieval system 100 sends the aforementioned document to the device and receives the data from the results of the morpheme analysis and dependency analysis performed by the device, and then sends the received data to the graph structure forming unit 102.

[0055] The similarity calculation unit 103 has the function of calculating the similarity between the first graph structure and the second graph structure. The first graph structure is graph structure 21. The second graph structure is one or more of a plurality of graph structures 11. That is, the similarity calculation unit 103 evaluates the similarity between the first document and the second document. The first document is document 20. The second document is one or more of a plurality of documents 10.

[0056] The output unit 104 has the function of supplying information. This information is related to the similarity results calculated by the similarity calculation unit 103. For example, this information refers to the document with the highest similarity to document 20 among multiple documents 10. In addition, this information refers to the result of rearranging the groups composed of the similarity between document 10_i and document 20 and document 10_i according to the similarity level. At this time, the number of groups is 2 or more and n or less.

[0057] The aforementioned information is supplied as visual or auditory information, such as strings, numbers, or charts. The output unit 104 is an output device such as a display or speaker.

[0058] The document retrieval system 100 may also have the function of converting text data into audio data. For example, the document retrieval system 100 may also include a text-to-audio conversion unit with this function.

[0059] The structure of the document retrieval system 100 has been described above. By using the document retrieval system, which is one aspect of the present invention, documents conceptually similar to document 20 can be retrieved from a plurality of documents 10. Furthermore, a list of documents conceptually similar to document 20 can also be retrieved from the plurality of documents 10.

[0060] According to one aspect of the present invention, a document retrieval system that takes into account the concept of a document can be provided.

[0061] <Document Retrieval Methods>

[0062] Figure 2 This is a flowchart illustrating the processing flow performed by the document retrieval system 100. In other words, Figure 2 This is a flowchart illustrating an example of a document retrieval method as one aspect of the present invention.

[0063] In one aspect of the document retrieval method of the present invention, documents are retrieved by comparing the similarity of the graph structures using methods such as the Weisfeiler-Lehman (WL) kernel after analyzing the documents to form a graph structure.

[0064] In step S001, multiple documents 10 are acquired. The multiple documents 10 are documents stored in the storage unit 105. The multiple documents 10 are stored in the storage unit 105 via the input unit 101, storage medium, communication, etc.

[0065] When multiple documents 10 constitute the claims of a patent application, document cleaning can be performed on each document 10 separately before proceeding to step S002. Document cleaning refers to, for example, deleting semicolons or replacing colons with commas. By performing document cleaning, the accuracy of morpheme analysis can be improved.

[0066] Furthermore, if multiple documents 10 are not within the scope (claims) of a patent application, the aforementioned document cleanup can be performed as needed. Additionally, the multiple documents 10 can be stored in the storage unit 105 after the aforementioned document cleanup.

[0067] In step S002, morpheme analysis is performed on each of the multiple documents 10 in the graph structure forming unit 102. As a result, each of the multiple documents 10 is segmented into morphemes. In this specification, the segmented morphemes are sometimes referred to as symbols.

[0068] In step S002, it is preferable to determine the word class of each of the segmented morphemes (markers) to associate the word class labels. By associating morphemes (markers) with word class labels, the accuracy of dependency analysis can be improved. Furthermore, in this specification, "associating morphemes (markers) with word class labels" can be replaced with "assigning word classes to morphemes (markers)".

[0069] Even if the graph structure forming unit 102 does not have the function of performing morpheme analysis, a morpheme analysis program (also called a morpheme analyzer) assembled in a device different from the document retrieval system can be used to perform morpheme analysis on multiple documents 10 separately. In this case, in step S002, multiple documents 10 are sent to the device, where morpheme analysis is performed and the data of the morpheme analysis results are received.

[0070] In step S003, dependency analysis is performed in the graph structure forming unit 102. That is, parts of multiple symbols are combined according to the dependency relationships of the segmented morphemes (symbols). For example, if symbols satisfy certain conditions, symbols that satisfy the conditions are combined with each other to generate new symbols.

[0071] When Japanese is used in a document, specifically, if the j-th (j is an integer of 2 or higher) symbol is a noun and the symbol immediately preceding the j-th symbol (called the (j-1)-th symbol) is an adjective, a new symbol is generated by combining the (j-1)-th symbol and the j-th symbol. Furthermore, if the j-th symbol is a noun and the symbol immediately following the j-th symbol (called the (j+1)-th symbol) is a noun, a new symbol is generated by combining the j-th symbol and the (j+1)-th symbol.

[0072] Note that the above conditions should preferably be set appropriately according to the language used in the document.

[0073] The aforementioned dependency analysis preferably includes compound word analysis. By performing this dependency analysis, a compound word is generated as a new token by combining a portion of multiple tokens. Thus, even if the document contains compound words not registered in the concept dictionary 112, the document can be segmented into tokens with high accuracy.

[0074] If the graph structure forming unit 102 does not have the function of performing dependency analysis, dependency analysis can also be performed using a dependency analysis program (also called a dependency analyzer) assembled in a device different from the document retrieval system. In this case, in step S003, the segmented morphemes (symbols) are sent to the device, dependency analysis is performed in the device, and the data of the dependency analysis results are received.

[0075] In step S004, the symbol is abstracted in the graph structure forming unit 102. For example, a representative word is obtained by analyzing the words contained in the symbol. Furthermore, if the representative word contains a superordinate word, the superordinate word is obtained. Then, the symbol is replaced with the obtained representative word or superordinate word. Here, the representative word refers to the headword (also called a term) in a synonym group. Furthermore, the superordinate word refers to a representative word that is equivalent to the superordinate concept of the representative word. That is, abstracting the symbol means replacing the symbol with a representative word or a superordinate word. Furthermore, if the symbol is a representative word or a superordinate word, it is not necessary to replace the symbol.

[0076] The upper limit of the superordinate term level to be replaced is preferably 1 or higher and 2 or lower, more preferably 1. Alternatively, the upper limit of the superordinate term level to be replaced can be specified. This helps to prevent the token from being over-converted into a superordinate concept.

[0077] The appropriate level of abstraction for notations varies depending on the domain. Therefore, it is preferable to abstract the notations by performing domain-specific machine learning. For example, notations can be abstracted by vectorizing the morphemes they contain and classifying them using a classifier. Algorithms such as decision trees, support vector machines, random forests, and multilayer perceptrons can be used as this classifier. Specifically, "oxide semiconductor," "amorphous semiconductor," "silicon semiconductor," and "GaAs semiconductor" can be classified as "semiconductor." Furthermore, "oxide semiconductor layer," "oxide semiconductor film," "amorphous semiconductor layer," "amorphous semiconductor film," "silicon semiconductor layer," "silicon semiconductor film," "GaAs semiconductor layer," and "GaAs semiconductor film" can also be classified as "semiconductor."

[0078] Alternatively, a classifier can be used to determine whether to extract the morphemes contained in the symbol. For example, when abstracting the symbol "oxide semiconductor layer," the symbol is further decomposed into morphemes, and the decomposed morphemes "oxide," "object," "semiconductor," and "layer" are input into a classifier. As a result of the input to the classifier, if it is classified as "semiconductor," the symbol is replaced with "semiconductor." This allows for the abstraction of the symbol.

[0079] In addition to the machine learning algorithms mentioned above, Conditional Random Fields (CRFs) can also be used. Furthermore, CRFs can be combined with the methods described above.

[0080] By abstracting the symbols, a document can be grasped conceptually. Therefore, a document can be retrieved based on conceptual elements, rather than being easily influenced by the document's structure or presentation.

[0081] To obtain representative terms and hypernyms, a concept dictionary can be used, or classification can be performed through machine learning. This concept dictionary can be either a concept dictionary 112 located in a device different from the document retrieval system 100, or a concept dictionary possessed by the document retrieval system 100.

[0082] In step S005, multiple graph structures 11 are formed in the graph structure forming unit 102. That is, the symbols prepared in step S004 are used as nodes or edges to form a graph structure. Specifically, if the document includes a first symbol and a second symbol as nouns, and includes a third symbol indicating the relationship between the first and second symbols, the first and second symbols are each used as nodes and their labels, and the third symbol is used as edges and their labels, thereby forming a graph structure. In other words, the node labels and edge labels are composed of the symbols prepared in step S004.

[0083] For example, when the document is the scope (claims) of a patent application, the nodes of the graph structure represent the constituent elements, and the edges of the graph structure represent the relationships between the constituent elements. Furthermore, when the document is a contract, etc., the nodes of the graph structure represent Party A and Party B, and the edges of the graph structure represent detailed conditions.

[0084] Graph structures can also be formed according to rules based on the dependencies between tokens. Furthermore, CRF can be used for machine learning to label nodes and edges based on a token list. Thus, nodes and edges can be labeled based on the token list. Additionally, Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), and similar technologies can be used to learn Seq2Seq models that output the directions of nodes and edges from an input token list. Thus, the directions of nodes and edges can be output based on the token list.

[0085] The graph structure forming unit 102 may also have the function of reversing the direction of an edge and replacing the label of that edge with the antonym of that edge's label. For example, if the graph structure includes a first edge and a second edge with a label that is given as the antonym of the label of the first edge, a new graph structure can be formed by performing the process of reversing the direction of the second edge and replacing the label of the second edge with the antonym of the label of the second edge (i.e., the label of the first edge). Thus, structures that are conceptually the same can be included. Therefore, documents can be retrieved based on conceptual elements, and are less susceptible to the influence of the document's structure or presentation.

[0086] Furthermore, it is preferable to perform the above processing on the edge that appears less frequently in the document. That is, when the frequency of the second edge is lower than or equal to the frequency of the first edge, it is preferable to reverse the direction of the second edge and replace the label of the second edge with the antonym of the label of the second edge (i.e., the label of the first edge). This can reduce the frequency of forming new graph structures.

[0087] The order of steps S004 and S005 can also be replaced. When the order of steps S004 and S005 is replaced, the nodes and edges contained in the graph structure are abstracted after the graph structure is formed. Therefore, even when the order of steps S004 and S005 is replaced, an abstracted graph structure can still be formed using documentation.

[0088] By performing steps S001 to S005, multiple graph structures 11 can be formed using multiple documents 10. Furthermore, it is preferable to perform steps S001 to S005 before calculating the similarity. By pre-forming multiple graph structures 11, the time required for document retrieval can be reduced.

[0089] In step S011, document 20 is obtained. Document 20 is the document input into input unit 101. Furthermore, if document 20 is audio data or image data, i.e., data other than text data, document 20 is converted into text data before proceeding to step S012. To convert document 20 into text data, it is preferable to utilize the function of converting audio data into text data provided by the graph structure forming unit 102, the audio-to-text conversion unit, or the optical character recognition (OCR) function or character recognition unit provided by the graph structure forming unit 102.

[0090] If document 20 falls within the scope (claims) of a patent application, the aforementioned document cleaning can be performed on document 20 before proceeding to step S012. Document cleaning improves the accuracy of morpheme analysis. Furthermore, if document 20 is not within the scope (claims) of a patent application, the aforementioned document cleaning can be performed as needed.

[0091] In step S012, morpheme analysis is performed on document 20 in the graph structure forming unit 102. Since step S012 is the same as step S002, the explanation of step S002 can be referred to.

[0092] In step S013, dependency analysis is performed in the diagram structure forming section 102. Since step S013 is the same as step S003, the explanation of step S003 can be referred to.

[0093] In step S014, the symbols are abstracted in the diagram structure forming section 102. Since step S014 is the same as step S004, the explanation of step S004 can be referred to.

[0094] In step S015, a graph structure 21 is formed in the graph structure forming section 102. Since step S015 is the same as step S005, the explanation of step S005 can be referred to.

[0095] In step S016, the similarity calculation unit 103 evaluates the similarity between document 20 and each of the plurality of documents 10. Specifically, by using the WL kernel to vectorize graph structure 21 and the plurality of graph structures 11, the vector similarity between the vectorized graph structure 21 and each of the vectorized graph structures 11 is evaluated.

[0096] In step S017, information is output to the output unit 104. This information refers to information about the similarity result calculated in the similarity calculation unit 103.

[0097] The above describes the document retrieval method. By using the document retrieval method as an embodiment of the present invention, documents conceptually similar to the documents specified for retrieval can be retrieved. Furthermore, documents conceptually similar to the documents specified for retrieval can be retrieved in a sorted state. Moreover, documents can be retrieved based on conceptual elements, making them less susceptible to the influence of document structure or presentation.

[0098] According to one aspect of the present invention, a document retrieval method that takes into account the concept of a document can be provided.

[0099] <<Examples of using documents to form graph structures>>

[0100] Reference Figures 3A to 6C This section describes an example of using a document-forming graph structure in the document retrieval method described above.

[0101] First, refer to the document written in Japanese, namely, "Oxide semiconductor layer on top of insulating layer (SANKABUTSUHANDOUTAISOU HA ZETSUENTAISOU NO JOUHOU NI ARU)" (see reference). Figure 3A Let's take an example to illustrate. Furthermore, Figure 3B , Figure 3C and Figure 4A The rounded quadrilateral shown is a symbol, and the word class assigned to the symbol is attached below the rounded quadrilateral.

[0102] First, by performing morpheme analysis on the above documents, the documents are segmented into tokens, and each token is assigned a word class ( Figure 2The steps S002 and S012 shown are illustrated. The result is that... Figure 3B The results are shown. Specifically, the above document was segmented into symbols, namely "Oxidation (SANKA)" (noun) | "Matter (BUTSU)" (noun) | "Semiconductor (HANDOUTAI)" (noun) | "Layer (SOU)" (noun) | "(HA)" (particle) | "Insulation (ZETSUEN)" (noun) | "Body (TAI)" (noun) | "Layer (SOU)" (noun) | "(NO)" (particle) | "Above (JOUHOU)" (noun) | "(NI)" (particle) | "In (ARU)" (verb), and each symbol was assigned a word class.

[0103] Next, dependency analysis is performed. Figure 2 The steps S003 and S013 shown are illustrated. The result is that... Figure 3C The results are shown. Specifically, "Oxide (SANKA)" and "Material (BUTSU)," "Material (BUTSU)" and "Semiconductor (HANDOUTAI)," and "Semiconductor (HANDOUTAI)" and "Layer (SOU)" all satisfy the conditions mentioned in step S003 above. Therefore, the four symbols ("Oxide (SANKA)," "Material (BUTSU)," "Semiconductor (HANDOUTAI)," and "Layer (SOU)") are combined and replaced by one symbol ("Oxide Semiconductor Layer (SANKABUTSUHANDOUTAISOU)"). Furthermore, "Insulator (ZETSUEN)" and "Bulk (TAI)," and "Bulk (TAI)" and "Layer (SOU)" all satisfy the conditions mentioned in step S003 above. Therefore, the three symbols ("Insulator (ZETSUEN)," "Bulk (TAI)," and "Layer (SOU)") are combined and replaced by one symbol ("Insulator Layer (ZETSUENTAISOU)"). Thus, the above document becomes “Oxide semiconductor layer (SANKABUTSUHANDOUTAISOU)” (noun) | “(HA)” (particle) | “Insulator layer (ZETSUENTAISOU)” (noun) | “(NO)” (particle) | “Above (JOUHOU)” (noun) | “(NI)” (particle) | “In (ARU)” (verb)

[0104] Next, abstract the notation ( Figure 2 The steps S004 and S014 are shown. The result is that... Figure 4AThe results are shown below. Specifically, "oxide semiconductor layer (SANKABUTSUHANDOUTAISOU)" is replaced with the superordinate term "semiconductor (HANDOUTAI)". Furthermore, "insulator layer (ZETSUENTAISOU)" is replaced with the superordinate term "insulator (ZETSUENTAI)". Additionally, "above (JOUHOU)" is replaced with the representative term "on (UE)". Thus, the above document is abstracted into "semiconductor (HANDOUTAI)" (noun) | "(HA)" (particle) | "insulator (ZETSUENTAI)" (noun) | "(NO)" (particle) | "on (UE)" (noun) | "(NI)" (particle) | "at (ARU)" (verb)".

[0105] Next, a graph structure is formed ( Figure 2 The steps S005 and S015 are shown. The result is that... Figure 4B The results are shown. Specifically, "Semiconductor (HANDOUTAI)" and "Insulator (ZETSUENTAI)" become nodes of the graph structure and their labels, and "On (UE)" becomes an edge of the graph structure and its label.

[0106] Here, the antonym of "up (UE)" is "down (SHITA)". Therefore, it is also possible to... Figure 4B The arrows in the diagram structure shown are reversed, and... Figure 4B The edges of the graph structure shown, along with their labels, are replaced with "Down (SHITA)" to reshape the graph. Figure 4C The diagram structure shown can be used to encompass conceptually similar structures.

[0107] Figure 4B , Figure 4C The arrows shown indicate the following direction: from the first node appearing in the document (equivalent to "semiconductor (HANDOUTAI)" in the above document) to the last node appearing (equivalent to "insulator (ZETSUENTAI)" in the above document). That is, the arrow starts at the first node and ends at the last node. Note that this implementation is not limited to this. For example, the direction of the arrows can also be determined based on the semantic relationships between words, such as positional relationships. Specifically, a graph structure can also be formed where the starting point of the arrow is the node labeled "insulator (ZETSUENTAI)," the ending point is the node labeled "semiconductor (HANDOUTAI)," and the edges between these nodes and their labels are "up (UE)." This allows for an intuitive understanding of the graph structure. However, the method for determining the direction of the arrows needs to be standardized across document retrieval methods.

[0108] As mentioned above, the above documents can be used to form an abstract graph structure.

[0109] Next, the document uses English, namely, "A semiconductor device comprising: an oxide semiconductor layer over an insulator layer." (See reference...) Figure 5A Let's take an example to illustrate. Furthermore, Figure 5C , Figure 5D and Figure 6A The rounded quadrilaterals shown represent symbols. Here, examples are shown where symbols are not assigned a word class, but symbols can also be assigned a word class.

[0110] First, clean up the document. Here, remove the semicolons. The result is: Figure 5B The results are shown.

[0111] Next, by performing morpheme analysis on the above document, the document is segmented into tokens ( Figure 2 The steps S002 and S012 shown are illustrated. The result is that... Figure 5C The results are shown. Specifically, the above document becomes “A” | “semiconductor” | “device” | “comprising” | “an” | “oxide” | “semiconductor” | “layer” | “over” | “an” | “insulator” | “layer”.

[0112] Next, dependency analysis is performed. Figure 2 The steps S003 and S013 shown are illustrated. The result is that... Figure 5DThe results are shown. Specifically, the three symbols ("A", "semiconductor", "device") are combined and replaced with one symbol ("A semiconductor device"). Furthermore, the four symbols ("an", "oxide", "semiconductor", "layer") are combined and replaced with one symbol ("an oxide semiconductor layer"). Additionally, the three symbols ("an", "insulator", "layer") are combined and replaced with one symbol ("an insulator layer"). Thus, the above document becomes "A semiconductor device" | "comprising" | "an oxide semiconductor layer" | "over" | "an insulator layer".

[0113] Next, abstract the notation ( Figure 2 The steps S004 and S014 are shown. The result is that... Figure 6A The results are shown below. Specifically, "A semiconductor device" is replaced with the superordinate term "device". Furthermore, "anoxide semiconductor layer" is replaced with the superordinate term "a semiconductor". Additionally, "aninsulator layer" is replaced with the superordinate term "an insulator". Thus, the above document is abstracted into "device" | "comprising" | "a semiconductor" | "over" | "an insulator".

[0114] Next, a graph structure is formed ( Figure 2 The steps S005 and S015 are shown. The result is that... Figure 6B The results are shown. Specifically, "device", "semiconductor" and "insulator" become nodes of the graph structure and their labels, respectively, and "comprising" and "over" become edges of the graph structure and their labels, respectively.

[0115] Here, the antonym of "over" is "under". Therefore, it can also be made... Figure 6B The arrows in the diagram structure shown are reversed, and... Figure 6B The edges of the graph structure shown, along with their labels, are replaced with "under" to reshape the graph. Figure 6CThe diagram structure shown can be used to encompass conceptually similar structures.

[0116] Figure 6B , Figure 6C The arrows shown indicate the following direction: from the first node appearing in the document (equivalent to "semiconductor" in the above document) to the second node appearing (equivalent to "insulator" in the above document). That is, the arrow starts at the first node and ends at the second node. Note that this implementation is not limited to this. For example, the direction of the arrows can also be determined based on the semantic relationships between words, such as positional relationships. Specifically, a graph structure can also be formed where the arrows start at the node labeled "insulator" and end at the node labeled "semiconductor," with the edges between these nodes and their labels being "over." This allows for an intuitive understanding of the graph structure. However, the method for determining the direction of the arrows needs to be standardized across document retrieval methods.

[0117] As mentioned above, the above documents can be used to form an abstract graph structure.

[0118] Furthermore, although the steps for creating a graph structure using documents in Japanese and English are used as examples, the language of the documents is not limited to Japanese and English. The same steps can also be used to create graph structures using documents in languages ​​such as Chinese, Korean, German, French, Russian, or Amagi script.

[0119] This embodiment can be appropriately combined with other embodiments. Furthermore, in this specification, where multiple structural examples are shown in one embodiment, these structural examples can be appropriately combined.

[0120] (Implementation Method 2)

[0121] In this embodiment, for a document retrieval system according to one aspect of the present invention, refer to... Figure 7 and Figure 8 Please provide an explanation.

[0122] The document retrieval system of this embodiment can easily retrieve the required information from documents using the document retrieval method shown in Embodiment 1.

[0123] <Example 1 of the structure of a document retrieval system>

[0124] Figure 7A block diagram of the document retrieval system 200 is shown. Note that the accompanying drawings in this specification show the components categorized according to their function in separate boxes; however, the actual components are difficult to clearly distinguish according to function, and a component may sometimes have multiple functions. Furthermore, a function may sometimes involve multiple components; for example, the various processes performed in the processing unit 202 may sometimes be performed on different servers.

[0125] The document retrieval system 200 includes at least a processing unit 202. Figure 7 The document retrieval system 200 shown also includes an input unit 201, a storage unit 203, a database 204, a display unit 205, and a transmission channel 206.

[0126] [Input Section 201]

[0127] The input unit 201 receives documents from outside the document retrieval system 200. These documents are specified by the user for retrieval, corresponding to document 20 shown in Embodiment 1. Furthermore, the input unit 201 may also receive multiple documents from outside the document retrieval system 200. These multiple documents are comparison documents used as the aforementioned documents, corresponding to multiple documents 10 shown in Embodiment 1. Both the multiple documents supplied to the input unit 201 and the aforementioned documents are supplied to the processing unit 202, storage unit 203, or database 204 via the transmission channel 206.

[0128] The aforementioned documents may be input as text data, audio data, or image data, for example. Preferably, the aforementioned documents may be input as text data.

[0129] Examples of input methods for the aforementioned documents include keyboard input using a keyboard or touch panel, audio input using a microphone, reading from a recording medium, image input using a scanner or camera, and acquisition using communication.

[0130] The document retrieval system 200 may also have the function of converting audio data into text data. For example, the processing unit 202 may also have this function. Alternatively, the document retrieval system 200 may also include an audio conversion unit with this function.

[0131] The document retrieval system 200 may also have optical character recognition (OCR) functionality. This allows it to recognize text within image data and generate text data. For example, the processing unit 202 may also have this functionality. Alternatively, the document retrieval system 200 may also include an OCR unit with this functionality.

[0132] [Processing Department 202]

[0133] The processing unit 202 has the function of performing calculations using data supplied from the input unit 201, storage unit 203, database 204, etc. The processing unit 202 can supply the calculation results to the storage unit 203, database 204, display unit 205, etc.

[0134] The processing unit 202 includes the graph structure forming unit 102 and the similarity calculation unit 103 shown in Embodiment 1. That is, the processing unit 202 has the functions of performing morpheme analysis, performing dependency analysis, performing abstraction, and forming graph structures.

[0135] The processing unit 202 may also use a transistor containing metal oxide in the channel formation region. Since the off-state current of this transistor is extremely small, by using it as a switch to maintain the charge (data) flowing into the capacitor used as a storage element, long-term data retention can be ensured. By applying this characteristic to at least one of the registers and cache memories included in the processing unit 202, the processing unit 202 can be operated only when necessary, while previously processed information is stored in the storage element in other situations, thus turning the processing unit 202 off. In other words, normally-off computing can be implemented, thereby achieving low power consumption in the document retrieval system 200.

[0136] Furthermore, in this specification and the like, transistors using oxide semiconductors in the channel formation region are referred to as Oxide Semiconductor transistors or OS transistors. The channel formation region of an OS transistor preferably comprises a metal oxide.

[0137] The metal oxide in the channel forming region preferably contains indium (In). When the metal oxide in the channel forming region contains indium, the carrier mobility (electron mobility) of the OS transistor is improved. Furthermore, the metal oxide in the channel forming region preferably contains element M. Element M is preferably aluminum (Al), gallium (Ga), or tin (Sn). Other elements that can be used as element M include boron (B), titanium (Ti), iron (Fe), nickel (Ni), germanium (Ge), yttrium (Y), zirconium (Zr), molybdenum (Mo), lanthanum (La), cerium (Ce), neodymium (Nd), hafnium (Hf), tantalum (Ta), and tungsten (W). Note that multiple of the above elements can sometimes be combined as element M. Element M is, for example, an element with a high bond energy with oxygen. Element M is, for example, an element with a higher bond energy with oxygen than indium. Furthermore, the metal oxide in the channel forming region preferably contains zinc (Zn). Metal oxides containing zinc are sometimes prone to crystallization.

[0138] The metal oxide in the channel forming region is not limited to indium-containing metal oxides. For example, the metal oxide in the channel forming region may also be zinc-tin oxide or gallium-tin oxide, which do not contain indium but contain zinc, gallium-containing metal oxides, or tin-containing metal oxides.

[0139] Alternatively, a transistor containing silicon in the channel forming region can be used in the processing unit 202.

[0140] Furthermore, as the processing unit 202, it is preferable to use a transistor that contains an oxide semiconductor in the channel formation region and a transistor that contains silicon in the channel formation region in combination.

[0141] The processing unit 202 may include, for example, arithmetic circuitry or a central processing unit (CPU).

[0142] The processing unit 202 may also include microprocessors such as DSPs (Digital Signal Processors) and GPUs (Graphics Processing Units). Microprocessors may also be implemented using PLDs (Programmable Logic Devices) such as FPGAs (Field-Programmable Gate Arrays) and FPAAs (Field-Programmable Analog Arrays). The processing unit 202 can perform various data processing and program control by interpreting and executing instructions from various programs. Programs executable by the processor are stored in at least one of the memory areas included in the processor and the storage unit 203.

[0143] The processing unit 202 may also include a main memory. The main memory includes at least one of volatile memory such as RAM and non-volatile memory such as ROM.

[0144] The RAM used is, for example, DRAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory). This RAM has virtual storage space allocated as the working space for the processing unit 202. The operating system, application programs, program modules, program data, and lookup tables stored in the storage unit 203 are loaded into the RAM during execution. The processing unit 202 directly accesses and operates on this data, program, and program modules loaded into the RAM.

[0145] ROM can store BIOS (Basic Input / Output System) and firmware that do not need to be rewritten. Examples of ROMs include mask ROM, OTPROM (One-Time Programmable Read Only Memory), and EPROM (Erasable Programmable Read Only Memory). Examples of EPROMs include UV-EPROM (Ultra-Violet Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), and flash memory, where the stored data can be erased by ultraviolet light.

[0146] [Storage Department 203]

[0147] The storage unit 203 has the function of storing the program executed by the processing unit 202. Furthermore, the storage unit 203 may also have the function of storing, for example, the calculation results generated by the processing unit 202 and the data input to the input unit 201. Specifically, it is preferable that the storage unit 203 has the function of storing the graph structure generated by the processing unit 202 (for example, the graph structure 21 shown in Embodiment 1), the calculated similarity results, etc.

[0148] Storage unit 203 includes at least one of volatile memory and non-volatile memory. Storage unit 203 may include, for example, volatile memory such as DRAM or SRAM. Storage unit 203 may also include, for example, non-volatile memory such as ReRAM (Resistive Random Access Memory), PRAM (Phase Change Random Access Memory), FeRAM (Ferroelectric Random Access Memory), MRAM (Magnetoresistive Random Access Memory), or flash memory. Furthermore, storage unit 203 may also include recording media drives such as hard disk drives (HDDs) and solid-state drives (SSDs).

[0149] [Database 204]

[0150] The document retrieval system 200 may also include a database 204. For example, the database 204 may have the function of storing multiple documents and multiple graph structures corresponding to the multiple documents respectively. For example, the document retrieval method of one aspect of the present invention may also be used with the multiple documents stored in the database 204 as the object. In addition, the database 204 may also store a concept dictionary.

[0151] Furthermore, the storage unit 203 and the database 204 may not be separate from each other. For example, the document retrieval system 200 may also include a storage unit that has the functions of both the storage unit 203 and the database 204.

[0152] Furthermore, the memory included in the processing unit 202, the storage unit 203, and the database 204 is referred to as an example of a non-transitory computer-readable storage medium.

[0153] [Display Unit 205]

[0154] The display unit 205 has the function of displaying the calculation results of the processing unit 202. Furthermore, the display unit 205 has the function of displaying the results of document comparison and similarity. Additionally, the display unit 205 may also have the function of displaying the document specified for the search.

[0155] In addition, the document retrieval system 200 may also include an output unit. The output unit has the function of supplying data to external systems.

[0156] [Teleportation Channel 206]

[0157] The transmission channel 206 has the function of transmitting various types of data. Data exchange between the input unit 201, processing unit 202, storage unit 203, database 204, and display unit 205 can be performed through the transmission channel 206. For example, data such as a document specified by the user for retrieval and the graph structure corresponding to the document's comparison documents are sent and received through the transmission channel 206.

[0158] <Example 2 of a document retrieval system structure>

[0159] Figure 8 A block diagram of a document retrieval system 210 is shown. The document retrieval system 210 includes a server 220 and a terminal 230 (such as a personal computer).

[0160] Server 220 includes a processing unit 202, a transmission channel 212, a storage unit 213, and a communication unit 217a. Although in Figure 8 Although not shown in the diagram, server 220 may also include input / output units, etc.

[0161] Terminal 230 includes an input unit 201, a storage unit 203, a display unit 205, a transmission channel 216, a communication unit 217b, and a processing unit 218. Although in Figure 8 Although not shown in the diagram, terminal 230 may also include databases, etc.

[0162] The user of the document retrieval system 210 inputs a document into the input section 201 of the terminal 230. This document is a document specified by the user for retrieval, equivalent to document 20 shown in Embodiment 1. The document is transmitted from the communication section 217b of the terminal 230 to the communication section 217a of the server 220.

[0163] The documents received by the communication unit 217a are stored in the storage unit 213 via the transmission channel 212. Alternatively, the documents may be supplied directly from the communication unit 217a to the processing unit 202.

[0164] The graph structure formation and similarity calculation described in Embodiment 1 both require high processing power. Compared to the processing unit 218 included in the terminal 230, the processing unit 202 included in the server 220 has higher processing power. Therefore, it is preferable that both graph structure formation and similarity calculation are performed in the processing unit 202.

[0165] Then, the processing unit 202 calculates the similarity. The similarity is stored in the storage unit 213 via the transmission channel 212. Alternatively, the similarity can be directly supplied from the processing unit 202 to the communication unit 217a. The similarity is transmitted from the communication unit 217a of the server 220 to the communication unit 217b of the terminal 230. The similarity is displayed on the display unit 205 of the terminal 230.

[0166] [Teleportation Channel 212 and Teleportation Channel 216]

[0167] Transmission channels 212 and 216 are functional for transmitting data. Data transmission and reception between processing unit 202, storage unit 213, and communication unit 217a can be performed via transmission channel 212. Data transmission and reception between input unit 201, storage unit 203, display unit 205, communication unit 217b, and processing unit 218 can be performed via transmission channel 216.

[0168] [Processing Department 202 and Processing Department 218]

[0169] Processing unit 202 has the function of performing calculations using data supplied from storage unit 213 and communication unit 217a, etc. Processing unit 218 has the function of performing calculations using data supplied from input unit 201, storage unit 203, display unit 205 and communication unit 217b, etc. The description of processing unit 202 and processing unit 218 can be referred to. The processing capability of processing unit 202 is preferably higher than that of processing unit 218.

[0170] [Storage Department 203]

[0171] The storage unit 203 has the function of storing the program executed by the processing unit 218. In addition, the storage unit 203 has the function of storing the calculation results generated by the processing unit 218, the data input to the communication unit 217b, and the data input to the input unit 201.

[0172] [Storage Department 213]

[0173] The storage unit 213 has the functions of storing multiple documents, graph structures corresponding to the multiple documents, calculation results generated by the processing unit 202, and data input to the communication unit 217a.

[0174] [Communications Department 217a and Communications Department 217b]

[0175] Data can be sent and received between server 220 and terminal 230 by using communication units 217a and 217b. Communication units 217a and 217b can be hubs, routers, modems, etc. Data can be sent and received via wired or wireless means (e.g., radio waves, infrared, etc.).

[0176] In addition, communication between server 220 and terminal 230 can be conducted through computer networks such as the Internet, intranet, extranet, PAN (Personal Area Network), LAN (Local Area Network), CAN (Campus Area Network), MAN (Metropolitan Area Network), WAN (Wide Area Network), and GAN (Global Area Network), which serve as the base for the World Wide Web (WWW).

[0177] This implementation method can be appropriately combined with other implementation methods.

[0178] [Symbol Explanation]

[0179] 10: Multiple documents, 10_1: Document, 10_i: Document, 10_n: Document, 11: Multiple graph structures, 11_1: Graph structure, 11_i: Graph structure, 11_n: Graph structure, 20: Document, 21: Graph structure, 100: Document retrieval system, 101: Input unit, 102: Graph structure forming unit, 103: Similarity calculation unit, 104: Output unit, 105: Storage unit, 112: Concept dictionary, 200: Document retrieval system, 201: Input unit, 202: Processing unit, 203: Storage unit, 204: Database, 205: Display unit, 206: Transmission channel, 210: Document retrieval system, 212: Transmission channel, 213: Storage unit, 216: Transmission channel, 217a: Communication unit, 217b: Communication unit, 218: Processing unit, 220: Server, 230: Terminal

Claims

1. A document retrieval system, comprising: Input section; First Processing Department; Storage Department; Second processing department; as well as Output section The input section has the function of inputting a first document. The first processing unit has the function of forming a first diagram structure using the first document. The storage unit has the function of storing the structure of the second diagram. The second processing unit has the function of calculating the similarity between the first graph structure and the second graph structure. The output unit has the function of supplying information. The first processing unit has the function of dividing the first document into multiple tokens. The nodes and edges of the first graph structure are labeled. The label is composed of the plurality of symbols. Furthermore, if the labels of the edges in the first graph structure have antonyms, the first processing unit has the function of generating a new graph structure by reversing the direction of the edges of the first graph structure and replacing the labels of the edges with the antonyms.

2. The document retrieval system according to claim 1, The first processing unit has the function of assigning word classes to the tokens.

3. The document retrieval system according to claim 1, The first processing unit has the function of performing dependency analysis. Furthermore, the first processing unit has the function of linking partial tokens based on the result of the dependency analysis.

4. The document retrieval system according to claim 1, The first processing unit has the function of replacing a symbol including a representative word or a superordinate word with the representative word or the superordinate word.

5. The document retrieval system according to claim 1, The first processing unit uses the second document to form the second graph structure.

6. The document retrieval system according to claim 1, The second processing unit has the following functions: vectorizing the first graph structure and the second graph structure, and evaluating the vector similarity between the vectorized first graph structure and the vectorized second graph structure.

7. The document retrieval system according to claim 6, The second processing unit has the function of vectorizing the first graph structure and the second graph structure using the Weisfeiler-Lehman kernel.

8. The document retrieval system according to claim 2, In the case where the word class assigned to the first symbol is a noun and the word class assigned to the second symbol located directly before the first symbol is an adjective, the first processing unit has the function of connecting the second symbol and the first symbol.

9. The document retrieval system according to claim 2, When both the word class assigned to the third symbol and the word class assigned to the fourth symbol immediately following the third symbol are nouns, the first processing unit has the function of linking the third symbol and the fourth symbol.

Images