Generation apparatus, generation method, and generation program

Abstract

To provide a generation device, generation method, and generation program, which enable accurate automatic generation of an abstracted graph pattern and enhance extraction performance of the graph pattern for extracting similar sentences with high accuracy.SOLUTION: A generation device 100 is provided, comprising an acquisition unit configured to acquire graph structure data indicative of a dependency relationship between nodes representing attribute information including words and phrases within a sentence and information on parts of speech thereof, and an abstraction unit 113 configured to abstract the graph structure data acquired by the acquisition unit on the basis of the attribute information within the nodes.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a generation device, a generation method, and a generation program for generating data.

Background Art

[0002] Regarding graph structure data indicating the relationships between words in a sentence, it is possible to efficiently extract similar sentences by creating a graph pattern that abstracts its attributes and performing pattern matching. The technology for extracting similar sentences is important from the viewpoints of document search and text mining. As the background art of the graph pattern generation technology, there is Patent Document 1 below.

[0003] Patent Document 1 below describes that "in a text mining device, a synonymous expression identification means identifies whether a subtree (matched subtree) that matches the dependency structure tree (synonymous expression dependency structure tree) of an expression registered in a synonymous expression dictionary is included in the dependency structure tree (target sentence dependency structure tree) of the sentence to be text mined", and the synonymous expression dictionary is utilized by the synonymous expression identification means. Further, in Patent Document 1 below, it is described that "the node replacement means replaces the matched subtree with a special node (synonymous expression node) indicating the group to which the synonymous expression belongs, and the characteristic subtree extraction means extracts a characteristic subtree from the target sentence dependency structure tree after replacement", and the information of the synonymous expression dictionary is utilized by the node replacement means.

[0004] Furthermore, Patent Document 2 states that, "In a method for creating a document digest, when extracting a subgraph corresponding to the context from the document graph, the information processing device calculates an indirect relevance score representing the strength of the relationship between each word in the context and each word on the document graph, creates an indirect relevance vector for each word in the context by listing the indirect relevance scores between that word and all the words on the document graph, obtains a set of related words consisting of words strongly associated with the context from the importance of the words in the context and the indirect relevance vector, and creates the subgraph by extracting only the nodes containing words included in the set of related words from the document graph." [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2007-041767 [Patent Document 2] Japanese Patent Publication No. 2001-249935 [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] Graph patterns are created by abstracting the graph structure data corresponding to example sentences. In creating graph patterns, it is important to explicitly identify or estimate important parts of the graph structure data during the abstraction process, and to reflect this important information in the degree of abstraction of each node.

[0007] The graph structure data abstraction techniques described in Patent Documents 1 and 2 implicitly identify important sections within the algorithm, and do not anticipate cases where important sections are explicitly accepted as input from the user. Therefore, users cannot explicitly control the graph pattern abstraction process. Consequently, similar sentences that users consider important may not be considered important within the algorithm and may not be extracted.

[0008] The present invention aims to automatically generate abstracted graph patterns with high accuracy. [Means for solving the problem]

[0009] A generating device that represents one aspect of the invention disclosed in this application is a lemma and part-of-speech attribute information indicating the attributes of its part of speech. and Information indicating whether the aforementioned lemma is an important point Report and The present invention is characterized by having: a storage unit that stores attribute branching conditions including the above, attribute omission information that defines for each attribute branching condition whether the words and phrases corresponding to the lemma and the part-of-speech attribute information can be omitted; an acquisition unit that acquires graph structure data consisting of nodes containing words and phrases in a sentence and their part-of-speech attribute information and dependencies between the nodes, with the important location information attached to some of the nodes; and an abstraction unit that outputs a graph pattern abstracting the sentence by deleting the words and phrases in the sentence corresponding to the lemma that can be omitted by the attribute omission information and the part-of-speech attribute information that can be omitted by the attribute omission information from the nodes constituting the graph structure data acquired by the acquisition unit that correspond to the attribute branching conditions. [Effects of the Invention]

[0010] According to a typical embodiment of the present invention, abstracted graph patterns can be automatically generated with high accuracy. Other problems, configurations, and effects will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0011] [Figure 1] FIG. 1 is a block diagram showing a configuration example of the generation device according to Embodiment 1. [Figure 2] FIG. 2 is an explanatory diagram showing an example of important location information. [Figure 3] FIG. 3 is an explanatory diagram showing an example of graph structure data. [Figure 4] FIG. 4 is a flowchart showing a detailed processing procedure example of the abstraction processing by the abstraction unit. [Figure 5] FIG. 5 is an explanatory diagram showing an example of the omission determination process (step S401) of attribute information. [Figure 6] FIG. 6 is an explanatory diagram showing an example of attribute branch condition information. [Figure 7] FIG. 7 is an explanatory diagram showing an example of attribute omission information. [Figure 8] FIG. 8 is an explanatory diagram showing Example 1 of the node substitution process without attributes (step S402). [Figure 9] FIG. 9 is an explanatory diagram showing Example 2 of the node substitution process without attributes (step S402). [Figure 10] FIG. 10 is an explanatory diagram showing an example of the first addition method of the quantifier node. [Figure 11] FIG. 11 is an explanatory diagram showing an example of a list of quantifier node insertion conditions. [Figure 12] FIG. 12 is an explanatory diagram showing an example of the exclusion condition setting process (step S404). [Figure 13] FIG. 13 is an explanatory diagram showing an example of the confirmation screen of the graph pattern 131 of Embodiment 1. [Figure 14] FIG. 14 is a block diagram showing a configuration example of the generation device according to Embodiment 2. [Figure 15] FIG. 15 is a block diagram showing a configuration example of the generation device according to Embodiment 3. [Figure 16] FIG. 16 is an explanatory diagram showing an example of a lemma dictionary. [Figure 17] FIG. 17 is a flowchart showing a detailed processing procedure example of the abstraction processing by the abstraction unit according to Embodiment 3. [Figure 18] FIG. 18 is an explanatory diagram showing an example of important part information according to Example 4. [Figure 19] FIG. 19 is an explanatory diagram showing an example of a screen output of English graph structure data when English text is input to the graph structure conversion unit. [Figure 20] FIG. 20 is an explanatory diagram showing an example of attribute branch condition information according to Example 4. [Figure 21] FIG. 21 is an explanatory diagram showing an example of attribute omission information according to Example 4. [Figure 22] FIG. 22 is an explanatory diagram showing an example of a lemma dictionary according to Example 4. [Figure 23] FIG. 23 is an explanatory diagram showing an example of a graph pattern according to Example 4. [Figure 24] FIG. 24 is a block diagram showing an example of the hardware configuration of the generation device shown in Examples 1 to 4.

BEST MODE FOR CARRYING OUT THE INVENTION

[0012] Hereinafter, examples will be described with reference to the accompanying drawings. In the following, each example and each modification can be partially or wholly combined without departing from the spirit of the present invention. In the following description, the generation device is a computer that inputs text and data indicating important phrases included in the text, and presents a user with a graph pattern that is graph structure data abstracted by an abstraction device.

EXAMPLE

[0013] FIG. 1 is a block diagram showing a configuration example of the generation device according to Example 1. The generation device 100 inputs text 101, important part information 102 of text 101, attribute branch condition information 121, attribute omission information 122, and a list 123 of quantifier node insertion conditions, and generates an abstracted graph pattern 131 of text 101. The generation device 100 can handle any language, but in Example 1, an example of handling Japanese text 101 as an input will be described.

[0014] Text 101 is the data from which graph pattern 131 is generated, and it is the string to be extracted for similar sentences. For example, text 101 is "This command is a subcommand called from stopping the Database instance."

[0015] Furthermore, the important section information 102 is data that indicates which parts are important for determining similarity when extracting similar sentences. Note that the expressions included in text 101 and important section information 102 do not necessarily have to be from a single sentence, but may be from one or more sentences. In the following Example 1, for convenience, we will describe an example in which the important section information 102 consists of words and phrases extracted from text 101.

[0016] Figure 2 is an explanatory diagram showing an example of important section information 102. Important section information 102 includes the focus section ID 201 and the word / phrase 202. The focus section ID 201 is identification information that uniquely identifies the focus section of the string targeted for similar sentence extraction (in this example, text 101). For example, "a0" indicates a predicate, "a1" indicates a subject, and "a2" indicates an object.

[0017] Words and phrases 202 are words or phrases that classify the strings targeted for similar sentence extraction according to the focus point indicated by focus point ID 201. For example, "a0" is a predicate, so it indicates "is called.", "a1" is a subject, so it indicates "this command is", and "a2" is an object, so it indicates "from stopping the Database instance".

[0018] In Example 1, the input data used by the generation device 100 can be represented in any data structure, regardless of the data structure. For example, information can be stored in data structures such as lists, databases, queues, stacks, or tables. Figure 2 shows an example where the important location information 102 is held in table data.

[0019] Returning to Figure 1, the generation device 100 includes a graph structure conversion unit 111, an addition unit 112, and an abstraction unit 113. The graph structure conversion unit 111 analyzes the text 101 using a known algorithm and converts the text 101 into graph structure data that shows the relationships within the text 101. For example, the graph structure conversion unit 111 converts the text 101 into graph structure data using a conversion method that uses a Japanese dependency parser such as CaboCha, or a conversion method that uses the GiNZA Japanese parser based on Universal Dependencies.

[0020] In Example 1, the graph structure transformation unit 111 utilizes a known Japanese parser. Each node of the transformed graph structure data consists of words (including stems and endings) and their part-of-speech attribute information from the text 101, obtained from the analysis results of the Japanese parser.

[0021] Part-of-speech attribute information is information that indicates the attributes of a part of speech, specifically, for example, the part of speech of a lemma, the part of speech of a particle, and tense. By considering part-of-speech attribute information, a finer level of abstraction can be performed that does not rely solely on words and phrases, reducing noise in similar sentence extraction and improving extraction performance compared to constructing graph patterns using only words and phrases.

[0022] Graph structure data, like the input data mentioned above, can be represented in any data structure, regardless of the data structure itself. However, for convenience, in Example 1, it is assumed to be stored in list format.

[0023] Figure 3 is an explanatory diagram showing an example of graph structure data. The graph structure data 300 is the result of the graph structure conversion unit 111 converting the text 101. In the graph structure data 300, one node is represented by a string of characters (word phrase and part-of-speech attribute information) from the opening parenthesis "(" to the closing parenthesis ")", and the dependency between nodes is indicated by indentation, however, the method of representation is not limited to parentheses or indentation. The string of characters (word phrase and part-of-speech attribute information) between nodes is referred to as the attribute information of that node.

[0024] Furthermore, in graph structure data 300, the part-of-speech attribute ".lemma" indicates a lemma (headword). Lemmas may include content words such as nouns, verbs, adjectives, adjectival nouns, and attributive adjectives, as well as function words such as auxiliary verbs, conjunctions, and articles.

[0025] ".POS" (Part Of Speech) indicates the part of speech information of the lemma. ".POS2" indicates the second part of speech attribute information, which is a detailed classification of POS. ".casePOS" indicates a particle. ".casePOS2" indicates the second particle attribute information, which is a detailed classification of particles. ".suf" is used to distinguish auxiliary verbs, some particles, auxiliary verbs, auxiliary adjectival verbs, etc., mainly to distinguish word endings, negative forms, and interrogative forms. In addition, "&" indicates an AND condition for part of speech attribute information, but the symbol is not limited to "&". Note that the part of speech attribute information handled in Example 1 is limited to the above, but it is not necessarily limited to the above part of speech attribute information, and the addition or deletion of attributes may be considered depending on the output of the graph structure data conversion method.

[0026] In Figure 3 and subsequent figures, the sign of a node is represented by n# (where # is a number of one or more digits). Furthermore, because there are dependencies between nodes in the graph structure data 300, the node groups are hierarchically structured. Specifically, for example, the graph structure data 300 is hierarchically composed of nodes n0, n01, n011, n02, n021, and n0211 (if these are not distinguished, they are simply referred to as node n). In Figure 3 and subsequent figures, between two dependent nodes n, the sign of the parent node n is represented as being included in the sign of the child node. For example, the parent node of node n01 is node n0, and the parent node of node n011 is node n01.

[0027] Although the graph structure conversion unit 111 converts the text 101 into graph structure data 300, it may also function as an acquisition unit to acquire already converted graph structure data 300.

[0028] Returning to Figure 1, the addition unit 112 applies pattern matching to the graph structure data 300 of the text 101 obtained from the graph structure conversion unit 111, assigning important location information 102 to the corresponding nodes. In this case, each element of the important location information 102 may span multiple nodes; however, in such cases, the addition unit 112 assigns the important location information 102 to the node corresponding to the higher level of dependency. If the graph structure data 300 does not have dependencies, the addition unit 112 can simply assign the important location information 102 to the relevant multiple nodes. In the following example 1, for convenience, we will assume graph structure data that allows for dependencies between nodes.

[0029] The addition unit 112 assigns the relevant node in the graph structure data 300 a focus point ID 201 ("#a1", "#a2", "#a3") as important point information 102, resulting in the graph structure data 300a shown in Figure 3. Note that "#" is a convenient representation in Example 1, and the representation method is not limited to these.

[0030] The suffix "a" is added to the code of node n to which the important location information 102 is assigned. For example, in graph structure data 300a, nodes n0, n011, and n0211 do not have "a" at the end, so nodes n0, n011, and n0211 do not have the important location information 102 assigned to them. On the other hand, nodes n01a, n02a, and n021a have "a" at the end, so nodes n01a, n02a, and n021a have the important location information 102 assigned to them.

[0031] The abstraction unit 113 abstracts the graph structure data 300a of the text 101 obtained from the addition unit 112, preserving the parts considered important while abstracting the parts that are not considered important, and outputs a graph pattern 131. Abstraction includes operations such as deleting some of the part-of-speech attribute information attached to a node, deleting a node if all of the part-of-speech attribute information attached to a node is deleted during the abstraction process, inserting quantifier nodes, and setting exclusion conditions that do not allow specified part-of-speech attribute information to be placed in a node. A quantifier node is a node that allows arbitrary subgraphs to be inserted between certain nodes.

[0032] Figure 4 is a flowchart showing a detailed example of the abstraction process performed by the abstraction unit 113. First, the abstraction unit 113 receives the graph structure data 300a of the text 101 via the addition unit 112 and performs a process to determine whether to omit attribute information attached to each node in the graph structure data 300a (step S401).

[0033] Figure 5 is an explanatory diagram showing an example of the attribute information omission determination process (step S401). The graph structure data 300ad is a graph structure data abstraction of the graph structure data 300a to which important location information 102 has been attached. The graph structure data 300ad is the graph structure data to which the attribute information omission determination process (step S401) has been performed. In the graph structure data 300ad, "(_)" means that all word and part-of-speech attribute information within node n (for example, nodes n011d, n0211d) has been deleted. Note that "_" is a convenient representation in Example 1, and the representation method is not limited to these.

[0034] In the attribute information omission determination process (step S401), the abstraction unit 113 performs binary identification on the graph structure data 300a to determine whether to delete attributes attached to each node, and removes the attribute information that needs to be deleted. Possible methods for binary identification include, for example, a method that determines attribute deletion based on a predetermined rule, and a method that probabilistically determines attribute deletion from surrounding information of a node using a machine learning model. When a machine learning model is used as a means of determining attribute deletion, it will be described later in Example 2.

[0035] In attribute omission determination based on default rules, the abstraction unit 113 refers to the attribute branching condition information 121 and attribute omission information 122 and performs node attribute information omission determination by using multiple conditional branches with If-else statements for the attributes assigned to each node. The attribute branching condition information 121 and attribute omission information 122 can be stored as data in advance in a tabular format as shown in Figures 6 and 7, respectively, but the data format is not limited to a tabular format and may be expressed in any data format.

[0036] Figure 6 is an explanatory diagram showing an example of attribute branching condition information 121. Attribute branching condition information 121 has three fields: rule ID 601, important location assignment flag 602, and part-of-speech attribute information 603. A combination of values ​​for each field in the same row defines an attribute branching condition which is a single rule. Rule ID 601 is identification information that uniquely identifies the default rule, which is the attribute branching condition. Important location assignment flag 602 indicates whether or not important location information 102 has been assigned by the attribute branching condition identified by rule ID 601 (specifically, the part-of-speech attribute information 603). "True" indicates that important location information 102 has been assigned, and "False" indicates that important location information 102 has not been assigned.

[0037] For example, an entry where rule ID 601 is "1" (hereinafter referred to as attribute branching condition 1) and rule ID 601 is " 2The entry for '' (hereinafter referred to as attribute branching condition 2) is defined as "True". On the other hand, the entry for rule ID 601 with "3" (hereinafter referred to as attribute branching condition 3) is defined as "False".

[0038] The part-of-speech attribute information 603 includes lemma 631, particle (case) 632, part-of-speech information (POS) 633 of lemma 631, second part-of-speech attribute information (POS) 634, particle attribute information (casePOS) 635 of the particle, and second particle attribute information (casePOS2) 636. For example, attribute branching condition 1 specifies that the value of part-of-speech information (POS) 633 of lemma 631 is "noun" and the value of second particle attribute information (casePOS2) 636 is "conjunctive particle".

[0039] Therefore, graph structure data 30 0 In this case, a node that has important location information 102 attached, where the part-of-speech information (POS) 633 of lemma 631 is "noun" and the second particle attribute information (casePOS2) 636 is "conjunction" will meet attribute branching condition 1.

[0040] For example, in graph structure data 300a, node n01a has the focus ID 201 "#a1" attached as important location information 102, but the part-of-speech information (POS) 633 is "noun" and the second particle attribute information (casePOS2) 636 is "particle". Therefore, node n01a does not meet attribute branching condition 1. The same applies to nodes n02a and n021a.

[0041] On the other hand, for attribute branching condition 2, the value of the important location assignment flag 602 is "True," and the part-of-speech information (POS) 633 of lemma 631 is defined as "noun," and the second part-of-speech attribute information (POS) 634 is defined as "general." Node n01a has the focus location ID 201 "#a1" attached as important location information 102, and the part-of-speech information (POS) 633 is "noun" and the second part-of-speech attribute information (POS) 634 is "general." Therefore, node n01a meets attribute branching condition 2. Nodes n02a and n021a do not meet attribute branching condition 2.

[0042] Furthermore, for attribute branching condition 3, the value of the important location assignment flag 602 is "False", lemma 631 is defined as "this", and part-of-speech attribute information (POS) 633 is defined as "adjective". Child node of node n01a n0 11 does not have important location information 102 attached, lemma 631 is "this", and part-of-speech attribute information (POS) 633 is "adjective". Therefore, child node n0 11 corresponds to attribute branching condition 3.

[0043] Figure 7 is an explanatory diagram showing an example of attribute omission information 122. Attribute omission information 122 has two fields: rule ID 601 and part-of-speech attribute information 703. The combination of values ​​for each field in the same row defines attribute omission in the attribute branching condition shown in Figure 6. That is, if a node satisfying attribute branching condition # (where # is the value of rule ID 601) exists in the graph structure data 300a, attribute omission # is applied to that node.

[0044] Part-of-speech attribute information 703, like part-of-speech attribute information 603, has the following as omitted attributes: lemma 731, particle (case) 732, part-of-speech information (POS) 733 for lemma 731, second part-of-speech attribute information (POS2) 734, particle attribute information (casePOS) 735 for particle, and second particle attribute information (casePOS2) 736. Part-of-speech attribute information 703 holds either "Hold" or "Delete" as its value. "Hold" indicates that the omitted attribute will be retained, while "Delete" indicates that the omitted attribute will be omitted.

[0045] For example, in Figure 5, since no node satisfies attribute branching condition 1 in graph structure data 300a, no node to which attribute omission 1 applies exists in graph structure data 300a.

[0046] On the other hand, node n01a satisfies attribute branching condition 2, where rule ID 601 in Figure 6 is entry "2". Therefore, attribute omission 2, where rule ID 601 in Figure 7 is "2", is applied to node n01a. In attribute omission 2, lemma 731 and second part of speech attribute information (POS2) 734 are "Delete". Therefore, ".lemma=command&" and ".POS2=general&" are removed from node n01a, resulting in node n01ad.

[0047] Furthermore, node n011 satisfies attribute branching condition 3, where rule ID 601 in Figure 6 is entry "3". Therefore, attribute omission 3, where rule ID 601 in Figure 7 is "3", is applied to node n011. In attribute omission 3, lemma 731 and part-of-speech attribute information (POS) 733 are "Delete". Therefore, ".lemma=this&" and ".POS=adjective" are deleted from node n011, resulting in node n011d.

[0048] Thus, when word and part-of-speech attribute information is removed from node n#, the letter 'd' is appended to the end, resulting in node n#d.

[0049] The part-of-speech attribute information 703 includes a lemma 731, a particle 732, part-of-speech information 733 of the lemma 731, second part-of-speech attribute information 734, a particle 735, and second particle attribute information 736. For example, in an entry where rule ID 701 is "1" (hereinafter referred to as attribute omission condition 1), the value of the lemma 731 is defined as "Delete", the value of the particle 732 is "Delete", the value of the part-of-speech information 633 of the lemma 731 is "Hold", the value of the second part-of-speech attribute information 734 is "Hold", the value of the particle 735 is "Hold", and the value of the second particle attribute information 736 is "Hold". Therefore, in the graph structure data 302, if the part-of-speech information 633 of the lemma 631 is "noun" and the second particle attribute information 636 is "conjunction," then attribute branching condition 1 is met.

[0050] Returning to Figure 4, the abstraction unit 113 performs the attribute-less node replacement process (step S402). In the attribute-less node replacement process (step S402), the abstraction unit 113 replaces nodes that do not hold any attribute information (hereinafter referred to as attribute-less nodes) with quantifier nodes in the graph structure data 300ad received from the attribute information omission determination process (step S401) after the attribute omission determination. A quantifier node is a node that allows the insertion of any subgraph (and the attribute information that makes it a subgraph).

[0051] A quantifier node is, for example, represented by "*", and is an exception that is not enclosed in opening parentheses "(" and closing parentheses ")". Therefore, a quantifier node is contained within a parent node that had a dependency on the attribute-less node before substitution. That is, an "*" indicating the quantifier node is appended to the end of the parent node.

[0052] Figure 8 is an explanatory diagram showing Example 1 of the attributeless node replacement process (step S402). In the graph structure data 300ad, the attributeless nodes are nodes n011d and n0211d. The graph structure data 300adq is the graph structure data in which the attributeless nodes in the graph structure data 300ad have been replaced with quantifier nodes by the attributeless node replacement process (step S402).

[0053] In graph structure data 300ad, quantifier node 801 is replaced from attributeless node n011d and taken over by its parent node n01ad, which had a dependency on attributeless node n011d. As a result, node n01ad becomes node n01adq.

[0054] Furthermore, the quantifier node 802 is replaced by the attributeless node n011d and taken over by the parent node n01ad, which was dependent on the attributeless node n011d. As a result, node n011d becomes node n011dq.

[0055] Figure 9 is an explanatory diagram showing Example 2 of the attribute-less node replacement process (step S402). Figure 9 is an example in which graph structure data 900ad is used as the source for replacement instead of graph structure data 300ad. Graph structure data 900ad has a node nxd that does not have a parent node and a node nxyd whose parent node is an attribute-less node in graph structure data 300ad.

[0056] In the attribute-less node replacement process (step S402), such attribute-less nodes nxd and nxyd are deleted without being replaced by quantifier nodes, and graph structure data 300adq is output.

[0057] By applying the graph structure data 300adq shown in Figures 8 and 9, it becomes possible to extract not only text 101 but also similar sentences such as "These programs are setup programs that are called from the startup of the operating system."

[0058] Returning to Figure 4, the abstraction unit 113 executes the quantifier node insertion process (step S403). In the quantifier node insertion process (step S403), the abstraction unit 113 adds quantifier nodes to the graph structure data 300adq output from the attributeless node replacement process (step S402) between nodes where it is determined that quantifier node insertion is necessary.

[0059] Specifically, there are three methods for adding quantifier nodes: a first method which inserts quantifier nodes between all dependent nodes n within the graph structure data 300adq; a second method which determines the location of the quantifier node based on node information around the target location using a default rule (multiple conditional branching using If-else statements); and a third method which determines the location of the quantifier node using a machine learning model.

[0060] Figure 10 is an explanatory diagram showing an example of the first method for adding quantifier nodes. Using this first method, the abstraction unit 113 adds quantifier nodes 1001 to 1005 between nodes n that do not already have quantifier nodes. Note that the addition of quantifier node 1003 causes node n02a to become node n02aq. The graph structure data 300adq, after the quantifier node insertion process (step S403) has been performed, is defined as graph structure data 300adqi.

[0061] Furthermore, in the case of the second addition method, the abstraction unit 113 uses the quantifier node insertion condition list 123. The third addition method will be described later in Example 2.

[0062] Figure 11 is an explanatory diagram showing an example of a quantifier node insertion condition list 123. The quantifier node insertion condition list 123 has the following fields: additional rule ID 1101, parent node part-of-speech attribute information 1102, and child node part-of-speech attribute information 1103. The additional rule ID 1101 is identification information that uniquely identifies the additional rule. When the additional rule ID 1101 is "#" (# is an ascending number starting from 1), it is written as insertion condition #.

[0063] The parent node's part-of-speech attribute information 1102 includes the parent node's lemma 1131 (which is the higher-level node among two dependent nodes), its particle (case) 1132, the part-of-speech information (POS) 1133 of the lemma 1131, the second part-of-speech attribute information (POS) 1134, the particle attribute information (casePOS) 1135 of the particle, and the second particle attribute information (casePOS2) 1136.

[0064] For example, the part-of-speech attribute information 1102 value "P1" (hereinafter referred to as part-of-speech attribute information P1) of the parent node of insertion condition 1 is defined as having the part-of-speech information (POS) 1133 value of lemma 1131 as "noun" and the second particle attribute information (casePOS2) 1136 value as "conjunctive particle".

[0065] The part-of-speech attribute information 1103 of the child node includes the lemma 1131, the particle (case) 1132, the part-of-speech information (POS) 1133 of the lemma 1131, the second part-of-speech attribute information (POS) 1134, the particle attribute information (casePOS) 1135 of the particle, and the second particle attribute information (casePOS2) 1136.

[0066] For example, the value "C1" (hereinafter referred to as "part-of-speech attribute information C1") of the child node of insertion condition 1 has the part-of-speech information (POS) 1133 of lemma 1131 defined as "noun" and the value of the second part-of-speech attribute information (POS2) 1134 as "general".

[0067] For example, if two dependent nodes satisfy insertion condition 1, that is, if the parent node satisfies part-of-speech attribute information P1 and the child node satisfies part-of-speech attribute information C1, then a quantifier child node "*" is added between those two nodes.

[0068] By applying the quantifier node insertion process (step S403), the likelihood of extracting a graph pattern 131 corresponding to text 101 as a similar sentence is improved, even if the similar sentence has more modifiers than text 101.

[0069] Returning to Figure 4, the abstraction unit 113 executes the exclusion condition setting process (step S404). In the exclusion condition setting process (step S404), the abstraction unit 113 sets extraction exclusion conditions for the quantifier nodes in the graph structure data 300adqi output from the quantifier node insertion process (step S403) after quantifier node insertion.

[0070] Specifically, there are two methods for setting the extraction exclusion conditions: a first setting method in which the extraction exclusion conditions are set for all quantifier nodes in the graph structure data 300adqi, and a second setting method in which the extraction exclusion conditions are set for quantifier nodes located below node n to which important location information 102 is attached (hereinafter referred to as the lowest nearest neighbor quantifier node). The exclusion condition setting process (step S404) reduces the mis-extraction of similar sentences. The second setting method will be explained using Figure 12.

[0071] Figure 12 is an explanatory diagram showing an example of the exclusion condition setting process (step S404). Figure 12 is an example in which, using the second setting method, extraction exclusion conditions are set for the nearest neighbor quantifier node of node n to which important location information 102 is attached, based on the part-of-speech attribute information of node n to which important location information 102 is attached. In Figure 12, if node n to which important location information 102 is attached is node n01adq, then its nearest neighbor quantifier node is node 1002. The abstraction unit 113 sets the extraction exclusion condition 1200 for this node 1002 to "case!=は".

[0072] "YYY!=zzz" (where YYY is part-of-speech attribute information and zzz is an arbitrary string) is a conditional that indicates that the string zzz, whose part-of-speech attribute information is YYY, does not match the condition following node n, which is assigned important location information 102. For example, "case!=wa" means that the particle "wa" and strings indicating its nominative case will not be extracted within that node. For example, since node 1002 is a quantifier node, strings excluding "○○wa" which corresponds to the extraction exclusion condition 1200 will be extracted.

[0073] Returning to Figure 2, the graph pattern 131 is the output of the generation device 100, and is data that abstracts the graph structure data of the text 101 based on the important location information 102. Specifically, for example, if the abstraction unit 113 executes only the attribute information omission determination process (step S401) in Figure 4, the graph pattern 131 becomes the graph structure data 300ad.

[0074] Furthermore, if the abstraction unit 113 executes only the attribute information omission determination process (step S401) and the attributeless node replacement process (step S402), the graph pattern 131 becomes graph structure data 300adq. Furthermore, if the abstraction unit 113 executes only the attribute information omission determination process (step S401) to the quantifier node insertion process (step S403), the graph pattern 131 becomes graph structure data 300adqi. Furthermore, if the abstraction unit 113 executes the attribute information omission determination process (step S401) to the exclusion condition setting process (step S404), the graph pattern 131 becomes graph structure data 300adqis.

[0075] Using the graph pattern 131 (graph structure data 300adqis) obtained from text 101 and important information 102, similar sentences such as "This command is the main command called from the start of the update process." can be extracted.

[0076] Figure 13 is an explanatory diagram showing an example of the confirmation screen for the graph pattern 131 of Embodiment 1. The confirmation screen 1300 may be displayed simultaneously with the output of the generation device 100, but the graph pattern 131 can be read by pressing the read button 1303 as needed.

[0077] The graph pattern 131 is displayed in the pattern display / editing area 1306. The original text 101 of the graph pattern 131 is displayed in the input text display area 1304. The graph structure data 300ad with important section information 102 is displayed in the important section display area 1305. The user can edit the graph pattern 131 in the pattern display / editing area 1306 by pressing the edit button 1301. The user can overwrite and save the edited graph pattern 131 by pressing the save button 1302. The confirmation screen 1300, which allows confirmation and editing of the graph pattern 131 as shown in Figure 13, enables manual pattern correction and contributes to improving the accuracy of similar sentence extraction.

[0078] Thus, according to Embodiment 1, the generation device 100 accepts text 101 and its important section information 102 as input, converts the text 101 into graph structure data 300, assigns the important section information 102 to the graph structure data 300, and the abstraction unit 113 not only focuses on lemmas and phrases but also utilizes all part-of-speech attribute information assigned to each node n to generate a highly accurate graph pattern 131. [Examples]

[0079] Example 2 will now be described. Example 2 is an example in which the attribute branching condition information 121, attribute omission information 122, and quantifier node insertion condition list 123, which are inputs to the abstraction unit 113 shown in Example 1, are changed to a machine learning model 1400. This makes it possible to generate a graph pattern 131 with higher accuracy. Note that in Example 2, the explanation will focus on the differences from Example 1, so the explanation of the same configuration as Example 1 will be omitted.

[0080] Figure 14 is a block diagram showing an example configuration of the generation apparatus according to Example 2. The generation apparatus 100 of Example 2 uses text 101, its important section information 102, and a machine learning model 1400 to generate a graph pattern 131 that is an abstraction of the text 101.

[0081] In Example 2, the abstraction unit 113 receives graph structure data 300a of text 101 to which important location information 102 has been added from the addition unit 112, and executes attribute information omission determination processing (step S401), attribute-less node replacement processing (step S402), quantifier node insertion processing (step S403), and exclusion condition setting processing (step S404).

[0082] In this process, the attribute omission determination process (step S401) and the quantifier node insertion process (step S403) are performed using the machine learning model 1400. The machine learning model 1400 is pre-trained, for example, using text 101, important section information 102, and the graph pattern 131 which is an abstraction result.

[0083] For example, for the attribute omission judgment process (step S401), a machine learning model 1400 usable in the attribute omission judgment process (step S401) is generated by training with the part-of-speech attribute information 603 shown in Figure 6 as the explanatory variable (training data) and the part-of-speech attribute information 703 shown in Figure 7 as the target variable (ground truth data).

[0084] For example, the abstraction unit 113 uses a machine learning model 1400, which has been trained based on the attribute information of a specific node n that matches one of the attribute branching conditions of the part-of-speech attribute information 603 shown in Figure 6, and the attribute omissions of the part-of-speech attribute information 703 shown in Figure 7 that correspond to the attribute branching conditions, to delete the attribute information that is the target of attribute omission ("Delete") as a result of inputting the attribute information of a specific node into the machine learning model 1400.

[0085] Furthermore, for the quantifier node insertion process (step S403), the part-of-speech attribute information 1102 of the parent node and the part-of-speech attribute information 1103 of the child node shown in Figure 11 are used as explanatory variables (training data), and the question of whether to add or not is used as the target variable (ground truth data). This generates a machine learning model 1400 that can be used in the quantifier node insertion process (step S403).

[0086] For example, the abstraction unit 113 uses a machine learning model 1400, which has been trained based on the part-of-speech attribute information 1102 of the parent node and the part-of-speech attribute information 1103 of the child node shown in Figure 11, to delete the part-of-speech attribute information 1103 of the child node that is output as a result of inputting the part-of-speech attribute information 1102 of the parent node into the machine learning model 1400.

[0087] Furthermore, machine learning model 1400 can utilize known models such as random forests, graph attention networks, and neural networks. Machine learning model 1400 can also accept either only the attribute information of the node of interest as input, or it can also accept the parent and child nodes of the node of interest as input. For feature extraction of attribute information, known feature extraction algorithms can be used for words, phrases, and particles.

[0088] For example, one could utilize Term Frequency Inverse Document Frequency (TF-IDF) or GloVe, a word embedding representation, but the specific method is not limited. Furthermore, each part-of-speech condition and important location information could be feature-quantified using one-hot encoding.

[0089] Thus, according to Example 2, by accepting text 101 and important information 102 as input, and receiving a machine learning model 1400 as additional input, the abstraction unit 113 can perform more accurate classification and generate a highly accurate graph pattern 131. [Examples]

[0090] Example 3 will now be described. In Example 3, the generation device 100 additionally inputs the lemma dictionary 1500 to the abstraction unit 113 shown in Example 1, and the abstraction unit 113 performs dictionary application processing as shown in Figure 16. This improves the extraction accuracy of the graph pattern 131. Note that since Example 3 will focus on the differences from Example 1, the description of the same configuration as Example 1 will be omitted.

[0091] Figure 15 is a block diagram showing an example configuration of the generation device according to Example 3. The generation device 100 according to Example 3 takes text 101, important location information 102, attribute branching condition information 121, attribute omission information 122, quantifier node insertion condition list 123, and lemma dictionary 1500 as input and generates a graph pattern 131 in which the text 101 is abstracted.

[0092] The lemma dictionary 1500 is constructed by forming semantic groups about lemmas, either manually or according to a known algorithm. For example, it may be constructed by referring to a thesaurus to form semantic groups for a given lemma, but any method for forming groups about lemmas is acceptable.

[0093] Figure 16 is an explanatory diagram showing an example of a lemma dictionary 1500. The lemma dictionary 1500 has the fields Group ID 1601 and Lemma Group 1602. A combination of values ​​for each field in the same row constitutes a group of lemmas. Group ID 1601 is identification information that uniquely identifies a group of lemmas. Lemma Group 1602 is one or more lemmas belonging to the group identified by Group ID 1601.

[0094] Figure 17 is a flowchart showing a detailed example of the processing procedure of the abstraction process performed by the abstraction unit 113 according to Embodiment 3. After the exclusion condition setting process (step S404), the abstraction unit 113 executes the lemma dictionary application process (step S1705). In the lemma dictionary application process (step S1705), the abstraction unit 113 receives the graph structure data 300adqis output in the lemma dictionary application process (step S1705) and updates the graph pattern 131 to allow all lemma groups 1602 within the dictionary group to which a certain lemma in the lemma dictionary 1500 belongs.

[0095] For example, if we apply the lemma dictionary 1500 shown in Figure 16 to the graph structure data 300adqis in Figure 12, not only "lemma=call" for node n02aq, but also "call" and "invite" from the lemma group 1602, which belongs to the same group as "call," will be allowed as lemma conditions. In other words, "lemma=call" for node n02aq will be updated to "lemma=Call_verb".

[0096] Thus, according to Example 3, by accepting inputs of text 101, important section information 102, attribute branching condition information 121, attribute omission information 122, and quantifier node insertion condition list 123, and receiving a lemma dictionary 1500 as additional input, it becomes possible to generate a graph pattern 131 capable of extracting a wider range of similar sentences. [Examples]

[0097] Next, we will describe Example 4. In Examples 1 to 3, the description language of text 101 was Japanese, but in Example 4, the description language of text 101 is English, and the generation device 100 outputs an English graph pattern 131. Since Example 4 will mainly be described in terms of the differences from Examples 1 to 3, the explanations of Examples 1 to 3 will be omitted.

[0098] In Example 4, the example of English text 101 is "This command is a subcommand called when the Database instance is shutdown."

[0099] Figure 18 is an explanatory diagram showing an example of important location information 1800 according to Embodiment 4. Important location information 1800 is data in which important location information 102 is written in English. Since the English text 101 is written with spaces between words, as shown in Figure 18, important location information 1800 may consist of multiple words, such as entries with ID 201 "a1" and "a2".

[0100] Figure 19 is an explanatory diagram showing an example of screen output of English graph structure data when English text 101 is input to the graph structure conversion unit 111. The English graph structure data 1900 corresponds to the graph structure data 300 shown in Figure 3. When the text 101 is in English, the graph structure conversion unit 111 can utilize, for example, spaCy or Stanford CoreNLP as known graph structure conversion algorithms.

[0101] In the English graph structure data 1900, the part-of-speech attribute ".type" indicates the type of dependency, ".POS" represents part-of-speech information, and ".lemma" indicates the content word. For example, the first line, ".type=root&.POS=NN&.lemma=subcommand", specifies that the dependency type is root, the part of speech is a singular noun, and its content word is "subcommand".

[0102] Although not shown in the figures, the addition section 112 adds important location information 1800 to the English graph structure data 1900, similar to Example 1. To distinguish the English graph structure data 1900 with the important location information 1800 added from the English graph structure data 1900 without the important location information 1800, it is denoted as English graph structure data 1900a.

[0103] The abstraction unit 113 receives English graph structure data 1900a from the addition unit 112 and executes attribute omission determination processing (step S401).

[0104] Figure 20 is an explanatory diagram showing an example of attribute branching condition information according to Embodiment 4. Similar to attribute branching condition information 121, attribute branching condition information 2000 has the following fields: rule ID 601, important location assignment flag 602, and part-of-speech attribute information 603. It defines an attribute branching condition where the combination of values ​​of each field in the same row constitutes a single rule.

[0105] The part-of-speech attribute information 603 includes a lemma 631, part-of-speech information (POS) 633 of the lemma 631, and a node-to-node relationship (type) 2001. The node-to-node relationship (type) 2001 is defined by the dependency relationships of Universal Dependencies.

[0106] Figure 21 is an explanatory diagram showing an example of attribute omission information according to Embodiment 4. Similar to attribute omission information 122, attribute omission information 2100 has a rule ID 601 and part-of-speech attribute information 703 as fields. It defines an attribute branching condition where the combination of values ​​of each field in the same row constitutes a single rule.

[0107] The part-of-speech attribute information 703 has a lemma 731, the part-of-speech information (POS) 733 of the lemma 731, and a node-to-node relationship (type) 2101. The node-to-node relationship (type) 2101, like the node-to-node relationship (type) 2001, is defined by the dependency relations of Universal Dependencies.

[0108] Furthermore, when performing machine learning-based judgment in steps S401 and S403, for feature extraction of attribute information, for example, one-hot representations may be used for dependency types, part-of-speech information, and important location information, while TF-IDF or word embedding representations such as GloVe or Word2Vec may be used for content words.

[0109] Figure 22 is an explanatory diagram showing an example of a lemma dictionary according to Embodiment 4. The lemma dictionary 2200 has as fields a group ID 2201 and a lemma group 2202. A combination of values ​​for each field in the same row constitutes a group of lemmas. The group ID 2201 is identification information that uniquely identifies an internal term group. The lemma group 2202 is one or more internal terms belonging to the internal term group identified by the group ID 1601.

[0110] For example, when the lemma dictionary 2200 is applied to graph structure data 1900, a node specifying "call" as a content word in the 6th row will allow other content words such as "cause" and "activate" in addition to "call".

[0111] Figure 23 is an explanatory diagram showing an example of a graph pattern according to Example 4. Using graph pattern 131, for example, similar sentences to the English text 101, such as "This command is a main command called when the update process is started," can be extracted. Here, "|" is a symbol indicating an OR condition, but "|" is a convenient representation in Example 4 and is not limited to this.

[0112] Thus, according to Example 4, a graph pattern 131 can be generated from the English text 101 and the corresponding important section information 1800, which can extract similar English sentences.

[0113] The graph pattern 131 generated by the generation device 100 described in Examples 1 to 4 above enables pattern matching with similar sentences. Specifically, for example, the generation device 100 converts the similar sentence "This command is the main command called from the start of the update process." into graph structure data using the graph structure conversion unit 111 (hereinafter referred to as "similar sentence graph structure data"). The generation device 100 then performs pattern matching on the similar sentence graph structure data using the graph pattern 131. As a result, from the group of nodes constituting the similar sentence graph structure data, words corresponding to the graph pattern 131 are extracted, such as "called," "this command is," and "from the start of the update process."

[0114] Figure 24 is a block diagram showing an example hardware configuration of the generation device 100 shown in Examples 1 to 4. The generation device 100 is composed of a computer consisting of, for example, a processor (CPU) 2401, an auxiliary storage device 2402, a memory 2403, an input device 2404, an output device 2405, and a communication interface 2406. Components 2401 to 2406, which are components of the example hardware configuration of the computer 2400, are interconnected and can communicate as needed.

[0115] The processor 2401 executes the program stored in the memory 2403. The processor 2401 may consist of, for example, a single arithmetic unit and processing unit, or any number of arithmetic units and processing units. The memory 2403 includes a non-volatile memory element, ROM (Read Only Memory), and a volatile memory element, RAM (Random Access Memory). The ROM stores immutable programs, etc. The RAM temporarily stores the program executed by the processor 2401 and data used when the program is executed.

[0116] The auxiliary storage device 2402 is a high-capacity, non-volatile storage device such as a magnetic storage device (Hard Disk Drive) or flash memory (Solid State Drive). Furthermore, the auxiliary storage device 2402 stores the program executed by the processor 2401, and the data used during program execution. That is, the program is loaded from the auxiliary storage device 2402, read into memory 2403, and executed by the processor 2401.

[0117] Specifically, for example, each processing unit of the generation device 100 loads a non-temporary program stored in the auxiliary storage device 2402 into the memory 2403, and the processor 2401 executes the loaded program. In addition, the data used in each processing unit of Examples 1 to 4, such as text 101, important location information 102, 1800, attribute branching condition information 121, 2000, attribute omission information 122, 2100, quantifier node insertion condition list 123, machine learning model 1400, lemma dictionary 1500, 2200, and graph pattern 131, are stored in the auxiliary storage device 2402, for example.

[0118] The computer 2400 may have an input device 2404. The input device 2404 is a device on which a user inputs text 101 and important information 102, 1800 to the generator 100. The input device 2404 may be, for example, a keyboard or a mouse. The input from the input device 2404 may be stored in an auxiliary storage device 2402 or memory 2403.

[0119] The computer 2400 may have an output device 2405. The output device 2405 may be connected to, for example, a display or a printer, and will present the execution results output by the generation device 100 to the user. If the output device 2405 is, for example, a display or a printer, it can display the graph pattern 131. If the output device 2405 is a display, for example, it can display a screen for confirming the graph pattern 131.

[0120] The communication interface 2406 is a network interface device that controls communication with other devices according to a predetermined protocol. The communication interface 2406 includes, for example, a serial interface such as USB. The generation device 100 can send and receive data from any terminal via the network.

[0121] In the generation device 100, for example, the program executed by the processor 2401 may include an OS (Operating System) or arbitrary software. In this case, the OS or arbitrary software is stored, for example, in the auxiliary storage device 2402 and loaded into memory 2403 as needed.

[0122] Various forms are possible for the computer 2400. For example, the generation device 100 can be implemented on a single physical computer, or on a computer system consisting of multiple logically or physically configured computers. It may also operate on a virtual computer realized on multiple physical computer resources.

[0123] Furthermore, in the above-described embodiments 1 to 4, the abstraction unit 113 only needs to execute at least one of the following: attribute information omission determination process (step S401), attributeless node replacement process (step S402), quantifier node insertion process (step S403), exclusion condition setting process (step S404), and lemma dictionary application process (step S1705), as long as no inconsistencies arise. Also, the execution order of the attribute information omission determination process (step S401), attributeless node replacement process (step S402), quantifier node insertion process (step S403), exclusion condition setting process (step S404), and lemma dictionary application process (step S1705) may differ, as long as no inconsistencies arise.

[0124] For example, if an attribute-less node is found to have its attribute information omitted in the attribute information omission determination process (step S401), and an attribute-less node replacement process (step S402) is to be executed, the attribute information omission determination process (step S401) must always be executed before the attribute-less node replacement process (step S402).

[0125] Furthermore, the generating apparatus 100 according to the above-described Examples 1 to 4 can also be configured as shown in (1) to (15) below.

[0126] (1) The generation device 100 includes an acquisition unit (graph structure conversion unit 111) that acquires graph structure data 300 showing dependencies between nodes, with attribute information including information about words and phrases in a sentence and their parts of speech as nodes, and an abstraction unit 113 that abstracts the graph structure data 300 acquired by the acquisition unit based on the attribute information in the nodes.

[0127] (2) In the generation device 100 described in (1) above, the abstraction unit 113 abstracts the graph structure data 300a by deleting attribute information within a specific node and outputs the graph structure data 330ad.

[0128] (3) In the generation device 100 described in (2) above, the abstraction unit 113 abstracts the graph structure data 300a by deleting the attribute information of the specific node that corresponds to the first condition (attribute branching condition), and outputs the graph structure data 330ad.

[0129] (4) In the generation device 100 described in (3) above, the abstraction unit 113 abstracts the graph structure data 300a by deleting specific attribute information from the attribute information of the specific node that corresponds to the second condition (attribute branching condition and attribute omission of the same rule ID 601), and outputs the graph structure data 330ad.

[0130] (5) In the generation device 100 described in (3) above, the abstraction unit 113 abstracts the graph structure data 300a by using a machine learning model 1400 that has been trained based on the attribute information of the specific node that meets the first condition and the specific attribute information that meets the second condition, and by deleting the specific attribute information that is output as a result of inputting the attribute information of the specific node into the machine learning model 1400, thereby outputting the graph structure data 330ad.

[0131] (6) In the generation device 100 described in (1) above, the abstraction unit 113 abstracts the graph structure data 300ad by deleting nodes in the group of nodes in the graph structure data 300ad for which no attribute information exists, and outputs the graph structure data 330adq.

[0132] (7) In the generation device 100 described in (1) above, the abstraction unit 113 abstracts the graph structure data 300ad by replacing nodes in the group of nodes in the graph structure data 300ad that do not have attribute information with quantifier nodes (*) that allow the insertion of arbitrary attribute information, and outputs the graph structure data 330adq.

[0133] (8) In the generation device 100 described in (1) above, the abstraction unit 113 abstracts the graph structure data 300adq by replacing the attribute information of the lower node of the two dependent nodes with the quantifier node, and outputs the graph structure data 330adqi.

[0134] (9) In the generation device 100 described in (7) above, the abstraction unit 113 abstracts the graph structure data 300adq by replacing the attribute information of the lower-level node with the quantified child node when the attribute information of the lower-level node of the two dependent nodes is the first condition (part-of-speech attribute information of the parent node 1102) and the attribute information of the lower-level node of the two nodes is the second condition (part-of-speech attribute information of the child node 1103), thereby outputting the graph structure data 330adqi.

[0135] (10) In the generation device 100 of (7) above, the abstraction unit 113 uses a machine learning model 1400 that has been learned based on the attribute information of the upper node that meets the first condition and the attribute information of the lower node that meets the second condition of two dependent nodes to abstract the graph structure data 300adq and output graph structure data 330adqi by replacing the attribute information of the lower node that is output as a result of inputting the attribute information of the upper node into the machine learning model with the quantifier node.

[0136] (11) In the generation device 100 of (1) above, the abstraction unit 113 abstracts the graph structure data 300adqi by setting an extraction exclusion condition 1200 for any node of the graph structure data 300 that excludes the extraction of a predetermined word or phrase (for example, "is" and its nominative case), and outputs the graph structure data 330adqis.

[0137] (12) In the generation device 100 described in (11) above, the abstraction unit 113 abstracts the graph structure data 300adqi by setting an extraction exclusion condition 1200 that excludes the extraction of a predetermined phrase for a specific node n01adq of the graph structure data, and outputs the graph structure data 330adqis.

[0138] (13) In the generation device 100 of (1) above, the abstraction unit 113 abstracts the graph structure data 300 by referring to a lemma dictionary 1500 which classifies each of the word groups into semantic groups, and converting the words in the node into identification information (group ID 1601) of the group containing the words.

[0139] (14) In the generation device 100 of (1) above, the acquisition unit acquires the sentence and converts the sentence into graph structure data.

[0140] (15) The generation device 100 of (1) above further has an addition unit 112 that adds the identification information (location ID 201) to a node containing the phrase 202 by referring to important location information 102 which has the phrase 202 and its identification information (location ID 201).

[0141] It should be noted that the present invention is not limited to the embodiments described above, but includes various modifications and equivalent configurations within the spirit of the attached claims. For example, the embodiments described above are described in detail to make the present invention easier to understand, and the present invention is not necessarily limited to having all of the described configurations. Furthermore, some of the configurations of one embodiment may be replaced with those of another embodiment. Furthermore, some of the configurations of one embodiment may be added to those of another embodiment. Furthermore, some of the configurations of each embodiment may be added, deleted, or replaced with other configurations.

[0142] Furthermore, each of the aforementioned configurations, functions, processing units, and processing means may be implemented in hardware, for example, by designing them as integrated circuits, or they may be implemented in software by having a processor interpret and execute programs that realize each function.

[0143] Information such as programs, tables, and files that implement each function can be stored in memory, hard disks, SSDs (Solid State Drives), or on recording media such as IC (Integrated Circuit) cards, SD cards, and DVDs (Digital Versatile Discs).

[0144] Furthermore, the control lines and information lines shown are those deemed necessary for explanation purposes and do not necessarily represent all control lines and information lines required for implementation. In reality, it can be assumed that almost all components are interconnected. [Explanation of symbols]

[0145] 100 generator 101 Text 102 Important Information 111 Graph Structure Transformation Unit 112 Addition part 113 Abstraction part 121 Attribute Branching Condition Information 122 Attribute abbreviation information 123 List of Quantifier Node Insertion Conditions 131 Graph Patterns 202 words 300 graph structure data 1400 Machine Learning Models 1500 Lemma Dictionary

Claims

1. A storage unit that stores attribute branching conditions including part-of-speech attribute information indicating the attributes of a lemma and its part of speech, and important location information indicating whether the lemma is an important location, and attribute omission information that defines for each attribute branching condition whether the words and phrases corresponding to the lemma and the part-of-speech attribute information can be omitted, An acquisition unit acquires graph structure data consisting of nodes containing words and phrases within a sentence and their part-of-speech attribute information, and dependencies between said nodes, with the important location information assigned to some of the nodes. An abstraction unit outputs a graph pattern abstracted from the sentence by removing, from the nodes constituting the graph structure data acquired by the acquisition unit, the words and phrases in the sentence that correspond to the lemma that can be omitted by the attribute omission information and the part-of-speech attribute information that can be omitted by the attribute omission information, among the nodes that correspond to the attribute branching condition. A generating apparatus characterized by having the following features.

2. The generating apparatus according to claim 1, The abstraction unit replaces specific nodes among the nodes constituting the graph structure data that do not contain the word and part-of-speech attribute information with quantifier nodes that allow inclusion in parent nodes that have a dependency on the specific node. A generating apparatus characterized by the following features.

3. The generating apparatus according to claim 2, The abstraction unit, if there is no parent node that has a dependency on the specific node, deletes the specific node without replacing it with the quantifier node. A generating apparatus characterized by the following features.

4. The generating apparatus according to claim 2, The abstraction unit, if the parent node that has a dependency on the specific node also does not contain the phrase and the part-of-speech attribute information, deletes the specific node without replacing it with the quantifier node. A generating apparatus characterized by the following features.

5. The generating apparatus according to claim 2, The abstraction unit inserts the quantifier nodes between the nodes with dependencies. A generating apparatus characterized by the following features.

6. The generating apparatus according to claim 5, The storage unit stores the additional rules for the quantification child node, which are defined by the parent node and the part-of-speech attribute information of the child node that has a dependency on the part-of-speech attribute information of the parent node. The abstraction unit inserts the quantifier node between the dependent nodes if the additional rule applies between the dependent nodes. A generating apparatus characterized by the following features.

7. The generating apparatus according to claim 5, The abstraction unit sets an exclusion condition in the quantifier node, which includes a phrase specifying the target to be excluded from extraction and its part-of-speech attribute information. A generating apparatus characterized by the following features.

8. The generating apparatus according to claim 7, The abstraction unit sets the extraction exclusion conditions on the quantifier node that has been replaced from the node to which the important location information has been assigned. A generating apparatus characterized by the following features.

9. The generating apparatus according to claim 1, The storage unit stores identification information that identifies a group of multiple lemmas, and a lemma dictionary which has multiple lemmas included in the group for each group. The abstraction unit converts the phrases within the node into the identification information of the group that contains the lemma corresponding to the phrase. A generating apparatus characterized by the following features.

10. The generating apparatus according to claim 1, The storage unit stores a machine learning model that has been trained using the attribute branching conditions as training data and the attribute omission information as ground truth data, instead of the attribute branching conditions and the attribute omission information. The abstraction unit outputs an abstract graph pattern of the sentence by removing the words corresponding to the lemma that can be omitted by the attribute omission information and the part-of-speech attribute information that can be omitted by the attribute omission information from the nodes constituting the graph structure data that meet the attribute branching conditions, based on the machine learning model. A generating apparatus characterized by having the following features.

11. The generating apparatus according to claim 6, The memory unit stores a machine learning model in which the addition rules for the quantifier node are learned, defined by the parent node and the part-of-speech attribute information of the child nodes that have a dependency relationship with the part-of-speech attribute information of the parent node, using the part-of-speech attribute information as training data and the presence or absence of insertion of the quantifier node as ground truth data. The abstraction unit inserts the quantifier nodes between the dependent nodes if the additional rule applies between the dependent nodes based on the machine learning model. A generating apparatus characterized by the following features.

12. A generation method using a generation apparatus having a processor for executing a program and a storage device for storing the program, The storage device stores attribute branching conditions including part-of-speech attribute information indicating the lemma and its part-of-speech attributes, and important location information indicating whether the lemma is an important location, and attribute omission information that defines for each attribute branching condition whether the word or phrase corresponding to the lemma and the part-of-speech attribute information can be omitted. The aforementioned processor, An acquisition process to acquire graph structure data consisting of nodes containing words and phrases within a sentence and their part-of-speech attribute information, and dependencies between said nodes, with the important location information attached to some of the nodes; An abstraction process outputs a graph pattern that abstracts the sentence by removing, from the nodes constituting the graph structure data obtained by the acquisition process, the words and phrases in the sentence that correspond to the lemma that can be omitted by the attribute omission information and the part-of-speech attribute information that can be omitted by the attribute omission information, among the nodes that meet the attribute branching conditions. A generation method characterized by performing the following.

13. A generation program that is executed by the processor, The processor is able to refer to a storage device that stores attribute branching conditions including part-of-speech attribute information indicating the attributes of a lemma and its part of speech, and important location information indicating whether or not the lemma is an important location, and attribute omission information that defines for each attribute branching condition whether or not the words and phrases corresponding to the lemma and the part-of-speech attribute information can be omitted. The aforementioned processor, An acquisition process to acquire graph structure data consisting of nodes containing words and phrases within a sentence and their part-of-speech attribute information, and dependencies between said nodes, with the important location information attached to some of the nodes; An abstraction process outputs a graph pattern that abstracts the sentence by removing, from the nodes constituting the graph structure data obtained by the acquisition process, the words and phrases in the sentence that correspond to the lemma that can be omitted by the attribute omission information and the part-of-speech attribute information that can be omitted by the attribute omission information, among the nodes that meet the attribute branching conditions. A generation program characterized by causing the execution of a specific action.

Images