JP2024041369A - Similarity determination device, similarity determination system, and similarity determination method - Google Patents

Abstract

To provide similarity determination means capable of comparing sentence similarity at high speed and with high accuracy.SOLUTION: Provided is a similarity determination device, including: a keyword management database that stores a set of first domain keywords corresponding to a first domain; a domain graph database that stores a first domain graph generated based on the set of first domain keywords for a first sentence corresponding to the first domain; a graph creation part that generates a second domain graph for a second sentence corresponding to the first domain on the basis of the first set of domain keywords; and a similarity determination part that generates a comparison result indicating similarity between the first domain graph and the second domain graph by comparing the first domain graph with the second domain graph.SELECTED DRAWING: Figure 2

Description

The present invention relates to a similarity determination device, a similarity determination system, and a similarity determination method.

In recent years, cyber attacks have continued to evolve year by year due to the discovery of new vulnerabilities and the emergence of new attack methods. Under these circumstances, the question of whether information systems have the ability to defend against cyber-attacks and how far the damage will spread in the event of a cyber-attack has become an important concern for society and organizations.

Therefore, in order to understand the risks to business, organizations such as companies, for example, obtain various reports on newly discovered vulnerabilities and cyber attacks (hereinafter referred to as cyber events), and take appropriate actions based on these reports. We are required to take appropriate measures.

In order to take appropriate measures against cyber events that may pose a risk, it is desirable to determine whether the reported cyber event has already been dealt with or is currently being dealt with. If a reported cyber event is similar to (or identical to) a cyber event for which countermeasures have already been developed, there is no need to develop countermeasures again, thereby saving time and resources required for countermeasure development.

One way to determine whether a cyber event has already been responded to or is currently being addressed is to determine whether there are other reports that are highly similar to the report reporting vulnerabilities or cyber attacks. It will be done.

Several proposals have been made to compare sentences and determine the degree of similarity.
For example, Rettinger et al. (Non-Patent Document 1) states that "Evaluating the relevance of documents is the core of many applications, such as document retrieval and recommendation. Most similarity approaches work with word distribution-based document representations. Although these approaches are computationally fast, they have problems when the documents have different languages, vocabularies, or types, and they ignore the rich relational knowledge available in the knowledge graph. ,While graph-based document models can leverage valuable knowledge,about relationships between entities, graph manipulation tends to be,resource intensive, making similarity evaluation infeasible in,many applications. This paper presents an efficient semantic similarity approach that exploits explicit hierarchical and cross-cutting relationships. In our experiments, we demonstrate that (i) the similarity measure of our approach is more effective than the equivalent measure for humans; has a significantly higher correlation with document similarity perception, (ii) this also applies to short documents with fewer annotations, and (iii) document similarity is significantly higher than other graph traversal-based approaches. The technology is described as "We show that calculations can be performed efficiently."

Paul C., Rettinger A., Mogadala A., Knoblock C.A., Szekely P. (2016) Efficient Graph-Based Document Similarity. In: Sack H., Blomqvist E., d'Aquin M., Ghidini C., Ponzetto S ., Lange C. (eds) The Semantic Web. Latest Advances and New Domains. ESWC 2016. Lecture Notes in Computer Science, vol 9678. Springer, Cham. https://doi.org/10.1007/978-3-319- 34129-3_21

The above-mentioned Non-Patent Document 1 describes a means for determining the similarity of sentences by utilizing hierarchical and cross-sectional relationships in the sentences and using graph traversal means.

However, with the method described in Non-Patent Document 1, the entire sentence is expressed as a graph, so the longer the sentence, the larger the graph, and the computer required to store and process the graph. The problem is that it requires a lot of resources and time.
Therefore, when the method described in Non-Patent Document 1 is applied to a report reporting a cyber event, for example, depending on the size of the graph, the time until the comparison judgment result is output is delayed, and the result of the comparison is delayed. The formulation of countermeasures may be delayed.

Therefore, the present disclosure provides a similarity determination means that can perform high-speed and highly accurate sentence similarity comparisons by suppressing the scale of a graph representation corresponding to a sentence while maintaining relationship information between elements in a sentence. The purpose is to

In order to solve the above problems, one of the typical similarity determination devices of the present invention includes a keyword management database that stores a set of first domain keywords corresponding to a first domain, and a keyword management database that stores a set of first domain keywords corresponding to a first domain. A domain graph database that stores a first domain graph generated based on the first set of domain keywords for a first sentence corresponding to a domain; A graph generating unit that generates a second domain graph for a second sentence corresponding to the first domain, and comparing the first domain graph and the second domain graph, and a similarity determination unit that generates a comparison result indicating the degree of similarity between the graph and the second domain graph.

According to the present disclosure, there is provided a similarity determination means that can perform high-speed and highly accurate sentence similarity comparisons by suppressing the scale of a graph representation corresponding to a sentence while maintaining relational information between elements in a sentence. be able to.
Problems, configurations, and effects other than those described above will be made clear by the description in the detailed description below.

FIG. 1 is a diagram illustrating a computer system for implementing embodiments of the present disclosure. FIG. 2 is a diagram illustrating an example of a configuration of a similarity determination system according to an embodiment of the present disclosure. FIG. 3 is a diagram showing the flow of preliminary graph generation processing and the flow of graph comparison processing in the similarity determination processing according to the embodiment of the present disclosure. FIG. 4 is a diagram illustrating details of the flow of advance graph generation processing according to the embodiment of the present disclosure. FIG. 5 is a diagram illustrating an example of a domain graph and a normalized domain graph according to an embodiment of the present disclosure. FIG. 6 is a diagram illustrating details of the flow of graph comparison processing according to the embodiment of the present disclosure. FIG. 7 is a diagram illustrating a flow of update processing for updating domain keywords and domain graphs according to the disclosed embodiment. FIG. 8 is a diagram illustrating an example of a keyword management screen according to an embodiment of the present disclosure. FIG. 9 is a diagram illustrating an example of a text management screen according to an embodiment of the present disclosure. FIG. 10 is a diagram illustrating an example of the flow of multi-domain management processing according to the embodiment of the present disclosure. FIG. 11 is a diagram illustrating a specific example of similarity determination for two domain graphs according to an embodiment of the present disclosure.

Embodiments of the present invention will be described below with reference to the drawings. Note that the present invention is not limited to this embodiment. In addition, in the description of the drawings, the same parts are denoted by the same reference numerals.
Additionally, terms such as "first,""second," and "third" may be used in the present disclosure to describe various elements or components; It will be understood that there should be no limitation by the terms. These terms are only used to distinguish one element or component from another. Accordingly, a first element or component discussed below may also be referred to as a second element or component without departing from the teachings of the inventive concept.

First, with reference to FIG. 1, a computer system 100 for implementing an embodiment of the present disclosure will be described. The mechanisms and apparatus of the various embodiments disclosed herein may be applied to any suitable computing system. The main components of computer system 100 include one or more processors 102 , memory 104 , terminal interface 112 , storage interface 113 , I/O (input/output) device interface 114 , and network interface 115 . These components may be interconnected via memory bus 106, I/O bus 108, bus interface unit 109, and I/O bus interface unit 110.

Computer system 100 may include one or more general purpose programmable central processing units (CPUs) 102A and 102B, collectively referred to as processors 102. In some embodiments, computer system 100 may include multiple processors, and in other embodiments, computer system 100 may be a single CPU system. Each processor 102 executes instructions stored in memory 104 and may include onboard cache.

In some embodiments, memory 104 may include random access semiconductor memory, storage devices, or storage media (either volatile or nonvolatile) for storing data and programs. Memory 104 may store all or a portion of programs, modules, and data structures that perform the functions described herein. For example, the memory 104 may store a similarity determination application 150. In some embodiments, similarity determination application 150 may include instructions or writing to perform functions described below on processor 102.

In some embodiments, the similarity determination application 150 operates on semiconductor devices, chips, logic gates, circuits, circuit cards, and/or other physical hardware instead of or in addition to processor-based systems. It may also be implemented in hardware via a device. In some embodiments, similarity determination application 150 may include data other than instructions or descriptions. In some embodiments, cameras, sensors, or other data input devices (not shown) may be provided to communicate directly with bus interface unit 109, processor 102, or other hardware of computer system 100. .

Computer system 100 may include a bus interface unit 109 that provides communication between processor 102 , memory 104 , display system 124 , and I/O bus interface unit 110 . I/O bus interface unit 110 may be coupled to I/O bus 108 for transferring data to and from various I/O units. I/O bus interface unit 110 connects via I/O bus 108 to a plurality of I/O interface units 112, 113, 114, also known as I/O processors (IOPs) or I/O adapters (IOAs). and 115.

Display system 124 may include a display controller, display memory, or both. A display controller may provide video, audio, or both data to display device 126. Computer system 100 may also include devices, such as one or more sensors, configured to collect data and provide the data to processor 102.

For example, the computer system 100 may include a biometric sensor that collects heart rate data, stress level data, etc., an environmental sensor that collects humidity data, temperature data, pressure data, etc., and a motion sensor that collects acceleration data, exercise data, etc. May include. Other types of sensors can also be used. Display system 124 may be connected to a display device 126, such as a standalone display screen, a television, a tablet, or a handheld device.

The I/O interface unit has the ability to communicate with various storage or I/O devices. For example, the terminal interface unit 112 may include a user output device such as a video display device, a speaker television, or a user input device such as a keyboard, mouse, keypad, touchpad, trackball, buttons, light pen, or other pointing device. It is possible to attach user I/O devices 116 such as: Using the user interface, a user operates user input devices to input input data and instructions to user I/O device 116 and computer system 100, and to receive output data from computer system 100. Good too. The user interface may be displayed on a display device, played through a speaker, or printed through a printer, for example, via the user I/O device 116.

Storage interface 113 may include one or more disk drives or direct access storage devices 117 (typically magnetic disk drive storage devices, but also an array of disk drives or other storage devices configured to appear as a single disk drive). ) can be installed. In some embodiments, storage device 117 may be implemented as any secondary storage device. The contents of memory 104 are stored in storage device 117 and may be read from storage device 117 as needed. I/O device interface 114 may provide an interface to other I/O devices such as printers, fax machines, etc. Network interface 115 may provide a communication path so that computer system 100 and other devices can communicate with each other. This communication path may be, for example, network 130.

In some embodiments, computer system 100 is a device that receives requests from other computer systems (clients) that do not have a direct user interface, such as a multi-user mainframe computer system, a single-user system, or a server computer. There may be. In other embodiments, computer system 100 may be a desktop computer, a portable computer, a laptop, a tablet computer, a pocket computer, a telephone, a smart phone, or any other suitable electronic device.

Next, with reference to FIG. 2, a similarity determination system according to an embodiment of the present disclosure will be described.

FIG. 2 is a diagram illustrating an example of a configuration of a similarity determination system 200 according to an embodiment of the present disclosure. The similarity determination system 200 generates a comparison result indicating the similarity between multiple sentences,
This is a system for providing information to users. As shown in FIG. 2, the similarity determination system 200 includes a similarity determination device 210, a communication network 250, and a user terminal 260. Similarity determination device 210 and user terminal 260 may be connected to each other via communication network 250.

The similarity determination device 210 is a device for determining the similarity of sentences, and as shown in FIG. 2, mainly includes a memory 220, a storage section 230, a processor 244, and an input/output section 246.
In some embodiments, similarity determination device 210 may be implemented by computer system 100 shown in FIG. 1.

The memory 220 may be a memory for storing the similarity determination application 150 for implementing the function of the similarity determination means according to the embodiment of the present disclosure. As shown in FIG. 2, the similarity determination application 150 may include processing instructions for implementing the functions of software modules such as the graph generation section 221, the similarity determination section 222, and the update section 223.

The graph generation unit 221 is a functional unit for generating a domain graph corresponding to a target sentence. The "domain graph" in the present disclosure may be a directed acyclic graph in which unique domain keywords in a target sentence are represented as nodes, and domain keywords are represented as edges. Furthermore, the term "domain" in the present disclosure means a specific field or topic, and may include any topic such as "famous physicists,""automobiles,""DoSattacks," and the like. By comparing domain graphs generated for different sentences, it is possible to determine the degree of similarity between the sentences.
Note that the details of the function of the graph generation unit 221 will be described later, so a description thereof will be omitted here.

The similarity determination unit 222 is a functional unit that determines the degree of similarity between a plurality of sentences using the domain graph generated by the graph generation unit 221, and outputs a comparison result indicating the determined degree of similarity.
Note that the details of the function of the similarity determination unit 222 will be described later, so the description thereof will be omitted here.

The update unit 223 adds new domain keywords and deletes existing domain keywords to the domain keywords stored in the keyword management DB 231 (described later), and also adds new domain keywords to the domain keywords stored in the domain graph DB 232 based on changes in domain keywords. This is a functional unit for updating the domain graph.
Note that the details of the function of the update unit 223 will be described later, so the description thereof will be omitted here.

The storage unit 230 is a storage area that accommodates a database (hereinafter referred to as "DB") for storing various information according to the embodiment of the present disclosure, and as shown in FIG. May include.

The keyword management DB 231 is a database for storing a set of domain keywords according to the embodiment of the present disclosure. A "domain keyword" in the present disclosure is a word that is particularly important in a specific domain. These domain keywords may be selected by a user (for example, a user of the user terminal 260 described below) who requests a comparison result indicating the degree of similarity between sentences. As an example, in a domain with "particle physics", "Higgs boson" may be selected as a domain keyword. In some embodiments, keyword management DB 231 may store a set of domain keywords corresponding to various different domains. As described below, these domain keywords are used when generating a domain graph according to embodiments of the present disclosure.

The domain graph DB 232 is a database for storing domain graphs generated by the graph generation unit 221. As will be described later, a domain graph (for example, a first domain graph) stored in the domain graph DB 232 is used as a comparison target for a newly generated domain graph (for example, a second domain graph).

The processor 244 is a processing unit for executing processing instructions that define the functions of each functional unit of the similarity determination application 150 stored in the memory 220.

The input/output unit 246 is a functional unit that receives information input to the similarity determination device 210 and outputs information such as comparison results generated by the similarity determination device 210. In one embodiment, the input/output unit 246 may include, for example, a keyboard, a mouse, a display that displays a GUI (Graphical User Interface), and the like.

Communication network 250 may include, for example, a local area network (LAN), a wide area network (WAN), a satellite network, a cable network, a WiFi network, or any combination thereof.

The user terminal 260 is a terminal device that can be used by the user of the similarity determination device 210. By using the user terminal 260, the user can, for example, use the GUI provided by the input/output unit 246 to request a comparison result showing the similarity between sentences, select a domain keyword, and check the comparison result. You can do it. As an example, the user terminal 260 may include, for example, a smartphone, a smart watch, a tablet, a personal computer, etc., and is not particularly limited.
Note that in FIG. 2, for convenience of explanation, a configuration including one user terminal 260 is described as an example, but the number of user terminals 260 is not particularly limited.

According to the similarity determination device 210 described above, by suppressing the scale of the graph representation corresponding to the sentences while maintaining relationship information between elements in the sentences, the similarity determination device 210 enables high-speed and highly accurate sentence similarity comparison. It is possible to provide a degree determination means.

Next, with reference to FIG. 3, the overall flow of the similarity determination process according to the embodiment of the present disclosure will be described.

As described above, one aspect of the embodiment of the present disclosure relates to processing for determining the degree of similarity between multiple sentences. Further, this similarity determination process mainly includes a preliminary graph generation process 400 and a graph comparison process 600. FIG. 3 is a diagram showing the flow of preliminary graph generation processing 400 and the flow of graph comparison processing 600 in the similarity determination processing according to the embodiment of the present disclosure.

The preliminary graph generation process 400 generates in advance a domain graph (for example, a normalized first domain graph) that is a reference when comparing sentences based on an existing sentence (first sentence), and This is a process for storing in the graph DB 232. More specifically, as shown in FIG. 3, in the pre-graph generation process 400, the graph generation unit 221 generates a first sentence 412 corresponding to the first domain and the first sentence 412 that is stored in the keyword management DB 231. A normalized first domain graph 415 representing the first sentence 412 in a graph format is generated using the first domain keyword corresponding to the domain and stored in the domain graph DB 232.

Graph comparison processing 600 generates a new domain graph (for example, a normalized second domain graph) corresponding to the target sentence (second sentence), and generates a new domain graph (for example, the normalized second domain graph) corresponding to the target sentence (second sentence), and generates a new domain graph (for example, the normalized second domain graph) corresponding to the target sentence (second sentence), and The similarity between the normalized first domain graph and the normalized second domain graph is determined by comparing the normalized first domain graph generated by the preliminary graph generation process 400. This is the process for generating the comparison results shown.

More specifically, as shown in FIG. 3, in the graph comparison process 600, the graph generation unit 221 generates a second sentence 422 corresponding to the first domain and the first sentence stored in the keyword management DB 231. A normalized second domain graph 425 showing the second sentence 422 in a graph format is generated using the first domain keyword corresponding to the domain. Thereafter, the similarity determination unit 222 performs normalization by comparing the normalized first domain graph 415 stored in the domain graph DB 232 and the newly generated normalized second domain graph 425. A comparison result indicating the degree of similarity between the normalized first domain graph and the normalized second domain graph is generated.

Note that although the general flow of the preliminary graph generation process 400 and the graph comparison process 600 included in the similarity determination process in the embodiment of the present disclosure has been described above, the details of these processes will be described later, so they will not be described here. The explanation will be omitted.

Next, details of the preliminary graph generation process according to the embodiment of the present disclosure will be described with reference to FIG. 4.

FIG. 4 is a diagram illustrating details of the flow of preliminary graph generation processing 400 according to the embodiment of the present disclosure. As described above, the preliminary graph generation process 400 is a process for generating a normalized first domain graph in advance and storing it in the domain graph DB 232, and mainly generates a graph of the similarity determination device 210. This is performed by the generation unit 221.

First, in step S410, the graph generation unit 221 extracts a set of domain keywords corresponding to the domain of the first sentence 412 from the keyword management DB 231, and among the set of extracted domain keywords, the graph generation unit 221 extracts a set of domain keywords corresponding to the domain of the first sentence 412, Identify a subset of domain keywords that will be used.

As an example, if the first sentence 412 is related to the domain "physics", the graph generation unit 221 extracts a set of domain keywords corresponding to "physics" from the keyword management DB 231. If this set of domain keywords includes "atom," "gravity," "neutron," "charge," "electron," and "uncertainty principle," the graph generation unit 221 converts these domain keywords into the first Search and specify in the sentence 412 of. For example, if only "atom," "neutron," and "electron" from the above set of domain keywords appear in the first sentence 412, then "atom," "neutron," and "electron" are included as a subset of the domain keywords. Identify.

Next, in step S420, the graph generation unit 221 generates nodes based on the subset of domain keywords identified in step S410. A node here is a data structure for expressing a subset of domain keywords in a domain graph described later. In some embodiments, the graph generator 221 may create a node for each unique domain keyword in the subset of domain keywords identified in the first sentence 412. As an example, when "atom", "neutron", and "electron" are specified as a subset of domain keywords, the graph generation unit 221 generates a node containing nodes corresponding to "atom", "neutron", and "electron", respectively. may generate a set of

ここで、Ｃｏｕｎｔ(Ｎｉ)は、特定のノードＮに対応するドメインキーワードの特定の文章（例えば、第１の文章４１２）における出現回数であり、Ｔは、グラフにおけるノートの総数である。
なお、以上では、ノードスコアをドメインキーワードの文章における出現回数に基づいて計算した場合を一例として説明したが、本開示はこれに限定されず、例えば特定のドメインキーワードの意味的コサイン値に基づいてノードスコアを計算することも可能である。 Next, in step S430, the graph generation unit 221 calculates a node score for each node generated in step S420. A "node score" in the present disclosure is a quantitative measure indicating the relative importance of a domain keyword corresponding to a specific node in a sentence, and the higher the importance, the higher the node score. In one embodiment, the graph generation unit 221 may calculate the node score of a particular node based on the number of times a domain keyword corresponding to a particular node appears in the first sentence 412 (number of appearances). In this case, the graph generation unit 221 may calculate the node score N _i (N ₁ , N ₂ , . . . N _T ) of each node using Equation 1 shown below.

Here, Count(Ni) is the number of appearances of the domain keyword corresponding to the specific node N in a specific sentence (eg, first sentence 412), and T is the total number of notes in the graph.
Note that, although the above example describes a case where the node score is calculated based on the number of occurrences of the domain keyword in the sentence, the present disclosure is not limited to this, and for example, the node score is calculated based on the semantic cosine value of a specific domain keyword. It is also possible to calculate node scores.

Next, in step S440, the graph generation unit 221 generates edges indicating the relationship between the set of nodes generated in step S420. The edge here is a data structure for expressing relationships between nodes in a domain graph. In one embodiment, the graph generation unit 221 may create a knowledge graph corresponding to the first sentence in advance, and generate edges indicating relationships between sets of nodes based on the knowledge graph.
Generally, a knowledge graph is a data structure that systematically connects various pieces of knowledge and represents it in a graph structure. Here, the means for generating the knowledge graph is not particularly limited, and any existing means such as natural language processing or neural networks may be used.

ここで、Ｌ_ijは、知識グラフにおいて、ノードｉに対応するドメインキーワードとノードｊに対応するドメインキーワードの間の接続数である。知識グラフにおいて、２つのドメインキーワード間の接続数が少なければ少ない程、関連度が高いため、ここでは、説明の便宜上、接続数の逆（１/接続数）をエッジ重みＥ_ijとする。従って、エッジ重みＥ_ijは、２つのドメインキーワード間の関連度が高ければ高い程、高くなる。 Next, in step S450, the graph generation unit 221 calculates edge weights for each edge generated in step S440. "Edge weight" in this disclosure is a measure that quantitatively indicates the degree of association of domain keywords corresponding to two nodes connected by an edge in a domain graph, and the higher the degree of association of domain keywords, the higher the degree of association between domain keywords. Edge weight increases.
In some embodiments, the graph generation unit 221 may calculate edge weights based on a knowledge graph created in advance based on the first sentence. In this case, the graph generation unit 221 may calculate edge weights based on the shortest distance (number of connections) between domain keywords corresponding to two nodes in the knowledge graph. In one embodiment, the graph generation unit 221 may calculate the edge weight E _ij of the edge between node i and node j using Equation 2 shown below.

Here, L _ij is the number of connections between the domain keyword corresponding to node i and the domain keyword corresponding to node j in the knowledge graph. In the knowledge graph, the smaller the number of connections between two domain keywords, the higher the degree of association; therefore, for convenience of explanation, here, the inverse of the number of connections (1/number of connections) is assumed to be the edge weight E _ij . Therefore, the higher the degree of association between two domain keywords, the higher the edge weight E _ij becomes.

Next, in step S460, the graph generation unit 221 generates the set of nodes generated in step S420, the node scores calculated in step S430, the set of edges generated in step S440, and the edge weights calculated in step S450. to generate a first domain graph. This first domain graph is a directed acyclic graph in which the subset of domain keywords specified in the first sentence 412 is represented as nodes, and the spaces between domain keywords are represented as edges. Furthermore, each node in this first domain graph is associated with a node score that indicates the importance of the domain keyword corresponding to the node, and each edge is associated with an edge weight that indicates the degree of association between the nodes that the edge connects. can be mapped to.

Next, in step S470, the graph generation unit 221 generates a normalized first domain graph by normalizing the first domain graph 415 generated in step S460. More specifically, the graph generation unit 221 generates the normalized first domain graph by normalizing the node score of each node and the edge score of each edge in the normalized first domain graph. May be generated. Here, the graph generation unit 221 may calculate the normalized node score NS _i based on Equation 3 shown below, and calculate the normalized edge weight NE _ij based on Equation 4 shown below.

Next, in step S480, the graph generation unit 221 associates the normalized first domain graph generated in step S470 with the first sentence 412 and the subset of domain keywords identified in step S410. It is stored in the domain graph DB 232.

By performing the preliminary graph generation process 400 described above on various sentences, domain graphs corresponding to many different sentences can be prepared in advance and stored in the domain graph DB 232. As will be described later, the normalized first domain graph generated by the pre-graph generation process 400 is compared to the normalized second domain graph generated based on the second sentence in the graph comparison process 600, which will be described later. Used for comparison with domain graph. Therefore, by performing the preliminary graph generation process 400 on various sentences and preparing a large number of domain graphs for comparison, it is possible to perform a broader similarity determination, and it is possible to determine the degree of similarity in a wider range of ways. The first sentence becomes easier to identify.

Next, with reference to FIG. 5, a domain graph and a normalized domain graph according to an embodiment of the present disclosure will be described.

FIG. 5 is a diagram illustrating an example of a domain graph 510 and a normalized domain graph 520 according to an embodiment of the present disclosure.

As described above, the domain graph in the present disclosure may be a directed acyclic graph in which unique domain keywords in a target sentence are represented as nodes, and spaces between domain keywords are represented as edges.

As an example, as shown in FIG. 5, domain graph 510 may include a number of nodes, such as node "i," node "i+1," node "i+x," and so on. Further, each node in the domain graph is associated with a node score indicating the relative importance of the node, and each edge is associated with an edge weight indicating the degree of association of nodes connected by the edge. For example, a node "i" is associated with a node score "Ni", and an edge connecting node "i" and node "i+1" is associated with an edge weight of "E _i,i+1 ".

Further, by normalizing each node score and edge weight in the domain graph 510, a normalized domain graph 520 can be generated. As shown in FIG. 5, each node in the normalized domain graph 520 is associated with a normalized node score, and each edge is associated with a normalized edge weight. For example, node "i" is associated with the normalized node score "NSi", and the edge connecting node "i" and node "i+1" corresponds to the edge weight of "NE _i,i+1 ". Can be attached.

In this way, by representing a sentence as a domain graph, it becomes possible to determine similarity with higher accuracy.
Furthermore, as described above, each node in the domain graph is associated with a node score that indicates the importance of the domain keyword corresponding to the node in the sentence, and each edge in the domain graph is associated with the node connected by the edge. It can be associated with edge weights that indicate the degree of relevance of the corresponding domain keyword. This makes it possible to reduce the size of the domain graph while maintaining semantic information regarding particularly important keywords in the text, without having to express the entire text as a domain graph.
Furthermore, by generating a normalized domain graph in which each node score and edge weight in the domain graph are normalized, when comparing domain graphs, it is possible to avoid a decline in similarity judgment due to differences in the size of domain graphs. It can be prevented.

Next, details of the graph comparison process according to the embodiment of the present disclosure will be described with reference to FIG. 6.

FIG. 6 is a diagram illustrating details of the flow of graph comparison processing 600 according to the embodiment of the present disclosure. As described above, the graph comparison process 600 according to the embodiment of the present disclosure generates a new domain graph (for example, a normalized second domain graph) corresponding to the target sentence (second sentence). , by comparing it with an existing domain graph (for example, the normalized first domain graph generated by the pre-graph generation process 400 shown in FIG. This is a process for generating a comparison result 685 indicating the degree of similarity of the converted second domain graph. Further, the graph comparison process 600 is mainly performed by the graph generation unit 221 and the similarity determination unit 222.
In the graph comparison process 600 described below, the process of generating the normalized second domain graph corresponding to the second sentence (steps S610 to S670) is substantially the same as the preliminary graph generation process 400 shown in FIG. Since they are essentially the same, repetitive explanation will be omitted.

First, in step S605, the input/output unit 246 receives an input including the second text 422 and information indicating the domain name of the second text 422 from the user terminal 260. This second sentence 422 may be the same sentence as the first sentence 412 shown in FIG. 4, or may be a different sentence. In some embodiments, this second sentence 422 may be a sentence for which similarity determination with respect to an existing sentence is desired. The input/output unit 246 transfers the received second sentence 422 to the graph generation unit 221 and also transfers information indicating the domain name of the second sentence 422 to the keyword management DB 231 and domain graph DB 232.

Next, in step S610, the graph generation unit 221 extracts a set of domain keywords corresponding to the domain of the second sentence 422 from the keyword management DB 231, and extracts a set of domain keywords corresponding to the domain of the second sentence 422 from the set of extracted domain keywords. A subset of included domain keywords (a first subset of domain keywords) is identified.

Next, in step S620, the graph generation unit 221 generates nodes (first node set) based on the subset of domain keywords identified in step S610.

Next, in step S630, the graph generation unit 221 calculates a node score for each node generated in step S620. Here, in order to calculate the node score, Equation 1 described above may be used.

Next, in step S640, the graph generation unit 221 generates edges indicating the relationship between the set of nodes generated in step S620.

Next, in step S650, the graph generation unit 221 calculates edge weights for each edge generated in step S640. Here, in order to calculate the edge weight, the knowledge graph generated based on the second sentence 422 and Equation 2 described above may be used.

Next, in step S660, the graph generation unit 221 generates the set of nodes generated in step S620, the node scores calculated in step S630, the set of edges generated in step S640, and the edge weights calculated in step S650. to generate a second domain graph.

Next, in step S670, the graph generation unit 221 generates a normalized second domain graph by normalizing the second domain graph generated in step S660. Here, in order to normalize the second domain graph and generate a normalized second domain graph, the graph generation unit 221 may use Equation 3 and Equation 4 described above.

Next, in step S680, the similarity determination unit 222 acquires the normalized first domain graph generated by the advance graph generation process 400 shown in FIG. The degree of similarity between the first domain graph and the normalized second domain graph generated in step S670 is determined.

ここで、f₁は正規化済みの第１のドメイングラフであり、f_２は正規化済みの第２のドメイングラフであり、ＮＳ_iは正規化済みのノードスコアである。 Here, first, the similarity determination unit 222 determines the normalized first domain graph and the normalized domain graph based on the normalized first domain graph and the normalized second domain graph. A common node score (CNS) indicating the importance of a node common to the completed second domain graph is calculated using Equation 5 below.

Here, f ₁ is the normalized first domain graph, f ₂ is the normalized second domain graph, and NS _i is the normalized node score.

Next, the similarity determination unit 222 determines the normalized first domain graph and the normalized second domain graph based on the normalized first domain graph and the normalized second domain graph. A relationship score (Relationship score for common nodes; Rs) indicating the degree of association of edges between nodes that are common to the second domain graph is calculated using Equation 6 below.

Next, the similarity determination unit 222 uses the calculated CNS and Rs to calculate the similarity between the normalized first domain graph and the normalized second domain graph using the following formula 7. calculate. This similarity may be expressed as a percentage. Therefore, "100%" similarity is between the normalized first domain graph and the normalized second domain graph (and between the first sentence and second sentence that are the source of the graph). ) means that they are the same.

Next, the similarity determination unit 222 generates a comparison result 685 indicating the similarity calculated here, and transmits it to the user terminal 260.

According to the graph comparison processing 600 described above, by maintaining the relational information between elements in a sentence and suppressing the scale of the graph representation corresponding to the sentence, the similarity allows for high-speed and highly accurate sentence similarity comparison. Determination means can be provided.

Next, with reference to FIG. 7, processing for updating domain keywords and domain graphs according to an embodiment of the present disclosure will be described.

As described above, the keyword management DB 231 according to the embodiment of the present disclosure can store sets of domain keywords corresponding to various different domains. Further, similarity determination according to the embodiment of the present disclosure is performed using a domain graph generated based on these domain keywords, so in order to promote highly accurate similarity determination, it is necessary to It may be desirable to update the set of domain keywords by adding new domain keywords, deleting existing domain keywords, etc. to the corresponding set of domain keywords.

Further, when a specific set of domain keywords stored in the keyword management DB 231 is updated, it is desirable to update a domain graph generated based on the set of domain keywords and stored in the domain graph DB 232.
Therefore, one aspect of the present disclosure updates the set of domain keywords stored in the keyword management DB 231 and the domain graph stored in the domain graph DB 232 based on an update request input by the user via the user terminal 260. Concerning what to do. FIG. 7 is a diagram illustrating a flow of an update process 700 for updating domain keywords and domain graphs according to a disclosed embodiment.

First, the input/output unit 246 receives an update request from the user terminal 260 requesting an update to a specific set of domain keywords (for example, the first set of domain keywords) stored in the domain graph DB 232 . The update request here may be a user input requesting addition of a new domain keyword, deletion of an existing domain keyword, or modification of an existing domain keyword for a particular set of domain keywords. In some embodiments, this update request may be input via a user interface presented to the user by input/output unit 246.

Next, the update unit 223 updates the specified set of domain keywords in the keyword management DB 231 based on the update request obtained from the user terminal 260. As an example, when an update request specifies addition of a domain keyword of "photon" to a set of domain keywords corresponding to a domain of "physics", the update unit 223 stores the domain keyword of "photon" in the keyword management DB 231. The domain keyword "photon" may be added to the set of domain keywords corresponding to the domain "physics".

Thereafter, the updating unit 223 updates the domain graph corresponding to the domain whose set of domain keywords has been updated among the domain graphs stored in the domain graph DB 232 based on the updated set of domain keywords. The generation unit 221 is instructed. Here, the graph generation unit 221 adds new nodes and edges to the domain graph, deletes existing nodes and edges, and calculates the node score of each node and each edge based on the updated set of domain keywords. The edge score may be calculated again.

As explained above, according to the update process 700 according to the embodiment of the present disclosure, the set of domain keywords stored in the keyword management DB 231 is updated, and the domain graph corresponding to the set of domain keywords is updated. be able to. This makes it possible to reflect changes in domain keywords due to progress in a particular domain and to set domain keywords that are appropriate for the user's purpose and values, thereby improving the accuracy of similarity determination.

Next, with reference to FIG. 8, a keyword management screen according to an embodiment of the present disclosure will be described.

FIG. 8 is a diagram illustrating an example of a keyword management screen 800 according to an embodiment of the present disclosure. According to this keyword management screen 800, the user can register or update domain keywords stored in the keyword management unit DB 231 of the similarity determination device 210. In some embodiments, the keyword management screen 800 may be a user interface screen generated by the input/output unit 246 and presented to the user terminal 260, for example.

As shown in FIG. 8, the keyword management screen 800 may include a new domain registration window 810 and a domain keyword update window 820.

In the new domain registration window 810, the user can register a new domain in the keyword management section DB 231. For example, the user enters the domain name "architecture" in the input area 811 of the new domain registration window 810 and presses the confirm button 812, thereby inputting the new domain name "architecture" into the keyword management unit DB 231. can be registered.

In the domain keyword update window 820, the user can add a new domain keyword, delete an existing domain keyword, change an existing domain keyword, etc. for a registered domain (for example, a domain registered in the new domain registration window 810). It can be performed. For example, after selecting a particular domain name in the domain selection window 821, the user can input added, deleted, or changed domain keywords for the domain in the domain keyword input window 822. Thereafter, by pressing the confirm button 823, the user can select an action (add, delete, or change) that he or she wants to perform on the domain keyword input in the domain keyword input window 822. Furthermore, by pressing the domain graph update 824, the user can input an update request and implement the update process shown in FIG. 7.

On the keyword management screen 800 described above, the user can easily manage domain keywords stored in the keyword management unit DB 231.

Next, with reference to FIG. 9, a text management screen according to an embodiment of the present disclosure will be described.

FIG. 9 is a diagram illustrating an example of a text management screen 900 according to an embodiment of the present disclosure. According to this text management screen 900, the user can search for other texts that are similar to a specific text, or compare two texts (determine the degree of similarity). In one embodiment, the text management screen 900 may be a user interface screen generated by the input/output unit 246 and presented to the user terminal 260, for example.

As shown in FIG. 9, the text management screen 900 may include a similar text search window 910 and a text comparison window 920.

Similar text search window 910 allows the user to search for other texts that are similar to a specific text. For example, the user enters the domain name "physics" in the domain selection window 911 of the similar text search window 910, enters the text "laws of physics" in the file input window 912, and presses the confirm button 913. By pressing the button, the graph comparison process 600 according to the embodiment of the present disclosure is executed, and a sentence having a high degree of similarity to "laws of physics" can be searched from the domain graph DB 232.

Text comparison window 920 allows a user to compare two particular texts. For example, the user inputs the domain name "architecture" in the domain selection window 921 of the sentence comparison window 920, inputs the sentence "ancient Roman architecture" in the first file input window 922, and inputs the sentence "architecture of ancient Rome" in the first file input window 922. By inputting the text "architecture of ancient Rome" in the second file input window 923 and pressing the comparison button 924, the graph comparison process 600 according to the embodiment of the present disclosure is executed, and the text "architecture of ancient Rome" and "architecture of ancient It is possible to judge the degree of similarity with "Greek architecture".

On the text management screen 900 described above, the user can easily determine the degree of similarity between predetermined texts.

Next, with reference to FIG. 10, multi-domain management processing according to an embodiment of the present disclosure will be described.

Generally, there are texts related to several different domains. Therefore, in order to provide highly accurate similarity determination results even for sentences related to a plurality of different domains, one aspect of the present disclosure is to determine a domain that has a high degree of correspondence with a predetermined sentence, This invention relates to multi-domain management processing for generating a domain graph using domain keywords. FIG. 10 is a diagram illustrating an example of the flow of multi-domain management processing 1000 according to the embodiment of the present disclosure. This multi-domain management process 1000 is mainly performed by the graph generation unit 221 of the similarity determination device 210.

First, in step S1010, the graph generation unit 221 receives an input of a second sentence for which similarity determination is desired, and then determines a domain that has high correspondence with the second sentence.
More specifically, the graph generation unit 221 calculates a domain score indicating the degree of association of the second sentence with each set of domain keywords stored in the keyword management DB 231. In an embodiment, the graph generation unit 221 may calculate the domain score based on the proportion of domain keywords included in the second sentence among each set of domain keywords.

As an example, the keyword management DB 231 includes a set of first domain keywords corresponding to a first domain of "physicist" and a set of second domain keywords corresponding to a second domain of "chemist". Suppose that you are storing . It is also assumed that the graph generation unit 221 receives an input of a second sentence with the phrase "famous scientist" regarding chemists in various scientific fields, such as physicists, chemists, and biologists.
In this case, the graph generation unit 221 generates a first domain score indicating the degree of association of the second sentence with “famous scientist” with the first domain of “physicist” and a first domain score with “chemist”. A second domain score indicating the degree of relevance of the domain to the second domain is calculated.

For example, if 70% of the domain keywords in the set of first domain keywords corresponding to the first domain with "physicist" are included in the second sentence with "famous scientist", The graph generation unit 221 may set the first domain score of the second sentence "famous scientist" to "70%" for the first domain "physicist".
Furthermore, if 30% of the domain keywords of the set of second domain keywords corresponding to the second domain with "chemist" are included in the second sentence with "famous scientist", The graph generation unit 221 may set the second domain score of the second sentence "famous scientist" to "30%" for the second domain "chemist".
In this way, the graph generation unit 221 calculates the proportion of each domain keyword set stored in the keyword management DB 231 in the second sentence based on the ratio of each domain keyword set included in the second sentence. A domain score indicating relevance can be calculated.

Next, in step S1020, the graph generation unit 221 obtains an appropriate set of domain keywords from the keyword management DB 231 based on the domain score calculated in step S1010. In one embodiment, the graph generation unit 221 may obtain an appropriate set of domain keywords based on the domain score relationship calculated for each domain and a predetermined domain score threshold.
More specifically, the graph generation unit 221 determines that a first domain score indicating the degree of association of the second sentence with the first domain is a second domain score indicating the degree of association of the second sentence with the second domain. If the domain score is exceeded and a predetermined domain score threshold (for example, 50%) is satisfied, a set of first domain keywords corresponding to the first domain may be acquired from the keyword management DB 231.
On the other hand, the graph generation unit 221 determines that the second domain score indicating the degree of association of the second sentence with the second domain exceeds the first domain score indicating the degree of association of the second sentence with the first domain. , and if a predetermined domain score threshold (for example, 50%) is satisfied, a set of second domain keywords corresponding to the second domain may be acquired from the keyword management DB 231.

Next, in step S1030, the graph generation unit 221 uses the domain keyword acquired in step S1020 to generate a normalized second domain graph corresponding to the second sentence.
Note that the details of the process of generating the normalized second domain graph have been described with reference to FIG. 6, so the description thereof will be omitted here.

According to the multi-domain management process 1000 described above, even when a given text is related to multiple domains, it is possible to determine an appropriate domain keyword for generating a domain graph corresponding to the text. . Moreover, this makes it possible to improve the accuracy of similarity determination.

Next, a specific example of similarity determination according to the embodiment of the present disclosure will be described with reference to FIG. 11.

FIG. 11 is a diagram illustrating a specific example of similarity determination for two domain graphs according to an embodiment of the present disclosure. FIG. 11 shows parameters such as node scores, normalized node scores, edge weights, and normalized edge weights for two domain graphs, domain graph A and domain graph B.

More specifically, nodes in domain graph A are associated with node scores of f _A (N)={3, 2, 1, 4}. By dividing each node score by the sum of the node scores of the domain graph (3+2+1+4=10), the normalized node score f _A (NS)={0.3, 0.2, 0.1, 0. 4} can be obtained.
Similarly, nodes in domain graph B are associated with node scores of f _B (N)={3, 2, 1, 4}. By dividing each node score by the sum of the node scores of the domain graph (3+2+1+4=10), the normalized node score f _B (NS)={0.3, 0.2, 0.1, 0. 4} can be obtained.

Further, in FIG. 11, the number of connections f _A (L) between each node in domain graph A is shown in table 1105, and the number of connections f _B (L) between each node in domain graph B is shown in table 1110. By taking the inverse of these connections, the edge weights E _ij of domain graph A and domain graph B can be calculated.
Further, by using Equation 3 described above, the normalized edge weight NE _ij of the domain graph A can be calculated. The normalized edge weights for domain graph A are shown in table 1115 and the normalized edge weights for domain graph B are shown in table 1120.

By substituting the normalized node score f _A (NS) of domain graph A and the normalized node score f _B (NS) of domain graph B explained above into Equation 5, domain graph A and domain graph B can be The CNS of can be calculated. In this case, since the normalized node score f _A (NS) of domain graph A and the normalized node score f _B (NS) of domain graph B are the same, Equation 5 can be expressed as (1-0= 1).

Further, by substituting the normalized edge weights of domain graph A and domain graph B described above into Equation 6, the RS of domain graph A and domain graph B can be calculated. In this case, Equation 6 becomes (1-0.2163=0.7837).

Then, by substituting the calculated CNS and RS into Equation 7, the degree of similarity between domain graph A and domain graph B can be calculated. In this case, as a result of calculation using Equation 7, the degree of similarity between domain graph A and domain graph B is "89.185%."

In this way, by representing sentences as domain graphs and comparing these domain graphs, it becomes possible to determine the degree of similarity between two sentences with high precision and at high speed.

The similarity determination means according to the embodiment of the present disclosure has been described above.
As mentioned above, one aspect of the present disclosure relates to representing sentences as domain graphs. This makes it possible to express the semantic relationships between the words of the sentences compared to the case where the texts of the sentences are compared as they are, thereby improving the accuracy of similarity determination.

However, as mentioned above, in the conventional means of representing the entire sentence as a domain graph, all words included in the sentence are represented as nodes, so if the sentence is long, the number of nodes and edges becomes enormous, and the graph There is a problem that processing such as comparisons between the two files is delayed.
Therefore, one aspect of the present disclosure relates to generating a domain graph that expresses only specific keywords in a sentence as nodes instead of the entire sentence. This makes it possible to reduce the size of the graph compared to conventional methods that express the entire chapter as a domain graph. In addition, by reducing the scale of the graph, the time required for processing such as comparing the graphs can be shortened, and high-speed similarity determination can be made.

However, if a domain graph is generated based only on specific keywords in the text rather than the entire text, information loss may occur compared to when the entire text is represented as a domain graph.
Therefore, in one aspect of the present disclosure, each node in the domain graph is associated with a node score that indicates the importance of the domain keyword corresponding to the node, and each edge is associated with a degree of association between the nodes that the edge connects. is associated with the edge weight shown.
As a result, even if the entire text is not expressed as a domain graph, it is possible to maintain semantic information regarding particularly important keywords in the text and suppress information loss.

Further, one aspect of the present disclosure relates to generating a normalized domain graph in which each node score and edge weight in the domain graph are normalized. By normalizing domain graphs in this way, when comparing domain graphs, it is possible to prevent a decrease in similarity determination due to a difference in the size of domain graphs.

When the similarity determination means according to the embodiment of the present disclosure described above is applied to, for example, a report regarding a cyber event, it is possible to specify an existing report that has a high degree of similarity to a specific report. If there is an existing report that satisfies the predetermined similarity criteria with respect to the new report, the cyber event reported in that report will be considered to have already been responded to or is being responded to, and measures will be taken to formulate the countermeasures again. It can save you time and resources. On the other hand, if there is no existing report that satisfies the predetermined similarity criteria with respect to the new report, it is determined that the cyber event reported in the report has not been addressed and new countermeasures need to be developed. can do.
In this way, it is possible to determine with high accuracy whether the cyber event reported in the report has already been handled, is currently being handled, or has not been handled yet. Furthermore, since the domain graph used in determining the degree of similarity according to the embodiment of the present disclosure is a domain graph based only on specific predetermined keywords, comparison results indicating the degree of similarity of reports can be generated at high speed. be able to. This makes it possible to quickly formulate countermeasures against cyber events.

As described above, according to the present disclosure, similarity determination enables high-speed and highly accurate text similarity comparison by suppressing the scale of the graph representation corresponding to the text while maintaining relationship information between elements in the text. means can be provided.

Above, the case where the similarity determination means according to the embodiment of the present disclosure is implemented as an apparatus, system, and method has been described as an example, but the present disclosure is not limited to this, and may be implemented as a computer program, for example. . This computer program may be introduced into a computer system implementing the similarity determination means according to the embodiment of the present disclosure from a storage medium of an external device via a network and/or a portable storage medium.

For example, one aspect according to an embodiment of the present disclosure is a computer program for determining similarity, in a computer system including a memory for storing processing instructions and a processor, in which the processing instructions stored in the memory include: obtaining a first set of domain keywords corresponding to a first domain; and normalizing a first sentence corresponding to the first domain based on the first set of domain keywords. and identifying a subset of first domain keywords included in a second sentence corresponding to the first domain based on the set of first domain keywords. determining the number of occurrences of the identified subset of first domain keywords in the second sentence; and calculating a node score indicating the importance of each node included in the first node set in the second sentence based on the determined number of occurrences. a step of generating a knowledge graph indicating a semantic relationship between each node included in the first node set based on the second sentence; , a step of determining the shortest distance in the knowledge graph; a step of calculating an edge weight indicating the degree of association between the first node and the second node based on the determined shortest distance; the first node set, the node score, the edge weight, and the edge; a step of generating a second domain graph corresponding to the second sentence based on the second sentence; a step of generating a normalized second domain graph in which the node score and the edge weight are normalized; By comparing the normalized first domain graph and the normalized second domain graph, the difference between the normalized first domain graph and the normalized second domain graph is determined. The computer program for determining similarity is characterized by causing the processor to execute the steps of generating and outputting a comparison result indicating the degree of similarity.

Although the embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above, and various changes can be made without departing from the gist of the present invention.

150 Similarity determination application 200 Similarity determination system 210 Similarity determination device 220 Memory 221 Graph generation section 222 Similarity determination section 223 Update section 230 Storage section 231 Keyword management DB
232 Domain graph DB
244 Processor 246 Input/output unit 250 Communication network 260 User terminal

Claims (8)

A similarity determination device,
Equipped with a processor and memory,
The memory is
a keyword management database storing a set of first domain keywords corresponding to the first domain;
a domain graph database that stores a first domain graph generated based on the first domain keyword set for a first sentence corresponding to the first domain;
a graph generation unit that generates a second domain graph for a second sentence corresponding to the first domain based on the first domain keyword set;
a similarity determination unit that generates a comparison result indicating the degree of similarity between the first domain graph and the second domain graph by comparing the first domain graph and the second domain graph;
A similarity determination device comprising a processing instruction for causing the processor to function as a similarity determination device. The graph generation unit is
identifying a subset of first domain keywords included in the second sentence based on the first set of domain keywords;
determining the number of occurrences of the identified subset of first domain keywords in the second sentence;
generating a first node set including at least a first node and a second node based on the first subset of domain keywords;
calculating a node score indicating the importance of each node included in the first node set in the second sentence based on the determined number of occurrences;
The similarity determination device according to claim 1, characterized in that: The graph generation unit is
generating a knowledge graph indicating a semantic relationship between each node included in the first node set based on the second sentence;
determining the shortest distance between the first node and the second node in the knowledge graph;
Based on the determined shortest distance, calculate an edge weight indicating the degree of association between the first node and the second node;
indicating a relationship between the first node and the second node, and generating an edge associated with the edge weight;
The similarity determination device according to claim 2, characterized in that: generating the second domain graph corresponding to the second sentence based on the first node set, the node score, the edge weight, and the edge;
generating a normalized second domain graph in which the node scores and the edge weights are normalized;
The similarity determination device according to claim 3, characterized in that: The keyword management database is
a second set of domain keywords corresponding to a second domain in addition to the first set of domain keywords corresponding to the first domain;
The graph generation unit is
calculating a first domain score indicating the degree of relevance of the second sentence to the first domain;
calculating a second domain score indicating the degree of relevance of the second sentence to the second domain;
if the first domain score exceeds the second domain score and satisfies a predetermined domain score threshold, generating the second domain graph based on the first set of domain keywords;
generating the second domain graph based on the set of second domain keywords if the second domain score exceeds the first domain score and satisfies a predetermined domain score threshold;
The similarity determination device according to claim 1, characterized in that: When an update request requesting addition or deletion of a domain keyword to the first set of domain keywords is received from a user,
further comprising an update unit that updates the second domain graph based on addition or deletion of the domain keyword indicated in the update request;
The similarity determination device according to claim 1, characterized in that: In a similarity determination system in which a user terminal and a similarity determination device are connected via a communication network,
The similarity determination device is
Equipped with a processor and memory,
The memory is
a keyword management database storing a set of first domain keywords corresponding to the first domain;
a domain graph database that stores a first domain graph generated based on the first domain keyword set for a first sentence corresponding to the first domain;
an input/output unit that acquires a second sentence corresponding to the first domain from the user terminal;
a graph generation unit that generates a second domain graph for the second sentence based on the first set of domain keywords;
By comparing the first domain graph and the second domain graph, a comparison result indicating the degree of similarity between the first domain graph and the second domain graph is generated and transmitted to the user terminal. a similarity determination unit that
A similarity determination system comprising processing instructions for causing the processor to function as a system. A similarity determination method,
obtaining a set of first domain keywords corresponding to the first domain;
generating a normalized first domain graph generated based on the set of first domain keywords for a first sentence corresponding to the first domain;
identifying a subset of first domain keywords included in a second sentence corresponding to the first domain based on the first set of domain keywords;
determining the number of occurrences of the identified subset of first domain keywords in the second sentence;
generating a first node set including at least a first node and a second node based on the first subset of domain keywords;
Calculating a node score indicating the importance of each node included in the first node set in the second sentence based on the determined number of occurrences;
generating a knowledge graph indicating a semantic relationship between each node included in the first node set based on the second sentence;
determining the shortest distance between the first node and the second node in the knowledge graph;
calculating an edge weight indicating a degree of association between the first node and the second node based on the determined shortest distance;
a step of indicating a relationship between the first node and the second node and generating an edge associated with the edge weight;
generating a second domain graph corresponding to the second sentence based on the first node set, the node score, the edge weight, and the edge;
generating a normalized second domain graph in which the node scores and the edge weights are normalized;
By comparing the normalized first domain graph and the normalized second domain graph, the normalized first domain graph and the normalized second domain graph are a step of generating and outputting a comparison result indicating the degree of similarity;
A similarity determination method characterized by comprising:

Images