CN112163097A

CN112163097A - Military knowledge graph construction method and system

Info

Publication number: CN112163097A
Application number: CN202011008095.3A
Authority: CN
Inventors: 黄茗; 杨军; 王滨; 钟晨; 张鑫
Original assignee: CETC 15 Research Institute
Current assignee: CETC 15 Research Institute
Priority date: 2020-09-23
Filing date: 2020-09-23
Publication date: 2021-01-01

Abstract

The invention relates to a military knowledge graph construction method and a military knowledge graph construction system. The method comprises the following steps: acquiring military data; defining an entity in military data to obtain an ontology corresponding to the entity and an ontology relation; marking entities in military data according to the ontology and the ontology relationship to obtain marked entity relationships; training the deep learning model according to military data and the marked entity relationship to obtain a trained deep learning model; inputting military data to be extracted into the trained deep learning model to perform entity relationship extraction to obtain a plurality of extracted entity relationships; and storing the extracted entity relations into a graph database to obtain a military knowledge graph. The invention can effectively utilize military data in the network, combines a semi-automatic labeling algorithm and a deep learning algorithm, constructs a military knowledge map under the condition of reducing certain labeling manpower, and provides accurate and comprehensive military information for military workers.

Description

Military knowledge graph construction method and system

Technical Field

The invention relates to the technical field of military knowledge graph construction, in particular to a military knowledge graph construction method and a military knowledge graph construction system.

Background

Knowledge-graphs describe concepts, entities and their relationships in the objective world in a structured form. An entity is a thing in the objective world, and a concept is a generalization and abstraction of things having the same attributes. Google proposed a knowledge graph in 2012 and successfully applied in semantic searches. The knowledge graph can be regarded as a large-scale application of ontology knowledge representation, the knowledge representation structure of the Google knowledge graph mainly describes the relationship between objective existing entities and entities, and for each concept, a certain attribute set for describing the concept is provided. A domain knowledge graph refers to a knowledge graph focused on a particular vertical domain, most of which are entities and concepts related to that domain.

With the development of information technology, the text data of the internet is also growing explosively. Relevant knowledge is extracted from the network text data quickly and accurately, a knowledge graph is constructed, and discrete knowledge can be connected through nodes and lines. The knowledge graph is used as a knowledge storage management mode based on a graph structure, and has the advantages of clear relationship, flexible structure and high visualization degree. Information contained in open source data such as military news and the like plays an important role in some military fields, corresponding knowledge is mined from the open source data to construct a knowledge map, and the knowledge map can be used for knowledge visualization and knowledge question and answer and assist workers in quickly searching and associating knowledge.

The current knowledge graph construction mainly comprises a domain knowledge graph and a general knowledge graph. The general knowledge graph has wider coverage, but the depth of the general knowledge graph is not as deep as the domain knowledge graph, and meanwhile, the requirement of the domain knowledge graph on the precision is higher. At present, a plurality of domain knowledge maps are constructed in China, such as medical treatment, electronic commerce, finance, military industry, electric power, education and the like. The knowledge graph in the military field is few, and is mainly based on the disclosed semi-structured encyclopedia data, the data source is single, and most of the data sources are based on a rule mode. For unstructured military news and other data, it is difficult to extract more information from the data by means of a method that only relies on specifications.

Disclosure of Invention

The invention aims to provide a military knowledge graph construction method and a military knowledge graph construction system, which can effectively utilize military data in a network, combine a semi-automatic labeling algorithm and a deep learning algorithm, construct a military knowledge graph under the condition of reducing certain labeling manpower, and provide accurate and comprehensive military information for military workers.

In order to achieve the purpose, the invention provides the following scheme:

a military knowledge graph construction method comprises the following steps:

acquiring military data;

defining the entity in the military data to obtain an ontology corresponding to the entity and an ontology relation;

marking entities in the military data according to the ontology and the ontology relationship to obtain marked entity relationship;

training a deep learning model according to the military data and the marked entity relationship to obtain a trained deep learning model;

inputting military data to be extracted into the trained deep learning model to perform entity relationship extraction to obtain a plurality of extracted entity relationships;

and storing each extracted entity relationship into a graph database to obtain a military knowledge graph.

Optionally, the acquiring military data specifically includes:

military news data and military encyclopedia data are obtained.

Optionally, the defining the entity in the military data to obtain the ontology and the ontology relationship corresponding to the entity specifically includes:

and defining the entity in the military data through a military dictionary and expert knowledge to obtain an ontology corresponding to the entity and an ontology relation.

Optionally, the labeling the entity in the military data according to the ontology and the ontology relationship to obtain a labeled entity relationship specifically includes:

and manually labeling the entities in the military data by using a Brat tool according to the ontology and the ontology relationship, and adding rule labeling and part-of-speech labeling in the labeling process to obtain the labeled entity relationship.

Optionally, the training of the deep learning model according to the military data and the labeled entity relationship to obtain the trained deep learning model specifically includes:

and training the deep learning model based on the Bert by taking the military data as input and the labeled entity relationship as output to obtain the trained deep learning model.

A military knowledge graph construction system, comprising:

the military data acquisition module is used for acquiring military data;

the definition module is used for defining the entity in the military data to obtain an ontology corresponding to the entity and an ontology relation;

the marking module is used for marking the entity in the military data according to the ontology and the ontology relationship to obtain a marked entity relationship;

the training module is used for training the deep learning model according to the military data and the marked entity relationship to obtain a trained deep learning model;

the entity relationship extraction module is used for inputting military data to be extracted into the trained deep learning model to perform entity relationship extraction so as to obtain a plurality of extracted entity relationships;

and the military knowledge map construction module is used for storing the extracted entity relationships into a map database to obtain the military knowledge map.

Optionally, the military data acquisition module specifically includes:

and the military data acquisition unit is used for acquiring military news data and military encyclopedia data.

Optionally, the definition module specifically includes:

and the definition unit is used for defining the entity in the military data through a military dictionary and expert knowledge to obtain the ontology and the ontology relationship corresponding to the entity.

Optionally, the labeling module specifically includes:

and the labeling unit is used for manually labeling the entity in the military data by using a Brat tool according to the ontology and the ontology relationship, and adding rule labeling and part-of-speech labeling in the labeling process to obtain the labeled entity relationship.

Optionally, the training module specifically includes:

and the training unit is used for training the deep learning model based on the Bert by taking the military data as input and the labeled entity relationship as output to obtain the trained deep learning model.

According to the specific embodiment provided by the invention, the invention discloses the following technical effects:

the method constructs the military knowledge map from the obtained data through the technologies of semi-automatic labeling, deep learning and the like, and can display events in military news, including information of characters, countries, weaponry and the like. By means of the military atlas, military workers can know the information of the network through the military knowledge atlas and can be used as a reference for decision judgment. According to the technical scheme of the invention, military related knowledge resources can be effectively collected, a military knowledge map is established, and accurate and comprehensive military knowledge is provided for workers.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.

FIG. 1 is a flow chart of a military knowledge graph construction method of the present invention;

FIG. 2 is a schematic diagram of a data acquisition process;

FIG. 3 is a schematic diagram of a data annotation process;

FIG. 4 is a diagram of an entity relationship extraction model structure based on Bert;

FIG. 5 is a diagram of a military knowledge graph construction system of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

With the rapid development of artificial intelligence, information extraction based on deep learning is gradually used for constructing knowledge graphs. The extraction of entity relations is used as a core task of information extraction, and the semantic relations between entity pairs are automatically extracted from a large amount of unstructured data mainly by modeling text information, so that effective semantic knowledge is extracted.

Based on the reasons, the invention provides a military knowledge graph construction method and a military knowledge graph construction system. Constructing a corresponding military ontology by acquiring the military-related news on the network, and carrying out data annotation based on ontology knowledge; carrying out deep learning model training by using the labeled data; and extracting entity relations of the news data by using a deep learning model, and constructing a knowledge graph. The invention formulates rules according to the data characteristics and utilizes part of speech to assist in tagging in the tagging process, thereby reducing the workload of workers to a certain extent. FIG. 1 is a flow chart of a military knowledge graph construction method of the present invention. As shown in fig. 1, a military knowledge graph construction method includes:

step 101: acquiring military data, which specifically comprises the following steps:

military news data and military encyclopedia data are obtained.

The method mainly aims at military news, so that military news data such as Xinlang, West road, China network and the like are obtained and stored in a classified mode; the method comprises the steps of acquiring military encyclopedia data, wherein the military encyclopedia mainly relates to weapon equipment, and the weapon equipment appears frequently in military news, so that weapon equipment information in a Chinese weapon warehouse and a global military network is acquired. The specific flow is shown in fig. 2.

Step 102: defining the entity in the military data to obtain an ontology and an ontology relationship corresponding to the entity, and specifically comprising:

Through a military dictionary, a plurality of military proper nouns and ontology information are defined in the military dictionary, for example, category hierarchies in weaponry are divided into a first-level category (aircrafts and the like) and a second-level category (such as second-level classification fighters, attackers, bombers, coaches and the like of the aircrafts), and a corresponding weaponry ontology library can be constructed by utilizing the military dictionary; meanwhile, weapon description information in military encyclopedia data can also complete and supplement the ontology base. The construction of the knowledge graph depends on the construction of an ontology base, in addition to weaponry, other types of entities such as characters, countries, organizations and the like can also appear in military news data, and the definition of the ontology needs military related experts for definition and guidance;

step 103: marking the entity in the military data according to the ontology and the ontology relationship to obtain the marked entity relationship, which specifically comprises:

Compared with an entity relation extraction method completely based on rules, the extraction accuracy and recall rate can be improved by using a deep learning model after data is labeled. According to the knowledge modeling in the step 102, data annotation is carried out by using a tool Brat, and the Brat is an open-source entity relation annotation tool and is friendly in interface.

On the basis, an automatic labeling module based on specification and part-of-speech labeling is added. After analyzing a large amount of weapon equipment, it is found that serial numbers such as "J-10" (referring to fighters of-10), or the alias of the weapon "KF-X" (referring to KF-X invisible fighters) are often used in military news. For the above situation, the regular expression is used for extraction, weapon aliases, numbers and the like in the military news are automatically extracted, and the annotation personnel are prompted; and meanwhile, information such as positions, countries and the like in news are extracted by using a part-of-speech tagging method. By combining rules and part-of-speech tagging, entities can be automatically extracted to a certain extent, the purpose of semi-automatic tagging is achieved, workload of tagging personnel is reduced, and missing and filling can be checked. The specific flow is shown in fig. 3.

Step 104: training a deep learning model according to the military data and the labeled entity relationship to obtain the trained deep learning model, which specifically comprises the following steps:

Step 105: inputting military data to be extracted into the trained deep learning model to perform entity relationship extraction to obtain a plurality of extracted entity relationships;

the Bert model uses a Transformer, a sequence model that discards the sequential structure of RNNs, and employs a completely attention-based approach. The Bert is a pre-training model, contains massive text information, obtains word vector representation of the text through the Bert, can be used for various natural language processing tasks, and has good universality. By analyzing military news data, it can be found that some data have a condition of entity overlapping, for example, the entity of "7.62 mm parallel machine gun" includes another entity of "7.62 mm", in this case, if the common BIO label is used, the two entities cannot be identified separately, so the deep learning model based on Bert is used in the invention.

The overall structure of the model is shown in fig. 4, and mainly comprises the following parts:

bert shared coding: and obtaining the vector characterization of each word through Bert, and using the output of Bert as a word vector.

b. Main body s label: this section is used to identify all possible subject objects. It judges whether it is the beginning or ending position of the entity by classifying the coded result of each position with two classifiers, i.e. the coding result of each position is classified into a single classifier

Wherein x is_iAnd outputting the ith word through the Bert code, wherein W and b are parameters of the full-connection layer classifier, and the activation function is sigmoid. For the case where there are multiple bodies in a sentence, the start pointer and the end pointer are paired by the near matching principle. With this approach, each character can be used as the beginning and end of an entity, which is better than having only one label in the BIO label.

c. Relationship-object tag: for each subject, a prediction of the object following it is required. As can be seen, it is basically consistent with the subject, and the main difference is that each relationship class shares a set of object classifiers. The input of this part adds, in addition to the Bert encoded result of the input sequence, the features of the body, which are an average pooling of the Bert tokens for each character of the body.

The main expression is as follows:

step 106: and storing each extracted entity relationship into a graph database to obtain a military knowledge graph.

And storing the extracted entity relationship triples into a graph database to form the triples. For each entity in the military encyclopedia, if an entity existing in the encyclopedia is identified, the weaponry attributes in the encyclopedia page are taken as attributes of the knowledge-graph entity.

And reading the triad of the graph database, and developing a page for displaying. Military staff can input entities and relations in a page and inquire related attributes; two weaponry entities may also be entered and their relationship paths obtained.

FIG. 5 is a diagram of a military knowledge graph construction system of the present invention. As shown in fig. 5, a military knowledge graph construction system includes:

and the military data acquisition module 201 is used for acquiring military data.

The defining module 202 is configured to define an entity in the military data to obtain an ontology and an ontology relationship corresponding to the entity.

And the labeling module 203 is used for labeling the entity in the military data according to the ontology and the ontology relationship to obtain a labeled entity relationship.

And the training module 204 is configured to train the deep learning model according to the military data and the labeled entity relationship, so as to obtain a trained deep learning model.

And the entity relationship extraction module 205 is configured to input military data to be extracted into the trained deep learning model to perform entity relationship extraction, so as to obtain a plurality of extracted entity relationships.

And the military knowledge graph construction module 206 is used for storing the extracted entity relationships into a graph database to obtain the military knowledge graph.

The military data acquisition module 201 specifically includes:

The definition module 202 specifically includes:

The labeling module 203 specifically includes:

The training module 204 specifically includes:

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.

The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.

Claims

1. A military knowledge graph construction method is characterized by comprising the following steps:

acquiring military data;

2. The military knowledge graph construction method according to claim 1, wherein the acquiring military data specifically comprises:

military news data and military encyclopedia data are obtained.

3. The military knowledge graph construction method according to claim 1, wherein the defining of the entity in the military data to obtain the ontology and the ontology relationship corresponding to the entity specifically comprises:

4. The military knowledge graph construction method according to claim 1, wherein the labeling of the entities in the military data according to the ontology and the ontology relationship to obtain a labeled entity relationship specifically comprises:

5. The military knowledge graph construction method according to claim 1, wherein the training of the deep learning model according to the military data and the labeled entity relationship to obtain the trained deep learning model specifically comprises:

6. A military knowledge graph construction system, comprising:

the military data acquisition module is used for acquiring military data;

7. The military knowledge graph building system of claim 6, wherein the military data acquisition module specifically comprises:

8. The military knowledge graph building system of claim 6, wherein the definition module specifically comprises:

9. The military knowledge graph building system of claim 6, wherein the labeling module specifically comprises:

10. The military knowledge graph building system of claim 6, wherein the training module specifically comprises: