A knowledge mining method based on cross-model and cross-language knowledge modeling

Abstract

The present application belongs to the technical field of knowledge mining, and particularly relates to a knowledge mining method based on cross-model and cross-language knowledge modeling. In the implementation process of the mining method, multi-language news knowledge is mined, different knowledge models constructed by the method have certain isolation and certain correlation, for different source language materials and knowledge models, a combination of a supervised knowledge extraction method and an unsupervised knowledge extraction method is adopted for knowledge extraction, and the knowledge obtained in the knowledge extraction process is mapped with the knowledge models, at the cross-language knowledge fusion level, a cross-language knowledge base is constructed based on an open source encyclopedia and an open source knowledge base, and knowledge fusion alignment is performed based on cross-language linking. Finally, cross-model and cross-language knowledge modeling and knowledge mining are completed, and a required cross-language news knowledge base is obtained.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of knowledge mining, and particularly relates to a knowledge mining method based on cross-model and cross-language knowledge modeling. BACKGROUND

[0002] With the explosive growth of open source news text data, traditional data retrieval and browsing methods have been unable to meet the research needs. As the research direction gradually subdivides, researchers urgently need cross-country model, cross-language knowledge graph technology to extract the target entities and the relationships between the entities from complex natural language, so as to mine potential high-value information.

[0003] Since Google proposed the knowledge graph in 2012, it has attracted widespread attention from academia and industry, and has been widely researched and applied in various fields. Currently, leading Internet companies such as Google, Baidu, Alibaba, Tencent, and emerging companies such as Minglve and Xinguan have established knowledge graph frameworks and technologies to improve data application capabilities using knowledge graphs. Traditionally, the process of building a knowledge graph includes knowledge modeling, data connection, knowledge mapping, knowledge extraction, and knowledge fusion, ultimately forming a knowledge graph to provide knowledge data support for upper-layer applications.

[0004] The Chinese patent "CN112199511A" cross-language multi-source vertical field knowledge graph construction method constructs a parallel corpus based on content and link analysis according to the input cross-language text, field dictionary, field terminology library, field material, and data, and realizes automatic translation of foreign language text based on a trained translation model on the basis of preprocessing, and combines vertical field translation data and actual scenarios to complete semantic feature extraction and entity relationship extraction based on deep learning. Different source knowledge is merged and fused through network equivalent entities to obtain a cross-language multi-source vertical field knowledge graph.

[0005] Traditional knowledge graph construction often targets a single knowledge model and a single language text, and cannot solve cross-language and cross-knowledge model knowledge mining. The cross-language knowledge graph construction in the Chinese patent "CN112199511A" is only for a single knowledge model, and the cross-language in the patent is realized by machine translation text, which is not a real cross-language knowledge graph construction, so it cannot meet the existing cross-model and cross-language knowledge graph construction application needs in the field of news research. SUMMARY

[0006] (I) Technical problems to be solved

[0007] The technical problem to be solved by the present application is: how to meet the needs of users for cross-model modeling and cross-language knowledge extraction, how to propose a cross-model and cross-language knowledge graph construction technology to facilitate researchers to model their own research direction, and to collect data in their own language for original knowledge extraction, and to realize knowledge fusion through cross-language knowledge linking.

[0008] (II) Technical solution

[0009] To solve the above technical problems, the present application provides a knowledge mining method based on cross-model and cross-language knowledge modeling, which comprises the following steps:

[0010] Step 1: Knowledge modeling is carried out for different countries, including multi-level entity concept modeling, entity attribute modeling and logical association relationship modeling; this step 1 comprises the following specific steps:

[0011] Step 11: In the knowledge modeling background system, different knowledge models are created for different countries, and the knowledge models corresponding to different countries are created and named by country name, and then multi-level entity concepts are created for each knowledge model according to the membership relationship constraint; the multi-level entity concepts include: first-level entity concepts and second-level entity concepts; wherein the first-level entity concepts include: organization, person, place and weapon equipment; the second-level entity concept is a subset of a first-level entity concept, and there is a containing and contained relationship between them in the membership relationship;

[0012] Step 12: For each specific entity concept, its corresponding entity attribute is created; wherein for an organization, its entity attribute includes: establishment time, headquarters location, size and number of people; for a person, its entity attribute includes: age, gender, position, birthplace and education; for a place, its entity attribute includes: climate and latitude and longitude; for weapon equipment, its entity attribute includes length, importance, combat radius and power;

[0013] Step 13: In the whole knowledge model, for each two entity concepts that can produce entity relationship, define them as a pair, and for each pair of entity concepts that can produce entity relationship, create the logical association relationship between them;

[0014] The logical association relationship includes:

[0015] The relationship between a person and an organization includes "employment" and "membership", the relationship between a person and a place includes "place of birth" and "place of death", the relationship between a person and a person includes "colleague" or "relative", the relationship between weapon equipment and organization includes "research unit", the relationship between organization and organization includes "affiliated organization", and the relationship between weapon equipment and weapon equipment includes "loading" and "alias".

[0016] Step 14: repeat steps 11 to 13 to create knowledge models of all countries involved in the research task and multi-level entity concepts, entity attributes and logical association relationships in each country knowledge model in turn, so as to form knowledge model data in data form; group the knowledge model data according to countries, and store them in the database table in the form of grouping;

[0017] Step 2: for the multi-level entity concepts, entity attributes and logical association relationships in the knowledge model data, take the official language of the country as the source language, model the knowledge model data in the source language, and form the source language models of the multi-level entity concepts, entity attributes and logical association relationships, i.e. the multi-level entity concept source language model, the entity attribute source language model and the logical association relationship source language model;

[0018] Then, according to the semantic relationship between the source language and Chinese, for the multi-level entity concept source language model, the entity attribute source language model and the logical association relationship source language model, their respective multi-level entity concept Chinese model, entity attribute Chinese model and logical association relationship Chinese model are established;

[0019] Step 3: knowledge extraction is performed on each news material;

[0020] The knowledge extraction is performed by combining the supervised knowledge extraction method and the unsupervised knowledge extraction method. This step 3 includes the following specific steps:

[0021] Step 31: a supervised knowledge extraction model is generated by supervised training through artificial annotation data by using a supervised deep learning method. The supervised knowledge extraction model is generated for each language respectively, including: a Chinese supervised knowledge extraction model, an English supervised knowledge extraction model, a Japanese supervised knowledge extraction model and a Russian supervised knowledge extraction model;

[0022] Step 32: dictionaries and rules are defined according to the multi-level entity concept model, entity attribute model and logical association relationship model established by the user;

[0023] For the multi-level entity concept model, when the user defines an entity concept "weapon equipment" in the knowledge model, a batch of "weapon equipment" related data of aircraft, ships and missiles are pre-processed as dictionaries, which are a mapping relationship between specific entity names and entity concepts. Once the data in the dictionary is matched in the news material, it is considered to belong to the "weapon equipment" concept;

[0024] At the same time, a rule is defined that any entity ending with "ship" or "aircraft" is considered to belong to the concept of "weapons and equipment". For the entity attribute model, when the user defines the entity attribute "length" of the concept of "weapons and equipment" in the knowledge model, then the rule is defined: once an entity under the concept of "weapons and equipment" and the keyword "length" are matched in the news material, "length" and the specific value corresponding to "length" will be used as the attribute of the "weapons and equipment".

[0025] For the logical association model, when a user defines the logical association "relatives" between the concept pairs "person" and "person" in the knowledge model, the rule is defined as follows: once the concept pair "person"-"person" and the keywords "father", "mother" or "relatives" are matched in the news material, then "relatives" will be used as the logical association between the concept pair "person"-"person".

[0026] Similarly, based on the multi-level entity concept model, entity attribute model, and logical relationship model, all rules and dictionaries are defined to form an unsupervised knowledge extraction model.

[0027] Similarly, the rules and dictionaries here are designed and generated separately for each language, including Chinese rule dictionaries, English rule dictionaries, Japanese rule dictionaries and Russian rule dictionaries, thus forming Chinese unsupervised knowledge extraction models, English unsupervised knowledge extraction models, Japanese unsupervised knowledge extraction models and Russian unsupervised knowledge extraction models;

[0028] Step 33: Extract knowledge from each news article. Based on the language of the news article itself, call the supervised knowledge extraction model and the unsupervised knowledge extraction model for the corresponding language respectively, and merge the results returned by the two.

[0029] Knowledge extraction encompasses three processes:

[0030] First, entity extraction is performed. The input is a single news article, and the output is all entity information contained in the article.

[0031] Next, attribute extraction is performed. The input is all the entity information in the material, and the output is the attribute information of each entity.

[0032] Finally, relationship extraction is performed. The input is all the entity information in the material, and the output is the logical relationship information between each pair of entities.

[0033] Through the above knowledge extraction process, knowledge is obtained from news materials;

[0034] Step 4: Map the knowledge obtained during the knowledge extraction process to the knowledge model. Since knowledge obtained from news materials in different languages ​​may not necessarily be mapped to the corresponding knowledge model for that language, it is necessary to make further judgments based on the semantic information of the news material's chapters and sentences to complete the mapping between knowledge and knowledge models. For example, the knowledge extracted from Japanese domestic news reported in English should be mapped to the Japanese knowledge model; the knowledge extracted from Indian domestic news reported in Japanese should be mapped to the Indian knowledge model.

[0035] Step 4 includes the following specific steps:

[0036] Step 41: Classify each sentence in the news material according to the knowledge model, specifically including the American knowledge model, the Japanese knowledge model, the Indian knowledge model, and the Russian knowledge model; if the confidence of the classification result is high, then the classification result is considered valid; map the knowledge extracted from the sentence to the classified knowledge model;

[0037] Step 42: If the credibility of the sentence classification result is low, then the sentence classification result is considered invalid; at this time, the entire news material is classified according to the knowledge model, specifically including the American knowledge model, the Japanese knowledge model, the Indian knowledge model, and the Russian knowledge model; the knowledge extracted from each sentence in the news material is mapped to the classified knowledge model;

[0038] Step 43: Through the mapping process in Steps 41 and 42 above, the output is a combination of knowledge and knowledge models, i.e., a knowledge base; here, the knowledge base obtained is a knowledge base in the source language form.

[0039] Step 44: For the knowledge base in the source language form, based on the corresponding Chinese models of the multi-level entity concept source language model, entity attribute source language model, and logical relationship source language model in Step 2, perform another mapping between the source language and Chinese to obtain the knowledge base in Chinese form.

[0040] Step 5: Since the same knowledge often has different forms of expression, it is necessary to further integrate cross-language knowledge in the Chinese knowledge base that is mapped to the knowledge model;

[0041] This step involves building a cross-language knowledge base based on open-source encyclopedias and knowledge repositories, and then integrating cross-language knowledge based on this knowledge base. The specific steps are as follows:

[0042] Step 51: Build a cross-language knowledge base using open-source encyclopedias and open-source knowledge bases;

[0043] The open source encyclopedia library and the open source knowledge base are sorted out and integrated to construct a unified cross-language knowledge base; the cross-language knowledge base contains the alias, attribute, description and label information of the same entity in different language dimensions;

[0044] Step 52 aligns the same knowledge in different languages by using the cross-language knowledge base, and completes the fusion alignment of the knowledge;

[0045] Therefore, cross-model and cross-language knowledge modeling and knowledge mining are finally completed, and the required cross-language news knowledge base is obtained.

[0046] In step 11, the respective knowledge models of different countries are created, including the US knowledge model, the Japanese knowledge model and the Indian knowledge model.

[0047] In step 11, in the case of the US knowledge model, first-level entity concepts are created, including organizations, persons, places and weapons.

[0048] According to the first-level entity concept, a second-level entity concept is created, and a second-level entity concept is created under the organization of the first-level entity concept, including a certain country's department, a certain country's oil association, a certain country's staff institute and a certain country's house of representatives.

[0049] In step 1, the knowledge models of different countries are isolated from each other in the display interface, physical storage and database to ensure the access rights of different users or user groups.

[0050] In step 14, different database open permissions are given to each group to realize the mutual isolation between groups, so that different users can only access the knowledge model data corresponding to their own permissions.

[0051] In step 32, the dictionary forms include {certain aircraft carrier: weapon equipment}, {certain person: person}, and {New York: place}.

[0052] In step 41, the high confidence of the classification result means that the confidence is higher than or equal to 0.7.

[0053] In step 42, the low confidence of the classification result means that the confidence is less than 0.7.

[0054] The open source encyclopedia library refers to the public encyclopedia knowledge base including Wikipedia, Baidu Encyclopedia and Interactive Encyclopedia.

[0055] The open source knowledge base refers to the public knowledge base including Baklib and Raneto.

[0056] (Three) beneficial effects

[0057] Compared with the prior art, the present application is based on a cross-language and cross-country knowledge mining method, and multi-language news knowledge is mined, and different knowledge models constructed have certain isolation and certain correlation. For different source language materials and knowledge models, the present application can automatically adapt the knowledge extraction algorithm to realize knowledge extraction of the real source language. In the cross-language knowledge fusion level, the construction of the cross-language knowledge base is realized based on the open source encyclopedia library, the open source knowledge base and the user experts, and the knowledge fusion alignment is realized based on the cross-language link.

[0058] Compared with the prior art, the present application is based on a cross-language and cross-country knowledge mining method, and multi-language news knowledge is mined, and different knowledge models constructed have certain isolation and certain correlation. For different source language materials and knowledge models, the present application can automatically adapt the knowledge extraction algorithm to realize knowledge extraction of the real source language. In the cross-language knowledge fusion level, the construction of the cross-language knowledge base is realized based on the open source encyclopedia library, the open source knowledge base and the user experts, and the knowledge fusion alignment is realized based on the cross-language link.

[0059] Moreover, the present application realizes the construction, retrieval and application of the cross-language and cross-model knowledge graph in the actual project. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 It is a multi-language knowledge extraction flowchart in the technical scheme of the present application.

[0061] Figure 2 It is a cross-language news knowledge graph construction flowchart in the technical scheme of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, content and advantages of the present application more clear, the specific embodiments of the present application are further described in detail below in combination with the drawings and examples.

[0063] In order to solve the problems of the prior art, the present application provides a knowledge mining method based on cross-model and cross-language knowledge modeling, as shown in the figure, the method comprises the following steps: Figure 2

[0064] Step 1: Different field news researchers pay attention to different news contents. Knowledge modeling is carried out for different countries, including multi-level entity concept modeling, entity attribute modeling and logical association relationship modeling; in the step 1, the knowledge models of different countries are isolated from each other in the display interface, physical storage and database to ensure the access permission of different users or user groups.

[0065] The step 1 comprises the following specific steps:

[0066] ​Step 11: After knowledge modeling, create a knowledge model for each country in the knowledge modeling background system, create and name the knowledge model corresponding to different countries according to the country name, and then create a multi-level entity concept according to the membership constraint for each knowledge model; The multi-level entity concept includes: first-level entity concept, second-level entity concept; Wherein, the first-level entity concept includes: organization, person, place and weapon equipment; The second-level entity concept is a subset of a first-level entity concept, and there is a containing and contained relationship between the two in the membership relationship;

[0067] In step 11, the knowledge model for each country includes: the United States knowledge model, the Japan knowledge model, and the India knowledge model.

[0068] In step 11, in the case of the United States knowledge model, create first-level entity concepts, including organization, person, place and weapon equipment;

[0069] Continue to create second-level entity concepts according to the first-level entity concept, create second-level entity concepts under the organization of the first-level entity concept, including a certain department of a certain country, a certain oil association of a certain country, a certain house of a certain country, and a certain house of a certain country.

[0070] Step 12: Create the corresponding entity attribute for each specific entity concept; Wherein, for the organization, the entity attribute includes: establishment time, headquarters location, size and number of people, etc.; For the person, the entity attribute includes: age, gender, position, place of birth and education, etc.; For the place, the entity attribute includes: climate and latitude and longitude, etc.; For the weapon equipment, the entity attribute includes length, importance, combat radius and power, etc.;

[0071] Step 13: In the entire knowledge model, for each two entity concepts that can produce entity relationship, define them as a pair, and create the logical association relationship between them for each pair of entity concepts that can produce entity relationship;

[0072] The logical association relationship includes: the relationship between the person and the organization includes "employment" and "membership", etc., the relationship between the person and the place includes "place of birth" and "place of death", etc., the relationship between the person and the person includes "colleague" or "relative", etc., the relationship between the weapon equipment and the organization includes "research unit", etc., the relationship between the organization and the organization includes "affiliated organization", etc., the relationship between the weapon equipment and the weapon equipment includes "loading" and "alias", etc.;

[0073] Step 14: repeating steps 11 to 13 to create knowledge models of all countries involved in the research task and multi-level entity concepts, entity attributes and logical association relationships in each country knowledge model in turn, thereby forming knowledge model data in data form; grouping the knowledge model data according to countries and storing it in the database table in the form of groups; in step 14, for each group, different database open permissions are assigned to achieve mutual isolation between groups, so that different users can only access the knowledge model data corresponding to their own permissions.

[0074] The knowledge base display mode is as follows:

[0075] American knowledge model

[0076] Person

[0077] Military figure

[0078] Ma Ma Ma

[0079] Bu Ma Ma

[0080] Scientist

[0081] Franklin

[0082] Edison

[0083] Weaponry

[0084] Aircraft carrier

[0085] A certain A aircraft carrier

[0086] A certain B aircraft carrier

[0087] Aircraft

[0088] A certain A fighter

[0089] A certain B fighter

[0090] Organization .......

[0091] Location ......

[0092] Japanese knowledge model

[0093] Indian knowledge model

[0094] Step 2: For multi-level entity concepts, entity attributes and logical association relationships in the knowledge model data, use the official language of the country as the source language to model the knowledge model data in the source language, forming source language models of multi-level entity concepts, entity attributes and logical association relationships, i.e. source language models of multi-level entity concepts, entity attributes and logical association relationships.

[0095] Then according to the contrastive semantic relationship between the source language and Chinese, for the multi-level entity concept source language model, entity attribute source language model and logical association relationship source language model, its respective corresponding multi-level entity concept Chinese model, entity attribute Chinese model and logical association relationship Chinese model are established;

[0096] (Such as the secondary entity concept "X country X parliament" under the U.S. knowledge model is stored in the modeling system as "UxxxxxSxxxxx Sxxxxx", and at the same time, the Chinese name is noted as "X country X parliament". The entity attribute "scale number" is stored in the modeling system as "Scale number", and at the same time, the Chinese name is noted as "scale number")

[0097] Step 3: Knowledge extraction is performed on each news material, and there are mainly two ways of knowledge extraction: one is general supervised knowledge extraction, and the other is non-supervised knowledge extraction based on knowledge model; The general supervised knowledge extraction adopts the current mainstream supervised deep learning method, and realizes knowledge extraction through supervised training of artificial labeled data, which has high accuracy, but needs a large amount of artificial labeled data, so the cycle is long, the cost is high, and it cannot be well adapted to various knowledge models; The non-supervised knowledge extraction based on knowledge model refers to defining dictionary and rule flexibly according to the multi-level entity concept model, entity attribute model and logical association relationship model established by the user, and obtaining knowledge through the syntax and semantic information of the text, which is simple and easy to understand, and can adapt to various knowledge models, but the accuracy is relatively low. Therefore, for the user's flexible configuration of the knowledge model, the combination of the supervised knowledge extraction method and the non-supervised knowledge extraction method is adopted, and this step 3 includes the following specific steps:

[0098] Step 31: Adopting the supervised deep learning method, the supervised training is performed through the artificial labeled data to generate the supervised knowledge extraction model; The supervised knowledge extraction model here is generated by training for each language respectively, including: Chinese supervised knowledge extraction model, English supervised knowledge extraction model, Japanese supervised knowledge extraction model and Russian supervised knowledge extraction model;

[0099] Step 32: According to the multi-level entity concept model, entity attribute model and logical association relationship model established by the user, the dictionary and rules are defined flexibly; for the multi-level entity concept model, when the user defines the entity concept "weapon equipment" in the knowledge model, a batch of "weapon equipment" related data of aircraft, ships and missiles are pre-processed as a dictionary, which refers to a mapping relationship between specific entity name and entity concept, such as {certain aircraft carrier: weapon equipment}, {certain person: person}, {certain place: place}; once the data in the dictionary is matched in the news material, it is considered to belong to the "weapon equipment" concept;

[0100] At the same time, the rules are defined, and the entity ending with "ship" or "machine" is considered to belong to the "weapon equipment" concept; for the entity attribute model, when the user defines the entity attribute "length" of the "weapon equipment" concept in the knowledge model, the rule is defined: once the entity under the "weapon equipment" concept and the keyword "length" are matched in the news material, "length" and the specific numerical value corresponding to "length" are taken as the attribute of the "weapon equipment";

[0101] For the logical association relationship model, when the user defines the logical association relationship "relative" between the concept pair "person" and "person" in the knowledge model, the rule is defined: once the concept pair "person"-"person" and the keywords "father", "mother" or "relative" are matched in the news material, "relative" is taken as the logical association relationship of the concept pair "person"-"person";

[0102] In this way, according to the multi-level entity concept model, entity attribute model and logical association relationship model, all rules and dictionaries are defined to form a non-supervised knowledge extraction model;

[0103] Similarly, the rules and dictionaries here are generated respectively for each language, including Chinese rule dictionary, English rule dictionary, Japanese rule dictionary and Russian rule dictionary, thereby forming Chinese non-supervised knowledge extraction model, English non-supervised knowledge extraction model, Japanese non-supervised knowledge extraction model and Russian non-supervised knowledge extraction model;

[0104] Step 33: Knowledge extraction is performed on each news material, and according to the language of the news material itself, the supervised knowledge extraction model and the non-supervised knowledge extraction model corresponding to the language are called respectively, and the results returned by the two are fused; for example, when the language of the news material is English, the English supervised knowledge extraction model and the English non-supervised knowledge extraction model are called respectively, and the results returned by the two are fused;

[0105] Knowledge extraction covers three processes:

[0106] First, entity extraction is performed, the input is a single news material, and the output is all entity information contained in the material;

[0107] Secondly, attribute extraction is performed, the input is all entity information in the material, and the output is attribute information possessed by each entity;

[0108] Finally, relationship extraction is performed, the input is all entity information in the material, and the output is logical association relationship information possessed by each group of entity pairs;

[0109] Figure 1 The process of multilingual knowledge extraction is specifically described.

[0110] Through the above knowledge extraction process, the knowledge in the news material is obtained;

[0111] Step 4: mapping the knowledge obtained by the knowledge extraction process and the knowledge model; since the knowledge obtained from different language news materials may not be mapped to the knowledge model corresponding to the language, it is necessary to judge again according to the chapter and sentence semantic information of the news material, so as to complete the mapping between the knowledge and the knowledge model; wherein, including the Japanese domestic news reported in English, the extracted knowledge should be mapped to the Japanese knowledge model; the Indian domestic news reported in Japanese, the extracted knowledge should be mapped to the Indian knowledge model;

[0112] This step 4 includes the following specific steps:

[0113] Step 41: classifying each sentence in the news material according to the knowledge model, the specific categories include the United States knowledge model, the Japanese knowledge model, the Indian knowledge model and the Russian knowledge model; if the reliability of the classification result is high (higher than or equal to 0.7), the classification result is considered effective; the knowledge extracted in the sentence is mapped to the knowledge model obtained by classification;

[0114] Among them, the classification algorithm is the most basic and most common algorithm in the industry, which is used to judge which category a sentence or a chapter belongs to, the input of the classification algorithm is a sentence or a chapter, and the output is each category and the corresponding reliability, such as {United States knowledge model: 0.7, Japanese knowledge model: 0.1, Indian knowledge model: 0.05, Russian knowledge model: 0.1}. The category with the highest reliability is taken as the final trusted category;

[0115] Step 42: if the reliability of the sentence classification result is low (less than 0.7), the sentence classification result is considered invalid; at this time, the whole news material is classified according to the knowledge model, the specific categories include the United States knowledge model, the Japanese knowledge model, the Indian knowledge model and the Russian knowledge model; the knowledge extracted from each sentence in the news material is mapped to the knowledge model obtained by classification;

[0116] Step 43: Through the mapping process of steps 41 and 42 above, the output is a combination of knowledge and knowledge model, that is, a knowledge base; here the knowledge base obtained is a knowledge base in the source language form;

[0117] Step 44: For the knowledge base in the source language form, according to the respective multi-level entity concept Chinese model, entity attribute Chinese model and logical association relationship Chinese model corresponding to the multi-level entity concept source language model, entity attribute source language model and logical association relationship source language model in step 2, the mapping of the source language and Chinese is performed again to obtain a knowledge base in Chinese form;

[0118] Step 5: Since the same knowledge often has different forms of expression, further cross-language knowledge fusion is needed for the knowledge in the Chinese form knowledge base that is mapped into the knowledge model;

[0119] This step is based on open source encyclopedia and open source knowledge base and user expert knowledge to realize cross-language knowledge base construction, and based on cross-language knowledge base to realize cross-language knowledge fusion; the specific steps are as follows:

[0120] Step 51: Use open source encyclopedia and open source knowledge base (which can further include user expert accumulated knowledge) to construct a cross-language knowledge base;

[0121] The open source encyclopedia refers to the public encyclopedia knowledge base including Wikipedia, Baidu Encyclopedia and Interactive Encyclopedia, and the open source knowledge base refers to the public knowledge base including Baklib and Raneto (user expert accumulation refers to the industry knowledge accumulated by users in work); these open source encyclopedia and open source knowledge base are combed and integrated to construct a unified cross-language knowledge base; the cross-language knowledge base contains alias, attribute, description and label information of the same entity in different language dimensions;

[0122] Step 52: Use the cross-language knowledge base to align the same knowledge in different languages, so as to complete the fusion and alignment of knowledge;

[0123] Thus, cross-model and cross-language knowledge modeling and knowledge mining are finally completed, and the required cross-language news knowledge base is obtained.

[0124] Among them, on the basis of the above technical solutions, it can further include:

[0125] Step 6: Knowledge correction and update, users can intervene in the review of the knowledge extracted by the knowledge extraction, guide the continuous optimization of the extraction model, and thus continuously improve the accuracy of knowledge extraction and knowledge fusion. The specific steps are as follows:

[0126] Step 61 For the knowledge obtained by the multilingual knowledge extraction model, the user can perform manual review and correction, including multi-dimensional correction of entity name, entity type, entity attribute and entity relationship, and finally obtain high-accuracy knowledge for storage and application.

[0127] Step 62 On the other hand, after the user correction, the high-accuracy knowledge can be used as high-quality training corpus of the knowledge extraction algorithm to guide the continuous optimization of the algorithm, thereby continuously improving the accuracy of knowledge extraction and knowledge fusion.

[0128] To sum up, the present application belongs to the technical field of knowledge mining, and specifically relates to a knowledge mining method based on cross-model and cross-language knowledge modeling. In the implementation process of the mining method, multilingual news knowledge is mined, different knowledge models constructed by the present application have certain isolation and certain correlation, for different source language materials and knowledge models, a combination of supervised knowledge extraction method and unsupervised knowledge extraction method is adopted for knowledge extraction, and the knowledge obtained in the knowledge extraction process is mapped with the knowledge model, at the cross-language knowledge fusion level, cross-language knowledge base construction is realized based on open source encyclopedia and open source knowledge base, and knowledge fusion alignment is realized based on cross-language linking. Finally, cross-model and cross-language knowledge modeling and knowledge mining are completed, and the required cross-language news knowledge base is obtained.

[0129] The above only describes the preferred embodiments of the present application, and it should be noted that for ordinary skilled persons in the technical field, several improvements and modifications can be made without departing from the technical principles of the present application, and these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1. A knowledge mining method based on cross-model, cross-language knowledge modeling, characterized in that, The method comprises the following steps: Step 1: knowledge modeling is carried out for different countries respectively, including multi-level entity concept modeling, entity attribute modeling and logical association relationship modeling; the step 1 comprises the following specific steps: Step 11: after knowledge modeling, knowledge models of different countries are respectively created in a knowledge modeling background system, the knowledge models corresponding to different countries are respectively created and named according to country names, and then multi-level entity concepts are created for each knowledge model according to the constraints of the membership relationship; The multi-level entity concept comprises a first-level entity concept and a second-level entity concept; wherein the first-level entity concept comprises an organization, a person, a place and a weapon equipment; the second-level entity concept is a subset of a first-level entity concept, and there is a containing and contained relationship between the two in the membership relationship; Step 12: entity attributes corresponding to each specific entity concept are created; wherein for an organization, the entity attributes comprise establishment time, headquarters location, scale and number of people; for a person, the entity attributes comprise age, gender, position, place of birth and education; for a place, the entity attributes comprise climate and latitude and longitude; for a weapon equipment, the entity attributes comprise length, importance, combat radius and power; Step 13: in the whole knowledge model, for each two entity concepts capable of producing entity relationship, the two are defined as a pair, and for each pair of entity concepts capable of producing entity relationship, a logical association relationship between the two is created; The logical association relationship comprises: The relationship between a person and an organization comprises "employment" and "membership", the relationship between a person and a place comprises "place of birth" and "place of death", the relationship between a person and a person comprises "colleague" or "relative", the relationship between a weapon equipment and an organization comprises "research unit", the relationship between an organization and an organization comprises "affiliated organization", and the relationship between a weapon equipment and a weapon equipment comprises "loading" and "alias"; Step 14: steps 11 to 13 are repeatedly implemented to sequentially create knowledge models of all countries involved in the research task and multi-level entity concepts, entity attributes and logical association relationships in the knowledge models of each country, so as to form knowledge model data in the form of data; the knowledge model data is grouped according to countries, and stored in a database table in the form of grouping; Step 2: for the multi-level entity concept, entity attribute and logical association relationship in the knowledge model data, the official language of the country is taken as the source language, the knowledge model data is modeled in the source language to form the source language model of the multi-level entity concept, the source language model of the entity attribute and the source language model of the logical association relationship, i.e. the source language model of the multi-level entity concept, the source language model of the entity attribute and the source language model of the logical association relationship; Then according to the contrast semantic relationship between the source language and Chinese, for the multi-level entity concept source language model, the entity attribute source language model and the logical association relationship source language model, the corresponding multi-level entity concept Chinese model, the entity attribute Chinese model and the logical association relationship Chinese model are established; Step 3: the multi-level entity concept Chinese model, the entity attribute Chinese model and the logical association relationship Chinese model are respectively taken as the source language and the target language to carry out machine translation, so as to form the multi-level entity concept Chinese model, the entity attribute Chinese model and the logical association relationship Chinese model in the target language; Step 3: knowledge extraction for each news material; The knowledge extraction is performed by combining the supervised knowledge extraction method and the unsupervised knowledge extraction method. Step 3 includes the following specific steps: Step 31: a supervised deep learning method is used to perform supervised training through artificial annotation of data to generate a supervised knowledge extraction model. The supervised knowledge extraction model is generated for each language, including a Chinese supervised knowledge extraction model, an English supervised knowledge extraction model, a Japanese supervised knowledge extraction model, and a Russian supervised knowledge extraction model; Step 32: dictionaries and rules are defined according to the multi-level entity concept model, the entity attribute model, and the logical association relationship model established by the user; For the multi-level entity concept model, when the user defines an entity concept "weapon equipment" in the knowledge model, a batch of data related to "weapon equipment" such as aircraft, ships, and missiles are pre-processed as dictionaries, which are a mapping relationship between specific entity names and entity concepts. Once the data in the dictionary is matched in the news material, it is considered to belong to the "weapon equipment" concept. Meanwhile, rules are defined, and entities ending with "ship" or "machine" are considered to belong to the "weapon equipment" concept. For the entity attribute model, when the user defines the entity attribute "length" of the "weapon equipment" concept in the knowledge model, the rule is defined: once the entity under the "weapon equipment" concept and the keyword "length" are matched in the news material, "length" and the specific numerical value corresponding to "length" are taken as the attribute of the "weapon equipment". For the logical association relationship model, when the user defines the logical association relationship "relative" between the concept pair "person" and "person" in the knowledge model, the rule is defined: once the concept pair "person"-"person" and the keywords "father", "mother", or "relative" are matched in the news material, "relative" is taken as the logical association relationship of the concept pair "person"-"person". Similarly, according to the multi-level entity concept model, the entity attribute model, and the logical association relationship model, all rules and dictionaries are defined to form an unsupervised knowledge extraction model. Similarly, the rules and dictionaries are designed and generated for each language, including Chinese rules and dictionaries, English rules and dictionaries, Japanese rules and dictionaries, and Russian rules and dictionaries. Thus, a Chinese unsupervised knowledge extraction model, an English unsupervised knowledge extraction model, a Japanese unsupervised knowledge extraction model, and a Russian unsupervised knowledge extraction model are formed. Step 33: knowledge extraction is performed for each news material. According to the language of the news material itself, the supervised knowledge extraction model and the unsupervised knowledge extraction model corresponding to the language are called respectively, and the results returned by the two are fused; Knowledge extraction covers three processes: Firstly, entity extraction is performed. The input is a single news material, and the output is all entity information contained in the material. Secondly, attribute extraction is performed. The input is all entity information in the material, and the output is attribute information possessed by each entity. Finally, relationship extraction is performed, the input is all entity information in the material, and the output is logical relationship information between each group of entity pairs; Through the above knowledge extraction process, the knowledge in the news material is obtained. Step 4: Map the knowledge obtained by the knowledge extraction process and the knowledge model; Since the knowledge obtained from news materials in different languages may not be mapped to the corresponding knowledge model of the language, it is necessary to judge again according to the chapter and sentence semantic information of the news material, so as to complete the mapping between the knowledge and the knowledge model; Among them, for example, Japanese domestic news reported in English, the extracted knowledge should be mapped to the Japanese knowledge model; Indian domestic news reported in Japanese, the extracted knowledge should be mapped to the Indian knowledge model; This step 4 includes the following specific steps: Step 41: Classify each sentence in the news material according to the knowledge model, and the specific categories include the United States knowledge model, the Japanese knowledge model, the Indian knowledge model, and the Russian knowledge model; If the reliability of the classification result is high, the classification result is considered valid; Map the knowledge extracted in the sentence to the knowledge model classified; Step 42: If the reliability of the sentence classification result is low, the sentence classification result is considered invalid; At this time, classify the entire news material according to the knowledge model, and the specific categories include the United States knowledge model, the Japanese knowledge model, the Indian knowledge model, and the Russian knowledge model; Map the knowledge extracted from each sentence in the news material to the knowledge model classified; Step 43: Through the mapping process of the above steps 41 and 42, generate the output as a combination of knowledge and knowledge model, that is, a knowledge base; Here, the obtained knowledge base is a knowledge base in the source language form; Step 44: For the knowledge base in the source language form, according to the multi-level entity concept Chinese model, entity attribute Chinese model and logical association relationship Chinese model corresponding to the multi-level entity concept source language model, entity attribute source language model and logical association relationship source language model in step 2 respectively, map the source language and Chinese again, and obtain a knowledge base in Chinese form; Step 5: Since the same knowledge often has different forms, it is necessary to further perform cross-language knowledge fusion on the knowledge mapped to the knowledge model in the Chinese form knowledge base; This step is based on open source encyclopedias and open source knowledge bases to realize cross-language knowledge base construction, and based on cross-language knowledge base to realize cross-language knowledge fusion; This step 5 includes the following specific steps: Step 51: Use open source encyclopedias and open source knowledge bases to build a cross-language knowledge base; Sort and integrate these open source encyclopedias and open source knowledge bases to build a unified cross-language knowledge base; The cross-language knowledge base contains alias, attribute, description and tag information of the same entity in different language dimensions; Step 52: Align the same knowledge in different languages using the cross-language knowledge base to complete the fusion and alignment of the knowledge; In this way, cross-model and cross-language knowledge modeling and knowledge mining are finally completed, and the required cross-language news knowledge base is obtained; The step 11 comprises: creating a knowledge model of the United States, a knowledge model of Japan and a knowledge model of India. The step 11 comprises: creating a first-level entity concept of the United States, including an organization, a person, a place and a weapon. The step 11 comprises: creating a second-level entity concept of the United States according to the first-level entity concept, including a certain department of a country, a certain oil association of a country, a certain house of a country and a certain house of a country. The step 1 comprises: isolating the knowledge models of different countries in a display interface, a physical storage and a database to ensure the contact permission of different users or user groups. The step 14 comprises: giving different database open permissions to each group to isolate the groups from each other, so that different users can only contact the knowledge model data corresponding to their own permissions. The step 32 comprises: the dictionary forms include: {a certain aircraft carrier: weapon}, {a certain person: person} and {a certain place: place}. 2.The cross-model and cross-language knowledge modeling based knowledge mining method according to claim 1, wherein, The step 41 comprises: the high confidence of the classification result refers to being higher than or equal to 0.

7. 3.The cross-model and cross-language knowledge modeling based knowledge mining method of claim 1, wherein, The step 42 comprises: the low confidence of the classification result refers to being lower than 0.

7. 4.The cross-model and cross-language knowledge modeling based knowledge mining method according to claim 1, wherein, The open source encyclopedia library refers to a public encyclopedia knowledge base including Wikipedia, Baidu Encyclopedia and Interactive Encyclopedia.

5. The cross-model, cross-lingual knowledge modeling based knowledge mining method of claim 1, wherein, The open source knowledge base refers to a public knowledge base including Baklib and Raneto.

Images