A method, device and system for fusing a knowledge graph and an event logic graph

Abstract

The present invention discloses a method for fusing a knowledge graph and an event logic graph, comprising: performing event extraction, event relationship extraction, event similarity calculation, and event generalization on a text corpus to form an event logic knowledge base; constructing a hypernym-hyponym concept system and an ontology to form an abstract knowledge graph; using entity recognition to match and generalize the concrete event entity words in the event logic knowledge base with the hyponyms in the hypernym-hyponym concept system into hypernyms, and using a visualization tool to construct an event logic graph; linking the event entities in the event logic graph to the corresponding knowledge graph through entity recognition and entity linking technologies to achieve the deep fusion of the knowledge graph and the event logic graph, and form a new fusion graph. The present invention can broaden the application scope of knowledge management, promote the application and iterative development of the fusion graph in the field, enrich the knowledge connotation, and expand the corpus resources for the domain knowledge base.

Description

Technical Field

[0001] The present invention relates to the technical field of common sense knowledge base construction, and specifically to a method, device and system for fusing a knowledge graph and a fact graph. Background Art

[0002] With the development of artificial intelligence and big data, knowledge graphs have become an important part of artificial intelligence and a product of the big data era. With its powerful semantic processing and open interconnection capabilities, knowledge graphs make the "knowledge network" vision of web 3.0 possible. Knowledge graphs are essentially a semantic network that describes entities or concepts in the real world and their relationships. Its basic component unit is the "entity-relationship-entity" or "entity-attribute-attribute value" triple. Entities are connected to each other through relationships to form a network of knowledge structures. It provides a better ability to organize, manage and understand the massive amount of information on the Internet. Typical application scenarios include intelligent semantic search, mobile personal assistants, and personalized recommendations.

[0003] "Event" is the core concept of human social activities. Exploring the evolution of the logic of events is of great significance to our prediction of the development trend of events and understanding the laws of social development and change. The event graph focuses on the state transition and relationship evolution between events. It takes events as entity nodes, closely integrates static knowledge and dynamic logical rules, and forms a logical link network with more powerful knowledge updating and reasoning capabilities. Therefore, the event graph can be applied to event prediction, development model mining, and decision-making knowledge reasoning.

[0004] The knowledge graph and the event graph are integrated, and the implementation technology includes event extraction, event relationship extraction, event generalization, entity recognition and entity linking. The ACE conference evaluation task defines that an event consists of an event trigger word and elements that describe the event structure. The event argument role is an entity that plays a certain role in the event, a value or time, an event participant, and an event-related attribute. The value is a non-entity event participant, for example, "detection time" and "detection location". Event extraction presents unstructured text containing event information in a structured form. There are two methods for event extraction: pattern matching-based methods and machine learning-based methods. Machine learning-based methods include extraction based on trigger words and identification of elements. The invention with patent number CN201911059404.7 proposes a method and device for constructing a domain event graph that integrates multiple types of event and entity knowledge. The method includes: extracting the relationship between facts and the elements of facts in the domain corpus to form an instance-fact logic knowledge base; constructing a hierarchical abstract concept knowledge base; using the abstract concept knowledge base to perform entity word abstraction and predicate abstraction on the instance events in the instance-fact logic knowledge base to form an abstract fact map; using the entity knowledge map, using the entity linking technology to perform entity linking on the instance events in the instance-fact logic knowledge base, integrating the event knowledge with the entity knowledge to form a fact knowledge map; merging the abstract fact map with the fact knowledge map to form a domain event map. The domain event map constructed by the invention can combine the advantages of both static entity knowledge and action event knowledge, can broaden the application scope of knowledge questions and answers, and can be used as a common sense knowledge base to expand the domain language resources. In the process of constructing the fact logic knowledge base, the invention is imperfect in that it only extracts the relationship between facts and the elements of facts in the corpus. The same event will have different expressions in different contexts. When multiple events point to the same natural event, it is considered that these events have a common reference relationship. It is necessary to calculate the similarity of the events and merge similar events into the same expression to further ensure the quality of the knowledge base. In addition, it is also a core task of event extraction to identify the event type system. Event generalization is a key step in building a causal graph. Some current generalization methods ignore the structural characteristics and dynamic changes of event composition, do not fully consider the generalization method of the noun entity word component in the event sentence, and fail to fully consider the use of the knowledge graph system to generalize the noun entity word component. In addition, the obtained node events and relationship data are visualized using graph database technology. The causal graph is a probabilistic directed graph. The uncertainty of knowledge logic makes the transition probability exist on the edge. Labeling the transition probability of events is also an important step in building the causal graph. Summary of the invention

[0005] Aiming at the deficiencies in the prior art, the present invention provides a method, device and system for fusing knowledge graphs and event graphs, which can integrate the characteristics of static entity knowledge and dynamic event knowledge to form a global knowledge base that combines static and dynamic elements. This can broaden the application scope of knowledge management, promote the application and iterative development of fused graphs in the field, enrich the knowledge connotation, and expand the corpus resources for the domain knowledge base, enabling a higher level of machine intelligence and being closer to human interpretable cognitive intelligence.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention proposes a method for fusing a knowledge graph and an event graph, the fusion method comprising the following steps:

[0008] S1. Obtain various types of source data from a data source, perform data cleaning and preprocessing on the source data to obtain structured data, and form different types of corpora;

[0009] S2. Use event extraction technology to perform event extraction, and then sequentially perform event relationship extraction, event similarity calculation, and event generalization processing to form an event logic knowledge base; the event relationships include the following five logical relationships: succession relationship, causal relationship, hyponymy relationship, conditional relationship, and reversal relationship;

[0010] S3. Construct a hyponymy concept system that simultaneously focuses on the hyponymy relationship of concepts and the related co-hyponyms; construct an ontology, which is used to identify and describe concepts in a related field and clarify the relationships between concepts from different levels of formalized models; integrate the hyponymy concept system and the ontology to form a complete abstract knowledge graph;

[0011] S4. Use entity recognition to match and generalize the concrete event entity words in the event logic knowledge base with the hyponyms in the hyponymy concept system into hypernyms, and use a visualization tool to construct an event graph;

[0012] S5. Link the event entities in the event graph to the corresponding abstract knowledge graph through entity recognition and entity linking technology to achieve the deep fusion of the abstract knowledge graph and the event graph, and form a new fused graph.

[0013] Further, in step S1, the data source includes news media, social media, and self-media.

[0014] Further, in step S2, the process of using event extraction technology to perform event extraction includes the following steps:

[0015] S21. Identify trigger words and classify them based on the method of integrating sequence element classification and dependency syntactic analysis, construct a correspondence table of "trigger word - event type", select seed trigger words in each event type and expand them based on the Word2vec method, and obtain the trigger words corresponding to each event category after screening and filtering the expansion results; Define the sentences containing trigger words as candidate event sentences, and identify the event type according to the expanded correspondence table of "trigger word - event type".

[0016] S22. Extract event elements and perform role annotation, where the roles include event time, event main body and object, event polarity, and event tense.

[0017] Further, in step S22, the process of extracting event elements and performing role annotation includes the following steps:

[0018] S221. Extract event time:

[0019] Use the IOB annotation form to annotate the time information in the training set, classify the training classifier, and find the words representing time in the sequence; Determine whether the found time words appear in the preset dictionary describing time; Map the time points described in the text to the time points in the real world; Form the event time expression through the manually set time expression template.

[0020] S222. Extract event main body and object:

[0021] Find the event word in an event sentence, and according to the dependency syntactic analysis, obtain the relationship category words and the words with parallel relationship that have a core relationship with the event word, and all are marked as event words.

[0022] Starting from the event word, find the different parts that have a subject-predicate relationship and an object-verb relationship with the event word, and perform dependency analysis on their sentence structures respectively; Combine the part that has a subject-predicate relationship with the event word with the combined attributive-modifying relationship, parallel relationship, and combined relationship as the event main body.

[0023] Combine the part that has an object-verb relationship with the event word with the combined attributive-modifying relationship, adverbial-modifying relationship, and verb-complement relationship as the event object.

[0024] S223. Extract event polarity:

[0025] Specify the possible polarity states of the event, including affirmative adverbs indicating affirmative mood, negative adverbs indicating negative mood, and possible adverbs indicating uncertainty; Construct a polarity library for different polar states and use frequent pattern mining and syntactic analysis to find the modifying relationship with the event main body and object; Match the text to be extracted with the polar state words in the polarity library, add them to the classified polarity library, and further determine the polarity of the event.

[0026] S224, extract event tense:

[0027] Construct a tense dictionary and formulate tense rules; match the text to be extracted with the tense words in the tense dictionary and add them to the classified tense dictionary library; the tenses of events include past, present, future, and other four tenses.

[0028] Furthermore, in step S2, the event generalization process includes the following steps:

[0029] The grounded theory method is used to conceptualize the event sentences of the corpus, extract the potential joint relationships between independent categories, and organize and reorganize them, including the following sub-steps:

[0030] Use selective coding to select core categories; create a clear storyline from the logical relationships between events; connect primary and secondary concept categories through the decoding paradigm model; expand faction types at the level of orientation; verify the relationship between concept categories in each faction through data; fill in categories that may need to be developed; compare the rationality of each independent logical chain, such as cause and effect, and sequential chain, generalize compatible events into an abstract high-level event, and combine related event chains;

[0031] The event components in the logic knowledge base of events are divided into noun entity words and verb event words, and part-of-speech marking is performed respectively; the entity words are matched with hyponyms in the hyponymous and hyponymous concept system, and they are corresponded to the superordinate concepts in the system to complete the abstraction of the noun entity words; the semantic similarity of the predicate event words is calculated, and the predicate event words with high semantic similarity are merged into the same predicate event word class, and the event words of this class are generalized using the constructed generalization framework to obtain a more abstract verb superordinate concept.

[0032] Furthermore, in step S3, the process of integrating the upper and lower concept systems with the ontology to form a complete abstract knowledge graph includes the following steps:

[0033] S31, using hownet as the basic component of the system architecture, combined with the domain vocabulary mining method of unsupervised learning, generates candidate phrases, obtains N word / word sequences at high frequency, and then uses the statistical feature calculation method PMI and C-value to calculate the candidate vocabulary to obtain statistical features, forming a hyponymous and hyponymous concept system;

[0034] S32, marking the polarity and corresponding weight or strength of different sentiment words, using dependency semantics-based, window-based sentiment calculation rules to complete sentiment determination at the relevant sentence and chapter level and extraction of series-level sentiment pairs, forming a semantic sentiment dictionary; mining synonyms and antonyms for text corpus to form a semantic relationship dictionary;

[0035] Combine the semantic sentiment dictionary and the semantic relation dictionary, summarize them to obtain the final semantic dictionary, perform upper-level abstraction, and form a hypernym-hyponym concept system;

[0036] S33. Obtain the usability of the structural path through the way of iterative upward abstraction to get the abstraction path of a certain word. Through this abstraction structure, complete the concept aggregation and concept expansion at different levels;

[0037] S34. Construct a finite set of semantic primitives and use it to store the hyponym relationships between concrete concepts and between attributes;

[0038] S35. Combine the abstraction thesaurus of the upper-level concepts in step S33 with the set of semantic primitives of the concrete concepts in step S34. The hypernym-hyponym concepts of each word and their descriptions form a triple, and then form a hypernym-hyponym concept system;

[0039] S36. Integrate different knowledge base concepts: Extract concept knowledge from the knowledge base constructed from Wikidata and merge it into the ConceptNet knowledge base in the form of alignment relationships, and retain the category relationships of the edges as much as possible; When fusing VisualGenome and FrameNet, allow the existence of heterogeneous nodes, explicitly or implicitly have a common correspondence, perform synonym annotation, and express the relationships between different entities in different data sources with the same edge type;

[0040] S37. Abstract the concept knowledge into different classification systems, fuse different classification systems to construct a hypernym-hyponym concept system, obtain a more accurate concept hierarchy system effect, and set the fused tree nodes as concepts;

[0041] S38. On the basis of the concept words generated in step S37, define the ontology for the connotations and extensions of concepts and abstract concepts; Ontology construction is to establish an ontology model for a specific application domain, clarify the concepts and mutual relationships in the domain; Starting from the above existing data, through the similarity and hypernym-hyponym analysis of the data, obtain a more intuitive category and hierarchy, and use the open-source software ontology construction Protege tool to form the ontology framework system of knowledge and obtain the mutual relationships between concepts;

[0042] S39. Combine the expression of the hierarchy and generalization of concepts and concept attributes in the hypernym-hyponym concept system with the clear definition of concept relationships in the ontology to form an abstract knowledge graph.

[0043] Furthermore, in step S32, the process of mining synonyms and antonyms from the text corpus to form a semantic relation dictionary includes the following steps:

[0044] S321, extracting mention words from the text, directly using the word segmentation method, and selecting some specific word segmentation results for synonym and antonym mining;

[0045] S321, for new words or different language expressions appearing in the text corpus, pattern mining, NER or noun phrase extraction are combined to obtain candidate words;

[0046] S321, using the limited synonym and antonym table constructed by manual enumeration as seed data, obtaining the features of all seed words and candidate words, considering the character and word level features from the local features; considering the semantic features of the paragraph and sentence where the target word is located from the global features;

[0047] S321, modeling is performed based on the distribution difference between the candidate words obtained in step S322 and the target synonym set to form a semantic relationship dictionary.

[0048] Furthermore, in step S4, the process of constructing a matter graph using a visualization tool includes the following steps:

[0049] The logic knowledge base built through a series of natural language processing technologies is visualized using the visualization technology Neo4j graph database, and finally forms a logic graph with an abstract form; the complete logic graph structure has three levels: basic network, abstract network and reasoning network;

[0050] The process of constructing the event graph includes the following steps:

[0051] All the node event data and event relationship data obtained are stored in the Neo4j graph database. The Neo4j graph database stores data in nodes and relationships. Through visualization operations, the stored event data information and the visualization graph formed by it can be viewed;

[0052] The event graph is formally described as: EEG = (V, E), where V = {v1, v2, v3, ..., v P} is a node set, where a node represents an event; E = {l1, l2, l3, ..., l Q} is an edge set, which represents the logical relationship between events. Each edge is a directed edge v i →v j ,w(v i |v j ) is the weight, indicating that in v i Under the conditions v j Likelihood of occurrence;

[0053] The weight is calculated by the following formula:

[0054]

[0055] Among them, count(v i , v j ) means (v i , v j ) appears in the event map.

[0056] Furthermore, in step S5, the event entities in the event graph are linked to the corresponding abstract knowledge graph through entity recognition and entity linking technology to achieve deep fusion of the abstract knowledge graph and the event graph. The process of forming a new fusion graph includes the following steps:

[0057] S51, the event nodes in the event graph have knowledge triples associated with entities in the knowledge graph, and the event entities that need to be linked are identified;

[0058] S52, align the event entity in step S51 with the entity node in the knowledge graph through entity linking, so as to build an edge, and combine the two through an edge, where the edge represents the associated entity, to achieve the connection from event to entity; through the logical relationship, point the single event as a whole to the next event as a whole, so that the entities in the event are associated, and combined with the association between the entities, further expansion and query are carried out to realize the linkage of the overall graph.

[0059] In a second aspect, an embodiment of the present invention proposes a fusion device for a knowledge graph and a fact graph, the fusion device comprising:

[0060] The corpus construction layer is used to obtain various types of source data from the data source, clean and preprocess the source data, obtain structured data, and form different types of corpora;

[0061] The logical knowledge base construction layer is used to apply event extraction technology to the corpus to extract events, extract event logical relations, calculate event similarity, and generalize events to form a logical knowledge base;

[0062] The abstract knowledge graph construction layer is used to construct a hierarchical and superior concept system with the same level and hierarchy, and integrate it with the ontology to form a complete abstract knowledge graph;

[0063] The event graph construction layer is used to use entity recognition to match the concrete event entity words in the event logic knowledge base with the hyponyms in the hyponym and hyponym concept system to generalize them into hypernyms, and use visualization tools to build the event graph.

[0064] The graph fusion layer is used to link the event entities in the matter graph to the corresponding abstract knowledge graph through entity recognition and entity linking technology, so as to achieve deep fusion of the abstract knowledge graph and the matter graph and form a new fusion graph.

[0065] In a third aspect, an embodiment of the present invention provides a fusion system of a knowledge graph and an event graph. The fusion system includes: a memory, a processor, and a program of a method for fusing a knowledge graph and an event graph running on the processor. When the program is executed by the processor, the steps of the method described above are implemented.

[0066] The beneficial effects of the present invention are:

[0067] The present invention discloses a method for fusing a knowledge graph and an event graph. The method includes: performing event extraction, event relationship extraction, event similarity calculation, and event generalization on a text corpus to form an event logic knowledge base; constructing a hypernym-hyponym concept system and an ontology to form an abstract knowledge graph; using entity recognition to match and generalize the concrete event entity words in the event logic knowledge base with the hyponyms in the hypernym-hyponym concept system into hypernyms, and using a visualization tool to construct an event graph; linking the event entities in the event graph to the corresponding knowledge graph through entity recognition and entity linking technology to achieve the deep fusion of the knowledge graph and the event graph, and form a new fusion graph. The fusion graph constructed by the present invention can integrate the characteristics of static entity knowledge and dynamic event knowledge to form a global knowledge base that combines static and dynamic elements, which can broaden the application scope of knowledge management, promote the application and iterative development of the fusion graph in the field, enrich the knowledge connotation, and expand the corpus resources for the domain knowledge base, making the machine intelligence level higher and closer to the interpretable cognitive intelligence of humans. Description of the Drawings

[0068] Figure 1 is the framework structure diagram of the fused knowledge graph and event graph of the present invention.

[0069] Figure 2 is the framework diagram for constructing the event logic knowledge base of the present invention.

[0070] Figure 3 The flow chart for constructing the event graph of the present invention.

[0071] Figure 4 is the framework diagram for constructing the abstract knowledge graph of the present invention. Detailed Embodiments

[0072] Now, the present invention will be further described in detail with reference to the accompanying drawings.

[0073] It should be noted that the terms such as "upper", "lower", "left", "right", "front", "rear", etc. cited in the invention are only for the convenience of description and are not used to limit the scope of implementation of the present invention. The change or adjustment of their relative relationships shall be regarded as the scope of implementation of the present invention when there is no substantial change in the technical content.

[0074] Embodiment 1

[0075] This embodiment mentions a framework structure diagram integrating a knowledge graph and an event logic graph. The integration method includes the following steps:

[0076] S1. Obtain various types of source data from the data source, perform data cleaning and preprocessing on the source data to obtain structured data, and form different types of corpora.

[0077] S2. Use event extraction technology to perform event extraction, and then perform event relationship extraction, event similarity calculation, and event generalization processing in sequence to form an event logic knowledge base; the event relationships include the following five logical relationships: succession relationship, causal relationship, superordinate-subordinate relationship, conditional relationship, and inversion relationship.

[0078] S3. Construct a superordinate-subordinate concept system that simultaneously pays attention to the superordinate-subordinate relationship of concepts and the related sibling concepts; construct an ontology, which is used to identify and describe the concepts in the relevant field and clarify the relationships between concepts from different levels of formalization patterns; integrate the superordinate-subordinate concept system and the ontology to form a complete abstract knowledge graph.

[0079] S4. Use entity recognition to match and generalize the concrete event entity words in the event logic knowledge base with the subordinate words in the superordinate-subordinate concept system into superordinate concepts, and use a visualization tool to construct an event logic graph.

[0080] S5. Link the event entities in the event logic graph to the corresponding abstract knowledge graph through entity recognition and entity linking technology to achieve the deep integration of the abstract knowledge graph and the event logic graph, and form a new integrated graph.

[0081] The implementation of the present invention mainly includes four steps: constructing an event logic knowledge base, constructing an event logic graph, constructing an abstract knowledge graph, and integrating the knowledge graph and the event logic graph. The process is as Figure 1 shown, and specifically includes the following steps:

[0082] I. Event Logic Knowledge Base

[0083] Figure 2 Construct a framework diagram for the event logic knowledge base, including steps such as event (element, trigger word) extraction, event relationship extraction, event similarity, and event generalization.

[0084] 1. Concept of the Event Logic Graph

[0085] The event logic graph (abbreviated as EEG) is essentially an event logic knowledge base, which describes the evolution laws and patterns between events. Structurally, it is a directed cyclic graph, where the nodes represent events, and the directed edges represent the event logic relationships such as succession, causation, condition, and superordinate-subordinate between events.

[0086] 2. Event Extraction

[0087] The core components of the event extraction task include: event type system recognition and event element extraction.

[0088] 2.1 Event Type System Recognition

[0089] (2.1.1) The above-mentioned event type recognition usually involves identifying trigger words and then judging the event type. The extraction method is to identify and classify trigger words based on the method that combines sequence element classification and dependency syntax analysis, construct a "trigger word - event type" correspondence table, select the seed trigger words in each event type and expand them based on the Word2vec method, and obtain the trigger words corresponding to each event category after screening and filtering the expansion results.

[0090] (2.1.2) Obtain candidate events, define the sentences containing trigger words as candidate event sentences, and identify the event type according to the "trigger word - event type" correspondence table expanded in 1).

[0091] 2.2 Event Element Extraction

[0092] Extract the above-mentioned event elements (time, main subject and object, polarity, tense) and perform role annotation. Event element extraction includes: extraction of event time, event main subject and object, event polarity, and tense.

[0093] (2.2.1) Event Time Extraction

[0094] The above-mentioned event time extraction refers to mapping the time point when the event described in the text occurs to the time point in the real world.

[0095] Specific extraction method: Use the method of supervised classification for sequence annotation tasks. Specifically, use the IOB annotation form to annotate the time information in the training set, classify the training classifier, and find the words representing time in the sequence. It is usually the words before and after this word or the words in the same window, as well as the basic phrase chunks of this word and the words in the same window; judge whether this time word appears in the preset dictionary describing time. Map the time point described in the text to the time point in the real world. Through the manually set time expression template, such as: month, day, year. Form the event time expression.

[0096] (2.2.2) Event Main Subject and Object Extraction

[0097] Event subject and object extraction is to extract the corresponding subjects and objects of events from text. The event subject and object extraction includes: the event subject, which refers to the doer around the event word; the event object, which refers to the recipient around the event word. The subject types of general events are personal names or company and institution names, etc. Due to the existence of syntactic structures in language expressions, such as subject-predicate-object and verb-object structures, etc., the relationships between multiple words can be analyzed based on these structures. Based on the dependency rules, the sentence is segmented, the part-of-speech is tagged, and the sentence structure is dependency-analyzed. For example, in the sentence "Young scholars such as Li Ming and Zhang Hong also participated in the academic conference", each word depends on another word. Among them, "participated" is the root node of the sentence and depends on the root. "Young" depends on "scholars", and the dependency relationship is a attributive-middle relationship; "Zhang Hong" depends on "Li Ming", and the dependency relationship is a coordinate relationship. The specific extraction method is as follows:

[0098] (201) Find the event word in an event sentence. This word is generally a verb or a verb phrase. According to the dependency syntactic analysis, obtain the relationship category word that has a core relationship with this word, mark this word as the event word, and continue to find the word that has a coordinate relationship with this word, and also take this word as the event word.

[0099] (202) Starting from the event word, find the different parts that have a subject-predicate relationship and a verb-object relationship with the event word, and respectively perform dependency analysis on the sentence structure of each part. Each part may generate an attributive-middle modification relationship, a coordinate relationship. Combine the part that has a subject-predicate relationship with the event word with the attributive-middle modification relationship, coordinate relationship, and combination relationship of the modification part as the event subject.

[0100] (203) Combine the part that has a verb-object relationship with the event word with the attributive-middle modification relationship, adverbial-middle relationship, and verb-complement relationship as the event object.

[0101] (2.2.3) Extraction of event polarity and tense

[0102] The extraction of the event polarity refers to extracting the polarity words contained in the event to determine the polarity of the event. The specific method is to first specify the possible polarity states of the event: affirmative adverbs indicating affirmative tone, negative adverbs indicating negative tone, and possible adverbs indicating uncertainty, etc. Build a polarity library for different polar states and use frequent pattern mining and syntactic analysis to find the modification relationship with the event subject and object. Match the text to be extracted with the polar state words in the polarity library and add them to the classified polarity library to further determine the polarity of the event.

[0103] The extraction of event tense refers to sorting out words from the text that express the tense of the event. The tense of an event is usually determined by constructing a tense dictionary and formulating rules for determining the tense. The text to be extracted is matched with the tense words in the tense dictionary and added to the classified tense dictionary library. For example, the World Health Organization has announced "Public Health Emergencies of International Concern" many times in the past. The event trigger word is announcement, the polarity is affirmative, and the tense is past tense. 3. Logical relationship of events in the event graph

[0104] The present invention divides the event logic relationship types into five types, including:

[0105] (3.1) Sequential relationship. Sequential relationship refers to the partial order relationship between two events that occur one after another in time. The sequential relationship in the event graph includes two situations: one is that the subsequent event occurs immediately after the preceding event ends; for example: the doctor finds that the patient faints and falls to the ground, and immediately treats the patient; the other is that the subsequent event occurs some time after the preceding event ends. (3.2) Causal relationship. Causal relationship refers to the relationship between two events, where the occurrence of the preceding event (cause) leads to the occurrence of the subsequent event (result). The causal relationship satisfies the partial order relationship in which the cause event comes first and the result event comes later. (3.3) Conditional relationship. The conditional relationship means that the preceding event is the condition for the occurrence of the subsequent event. For example, "If there are many people in the queue, then the food is delicious."

[0106] (3.4) Hierarchical relationship. There are two types of hierarchical relationship: nominal hierarchical relationship and verbal hierarchical relationship. For example, the events "tomato price increase" and "vegetable price increase" constitute a nominal hierarchical relationship; the events "injury" and "stabbing" are verbal hierarchical relationships.

[0107] (3.5) Inverse relationship. Inverse relationship means that one event is opposite to another event. It is expressed in the form of "although A but B". For example, although it starts a little later, it takes the least time.

[0108] 4. Event logical relationship extraction method

[0109] The event relationship extraction refers to extracting the relationship between events from a text and storing it in a knowledge database. The logical relationship extraction method includes:

[0110] (4.1) Extraction methods of explicit logical relationship qualifiers and implicit syntactic analysis. When extracting event relations from text, the relation extraction task is generally divided into explicit relation extraction and implicit relation extraction according to whether the text contains relation words. For example, in the task of extracting explicit causal relations, cause and effect events are extracted based on the obvious causal association words such as "the reason" and "cause" in the text. In the extraction, the event relations can be further divided into one cause and one effect, one cause and multiple effects, etc. The extraction of implicit relations mainly adopts the method of analyzing syntactic patterns to extract specific event pairs and their relations. (4.2) The extraction method adopts word sequence. The short sentences are segmented, part-of-speech tagged, and stop words are used to form word sequences. The extracted events have the characteristics of rich semantics and shorter form than short sentences. The dependency syntactic analysis method is used to label the semantic roles of events to form subject phrases, verb-object phrases, and subject-predicate-object phrases, which have the characteristics of semantic solidification. 5. Event similarity calculation

[0111] Calculate the semantic similarity of events extracted from the text corpus. Similarity is to compare the similarity of two event pairs, usually combined with cosine similarity and Jaccard similarity to measure the similarity of event pairs.

[0112] (5.1) Jaccard similarity regards two events as sample sets, and divides the size of the intersection of the two events by the size of their union as the event similarity. The Jaccard coefficient only cares whether the common features between events are consistent. The larger the coefficient value, the higher the similarity, and the events are considered to be similar if they have common features.

[0113] (5.2) The idea of the cosine similarity method is that causal logical event pairs can be represented as a spatial vector model, and causal events are represented as fixed-dimensional vectors: The smaller the angle, the more similar the two vectors are, and thus the more similar the two events are.

[0114] 6. Event Generalization

[0115] The event generalization is to extract the logical relationship between specific events from the text, but the low-level specific events are very sparse, resulting in a small possibility that the two extracted event pairs contain the same event, so it is difficult to connect the event pairs into a dense network. Therefore, it is necessary to generalize the extracted specific events, from the logic of the specific events to the logic of the abstract events, so as to discover more general logic laws, and construct a logic map containing abstract events after generalizing the extracted events. Specific generalization methods include:

[0116] 6.1 The grounded theory method conceptualizes the event sentences of the corpus, extracts the potential joint relationship between independent categories, and organizes and reorganizes them. Using the selective coding method, after selecting the core category, it mainly includes 5 steps:

[0117] (6.1.1) Create a clear storyline from the logical relationship between events; 2) Connect the main and secondary concept classes through the decoding paradigm model; 3) Expand the faction type at the level of orientation; 4) Verify the relationship between the concept classes in each faction through data; 5) Fill in the classes that may need to be developed. Compare the rationality of each independent logical chain, such as cause and effect, and sequential chain, generalize compatible events into an abstract high-level event, and combine related event chains.

[0118] (6.1.2) The event components in the logic knowledge base are divided into nominal entity words and verb event words, and part-of-speech tags are performed separately. The entity words are matched with the hyponyms in the hyponymy concept system, and corresponded to the hypernym concept in the system to complete the abstraction of the nominal entity words. The semantic similarity of the predicate event words is calculated, and the predicate event words with high semantic similarity are merged into the same predicate event word class. The event words of this class are generalized using the constructed generalization framework to obtain a more abstract verb hypernym concept.

[0119] 2. Abstract Knowledge Graph

[0120] The abstract knowledge graph is a common sense knowledge base formed by combining the hierarchical and generalized expressions of concepts and concept attributes by the upper and lower concept system with the clear definition of concept relationships by the ontology. The description of concepts in the encyclopedia is unstructured. In order for humans to understand its concept information more clearly, a more abstract description is needed to explain it. Therefore, a higher and lower concept system is needed as the basic knowledge support, which not only focuses on the upper relationship of the concept but also focuses on the related concepts of the same level. The upper concept usually exists in the text in an unstructured form. For example, "broccoli and potatoes are both vegetables". Descriptive extraction algorithms can be used, including extraction methods based on grammatical rule templates; subject and subject description sequence label recognition methods based on BIO tags, to identify the explanation and obtain the descriptive triple of <broccoli, belongs to, vegetables>. Through the upper abstraction of the concept, some other types related to "broccoli" are obtained; "broccoli" is a "vegetable", and through the abstract concept of "vegetable", other vegetables such as "potato" and "tomato" are obtained. Combining abstract concepts with the descriptive explanation information of the concepts themselves can deepen our clear understanding of "broccoli". The knowledge system is highly abstract and general. Generality is to connect the essential attributes of some abstract things to form a more complete knowledge system. The upper and lower knowledge system is an important means to formalize the ability of abstraction and generalization by directly constructing concepts and concept attributes. Ontology construction is to establish an ontology model for specific application fields and clarify the relationship between concepts in the field. The previous step has generated a large number of upper and lower concepts. On the basis of conceptual terms, concepts and the connotation and extension of concepts can define the ontology. In specific implementation, follow Figure 4 The process shown includes the following steps:

[0121] 1) Hownet based on the Chinese version of wordnet. This knowledge base has built hundreds of hierarchical concepts and has described and organized the meanings of more than 60,000 Chinese words. With hownet as the basic component of the system architecture, combined with the domain vocabulary mining method of unsupervised learning, candidate phrases are generated, and high-frequency N-Gram (continuous N word / word sequences) is obtained. By using the statistical feature calculation method PMI (point mutual information) and C-value to calculate the statistical features of candidate words, a large-scale hyponymous and hyponymous concept system can be formed.

[0122] 2) Semantic Dictionary

[0123] 201) Semantic sentiment dictionary. The sentiment dictionary can form a semantic sentiment dictionary, mark the polarity and corresponding weight or strength of different sentiment words, and use dependency semantics-based and window-based sentiment calculation rules to complete the sentiment judgment at the relevant sentence and chapter level and the extraction of series-level sentiment pairs.

[0124] 202) Semantic relation dictionary. Pairs of synonyms and antonyms described in Framenet-like forms are used to mine synonyms and antonyms from the text corpus to form a semantic relation dictionary.

[0125] 2021) Extract mention words from the text. Directly use the word segmentation method, and select some specific word segmentation results for synonym and antonym mining.

[0126] 2022) For new words or different language expressions that appear in the text corpus, based on the method in 1), combine methods such as Pattern mining, NER, or noun phrase extraction to obtain candidate words.

[0127] 2023) Use the finite synonym and antonym list constructed by manual enumeration as seed data, obtain the features of all seed words and candidate words, and consider character and word-level features from local features; consider the semantic features of the paragraph and sentence where the target word is located from global features.

[0128] 2024) Model from the perspective of the distribution difference between the candidate words obtained in 2) and the target synonym set to form a semantic relation dictionary.

[0129] Combine the dictionaries obtained by the two methods above, summarize them jointly to obtain the final semantic dictionary, and perform upper-level abstraction to form a hypernym-hyponym concept system.

[0130] 3) Abstract word library. As an important resource for the knowledge ontology description vocabulary in the knowledge graph and semantic reasoning. Its abstractness determines that the description of specific instances tends to be upper-level types, providing good structural hierarchy and usability. Obtain the usability of the structural path through the iterative upward abstraction method to get the abstract path of a certain word. Through this abstract structure, complete the concept aggregation and concept extension at different levels.

[0131] 4) Sememe set. A sememe is the most basic and indivisible minimum unit of meaning. It can be a collection of various attributes and concrete concepts, and can be constructed into a finite sememe set. Use it to store and describe the hyponymy relationships between concrete concepts and between attributes.

[0132] Combine the abstract word library of the hypernym concepts in 3) with the sememe set of the concrete concepts in 4). The hypernym and hyponym concepts and their descriptions of each word form a triple, and then form a hypernym-hyponym concept system.

[0133] 5) Integrate different knowledge base concepts. Extract conceptual knowledge from the knowledge base constructed from Wikidata and merge it into the ConceptNet knowledge base in the form of alignment relationships, preserving the category relationships of the edges as much as possible. When fusing Visual Genome and FrameNet, the existence of heterogeneous nodes is allowed, with an explicit or implicit common correspondence for synonym annotation. The relationships between different entities in the data sources can be expressed using the same edge type.

[0134] 6) Baidu, Hudong, and Wikipedia focus on depicting concepts and abstract concept knowledge into different classification systems, such as tree and graph conceptual architectures. When fusing different systems to construct a hypernym-hyponym concept system, a more precise conceptual hierarchy system effect can be obtained, and the fused tree nodes are set as "concepts".

[0135] 7) Ontology construction

[0136] A large number of abstract concepts have been generated in the previous step. Based on the concept words, the connotations and extensions of concepts and abstract concepts can be defined for the ontology. Ontology construction is to establish an ontology model for a specific application domain, clarifying the concepts and their mutual relationships within the domain. Starting from the existing data above, through similarity and hypernym-hyponym analysis of the data, more intuitive categories and hierarchies are obtained. Using numerous plugins and functions in the open-source software ontology construction tool Protege, according to certain logical rules, a knowledge ontology framework system is formed to obtain the mutual relationships between concepts.

[0137] Combining the hierarchical and general expressions of concepts and concept attributes in the hypernym-hyponym concept system with the explicit definition of concept relationships in the ontology, the formed knowledge base is called the "abstract knowledge graph".

[0138] III. Event graph

[0139] The event graph refers to an event logic knowledge base constructed through a series of natural language processing technologies, which is visualized using the visualization technology Neo4j graph database to form an event graph in an abstract form. A complete event graph structure has three levels: the basic network, the abstract network, and the reasoning network. The specific method for graph construction is as follows:

[0140] Store all the node event data and event relationship data obtained in the previous steps into the Neo4j graph database respectively. The Neo4j graph database stores the data in nodes and relationships. After visualization operations, the stored event data information and the formed visualization graph can be viewed. The event graph is a directed cyclic graph, and its formal description: EEG = (V, E), where V = {v1, v2, v3,..., v P} is a node set, where a node represents an event; E = {l1, l2, l3, ..., l Q} is an edge set, which represents the logical relationship between events, such as cause and effect, sequence, etc. Each edge is a directed edge v i →v j ,w(v i |v j ) is the weight, indicating that in v i Under the conditions v j The probability of occurrence can be calculated by the following formula:

[0141]

[0142] Among them, count(v i , v j ) means (v i , v j ) appears in the event map.

[0143] IV. Graph Fusion

[0144] 4.1 Methods of fusion graphs

[0145] The construction of the fusion graph refers to the fusion of the knowledge graph and the event graph to form a global knowledge base that combines dynamic and static. The new fusion graph opens the connection between events and entities and forms an event and entity information network. The fusion process includes the following steps:

[0146] (4.1.1) Entity recognition. The event nodes in the event graph have some knowledge triples associated with entities in the knowledge graph, and the event entities that need to be linked are identified.

[0147] (4.1.2) Entity linking. Through entity linking, the event entity in 1) is aligned with the entity node in the knowledge graph to build an edge. The two are combined through an edge. The edge represents the associated entity, realizing the connection from event to entity. In this way, a fusion effect from logic to entity can be achieved. The knowledge graph contains a large amount of entity attribute information. The new knowledge form graph after the fusion of facts and knowledge, the complexity of the relationship between facts and knowledge will further expand the relationship between facts and knowledge. Events may occur in succession. A single event as a whole needs to point to the next event as a whole through a logical relationship. The entities in the event are associated. Combined with the relationship between entities, it can be further expanded and queried, thereby realizing the linkage of the overall graph.

[0148] Embodiment 2

[0149] This embodiment proposes a fusion device for knowledge graph and matter graph, and the fusion device includes the following modules:

[0150] (1) Corpus construction layer, which is used to obtain various types of source data from data sources, clean and preprocess the source data to obtain structured data, and form different types of corpora.

[0151] (2) Event logic knowledge base construction layer, which is used to perform event extraction, event logical relationship extraction, event similarity calculation, and event generalization on the corpus by using event extraction technology to form an event logic knowledge base.

[0152] (3) Abstract knowledge graph construction layer, which is used to construct a hierarchical concept system with peer and hierarchical levels, and integrate with the ontology to form a complete abstract knowledge graph.

[0153] (4) Event graph construction layer, which is used to match and generalize the concrete event entity words in the event logic knowledge base with the lower-level words in the hierarchical concept system by using entity recognition to form upper-level concepts, and use visualization tools to construct an event graph.

[0154] (5) Graph fusion layer, which is used to link the event entities in the event graph to the corresponding abstract knowledge graph through entity recognition and entity linking technology, realize the deep fusion of the abstract knowledge graph and the event graph, and form a new fusion graph.

[0155] Through the fusion device of the second embodiment of the present invention, the transmission object is determined by establishing the data inclusion relationship of the entire application, achieving the goal of fusing the knowledge graph and the event graph. The fusion device provided by the embodiments of the present invention can execute the fusion method of the knowledge graph and the event graph provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing the method.

[0156] Embodiment Three

[0157] The embodiment of the present application provides an electronic device, including a processor, a memory, an input device, and an output device; in the electronic device, the number of processors can be one or more; the processor, memory, input device, and output device in the electronic device can be connected through a bus or other means.

[0158] The memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the detection method in the embodiments of the present invention. The processor executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory, that is, realizes the fusion method of the knowledge graph and the event graph provided by the embodiments of the present invention.

[0159] The memory may mainly include a program storage area and a data storage area. Among them, the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some instances, the memory may further include a memory remotely provided with respect to the processor, and these remote memories may be connected to the electronic device through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0160] The input device can be used to receive input digital or character information, and generate key signal inputs related to the user settings and function control of the electronic device, and may include a keyboard, a mouse, etc. The output device may include a display device such as a display screen.

[0161] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.

Claims

1. A method for fusing a knowledge graph and an event logic graph, characterized in that, The fusion method comprises the following steps: S1, obtain various types of source data from the data source, clean and preprocess the source data, obtain structured data, and form different types of corpora; S2, using event extraction technology to extract events, and then sequentially extract event relations, calculate event similarity, and generalize events to form a logical knowledge base; event relations include the following five logical relations: sequential relations, causal relations, superior-subordinate relations, conditional relations, and inversion relations; S3, build a hierarchical concept system, which focuses on the hierarchical relationship of concepts and related concepts of the same level; build an ontology, which is used to identify and describe concepts in related fields and clarify the relationship between concepts in formal models at different levels; integrate the hierarchical concept system with the ontology to form a complete abstract knowledge graph; S4, using entity recognition to match concrete event entity words in the logic knowledge base with hyponyms in the hyponymy concept system to generalize them into hypernyms, and using visualization tools to construct a logic graph; S5, through entity recognition and entity linking technology, link the event entities in the event graph to the corresponding abstract knowledge graph, realize the deep fusion of the abstract knowledge graph and the event graph, and form a new fusion graph; In step S5, the event entities in the event graph are linked to the corresponding abstract knowledge graph through entity recognition and entity linking technology to achieve deep fusion of the abstract knowledge graph and the event graph. The process of forming a new fusion graph includes the following steps: S51, the event nodes in the event graph have knowledge triples associated with entities in the knowledge graph, and the event entities that need to be linked are identified; S52, align the event entity in step S51 with the entity node in the knowledge graph through entity linking, so as to build an edge, and combine the two through an edge, where the edge represents the associated entity, to achieve the connection from event to entity; through the logical relationship, point the single event as a whole to the next event as a whole, so that the entities in the event are associated, and combined with the association between the entities, further expansion and query are carried out to realize the linkage of the overall graph.

2. The fusion method of the knowledge graph and the event logic graph according to claim 1, characterized in that In step S2, the process of extracting events using event extraction technology includes the following steps: S21, identify and classify trigger words based on the method of integrating sequence element classification and dependency syntactic analysis, build a "trigger word-event type" correspondence table, select seed trigger words in each event type and expand them based on the Word2vec method, filter the expansion results to obtain the trigger words corresponding to each event category; define sentences containing trigger words as candidate event sentences, and identify event types based on the expanded "trigger word-event type" correspondence table; S22, extract event elements and label them with roles, which include event time, event subject and object, event polarity and event tense.

3. The method for fusing a knowledge graph and an event logic graph according to claim 2, wherein, In step S22, the process of extracting event elements and labeling roles includes the following steps: S221, extract event time: Use IOB annotation to annotate the time information in the training set, classify the training classifier, find the words that represent time in the sequence; determine whether the found time words appear in the preset dictionary describing time; correspond the time points described in the text to the time points in the real world; form event time expressions through manually set time expression templates; S222, extract event subject and object: Find an event word in an event sentence, and obtain the relational category words that have a core relationship with the event word and the words that have a parallel relationship based on dependency syntactic analysis, and mark them as event words; Starting from the event word, find the different parts that have subject-predicate relationship and verb-object relationship with the event word, and analyze the sentence structure of each part respectively; combine the part that has subject-predicate relationship with the event word with the attributive modification relationship, parallel relationship, and joint relationship as the event subject; The part that has the verb-object relationship with the event word is combined with the attributive-predicate-modification relationship, adverbial-predicate relationship, and verb-complement relationship as the event object; S223, extract event polarity: Specify the possible polarity states of the event, including affirmative adverbs, negative adverbs, and uncertain possible adverbs; construct polarity libraries for different polarities and use frequent pattern mining and syntactic analysis to find the modification relationship with the subject and object of the event; match the text to be extracted with the polarity words in the polarity library, add them to the classified polarity library, and further determine the polarity of the event; S224, extract event tense: Construct a tense dictionary and formulate tense rules; match the text to be extracted with the tense words in the tense dictionary and add them to the classified tense dictionary library; the tenses of events include past, present, future, and other four tenses.

4. The method for fusing a knowledge graph and an event logic graph according to claim 1, characterized in that In step S2, the event generalization process includes the following steps: The grounded theory method is used to conceptualize the event sentences of the corpus, extract the potential joint relationships between independent categories, and organize and reorganize them, including the following sub-steps: Use selective coding to select core categories; create a clear storyline from the logical relationships between events; connect primary and secondary concept categories through the decoding paradigm model; expand faction types at the level of orientation; verify the relationship between concept categories in each faction through data; fill in categories that may need to be developed; compare the rationality of each independent logical chain, such as cause and effect, and sequential chain, generalize compatible events into an abstract high-level event, and combine related event chains; The event components in the logic knowledge base of events are divided into noun entity words and verb event words, and part-of-speech marking is performed respectively; the entity words are matched with hyponyms in the hyponymous and hyponymous concept system, and they are corresponded to the superordinate concepts in the system to complete the abstraction of the noun entity words; the semantic similarity of the predicate event words is calculated, and the predicate event words with high semantic similarity are merged into the same predicate event word class, and the event words of this class are generalized using the constructed generalization framework to obtain a more abstract verb superordinate concept.

5. The method for fusing a knowledge graph and an event logic graph according to claim 1, characterized in that, In step S3, the process of integrating the upper and lower concept systems with the ontology to form a complete abstract knowledge graph includes the following steps: S31. Using the basic components of the HowNet architecture and combining the domain vocabulary mining method of unsupervised learning, generate candidate phrases, obtain N-character / word sequences with high frequency, and then use the statistical feature calculation methods PMI and C-value to calculate candidate words to obtain statistical features, forming a hypernym-hyponym concept system; S32. Mark the polarity and corresponding weights or intensities of different sentiment words, and use the sentiment calculation rules based on dependency semantics and window-based methods to complete the sentiment determination at the sentence and text levels and the extraction task of serial sentiment pairs, forming a semantic sentiment dictionary; mine synonyms and antonyms from the text corpus to form a semantic relation dictionary; Combine the semantic sentiment dictionary and the semantic relation dictionary, summarize them jointly to obtain the final semantic dictionary, and perform upper-level abstraction to form a hypernym-hyponym concept system; S33. Obtain the usability of the structural path by iteratively abstracting upward to get the abstract path of a certain word. Through this abstract structure, complete the concept aggregation and concept expansion at different levels; S34. Construct a finite set of sememes and use it to store the hyponym relationships between concrete concepts and between attributes; S35. Combine the abstract word library of the hypernym concepts in step S33 with the set of sememes of the concrete concepts in step S34. The hypernym-hyponym concepts and their descriptions of each word form a triple, and then form a hypernym-hyponym concept system; S36. Integrate different knowledge base concepts: Extract concept knowledge from the knowledge base constructed from Wikidata and merge it into the ConceptNet knowledge base in the form of alignment relationships, and try to retain the category relationships of the edges as much as possible; When fusing Visual Genome and FrameNet, allow the existence of heterogeneous nodes, explicitly or implicitly have a common correspondence, perform synonym annotation, and express the relationships between different entities in different data sources with the same edge type; S37. Abstract the concept knowledge into different classification systems, fuse different classification systems to construct a hypernym-hyponym concept system, obtain a more accurate concept hierarchy system effect, and set the fused tree nodes as concepts; S38. Based on the concept words generated in step S37, define the ontology of the concept and the connotation and extension of the abstract concept; Ontology construction is to establish an ontology model for a specific application domain, clarify the concepts and mutual relationships in the domain; Through the similarity and hypernym-hyponym analysis of the data, obtain a more intuitive category and hierarchy, and use the open-source software ontology construction tool Protege to form the ontology framework system of knowledge and obtain the mutual relationships between concepts; S39. Combine the hierarchical and general expressions of the hypernym-hyponym concept system for concepts and concept attributes with the clear definition of concept relationships by the ontology to form an abstract knowledge graph.

6. The fusion method of the knowledge graph and the event logic graph according to claim 5, wherein In step S32, the process of mining synonyms and antonyms from the text corpus to form a semantic relation dictionary includes the following steps: S321. Extract mention words from the text, directly use the word segmentation method, and select some specific word segmentation results for synonym and antonym mining; S321, for new words or different language expressions appearing in the text corpus, pattern mining, NER or noun phrase extraction are combined to obtain candidate words; S321, using the limited synonym and antonym table constructed by manual enumeration as seed data, obtaining the features of all seed words and candidate words, considering the character and word level features from the local features; considering the semantic features of the paragraph and sentence where the target word is located from the global features; S321, modeling is performed based on the distribution difference between the candidate words obtained in step S322 and the target synonym set to form a semantic relationship dictionary.

7. The method for fusing a knowledge graph and an event logic graph according to claim 1, wherein In step S4, the process of constructing a matter-of-fact graph using a visualization tool includes the following steps: The logic knowledge base built through a series of natural language processing technologies is visualized using the visualization technology Neo4j graph database, and finally forms a logic graph with an abstract form; the complete logic graph structure has three levels: basic network, abstract network and reasoning network; The process of constructing the event graph includes the following steps: All the node event data and event relationship data obtained are stored in the Neo4j graph database. The Neo4j graph database stores data in nodes and relationships. Through visualization operations, the stored event data information and the visualization graph formed by it can be viewed; The event graph is formally described as: EEG = (V, E), where V = {v1, v2, v3, ..., v P } is a node set, where a node represents an event; E = {l1, l2, l3, ..., l Q } is an edge set, which represents the logical relationship between events. Each edge is a directed edge v i →v j ,w(v i |v j ) is the weight, indicating that in v i Under the conditions v j Likelihood of occurrence; The weight is calculated by the following formula: Among them, count(v i , v j ) means (v i , v j ) appears in the event map.

8. A fusion device for a knowledge graph and an event logic graph based on the method described in any one of claims 1-7, characterized in that, The fusion device comprises: The corpus construction layer is used to obtain various types of source data from the data source, clean and preprocess the source data, obtain structured data, and form different types of corpora; The logical knowledge base construction layer is used to apply event extraction technology to the corpus to extract events, extract event logical relations, calculate event similarity, and generalize events to form a logical knowledge base; The abstract knowledge graph construction layer is used to construct a hierarchical and superior concept system with the same level and hierarchy, and integrate it with the ontology to form a complete abstract knowledge graph; The event graph construction layer is used to use entity recognition to match the concrete event entity words in the event logic knowledge base with the hyponyms in the hyponym concept system to generalize them into hypernyms, and use visualization tools to build the event graph; The graph fusion layer is used to link the event entities in the matter graph to the corresponding abstract knowledge graph through entity recognition and entity linking technology, so as to achieve deep fusion of the abstract knowledge graph and the matter graph and form a new fusion graph.

9. A fusion system of a knowledge graph and an event logic graph, characterized in that, The fusion system includes: a memory, a processor, and a program of a method for fusing a knowledge graph and a matter graph running on the processor. When the program is executed by the processor, the steps of the method as described in any one of claims 1 to 7 are implemented.

Images