Semi-open multi-language entity relationship extraction method and system without annotation data

Abstract

The invention belongs to the technical field of information extraction of natural languages, and relates to a semi-open multi-language entity relationship extraction method and system without annotation data. The method comprises the following steps: extracting a semantic connection relationship between vocabularies in a sentence by adopting a dependency analysis tool, and establishing a semantic undirected graph according to the semantic connection relationship; performing graph search of a specified depth on the head entity and the tail entity on the semantic undirected graph, and taking an intersection of search results as a relational word set; and obtaining the relationship between the entities according to the relationship word set. Most of existing relation extraction methods depend on a predetermined relation type system, and the process is complex and long in time period. On the basis of an existing mature natural language processing tool, the complicated model training process is avoided, the time cost is greatly reduced, and the method is suitable for multiple languages such as English and Chinese.

Description

Technical Field

[0001] This invention belongs to the field of natural language information extraction technology, and in particular relates to a semi-open multilingual entity relation extraction method and system that utilizes dependency analysis without requiring labeled data. Background Technology

[0002] The continuous advancement of network technology has made it easy for people to access massive amounts of text content, such as news, books, and documents. How to efficiently and accurately acquire information of interest to users has become a key research focus. To address this problem, the task of Information Extraction (IE) has emerged. Its main goal is to extract specific event or factual information from natural language text, automatically classifying, extracting, and reconstructing massive amounts of content. This information typically includes entities, relations, and events. Relation Extraction (RE) is one of the key tasks of Information Extraction, used to identify and classify semantic relationships between entities, such as binary relationships like children, spouses, and location relationships. This task has wide applications in various fields such as text mining, search engines, and knowledge graph construction.

[0003] Currently, with the continuous development of natural language processing technology, relation extraction methods based on Language Large Model (LLM) have become increasingly mainstream and have demonstrated good performance. However, these methods are typical data-driven approaches, requiring a large amount of initial data for initialization, which makes relation extraction tasks in entirely new domains always face the challenge of missing data. Therefore, it is essential to construct a relation extraction method that can start quickly without labeled data and is widely applicable.

[0004] 1. Semi-open entity relation extraction

[0005] Relation extraction can be categorized based on whether entity and relation categories are restricted: domain-bound, open, and semi-open relation extraction. Semi-open entity relation extraction is a method between domain-bound and open relation extraction. Domain-bound relation extraction means that the types of relations and entity categories are predefined, and the algorithm extracts relations only within a given domain. This method is suitable for domain-specific tasks, such as drug-disease relation extraction in medical literature. Because the relation types and entity categories are predetermined, domain-bound relation extraction can extract relation information more accurately within a specific domain, but it cannot extract relations outside the defined domain. Open relation extraction, on the other hand, has no restrictions on relation types and entity categories. The algorithm can automatically discover various relations between different entities from the text. This method typically uses unsupervised or semi-supervised learning techniques and can adapt to relation extraction tasks in different domains and corpora. The advantage of open relation extraction is that it can discover new relation types or perform transfer learning between different domains. However, its disadvantage is that due to the limited restrictions, only a portion of the results from open relation extraction may be of interest to users.

[0006] Semi-open entity relation extraction is a compromise between domain-bounded and open relation extraction. In semi-open relation extraction, entity categories are usually predefined, but relation types can be open, meaning unknown relation types can be automatically discovered. This approach balances accuracy and adaptability in relation extraction tasks. It is suitable for tasks requiring the extraction of various relations between specific entities within a specific domain, while retaining some flexibility to handle unknown relation types.

[0007] Currently, semi-open entity relation extraction is achieved using deep learning technology, but there is very little research on it. At the same time, as mentioned earlier, this method faces the challenge of data shortage.

[0008] 2. Relation extraction method based on dependency analysis

[0009] Dependency parsing (DP) is an important subtask in natural language processing, aiming to identify semantic dependencies between words in a sentence. Relation extraction methods based on dependency parsing utilize these dependencies to extract the association information between entities from text. Currently, mature dependency parsing tools exist for both Chinese and English, including the LTP language technology platform from Harbin Institute of Technology, the NLTK toolkit from the University of Pennsylvania, and the CoreNLP toolkit from Stanford University, among others. Figure 1 The analysis results are from the LTP platform of Harbin Institute of Technology. Here, HED represents the core relation, SBV represents the subject-predicate relation, DBL represents the pivotal verb, VOB represents the verb-object relation, ADV represents the adverbial-head structure, and WP represents punctuation.

[0010] In relation extraction tasks, classic dependency-based algorithms first perform dependency analysis on the text, labeling each word in the sentence with its corresponding dependency relation. These dependency relations can be subject-verb, verb-object, or modifier relations. Next, the algorithm identifies entities and the relationships between them based on these dependency relations. For example, for the sentence "John loves Mary," dependency analysis can obtain the subject-verb relation between "John" and "loves," and the verb-object relation between "loves" and "Mary." Through these dependency relations, it can be concluded that there exists a relation triple (John, loves, Mary) in the sentence.

[0011] This method is simple and direct, naturally integrating sentence structure information without high computational complexity. Previous relation extraction methods based on dependency parsing have followed this approach, using the analysis results to construct triples according to rules such as "subject-verb-object" and "verb-object structures." However, often the entities researchers are interested in are not the objects, but rather entities that may randomly exist in other semantic components of the sentence, such as modifiers or clauses. Therefore, rule-based methods alone cannot obtain the target triples. Summary of the Invention

[0012] The purpose of this invention is to provide a semi-open multilingual entity relation extraction method and system that utilizes dependency graphs without requiring labeled data, addressing the shortcomings of existing technologies. This method utilizes dependency analysis combined with word feature heuristic rules to achieve multilingual relation extraction.

[0013] The technical solution adopted in this invention is as follows:

[0014] A semi-open, multilingual entity relation extraction method that does not require labeled data includes the following steps:

[0015] Dependency analysis tools are used to extract semantic connections between words in a sentence, and a semantic undirected graph is constructed based on these semantic connections.

[0016] Perform graph search at a specified depth on the head and tail entities in a semantic undirected graph, and use the intersection of the search results as the set of relation words;

[0017] The relationships between entities are obtained from the set of relation words.

[0018] Furthermore, the extraction of semantic connections between words in a sentence using dependency analysis tools includes:

[0019] Each sentence is segmented, part-of-speech tagging is performed, named entity recognition is performed, and dependency analysis is performed to obtain sentences with segmentation, part-of-speech tagging, and dependency analysis information;

[0020] A sentence is defined as consisting of a head entity set, a tail entity set, a predicate verb, and any other words.

[0021] A triple is formed by connecting the head entity and the tail entity with the predicate verb as the relational word.

[0022] Furthermore, the categories of the part-of-speech tagging are defined as the set {F|′verb′,′noun′,′adv′,′conj′,′auxiliary′…}, where noun represents a noun, verb represents a verb, conj represents a conjunction, adv represents an adverb, and auxiliary represents an auxiliary word; the categories of the dependency analysis are defined as the set {L|′SBV′,′VOB′,′ATT′,′RAD′,′COO′…}, where SBV represents a subject-verb relationship, VOB represents a verb-object relationship, ATT represents a modification relationship, RAD represents a right-additive relationship, and COO represents a coordinate relationship.

[0023] Furthermore, the step of establishing a semantic undirected graph based on semantic connection relationships includes: establishing a semantic undirected graph G for each sentence, where each word is a node, forming a node set W; the part-of-speech tagging results of each word constitute the features of each node v; the dependency analysis results L constitute an edge set E, where each edge represents the semantic connection relationship between words; and the distance between nodes reflects the semantic association strength between nodes.

[0024] Furthermore, the graph search of the head and tail entities at a specified depth on the semantic undirected graph includes:

[0025] Starting from the head entity node h, a depth-first search is performed, with the search depth controlled by a preset parameter N, to obtain a set R of strongly related words associated with the head entity. H ;

[0026] Starting from the tail entity node t, a depth-first search is performed, with the search depth controlled by a preset parameter M, to obtain a set R of strongly related words associated with the tail entity. T ;

[0027] The set of strongly related words related to head entities R H R is a set of strongly related terms associated with tail entities. T Take the intersection to obtain the set R of relation words that have strong semantic associations with both the head and tail entities.

[0028] Furthermore, obtaining the relationships between entities based on the set of relation words includes: recording the path length between relation words and target entities during the graph search process; after obtaining the set of relation words, calculating the confidence of each word in the set of relation words based on the path length; and selecting the word with the highest confidence as the final relation word result.

[0029] Furthermore, the confidence level is calculated using the following formula:

[0030]

[0031] Where n is the number of nodes in the undirected graph, r degree The degree of the relation node r. d(h,r) represents the degree centrality of the relation node r. The larger the value, the greater the influence of the node in the graph. d(h,r) represents the node distance from the head entity node h to the relation node r, and d(t,r) represents the node distance from the tail entity node t to the relation node r. The smaller the node distance, the stronger the semantic association and the higher the confidence.

[0032] A semi-open, multilingual entity relation extraction system that does not require labeled data includes:

[0033] The semantic undirected graph construction module is used to extract semantic connections between words in a sentence using dependency analysis tools, and to build a semantic undirected graph based on these semantic connections.

[0034] The relation word set construction module is used to perform graph search of head and tail entities at a specified depth on a semantic undirected graph, and the intersection of the search results is used as the relation word set.

[0035] The relation extraction module is used to obtain the relationships between entities based on the set of relation terms.

[0036] The beneficial effects of this invention are reflected in:

[0037] Existing relation extraction methods mostly rely on a pre-defined relation type system, a complex and time-consuming process. This invention, based on existing mature natural language processing tools, avoids the cumbersome model training process, significantly reducing time costs.

[0038] This invention is applicable to multiple languages, including English and Chinese. Through testing and optimization on English and Chinese text, it can process text data in different languages, extract the required information, and has a wide range of applications.

[0039] This invention requires no large amount of initial data, avoiding the data dependency problem common in data-driven methods and reducing the need for large amounts of labeled data. It employs graph search technology to extract relationships, providing more flexible relationship discovery capabilities. Furthermore, by calculating confidence levels, it can more accurately select relational terms, improving the accuracy of relationship extraction. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the dependency analysis results;

[0041] Figure 2 This is the result of dependency analysis;

[0042] Figure 3 It is the undirected graph structure corresponding to the sentence;

[0043] Figure 4 This is a diagram illustrating the method for obtaining the set of relational terms;

[0044] Figure 5 This is a schematic diagram of the entity relationship extraction method based on dependency parsing graphs. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0046] This invention relates to a semi-open multilingual entity relation extraction method that does not require labeled data. The method includes: extracting semantic connections between words in a sentence using dependency parsing tools, and establishing an undirected graph based on semantic connections; using predicate verbs as expressions of relations; performing graph search at a specified depth on the undirected graph for both head and tail entities to obtain their respective sets of semantically strongly related words; finally, using the intersection of the search results as a set of strongly semantic relation words, and selecting relation words from the set of relation words based on their features to obtain the relationships between entities.

[0047] (1) Extracting a set of words strongly associated with entities from an undirected graph based on semantic connectivity

[0048] In this method, a language technology platform cloud is first used to perform word segmentation, part-of-speech tagging, named entity recognition and dependency analysis on each sentence to obtain sentences with word segmentation, part-of-speech tagging and dependency analysis information.

[0049] The categories for part-of-speech tagging are defined as the set {F|′verb′,′noun′,′adv′,′conj′,′auxiliary′…}, where “noun” represents a noun, “verb” represents a verb, “conj.” represents a conjunction, “adv” represents an adverb, and “auxiliary” represents an auxiliary word.

[0050] The categories for dependency parsing are defined as the set {L|′SBV′,′VOB′,′ATT′,′RAD′,′COO′...}, where “SBV” represents subject-verb relation, “VOB” represents verb-object relation, “ATT” represents modifier relation, “RAD” represents right appender relation, and “COO” represents coordinate relation. The specific categories can be changed depending on the situation.

[0051] Taking the sentence "Coptis chinensis has the effect of reducing NF-κB expression and enhancing PPAR-γ mRNA expression in the kidneys of diabetic nephropathy rats, and can alleviate kidney lesions" as an example, the head entity is "Coptis chinensis", and the tail entities are "NF-κB" and "PPAR-γ mRNA". The analysis results are as follows: Figure 2 As shown in the diagram. Word segmentation results are represented by boxes, part-of-speech tagging results are appended below the words, and dependency analysis results are represented by lines. The part-of-speech tagging categories are {F|′verb′,′noun′,′adv′,′conj′,′auxiliary′}, and the dependency analysis categories are {L|′SBV′,′VOB′,′ATT′,′RAD′,′COO′}. Figure 2 The text introduces symbolic representations, where triangles represent predicate verbs and circles represent any other words besides entities and predicate verbs. Figure 2 In this context, ROOT represents the root node of the sentence, and COO represents a parallel relationship.

[0052] This method defines a sentence X (whether in Chinese or English) as consisting of a head entity set {H|HeadEntities}, a tail entity set {T|TailEntities}, a predicate verb {V|Verbs}, and any other words {Others}, i.e.:

[0053] X={H}∪{T}∪{V}∪{Others} (1)

[0054] A triple is constructed by using a predicate verb {v|v∈V} as a relation to connect the head entity {h|h∈H} and the tail entity {t|t∈T}, that is:

[0055] (HeadEntity,Relation,TailEntity)=(h,v,t) (2)

[0056] Based on the aforementioned semantic connectivity, this method constructs an undirected graph G for each sentence, where each word is treated as a node, forming a node set W; the part-of-speech tagging results for each word constitute the features of each node v; and the dependency analysis results L form an edge set E, with each edge representing a semantic connectivity between words. The semantic connectivity between words is implicitly contained in the graph, and the distance between nodes reflects the strength of their semantic association, such as... Figure 3 As shown, ROOT represents the predicate verb identified during relation extraction.

[0057] Next, the head entity and tail entity are selected as the target entities for relation extraction, and a graph search of a specified depth is performed on the constructed undirected graph. Specifically, starting from the head entity node h, a depth-first search is performed, with the search depth controlled by a preset parameter N, ultimately obtaining a set R of strongly related terms associated with the head entity. H :

[0058] R H ={r|r∈V,d(r,h)≤N,F(r)=′verb′} (3)

[0059] Where d(r,h) represents the node distance from the head entity node h to the relation word node r, F(v) represents the node attribute, and 'verb' indicates that the attribute of node r is a verb.

[0060] Similarly, a graph search is also performed starting from the tail entity node. Starting from the tail entity node t, a depth-first search is conducted, with the search depth controlled by a preset parameter M, ultimately obtaining a set R of strongly related terms associated with the tail entity. T :

[0061] R T ={r|r∈V,d(r,t)≤M,F(r)=′verb′} (4)

[0062] Where d(r,t) represents the node distance from the tail entity node t to the relation word node r, and 'verb' indicates that the attribute of node r is a verb.

[0063] Finally, by taking the intersection of the sets of strongly related terms for the head entity and the tail entity, we obtain a set R of relational terms that can have strong semantic associations with both the head entity and the tail entity, such as... Figure 4 As shown:

[0064] R = R H ∩R T (5)

[0065] (2) Confidence measurement of relational terms

[0066] During the graph search process, the path length between relational terms and target entities is recorded. Finally, after obtaining the relational term set R, the confidence score of each term in R is calculated based on the path length. The term with the highest confidence score is selected as the final relational term result. The confidence score calculation formula is as follows:

[0067]

[0068] Where n is the number of nodes in the undirected graph, r degree This represents the degree of the relation node r. d represents the degree centrality of the relation node r; a larger value reflects the greater influence of the node in the graph. d(h,r) represents the node distance from the head entity node h to the relation node r, and d(t,r) represents the node distance from the tail entity node t to the relation node r. The smaller the node distance and the larger the reciprocal value, the stronger the semantic association and the higher the confidence.

[0069] The complete steps of the present invention are as follows: Figure 5 As shown, it includes the following steps:

[0070] Step 1: Clean the input text by removing redundant punctuation marks such as vertical lines, special characters, and extra spaces.

[0071] Step 2: Identify entities and filter text. Use any entity recognition method to search each sentence in the sentence set. If the sentence contains the entity of interest, retain the sentence and its corresponding entity information; otherwise, discard the sentence.

[0072] After step two (identifying entities and filtering text), the entities of interest in the sentence will be obtained. Since only one pair of entities can be processed at a time, two entities can be arbitrarily specified (named entity 1 and entity 2).

[0073] Step 3: Using natural language processing analysis tools, each sentence is segmented, part-of-speech tagging is performed, and dependency analysis is conducted to obtain sentence information with the results of segmentation, part-of-speech tagging, and dependency analysis.

[0074] Step 4: Based on the results of Step 3, transform the dependency analysis results of the sentences into an undirected graph structure. Each word is treated as a node, and the part-of-speech tagging result of each word constitutes the attributes of each node. Then, based on the semantic connections obtained from the dependency analysis results, establish edges between nodes.

[0075] Step 5: For the specified entity 1 and entity 2, perform a graph search on the undirected graph. Starting with entity 1, perform a depth-first search using the specified depth constraint N, adding nodes along the search path to the search results set for entity 1, and obtaining the set R of strongly related terms R associated with the head entity. H As shown in formula (3). Simultaneously, starting from entity 2, a search operation is performed using the specified depth limit M, adding nodes along the search path to the search results set of entity 2, thus obtaining a set R of strongly related terms R associated with the tail entity. T As shown in formula (4).

[0076] Step Six: Retrieve the search results R for Entity 1 and Entity 2. H and R TThe intersection of the sets yields a set R of relation words that have strong semantic associations with both the head and tail entities, as shown in formula (5). These relation words connect entity 1 and entity 2 in the undirected graph and have high semantic relevance.

[0077] Step 7: Calculate the confidence level of each candidate word in the relation word set R using formula (6), and select the word with the highest confidence level as the most suitable relation word. Finally, obtain the relation extraction results for Entity 1, Entity 2, and the relation word. For example, if Entity 1 and Entity 2 are the head entity of traditional Chinese medicine and the tail entity of a gene, respectively, then the method of this invention can be used to obtain the relationship between traditional Chinese medicine and genes. This can be used to guide research on the mechanism of action of traditional Chinese medicine, promote the formulation of personalized medical plans, strengthen the application of traditional Chinese medicine genomics, and promote the modernization and precision development of traditional Chinese medicine.

[0078] Step 8: Repeat steps 5-7 until you obtain all the entity relation extraction results for a sentence, then end.

[0079] The entity relation extraction method based on dependency graphs of this invention includes steps such as text cleaning, text filtering and named entity recognition, natural language processing analysis, graph search, and relation word selection. This method effectively extracts relationships between entities from text by utilizing dependency analysis tools to extract semantic connections between words and combining graph search and relation word selection strategies.

[0080] Example:

[0081] (1) Experimental conditions

[0082] The hardware platform for the simulation experiment of this invention is an i7-12700H processor.

[0083] The software platform for the simulation experiment of this invention is Windows 11 operating system and Python 3.8.

[0084] (2) Experiment content:

[0085] Chinese and English literature were obtained from CNKI and PubMed. For Chinese literature, the LTP language analysis platform tool from Harbin Institute of Technology was used for preprocessing, while for English literature, the Spacy natural language processing library was used.

[0086] Different search depth parameters are selected based on the text language: When dealing with Chinese documents, a depth-first search with a depth of N=4 is selected starting from the head entity node, and a depth of M=3 is selected for the tail entity; when dealing with English documents, a depth-first search with a depth of N=6 is selected starting from the head entity node, and a depth of M=4 is selected for the tail entity.

[0087] Comparative Examples: To verify the effectiveness of this invention, 100 Chinese and 100 English articles in the field of traditional Chinese medicine were selected for testing, attempting to extract drug-target relationships. The test results were compared with classic dependency analysis-based methods, as shown in Table 1. "Prior art" refers to the existing DSNF method, derived from the article "Chinese Open Relation Extraction and Knowledge Base Establishment"; "Invention Method" refers to the method proposed in this invention.

[0088]

[0089] Based on the comparative experimental results, the method of the present invention has a significant advantage over the traditional method in terms of accuracy P, and the average F1 index also shows good performance, which proves the practicality and effectiveness of the present invention in the field of relation extraction.

[0090] Table 1. Comparison of Relationship Extraction Results Indicators

[0091]

[0092] As can be seen from the above definition of recall, it reflects the number of results that can be identified, while precision reflects the quality of the identified results.

[0093] As shown in Table 1, when searching Chinese literature, the precision, recall, and average F1 score of the method of this invention are significantly higher than those of existing methods.

[0094] Although the recall rate of the method described in this invention was not higher than that of the DSNF method when searching English literature, indicating that the method of this invention could identify fewer results, the accuracy of the method of this invention was much higher than that of the DSNF method, indicating that the results of this method are of high quality and high reliability.

[0095] Another embodiment of the present invention provides a semi-open multilingual entity relation extraction system that does not require labeled data, comprising:

[0096] The semantic undirected graph construction module is used to extract semantic connections between words in a sentence using dependency analysis tools, and to build a semantic undirected graph based on these semantic connections.

[0097] The relation word set construction module is used to perform graph search of head and tail entities at a specified depth on a semantic undirected graph, and the intersection of the search results is used as the relation word set.

[0098] The relation extraction module is used to obtain the relationships between entities based on the set of relation terms.

[0099] For the specific implementation process of each module, please refer to the description of the method of the present invention above.

[0100] Another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the method of the present invention.

[0101] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) storing a computer program that, when executed by a computer, implements the various steps of the method of the present invention.

[0102] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A semi-open, multilingual entity relation extraction method that does not require labeled data, characterized in that, Includes the following steps: Dependency analysis tools are used to extract semantic connections between words in a sentence, and a semantic undirected graph is constructed based on these semantic connections. Perform graph search at a specified depth on the head and tail entities in a semantic undirected graph, and use the intersection of the search results as the set of relation words; The relationships between entities are obtained from the set of relation words.

2. The method according to claim 1, characterized in that, The extraction of semantic connections between words in a sentence using dependency parsing tools includes: Each sentence is segmented, part-of-speech tagging is performed, named entity recognition is performed, and dependency analysis is performed to obtain sentences with segmentation, part-of-speech tagging, and dependency analysis information; A sentence is defined as consisting of a head entity set, a tail entity set, a predicate verb, and any other words. A triple is formed by connecting the head entity and the tail entity with the predicate verb as the relational word.

3. The method according to claim 2, characterized in that, The categories of the part-of-speech tagging are defined as the set {F|′verb′,′noun′,′adv′,′conj′,′auxiliary′…}, where noun represents a noun, verb represents a verb, conj represents a conjunction, adv represents an adverb, and auxiliary represents an auxiliary word; the categories of the dependency parsing are defined as the set {L|′SBV′,′VOB′,′ATT′,′RAD′,′COO′…}, where SBV represents a subject-verb relation, and VOB represents a verb-object relation. ATT represents the modifying relationship, RAD represents the right appending relationship, and COO represents the parallel relationship.

4. The method according to claim 1, characterized in that, The step of establishing a semantic undirected graph based on semantic connection relationships includes: For each sentence, a semantic undirected graph G is constructed, where each word is a node, forming a node set W; the part-of-speech tagging results of each word constitute the features of each node v; the dependency analysis results L constitute an edge set E, where each edge represents the semantic connection relationship between words; the distance between nodes reflects the semantic association strength between nodes.

5. The method according to claim 1, characterized in that, The graph search of the head and tail entities at a specified depth on the semantic undirected graph includes: Starting from the head entity node h, a depth-first search is performed, with the search depth controlled by a preset parameter N, to obtain a set R of strongly related words associated with the head entity. H ; Starting from the tail entity node t, a depth-first search is performed, with the search depth controlled by a preset parameter M, to obtain a set R of strongly related words associated with the tail entity. T ; The set of strongly related words related to head entities R H R is a set of strongly related terms associated with tail entities. T Take the intersection to obtain the set R of relation words that have strong semantic associations with both the head and tail entities.

6. The method according to claim 1, characterized in that, The process of obtaining the relationships between entities based on the set of relation words includes: recording the path length between relation words and target entities during the graph search process; after obtaining the set of relation words, calculating the confidence of each word in the set of relation words based on the path length; and selecting the word with the highest confidence as the final relation word result.

7. The method according to claim 6, characterized in that, The confidence level is calculated using the following formula: Where n is the number of nodes in the undirected graph, r degree The degree of the relation node r. d(h,r) represents the degree centrality of the relation node r. The larger the value, the greater the influence of the node in the graph. d(h,r) represents the node distance from the head entity node h to the relation node r, and d(t,r) represents the node distance from the tail entity node t to the relation node r. The smaller the node distance, the stronger the semantic association and the higher the confidence.

8. A semi-open, multilingual entity relation extraction system that does not require labeled data, characterized in that, include: The semantic undirected graph construction module is used to extract semantic connections between words in a sentence using dependency analysis tools, and to build a semantic undirected graph based on these semantic connections. The relation word set construction module is used to perform graph search of head and tail entities at a specified depth on a semantic undirected graph, and the intersection of the search results is used as the relation word set. The relation extraction module is used to obtain the relationships between entities based on the set of relation terms.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 7.

Images