A method and device for quickly constructing a semantic relation extraction training corpus

Abstract

The application discloses a kind of semantic relation extraction training corpus quick construction method, comprising the following steps: obtaining geology and mineral resources field related corpus, including to be annotated geology and mineral resources field related corpus and structured geology and mineral resources field related corpus;Geology and mineral resources field semi-structured, structured data are obtained, and structured data in geology and mineral resources field related corpus is extracted, and geology and mineral resources field knowledge base is constructed;Based on to be annotated geology and mineral resources field related corpus and the constructed knowledge base, the quick construction of training corpus is realized.Compared with the method of artificial annotation, the time used for annotating semantic relation training corpus is greatly reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph, and in particular to a method and device for quickly constructing training corpus of semantic relation extraction. BACKGROUND

[0002] With the development and maturity of knowledge graph technology, more and more geologists apply it to the field of geology, which provides a new way and direction for intelligent prospecting, ore prediction and ore exploration. In addition to mapping from relational data, the main construction method of deposit field knowledge graph is to extract entities and semantic relations between entities from unstructured geological text data to construct field knowledge graph. Text data mining mainly includes two aspects of named entity recognition and relation extraction. Among them, relation extraction is to extract unknown relationship facts from pure text and add them to the knowledge graph, which is the key to automatically constructing large-scale knowledge graph.

[0003] The semantic relationship between entities can be extracted by various methods, including commonly used extraction methods based on rules, machine learning and deep learning. Based on rules, the extraction rules need to be defined manually, matched with corpus, and then the relationship is extracted. It is mainly divided into two parts: trigger (pattern) based relation extraction and dependency syntax analysis based relation extraction. The extraction of semantic relationship between entities based on machine learning and deep learning often needs a large amount of annotated corpus. Annotated corpus is time-consuming and labor-intensive, and it is often difficult to add new semantic relationships. It is not very good for new fields, and the generalization ability is weak. Semi-supervised methods are often used to obtain good model results through iteration based on a small amount of annotated corpus.

[0004] With the continuous application of computer computing power and deep learning models, deep learning is widely used in the extraction of entity semantic relationship in the field of geology, which requires a large amount of high-quality artificial annotation data. Through the pure artificial method to construct training corpus, not only the annotation personnel need to have good background knowledge in the field of geology and mineral resources, but also have the disadvantages of low efficiency and high cost; the present application proposes a method for quickly constructing training corpus of semantic relation extraction in the field of geology and mineral resources based on distant supervision learning, which realizes the rapid construction and expansion of training corpus of entity semantic relationship in the field of geology and mineral resources.

[0005] The Chinese invention patent (CN115130458A Financial Field Based on Remote Supervision Relationship Extraction Model Design and Optimization Method) extracts the relationship between entities in the financial field text based on the remote supervision relationship extraction model, thereby supplementing the knowledge base in the financial field. The Chinese invention patent (CN115221877A News Text Entity Relationship Extraction Sample Annotation Method Based on Remote Supervision Technology) discloses a news text entity relationship extraction sample annotation method based on remote supervision technology, wherein a remote supervision template is prepared using the classified news text category. The geological and mineral resources field includes numerous and complex entity semantic relationships, and due to the accumulation of a large amount of relationship data and domain background knowledge, the number of domain semantic relationships applied in the related technologies mentioned in the above patent is small and relatively simple, and the pain points of rapid construction of semantic relationship extraction training corpus in the geological and mineral resources field cannot be well solved. SUMMARY

[0006] In order to solve the above problems, the present application provides a semantic relationship extraction training corpus rapid construction method, comprising the following steps:

[0007] S1, obtaining geological and mineral resources field related corpus, including to-be-labeled geological and mineral resources field related corpus and structured geological and mineral resources field related corpus;

[0008] S2, obtaining semi-structured and structured data of the geological and mineral resources field, and extracting structured data in the geological and mineral resources field related corpus to construct a geological and mineral resources field knowledge base;

[0009] S3, based on the to-be-labeled geological and mineral resources field related corpus and the constructed knowledge base, realizing rapid construction of training corpus.

[0010] Further, the geological and mineral resources field entity in step S1 includes ore-bearing strata, tectonic environment, ore-controlling strata, alteration type, alteration zoning, ore-controlling structure, ore-controlling rock mass, deposit location, mineral, and element.

[0011] Further, step S2 is specifically:

[0012] S21, obtaining semi-structured and structured data of the geological and mineral resources field;

[0013] S22, using a rule-based method to extract structured data from the geological and mineral resources field related corpus;

[0014] S23, using a regular expression to convert the semi-structured data into structured data;

[0015] S24, knowledge fusion of the structured data in S21, S22 and S23 to construct a triple data and obtain a geological and mineral resources field knowledge base.

[0016] Further, step S3 is specifically:

[0017] S31, using regular expressions to filter out all Chinese characters in the geological and mineral resources field knowledge base entities, and converting them into dictionary format to obtain entity dictionary;

[0018] S32, removing characters other than Chinese characters and Chinese commonly used punctuation in the to-be-labeled geological and mineral resources related corpus, and splitting the multiple sentences in the to-be-labeled geological and mineral resources related corpus after removing the characters into list format;

[0019] S33, matching the entities in the list format sentences in S32 and the entities in the entity dictionary in S31 and adding relationships.

[0020] Further, step S33 is specifically:

[0021] S331, traversing each sentence in the list format sentence, saving all entities appearing in the sentence according to string matching, and filtering out sentences without entities or with only one entity appearing;

[0022] S332, designing semantic relationships, aligning and aggregating the relationships in the to-be-labeled geological and mineral resources related corpus that have not been aligned by using artificial rules;

[0023] S333, traversing each pair of entities obtained in S331, and annotating the relationships according to the entity dictionary and the designed semantic relationships.

[0024] Also proposed is a semantic relationship extraction training corpus rapid construction device, characterized in that the device comprises:

[0025] a processor;

[0026] a memory having a computer program stored thereon and executable on the processor;

[0027] wherein the computer program is executed by the processor to implement a semantic relationship extraction training corpus rapid construction method.

[0028] The technical solution provided by the present application has the following beneficial effects:

[0029] Using the remote supervision method, the triple data in the knowledge base is extracted to align the to-be-labeled corpus for labeling, thereby realizing the rapid construction of the relationship extraction training corpus. Compared with the manual labeling method, the time required for labeling the semantic relationship corpus is greatly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a flowchart of the technical scheme of the embodiment of the present application;

[0031] Figure 2 is a flow chart of the technical scheme of the embodiment of the present application;

[0032] Figure 3 is a flow chart of the remote supervision training corpus construction of the embodiment of the present application;

[0033] Figure 4 is the remote supervision hypothesis of the embodiment of the present application;

[0034] Figure 5 is a flow chart of the rule-based structured data extraction of the embodiment of the present application;

[0035] Figure 6 is a preliminary knowledge base part display (taking the Ximeng mining area as an example) of the embodiment of the present application;

[0036] Figure 7 is a remote supervision training corpus construction project of the embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be further described below with reference to the drawings.

[0038] The flow chart of the semantic relation extraction training corpus rapid construction method of the embodiment is as shown in Figure 1 The flow chart of the technical scheme of the embodiment of the present application is referred to Figure 2 and includes the following steps:

[0039] S1, acquiring the related corpus of the geological and mineral product field, including the geological and mineral product field related corpus to be labeled and the structured geological and mineral product field related corpus. The geological and mineral product field entity includes: ore-bearing stratum, tectonic environment, ore-controlling stratum, alteration type, alteration zoning, ore-controlling structure, ore-controlling rock mass, deposit location, mineral, element, etc.

[0040] S2, acquiring the semi-structured and structured data of the geological and mineral product field, and extracting the structured data in the geological and mineral product field related corpus to construct the geological and mineral product field knowledge base. Specifically:

[0041] S21, acquiring the semi-structured and structured data of the geological and mineral product field;

[0042] S22, using a rule-based method to extract the structured data from the geological and mineral product field related corpus;

[0043] S23, using a regular expression to convert the semi-structured data into structured data;

[0044] S24, fusing the structured data in S21, S22 and S23 to construct a triple data and obtain the geological and mineral product field knowledge base.

[0045] S3, based on the geological mineral field related corpus to be annotated and the constructed knowledge base, realizing rapid construction of training corpus.

[0046] Specifically,

[0047] S31, using a regular expression to screen out entities in the geological mineral field knowledge base whose data are all Chinese characters, and converting the entities into a dictionary format;

[0048] S32, removing characters other than Chinese characters and Chinese common punctuation marks in the unstructured data in the geological mineral field related corpus to be annotated, and splitting the multiple sentences of the unstructured data in the geological mineral field related corpus to be annotated after removing the characters into a list format;

[0049] S33, matching entities in the sentences split into a list format and the entities converted into a dictionary format in S31 and adding relationships.

[0050] Specifically,

[0051] S331, traversing each sentence in the sentences split into a list format, saving all entities appearing in the sentences according to string matching, and filtering out sentences without entities or with only one entity appearing;

[0052] S332, designing semantic relationships, aligning and aggregating relationships not aligned in the unstructured data in the geological mineral field related corpus to be annotated by using artificial rules;

[0053] S333, traversing each piece of data in the geological mineral field related corpus to be annotated, and annotating relationships according to the entity dictionary and the artificially defined relationship alignment table.

[0054] The remote supervision idea involved in the present application mainly refers to obtaining a triple R(E1, E2) (wherein R represents a relationship, and E1 and E2 represent two entities) from a knowledge graph, and the two entities jointly appear in a sentence S, then S expresses the relationship R between E1 and E2, and S is annotated as training corpus. Specifically, for a triple in an existing knowledge base, it is assumed that any sentence containing the pair of entities in the training corpus reflects the relationship to some extent, for details see Figure 4 Based on this assumption, the remote supervision algorithm can label the relationship tags of the sentences in the training corpus based on a small annotated knowledge base, which is equivalent to realizing automatic annotation of samples, and therefore is also a semi-supervised algorithm. This algorithm greatly saves the cost of manual annotation to a great extent.

[0055] The semantic relations designed in the application include "hasAlteration", "isControlledBy", "isLocatedIn", "isFormedIn", "hasMinerals", "hasElement", "isRelatedTo", "isFoundIn", "isAnalyzedBy", "isRevealedBy". The artificial relation alignment table in the application is shown in Table 1:

[0056] Table 1

[0057]

[0058] The application takes the tin mine in Yunnan Province as an example to carry out the corresponding experiment. First, the structured and semi-structured data in the national resource potential evaluation library and Baidu Encyclopedia are used to process the semi-structured data into structured data by means of regular expressions. The specific implementation steps are as follows:

[0059] 1. Collect the structured and semi-structured data related to the tin mine field in the national resource potential evaluation library and Baidu Encyclopedia, but before that, the specific entities related to the tin mine field need to be defined. The specific entities involved in this paper are mainly derived from the entities related to the tin mine field in the geological dictionary. After the entity is defined, the semi-structured data related to it is extracted from Baidu Encyclopedia. The following is part of the semi-structured data extracted from Baidu Encyclopedia in the embodiment of the application.

[0060] "Gejiu tin mine # Chinese name: Gejiu tin mine # Features: super large tin polymetallic deposit # Metal reserves: cumulative proven non-ferrous metal reserves 4.76 million tons # Location: western end of the southeast non-ferrous metal mineralization belt in Yunnan

[0061] Structure # Chinese name: structure # English name: structure # Common fields: mineral resources, hydrology, engineering geology, earthquake, geothermal # Discipline: geology # Type: fault, joint, fold, etc. # Research method: field geological survey

[0062] Fault # Chinese name: fault # English name: fault # Pinyin: duàn céng # Attribute: geological structure type # Main cause: tectonic movement # Basic type: normal fault, reverse fault, flat push fault

[0063] Mineral # Chinese name: mineral # English name: mineral # Principle: geological process # Property: natural element or compound # Classification: sulfide, oxide # Color: different colors

[0064] Metallic mineral

[0065] Non-metallic mineral

[0066] Clay mineral

[0067] Primary mineral

[0068] Secondary mineral

[0069] Primary mineral

[0070] Inclusion

[0071] Cleavage

[0072] Parting

[0073] 2、When the semi-structured data extraction is completed, regular expressions are used to process it to obtain more than 6000 triple information, and Table 2 and Table 3 are part of different encyclopedia triple data after regular expression of the embodiment of the application.

[0074] Table 2

[0075] Head entity Relationship Tail entity Spodumene Cleavage Prismatic crystals, tabular prismatic crystals Hornblende Cleavage Prismatic cleavage complete Orthoamphibole Crystal habit Prismatic or tabular, commonly prismatic and fibrous aggregates Argyropyrite Crystal form Prismatic and tabular crystals Melanocerine Shape Prismatic Aegirine Crystal form Prismatic Calc-silicate rock Property Prismatic granulitic texture

[0076] Table 3

[0077]

[0078]

[0079] 3、The data in the national resource potential evaluation library needs to be processed by MapGis, and according to the attribute information, more than 2000 triple information of the tin mine field related to Yunnan Province is extracted, and the specific results are shown in Table 4, and Table 4 is part of the related triple related to the tin mine field in Yunnan Province in the national resource potential evaluation library of the embodiment of the application.

[0080] Table 4

[0081] Head entity Relationship Tail entity Lailishan tin deposit in Hexi Township, Lianghe County Located in Pidanxi-Xiaolonghe (Yanshanian magmatic arc) Sn-T-Fe-Pb-Zn-Cu-Ag belt Silguangping tin deposit in Hexi Township, Lianghe County Located in Qipanxi-Xiaolonghe (Yanshanian magmatic arc) Sn-W-Fe-Pb-Zn-Cu-Ag ore belt Xiaolonghe tin deposit in Ruidian Township, Tengchong County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-As-Fe-Mn ore belt Laopingshan tin deposit in Zhonghe Township, Tengchong County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Tieyao mountain tin deposit in Gudong Township, Tengchong County Located in Donghe-Mingguang (Huaoshan magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Fenshuiling tin deposit in Ruidian Township, Tengchong County Located in Donghe-Mingguang (Songshan magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Hongyan head tin deposit in Mingguang Township, Tengchong County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Mulianghe tin deposit in Ruodian Township, Tengchong County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Zhudushi tin deposit in Ruodian Township, Tengchong County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Dalonghe tin deposit in Ruodian Township, Tengchong County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Guaocangshan tin deposit in Ruizhen Township, Tengchong County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Baili tin deposit in Peli, Gongshan County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Liwadi tin deposit in Gongshan County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Changwa tin deposit in Gongshan County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Luopo tin deposit in Gongshan County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Xiyuege tin deposit in Gongshan County Located in Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Dangzhu tin deposit in Gongshan County Located in ​ ​ ​ ​ ​ Donghe-Mingguang (Yanshanian magmatic arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt Luo'ao tin deposit in Gongshan County Located Donghe-Mingguang (Liaoshannenxingli arc) Sn-Cu-Pb-Zn-Ag-Fe-Mn ore belt

[0082] Secondly, the structured data in the existing corpus related to the tin mine field in Yunnan Province is extracted by using the rule-based method, and the structured data is supplemented, and more than 10000 triple data related to the tin mine field is obtained, and a knowledge base related to the tin mine field is constructed, and the specific implementation steps are as follows:

[0083] 1、Define the relationship set to be extracted, such as the relationship referring word in Table 1, including "located in", "mineralization", "wall rock alteration" and the like;

[0084] 2、Traverse each sentence of the article, and remove the words in the non-entity and non-relationship set in each sentence, wherein the entity type is mainly 22 entity types related to the tin mine field in the geological dictionary and the constraint of the named entity recognition result, and the relationship set is the relationship referring word defined in Table 1;

[0085] 3、From the second word of each sentence, when the word in the relationship set is encountered, the left and right nearest entities are selected, such as in the sentence "The accessory minerals of Lailishan tin mine mainly include zircon, apatite, magnetite and epidote", the relationship word is "accessory minerals", when the relationship word "accessory minerals" is encountered, the left and right nearest entities "Lailishan tin mine" and "zircon" are extracted, and finally the triple <Lailishan tin mine> accessory minerals <zircon> is formed. Based on this, the rule-based relationship extraction flowchart of the tin mine field in Yunnan Province is established, as shown in Figure 5 .

[0086] Then we collected and processed structured and semi-structured data from the national resource potential evaluation library and Baidu Encyclopedia, and obtained more than 6,000 triples related to tin mining. We used a rule-based method to extract structured data in the tin mining field in Yunnan Province, and finally obtained more than 2,000 triples related to the tin mining field in Yunnan Province. A total of more than 10,000 triples related to the tin mining field were obtained, which were used to build a preliminary knowledge base using remote supervision. Figure 6 is a partial display of the preliminary knowledge base using neo4j, Figure 6 The circles in represent different entities, and the text on the arrows represents the relationship. The entity where the arrow starts represents the head entity (such as the Ximeng mining area), and the entity where the arrow points represents the tail entity (such as silicification, granulite, sandstone, etc.).

[0087] Finally, we used the remote supervision method to extract triples from the knowledge base to align the unlabeled corpus for labeling, thereby achieving rapid construction of the relationship extraction training corpus in the tin mining field in Yunnan Province. The specific implementation steps are as follows:

[0088] 1. For triples data, this paper first uses regular expressions to filter out all Chinese characters and converts them to dictionary format.

[0089] 2. For unlabeled raw data, preprocessing work is done to remove all Chinese characters and common Chinese punctuation characters, and multiple sentences are split into list format.

[0090] 3. Traverse each sentence and traverse the entity set. Save all entities that appear in the sentence according to string matching, and filter out sentences with no entities or only one entity.

[0091] 4. According to the analysis of the knowledge of tin mining in Yunnan Province, 10 types of relationships are predefined, including "hasAlteration", "isControlledBy", "isLocatedIn", "isFormedIn", "hasMinerals", "hasElement", "isRelatedTo", "isFoundIn", "isAnalyzedBy", and "isRevealedBy". These relationship types are listed in the relationship label column in Table 1.

[0092] 5. For relationships in the tin mining field in Yunnan Province that are not aligned (such as tin, sandstone corresponding to the same relationship), we use artificial rules to align and aggregate them.

[0093] 6. Traverse each piece of data in the relevant corpus of tin mining field in Yunnan Province, and mark the relationship according to the entity dictionary and the manually defined relationship alignment table 1. The specific implementation process is shown in Figure 3 .

[0094] Finally, more than 12000 pieces of relationship extraction training corpus of tin mining field in Yunnan Province are obtained, and the results are shown in Table 5. Table 5 is part of the training corpus constructed based on the remote supervision idea in the embodiment of the present application. Compared with the manual annotation method, the time for annotating the semantic relationship corpus is greatly reduced.

[0095] Table 5

[0096]

[0097]

[0098] Figure 7 The remote supervision training corpus construction project is the embodiment of the present application.

[0099] The above description of disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for quickly constructing a training corpus for semantic relation extraction, characterized in that, The method comprises the following steps: S1, obtaining geological and mineral resources field related corpus, including to-be-labeled geological and mineral resources field related corpus and structured geological and mineral resources field related corpus; S2, obtaining semi-structured and structured data of the geological and mineral resources field, and extracting structured data in the geological and mineral resources field related corpus to construct a geological and mineral resources field knowledge base; S3, based on the to-be-labeled geological and mineral resources field related corpus and the constructed knowledge base, realizing rapid construction of training corpus; Step S2 is specifically: S21, obtaining semi-structured and structured data of the geological and mineral resources field; S22, using a rule-based method to extract structured data from the geological and mineral resources field related corpus; S23, using a regular expression to convert the semi-structured data into structured data; S24, knowledge fusion of the structured data in S21, S22 and S23, construction of a triple data, and obtaining a geological and mineral resources field knowledge base; Step S3 is specifically: S31, using a regular expression to filter out all Chinese characters in the geological and mineral resources field knowledge base, and converting them into a dictionary format to obtain an entity dictionary; S32, removing characters other than Chinese characters and Chinese common punctuation in the to-be-labeled geological and mineral resources field related corpus, and splitting multiple sentences in the to-be-labeled geological and mineral resources field related corpus after removing the characters into a list format; S33, matching the entities in the list format sentences in S32 with the entities in the entity dictionary in S31 and adding relationships; Step S33 is specifically: S331, traversing each sentence in the list format sentences, saving all entities appearing in the sentences according to string matching, and filtering out sentences with no entities or only one entity appearing; S332, designing semantic relationships, aligning and aggregating relationships in the to-be-labeled geological and mineral resources field related corpus by using artificial rules; S333, traversing each pair of entities obtained in S331, and annotating the relationships according to the entity dictionary and the designed semantic relationships. 2.The method of claim 1, wherein, The geological and mineral resources field entities in step S1 include ore-bearing strata, tectonic environment, ore-controlling strata, alteration types, alteration zoning, ore-controlling structures, ore-controlling rock bodies, deposit locations, minerals and elements.

3. A device for quickly constructing a semantic relation extraction training corpus, characterized in that, The device comprises: a processor; a memory having a computer program stored thereon and executable on the processor; wherein the computer program, when executed by the processor, implements a semantic relationship extraction training corpus rapid construction method according to any one of claims 1 to 2.

Images